D issues are now tracked on GitHub. This Bugzilla instance remains as a read-only archive.
Issue 3398 - Attributes inside a union screws data alignment
Summary: Attributes inside a union screws data alignment
Status: RESOLVED FIXED
Alias: None
Product: D
Classification: Unclassified
Component: dmd (show other issues)
Version: D2
Hardware: Other Windows
: P2 critical
Assignee: No Owner
URL:
Keywords: patch, wrong-code
Depends on:
Blocks:
 
Reported: 2009-10-14 07:23 UTC by Koroskin Denis
Modified: 2015-06-09 01:28 UTC (History)
3 users (show)

See Also:


Attachments

Note You need to log in before you can comment on or make changes to this issue.
Description Koroskin Denis 2009-10-14 07:23:40 UTC
Here is a test case:

union Foo1
{
    int a;
    float b;
}

version = Test;

union Foo2
{
    version (Test) {
        int a;
        float b;
    }
}

pragma(msg, Foo1.sizeof.stringof); // 4u (okay)
pragma(msg, Foo2.sizeof.stringof); // 8u (wrong)
Comment 1 Don 2010-04-21 01:14:12 UTC
This applies not just to version declarations; it applies to any attribute.

union Foo1 {
    double[2] a;
    double[2] b;        
}

union Foo2 {
public:
     double[2] a;
     double[2] b;        
}

static assert(Foo1.sizeof == Foo2.sizeof);
---
bug.d(32): Error: static assert  (16u == 32u) is false

Subtle and disastrous bug (especially when dealing with C libraries), escalating severity.

Here's a mitigation to create an error message, rather than generate wrong code.

MITIGATION: struct.c StructDeclaration::semantic() line 384.
The problem is that AttribDeclaration::semantic has it's own loop for
doing the semantic on members, which does NOT include setting the offset to 0
if it's a union. Fixing this would not be difficult, but would require a minor structural change in the compiler.

---

    for (int i = 0; i < members_dim; i++)
    {
        Dsymbol *s = (Dsymbol *)members->data[i];
+        if (isUnionDeclaration() && s->isAttribDeclaration())
+            error("Attributes don't yet work inside unions, see Bugzilla 3398");
        s->semantic(sc2);
        if (isUnionDeclaration())
            sc2->offset = 0;
-------
Comment 2 Don 2010-04-22 06:28:52 UTC
One instance where wrong code is currently generated is in 
std.c.windows.winsock.d,  in_addr and in6_addr.

My mitigation would break existing code, though, in cases where the AttribDeclaration is actually an AnonDeclaration.
Comment 3 Don 2010-06-09 00:31:56 UTC
PATCH: Currently, after processing every field, sc->offset gets changed, and StructDeclaration::semantic() sets the offset back to zero if it was a union.
This patch moves the processing to AggregateDeclaration::addField() instead. If it's a union, don't change the offset in the first place.

Index: struct.c
===================================================================
--- struct.c	(revision 527)
+++ struct.c	(working copy)
@@ -194,20 +194,23 @@
         sizeok = 2;             // cannot finish; flag as forward referenced
         return;
     }
+    size_t ofs = sc->offset;
 
     memsize = t->size(loc);
     memalignsize = t->alignsize();
     xalign = t->memalign(sc->structalign);
-    alignmember(xalign, memalignsize, &sc->offset);
-    v->offset = sc->offset;
-    sc->offset += memsize;
-    if (sc->offset > structsize)
-        structsize = sc->offset;
+    alignmember(xalign, memalignsize, &ofs);
+    v->offset = ofs;
+    ofs += memsize;
+    if (ofs > structsize)
+        structsize = ofs;
     if (sc->structalign < memalignsize)
         memalignsize = sc->structalign;
     if (alignsize < memalignsize)
         alignsize = memalignsize;
     //printf("\talignsize = %d\n", alignsize);
+    if (!isUnionDeclaration())
+        sc->offset = ofs;
 
     v->storage_class |= STCfield;
     //printf(" addField '%s' to '%s' at offset %d, size = %d\n", v->toChars(), toChars(), v->offset, memsize);
@@ -386,8 +389,6 @@
     {
         Dsymbol *s = (Dsymbol *)members->data[i];
         s->semantic(sc2);
-        if (isUnionDeclaration())
-            sc2->offset = 0;
 #if 0
         if (sizeok == 2)
         {   //printf("forward reference\n");
Comment 4 Walter Bright 2010-06-09 16:37:24 UTC
http://www.dsource.org/projects/dmd/changeset/529