Hi!
As the testcase below shows, fold_nonarray_ctor_reference doesn't have a
correct check whether a field might overlap with the given offset, size
access (in this case a BIT_FIELD_REF). The bitfield ref wants to read
8 bits starting at offset 0, the bitfield defined there goes from bit 1
1 bit and as fold_nonarray_ctor_reference doesn't detect an overlap,
it says the result is 0 as if there was an initializer, but without any
overlapping fields. Fixed thusly, bootstrapped/regtested on x86_64-linux
and i686-linux, ok for trunk/4.6?
For trunk I think we could eventually do better and handle even multiple
fields overlapping the access and/or non-zero offsets within that,
provided the recursive call to fold_ctor_reference returned an INTEGER_CST
we could mask/shift/or them together.
2011-07-18 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/49768
* gimple-fold.c (fold_nonarray_ctor_reference): Return NULL
if offset is smaller than bitoffset, but offset+size is bigger
than bitoffset.
* gcc.c-torture/execute/pr49768.c: New test.
--- gcc/gimple-fold.c.jj 2011-07-15 20:46:49.000000000 +0200
+++ gcc/gimple-fold.c 2011-07-18 14:28:35.000000000 +0200
@@ -3231,7 +3231,7 @@ fold_nonarray_ctor_reference (tree type,
double_int bitoffset;
double_int byte_offset_cst = tree_to_double_int (byte_offset);
double_int bits_per_unit_cst = uhwi_to_double_int (BITS_PER_UNIT);
- double_int bitoffset_end;
+ double_int bitoffset_end, access_end;
/* Variable sized objects in static constructors makes no sense,
but field_size can be NULL for flexible array members. */
@@ -3252,14 +3252,16 @@ fold_nonarray_ctor_reference (tree type,
else
bitoffset_end = double_int_zero;
- /* Is OFFSET in the range (BITOFFSET, BITOFFSET_END)? */
- if (double_int_cmp (uhwi_to_double_int (offset), bitoffset, 0) >= 0
+ access_end = double_int_add (uhwi_to_double_int (offset),
+ uhwi_to_double_int (size));
+
+ /* Is there any overlap between [OFFSET, OFFSET+SIZE) and
+ [BITOFFSET, BITOFFSET_END)? */
+ if (double_int_cmp (access_end, bitoffset, 0) > 0
&& (field_size == NULL_TREE
|| double_int_cmp (uhwi_to_double_int (offset),
bitoffset_end, 0) < 0))
{
- double_int access_end = double_int_add (uhwi_to_double_int (offset),
- uhwi_to_double_int (size));
double_int inner_offset = double_int_sub (uhwi_to_double_int (offset),
bitoffset);
/* We do have overlap. Now see if field is large enough to
@@ -3267,6 +3269,8 @@ fold_nonarray_ctor_reference (tree type,
fields. */
if (double_int_cmp (access_end, bitoffset_end, 0) > 0)
return NULL_TREE;
+ if (double_int_cmp (uhwi_to_double_int (offset), bitoffset, 0) < 0)
+ return NULL_TREE;
return fold_ctor_reference (type, cval,
double_int_to_uhwi (inner_offset), size);
}
--- gcc/testsuite/gcc.c-torture/execute/pr49768.c.jj 2011-07-18
14:37:18.000000000 +0200
+++ gcc/testsuite/gcc.c-torture/execute/pr49768.c 2011-07-18
14:37:03.000000000 +0200
@@ -0,0 +1,12 @@
+/* PR tree-optimization/49768 */
+
+extern void abort (void);
+
+int
+main ()
+{
+ static struct { unsigned int : 1; unsigned int s : 1; } s = { .s = 1 };
+ if (s.s != 1)
+ abort ();
+ return 0;
+}
Jakub
