[PATCH] Fix PR94401 by considering reverse overrun

["Kewen.Lin" <linkw@linux.ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 710 bytes --]

Hi, 

The commit r10-7415 brings scalar type consideration
to eliminate epilogue peeling for gaps, but it exposed
one problem that the current handling doesn't consider
the memory access type VMAT_CONTIGUOUS_REVERSE, for
which the overrun happens on low address side.  This
patch is to make the code take care of it by updating
the offset and construction element order accordingly.

Bootstrapped/regtested on powerpc64le-linux-gnu P8
and aarch64-linux-gnu.

BR,
Kewen
-----------
gcc/ChangeLog

2020-04-02  Kewen Lin  <linkw@gcc.gnu.org>

	PR tree-optimization/94401
	* tree-vect-loop.c (vectorizable_load): Handle VMAT_CONTIGUOUS_REVERSE
	access type when loading halves of vector to avoid peeling for gaps.

[-- Attachment #2: PR94401_REVERSE.diff --]
[-- Type: text/plain, Size: 2277 bytes --]

diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
index 12beef6978c..3d27f59ba22 100644
--- a/gcc/tree-vect-stmts.c
+++ b/gcc/tree-vect-stmts.c
@@ -9590,11 +9590,20 @@ vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 			    if (new_vtype != NULL_TREE)
 			      ltype = half_vtype;
 			  }
+			tree offset = dataref_offset
+					? dataref_offset
+					: build_int_cst (ref_type, 0);
+			if (ltype != vectype
+			    && memory_access_type == VMAT_CONTIGUOUS_REVERSE)
+			  offset = size_binop (
+			    PLUS_EXPR,
+			    build_int_cst (ref_type,
+					   DR_GROUP_GAP (first_stmt_info)
+					     * tree_to_uhwi (
+					       TYPE_SIZE_UNIT (elem_type))),
+			    offset);
 			data_ref
-			  = fold_build2 (MEM_REF, ltype, dataref_ptr,
-					 dataref_offset
-					 ? dataref_offset
-					 : build_int_cst (ref_type, 0));
+			  = fold_build2 (MEM_REF, ltype, dataref_ptr, offset);
 			if (alignment_support_scheme == dr_aligned)
 			  ;
 			else if (DR_MISALIGNMENT (first_dr_info) == -1)
@@ -9607,16 +9616,27 @@ vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 						  TYPE_ALIGN (elem_type));
 			if (ltype != vectype)
 			  {
-			    vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr));
+			    vect_copy_ref_info (data_ref,
+						DR_REF (first_dr_info->dr));
 			    tree tem = make_ssa_name (ltype);
 			    new_stmt = gimple_build_assign (tem, data_ref);
-			    vect_finish_stmt_generation (stmt_info, new_stmt, gsi);
+			    vect_finish_stmt_generation (stmt_info, new_stmt,
+							 gsi);
 			    data_ref = NULL;
 			    vec<constructor_elt, va_gc> *v;
 			    vec_alloc (v, 2);
-			    CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem);
-			    CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
-						    build_zero_cst (ltype));
+			    if (memory_access_type == VMAT_CONTIGUOUS_REVERSE)
+			      {
+				CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
+							build_zero_cst (ltype));
+				CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem);
+			      }
+			    else
+			      {
+				CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem);
+				CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
+							build_zero_cst (ltype));
+			      }
 			    gcc_assert (new_vtype != NULL_TREE);
 			    if (new_vtype == vectype)
 			      new_stmt = gimple_build_assign (