From: Richard Biener <rguenther@suse.de>
Date: Mon, 6 Nov 2023 13:34:07 +0000 (+0100)
Subject: tree-optimization/112405 - SIMD clone calls with (loop) mask
X-Git-Tag: basepoints/gcc-15~4981
X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=9125969086bfa1bf804b246ea574a2329b06d2c5;p=thirdparty%2Fgcc.git

tree-optimization/112405 - SIMD clone calls with (loop) mask

The following fixes the mask argument generation for SIMD clone
calls under either loop masking or when the actual call is not
masked but only a inbranch simd clone is available.  The issue
was that we tried to directly convert the vector mask to the
call argument type but SIMD clone masks require 1 or 0 (which
could be even float) values for mask elements so we have to
resort to a VEC_COND_EXPR to generate them just like we do for
regular passing of the mask.

	PR tree-optimization/112405
	* tree-vect-stmts.cc (vectorizable_simd_clone_call):
	Properly handle invariant and/or loop mask passing.
---

diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index eefb1eec1eff..65883e04ad7f 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -4814,36 +4814,42 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
 	      else
 		mask = vect_build_all_ones_mask (vinfo, stmt_info, masktype);
 
-	      if (!useless_type_conversion_p (TREE_TYPE (mask), masktype))
+	      gassign *new_stmt;
+	      if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode))
 		{
-		  gassign *new_stmt;
-		  if (bestn->simdclone->mask_mode != VOIDmode)
-		    {
-		      /* This means we are dealing with integer mask modes.
-			 First convert to an integer type with the same size as
-			 the current vector type.  */
-		      unsigned HOST_WIDE_INT intermediate_size
-			= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (mask)));
-		      tree mid_int_type =
-			build_nonstandard_integer_type (intermediate_size, 1);
-		      mask = build1 (VIEW_CONVERT_EXPR, mid_int_type, mask);
-		      new_stmt
-			= gimple_build_assign (make_ssa_name (mid_int_type),
-					       mask);
-		      gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
-		      /* Then zero-extend to the mask mode.  */
-		      mask = fold_build1 (NOP_EXPR, masktype,
-					  gimple_get_lhs (new_stmt));
-		    }
-		  else
-		    mask = build1 (VIEW_CONVERT_EXPR, masktype, mask);
-
-		  new_stmt = gimple_build_assign (make_ssa_name (masktype),
-						  mask);
-		  vect_finish_stmt_generation (vinfo, stmt_info,
-					       new_stmt, gsi);
-		  mask = gimple_assign_lhs (new_stmt);
+		  /* This means we are dealing with integer mask modes.
+		     First convert to an integer type with the same size as
+		     the current vector type.  */
+		  unsigned HOST_WIDE_INT intermediate_size
+		      = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (mask)));
+		  tree mid_int_type =
+		      build_nonstandard_integer_type (intermediate_size, 1);
+		  mask = build1 (VIEW_CONVERT_EXPR, mid_int_type, mask);
+		  new_stmt
+		      = gimple_build_assign (make_ssa_name (mid_int_type),
+					     mask);
+		  gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
+		  /* Then zero-extend to the mask mode.  */
+		  mask = fold_build1 (NOP_EXPR, masktype,
+				      gimple_get_lhs (new_stmt));
+		}
+	      else if (bestn->simdclone->mask_mode == VOIDmode)
+		{
+		  tree one = fold_convert (TREE_TYPE (masktype),
+					   integer_one_node);
+		  tree zero = fold_convert (TREE_TYPE (masktype),
+					    integer_zero_node);
+		  mask = build3 (VEC_COND_EXPR, masktype, mask,
+				 build_vector_from_val (masktype, one),
+				 build_vector_from_val (masktype, zero));
 		}
+	      else
+		gcc_unreachable ();
+
+	      new_stmt = gimple_build_assign (make_ssa_name (masktype), mask);
+	      vect_finish_stmt_generation (vinfo, stmt_info,
+					   new_stmt, gsi);
+	      mask = gimple_assign_lhs (new_stmt);
 	      vargs.safe_push (mask);
 	    }
 	}