[PATCH, PR40351] When creating artificial (copy-propagation) sub-accesses, check that they make sense

[Martin Jambor]

Hi,

this  patch fixes  PR 40351, caused  by  the  fact  that we  propagate
sub-accesses across assignments, unions can now trick us into creating
some which  do not make  sense (they do  not exist in a  union field).
The  patch  below checks  that  does  not  happen before  creating  an
artificial access.

Bootstrapped and regression tested on x86_64-linux.  OK for trunk?

Thanks,

Martin

2009-06-05  Martin Jambor  <mjambor@suse.cz>

	PR tree-optimization/40351
	* tree-sra.c (propagate_subacesses_accross_link): Check that a refrence
	to a potential artifical subaccess can be constructed.

	* testsuite/gcc.c-torture/compile/pr40351.c: New file.

Index: gcc/testsuite/gcc.c-torture/compile/pr40351.c
===================================================================
--- gcc/testsuite/gcc.c-torture/compile/pr40351.c	(revision 0)
+++ gcc/testsuite/gcc.c-torture/compile/pr40351.c	(revision 0)
@@ -0,0 +1,22 @@
+/* PR tree-optimizations/40351 */
+
+struct IO_APIC_route_entry {
+    unsigned int vector : 8;
+    unsigned int delivery_mode : 1;
+    unsigned int mask : 1;
+    unsigned int __reserved_2 : 15;
+    unsigned int __reserved_3 : 8;
+} __attribute__ ((packed));
+union entry_union {
+    struct {
+        unsigned int w1, w2;
+    };
+    struct IO_APIC_route_entry entry;
+};
+unsigned int io_apic_read(void);
+struct IO_APIC_route_entry ioapic_read_entry(void)
+{
+  union entry_union eu;
+  eu.w1 = io_apic_read();
+  return eu.entry;
+}
Index: gcc/tree-sra.c
===================================================================
--- gcc/tree-sra.c	(revision 148205)
+++ gcc/tree-sra.c	(working copy)
@@ -1544,6 +1544,10 @@ propagate_subacesses_accross_link (struc
 	  continue;
 	}
 
+      if (!build_ref_for_offset (NULL, TREE_TYPE (lacc->base), norm_offset,
+				 rchild->type, false))
+	continue;
+
       new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
       if (racc->first_child)
 	propagate_subacesses_accross_link (new_acc, rchild);