tree-optimization/113630 - invalid code hoisting

The following avoids code hoisting (but also PRE insertion) of
expressions that got value-numbered to another one that are not
a valid replacement (but still compute the same value).  This time
because the access path ends in a structure with different size,
meaning we consider a related access as not trapping because of the
size of the base of the access.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

	PR tree-optimization/113630
	* tree-ssa-pre.cc (compute_avail): Avoid registering a
	reference with a representation with not matching base
	access size.

	* gcc.dg/torture/pr113630.c: New testcase.

gcc/testsuite/gcc.dg/torture/pr113630.c

@@ -0,0 +1,4 @@
+/* { dg-do run { target { { *-*-linux* *-*-gnu* *-*-uclinux* } && mmap } } } */
+/* { dg-additional-options "-fno-strict-aliasing" } */
+
+#include "pr110799.c"

gcc/tree-ssa-pre.cc

@@ -4221,6 +4221,20 @@ compute_avail (function *fun)
 			      = wide_int_to_tree (ptr_type_node,
 						  wi::to_wide (ref1->op2));
 			}
+		      /* We also need to make sure that the access path
+			 ends in an access of the same size as otherwise
+			 we might assume an access may not trap while in
+			 fact it might.  That's independent of whether
+			 TBAA is in effect.  */
+		      if (TYPE_SIZE (ref1->type) != TYPE_SIZE (ref2->type)
+			  && (! TYPE_SIZE (ref1->type)
+			      || ! TYPE_SIZE (ref2->type)
+			      || ! operand_equal_p (TYPE_SIZE (ref1->type),
+						    TYPE_SIZE (ref2->type))))
+			{
+			  operands.release ();
+			  continue;
+			}
 		      operands.release ();
 
 		      result = get_or_alloc_expr_for_reference