[llvm-gcc] Bitfields & large array don't mix well

[llvmbot]

Bugzilla Link	141
Resolution	FIXED
Resolved on	Feb 22, 2010 12:50
Version	1.0
OS	All
Reporter	LLVM Bugzilla Contributor

Extended Description

Whereas I thought that the inability of llvm-gcc to chew pine was due to bug 6,
apparently fixing it was not enough. Here's a test case reduced from addrbook.c
in pine-4.58:

struct pine {
unsigned read_predicted:1;
char cur_folder[4000];
int dlevel;
};

extern struct pine *ps_global;
void dump_some_debugging(char *message) {
ps_global->dlevel = 1;
}

I thought I already filed this as a bug, but I can't find it. If you do, feel
free to mark this as duplicate.

[llvmbot]

assigned to @lattner

[lattner]

I'm not terribly suprised. The bitfield handling code is quite fragile and due
for a rewrite. However, doing this will be a destabilizing change which I do
not want to do at this point. This bug should be fixable though. :)

-Chris

[lattner]

Another testcase:

struct face_cachel {
unsigned int reverse :1;
unsigned char font_specified[1];
};

void
ensure_face_cachel_contains_charset (struct face_cachel *cachel) {
cachel->font_specified[0] = 0;
}

[lattner]

The second testcase below is fixed by this patch:

$ diff -u llvm-expand.c~ llvm-expand.c
--- llvm-expand.c~ 2003-11-19 18:25:37.000000000 -0600
+++ llvm-expand.c 2003-11-20 12:59:12.000000000 -0600
@@ -4588,8 +4588,7 @@
llvm_type *Op0Ty = llvm_type_get_from_tree(TREE_TYPE(TREE_OPERAND(exp, 0)));

 if (llvm_type_is_composite(Op0Ty)) {

•  llvm_value *Op0 = llvm_expand_lvalue_expr(Fn, TREE_OPERAND(exp, 0),
  

•                                            BitStart, BitSize);
  

•  llvm_value *Op0 = llvm_expand_lvalue_expr(Fn, TREE_OPERAND(exp, 0), 0, 0);
   llvm_value *Op1 = llvm_expand_expr(Fn, TREE_OPERAND(exp, 1), 0);
   Op1 = cast_if_type_not_equal(Fn, Op1, LongTy);
   Result = append_inst(Fn, create_gep3(Op0, llvm_constant_long_0, Op1));
  

@@ -4655,7 +4654,7 @@
/* If this is accessing a union element, we can go ahead and cast the
* pointer to the desired type now.
*/

•  if (ActualOffset == Offset && Size > ActualSize) {
  

•  if (ActualOffset == Offset && ResultElTy != FieldDeclTy) {
     assert((ActualSize & 7) == 0 && (Size & 7) == 0 &&
            "Illegal union field reference!");
     Result = cast_if_type_not_equal(Fn, Op0,
  

... and is tested by CFrontend/2003-11-20-Bitfields.c

The first testcase in this patch is HELPED by this patch, but not fixed
completely. The rest of the fix falls under the guise of Bug 82.

-Chris