[gcc(refs/vendors/vrull/heads/slp-improvements)] Fix issues and improve vect_slp_optimize_permutes.

[Philipp Tomsich]

https://gcc.gnu.org/g:042460234612e9817e26d6f1c1dd41aa144a4c20

commit 042460234612e9817e26d6f1c1dd41aa144a4c20
Author: Manolis Tsamis <manolis.tsamis@vrull.eu>
Date:   Thu Nov 23 11:28:44 2023 +0100

    Fix issues and improve vect_slp_optimize_permutes.
    
     * Don't use tree_to_vec_perm_builder to get the elements of the
       permute constant vectors. This function uses
       vector_cst_encoded_nelts internally which may return fewer elements
       than VECTOR_CST_NELTS and in turn fewer elements than what is
       observed on the GIMPLE statement. Because we iterate based on NELTS
       this would cause memory corruption issues. The resulting code is
       also more clean and performant.
    
     * Call update_stmt after rewriting permutes because otherwise we may
       get SSA imm_use lists issues.
    
    Ref #346

Diff:
---
 gcc/tree-vect-slp.cc | 39 +++++++++++++++++++++++----------------
 1 file changed, 23 insertions(+), 16 deletions(-)

diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 68b745f8445..8dd9d85d9b8 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -8385,33 +8385,36 @@ vect_slp_optimize_permutes (function *fun)
 	  continue;
 
 	vec_perm_builder sel (nelts, nelts, 1);
-	vec_perm_builder sel1_1, sel1_2, sel2_1, sel2_2;
-	if (!tree_to_vec_perm_builder (&sel1_1, gimple_assign_rhs3 (src1_1))
-	    || !tree_to_vec_perm_builder (&sel1_2, gimple_assign_rhs3 (src1_2))
-	    || !tree_to_vec_perm_builder (&sel2_1, gimple_assign_rhs3 (src2_1))
-	    || !tree_to_vec_perm_builder (&sel2_2, gimple_assign_rhs3 (src2_2)))
-	  continue;
+	vec_perm_builder sel_1 (nelts, nelts, 1);
+	vec_perm_builder sel_2 (nelts, nelts, 1);
 
 	/* Rewrite the permutations based on MASK1_LANES/MASK2_LANES.  */
 	for (i = 0; i < nelts; i++)
 	  {
+	    /* Calculate new mask for STMT2.  */
 	    tree val = VECTOR_CST_ELT (mask2, i);
-	    unsigned HOST_WIDE_INT j
+	    unsigned HOST_WIDE_INT lane
 	      = TREE_INT_CST_LOW (val) & (2 * nelts - 1);
-	    unsigned off = (j < nelts) ? 0 : nelts;
-	    unsigned HOST_WIDE_INT new_lane = *mask2_lanes.get (j - off) + off;
+	    unsigned off = (lane < nelts) ? 0 : nelts;
+	    unsigned HOST_WIDE_INT new_lane
+	      = *mask2_lanes.get (lane - off) + off;
 	    sel.quick_push (new_lane);
 
-	    if (!mask1_lanes.get (i))
-	      {
-		sel1_1[i] = nelts + sel2_1[i];
-		sel1_2[i] = nelts + sel2_2[i];
-	      }
+	    /* Calculate new masks for SRC1_1 and SRC1_2.  */
+	    bool use_src1 = mask1_lanes.get (i);
+	    tree mask1 = gimple_assign_rhs3 (use_src1 ? src1_1 : src2_1);
+	    tree mask2 = gimple_assign_rhs3 (use_src1 ? src1_2 : src2_2);
+	    unsigned HOST_WIDE_INT lane1
+	      = TREE_INT_CST_LOW (VECTOR_CST_ELT (mask1, i)) & (nelts - 1);
+	    unsigned HOST_WIDE_INT lane2
+	      = TREE_INT_CST_LOW (VECTOR_CST_ELT (mask2, i)) & (nelts - 1);
+	    sel_1.quick_push (lane1 + (use_src1 ? 0 : nelts));
+	    sel_2.quick_push (lane2 + (use_src1 ? 0 : nelts));
 	  }
 
 	vec_perm_indices indices (sel, 2, nelts);
-	vec_perm_indices indices1_1 (sel1_1, 2, nelts);
-	vec_perm_indices indices1_2 (sel1_2, 2, nelts);
+	vec_perm_indices indices1_1 (sel_1, 2, nelts);
+	vec_perm_indices indices1_2 (sel_2, 2, nelts);
 
 	tree vectype = TREE_TYPE (gimple_assign_lhs (stmt2));
 	if (!can_vec_perm_const_p (TYPE_MODE (vectype), TYPE_MODE (vectype), indices)
@@ -8429,6 +8432,10 @@ vect_slp_optimize_permutes (function *fun)
 
 	gimple_assign_set_rhs3 (src1_1, vect_gen_perm_mask_checked (vectype, indices1_1));
 	gimple_assign_set_rhs3 (src1_2, vect_gen_perm_mask_checked (vectype, indices1_2));
+
+	update_stmt (stmt2);
+	update_stmt (src1_1);
+	update_stmt (src1_2);
       }
 }