[PR87054] fix unaligned access

Building an ADDR_EXPR uses the canonical type to build the pointer
type, but then, as we dereference it, we lose track of lax alignment
known to apply to the dereferenced object.  This might not be a
problem in general, but it is when the compiler implicitly introduces
address taking and dereferencing, as it does for asm statements, and
as it may do in some loop optimizations.

From: Richard Biener <rguenther@suse.de>
for  gcc/ChangeLog

	PR middle-end/87054
	* gimplify.c (gimplify_expr): Retain alignment of
	addressable lvalue in dereference.

From: Alexandre Oliva <oliva@adacore.com>
for  gcc/testsuite/ChangeLog

	PR middle-end/87054
	* gcc.dg/pr87054.c: New.

From-SVN: r264450

gcc/ChangeLog

+2018-09-20  Richard Biener <rguenther@suse.de>
+
+	PR middle-end/87054
+	* gimplify.c (gimplify_expr): Retain alignment of
+	addressable lvalue in dereference.
+
 2018-09-20  Alexandre Oliva <aoliva@redhat.com>
 
 	PR bootstrap/87013

gcc/gimplify.c

@@ -12538,9 +12538,15 @@ gimplify_expr (tree *expr_p, gimple_seq *pre_p, gimple_seq *post_p,
       /* An lvalue will do.  Take the address of the expression, store it
 	 in a temporary, and replace the expression with an INDIRECT_REF of
 	 that temporary.  */
+      tree ref_alias_type = reference_alias_ptr_type (*expr_p);
+      unsigned int ref_align = get_object_alignment (*expr_p);
+      tree ref_type = TREE_TYPE (*expr_p);
       tmp = build_fold_addr_expr_loc (input_location, *expr_p);
       gimplify_expr (&tmp, pre_p, post_p, is_gimple_reg, fb_rvalue);
-      *expr_p = build_simple_mem_ref (tmp);
+      if (TYPE_ALIGN (ref_type) != ref_align)
+	ref_type = build_aligned_type (ref_type, ref_align);
+      *expr_p = build2 (MEM_REF, ref_type,
+			tmp, build_zero_cst (ref_alias_type));
     }
   else if ((fallback & fb_rvalue) && is_gimple_reg_rhs_or_call (*expr_p))
     {

gcc/testsuite/ChangeLog

+2018-09-20  Alexandre Oliva <oliva@adacore.com>
+
+	PR middle-end/87054
+	* gcc.dg/pr87054.c: New.
+
 2018-09-20  Richard Sandiford  <richard.sandiford@arm.com>
 
 	PR tree-optimization/87288

gcc/testsuite/gcc.dg/pr87054.c

+// { dg-do run }
+// { dg-options "-O2" }
+
+#ifndef T
+# ifdef __SSE__
+#  define T __int128
+# else
+#  define T long
+# endif
+#endif
+#ifndef R
+# ifdef __SSE__
+#  define R "x"
+# else
+#  define R "r"
+# endif
+#endif
+
+
+typedef T A; // #define T to long or __int128
+struct B { char d; A c; } __attribute__((packed));
+struct B b[50]; // many elements to avoid loop unrolling
+
+int main () {
+  int i;
+  for (i = 0; i < sizeof(b) / sizeof(*b); i++) {
+    asm ("" : "+" R (b[i].c)); // #define R to "r" on ppc or "x" on x86_64
+  }
+}