PR25575: Make GCC 4.4+ comatible layout for packed bit-fileds of char type

[DmitryPolukhin]

Summary

This CL is for discussion how to better fix bit-filed layout compatibility issue with GCC (see PR25575 for test case and more details). Current clang behavior is compatible with GCC 4.1-4.3 series but it was fixed in 4.4+. Ignoring packed attribute looks very odd and because it was also fixed in GCC 4.4+, it makes sense also fix it in clang.

Open questions:

1. Should we add warning about changes in layout for packed bit-fileds of char type? GCC has it "note: offset of packed bit-field ‘b’ has changed in GCC 4.4" will add if needed.

2. This CL completely removes warning "packed attribute ignored for field of type char". Should we keep it but apply only to normal fields (i.e. not bit-fields)?

This CL removes the warning because IMHO it makes no sense if we fix bit-field case to match GCC 4.4+. Packed attribute was actually ignored only for bit-field. For all other cases it is not ignored but reported only when alignment is already fulfilled (i.e. attribute just is not required and doesn't change anything). But there are tons of cases when packed attribute doesn't change anything but warning is not emitted, for example:
struct S1 {

char a;
char b __attribute((packed)); // warning: 'packed' attribute ignored for field of type 'char'
char c;

};

extern int a[sizeof(struct S1) == 3 ? 1 : -1];
extern int b[__alignof(struct S1) == 1 ? 1 : -1];

struct S2 {

short a;
short b __attribute((packed)); // no warning
short c;

};

extern int c[sizeof(struct S2) == 6 ? 1 : -1];
extern int d[__alignof(struct S2) == 2 ? 1 : -1];

In both cases you can remove __attribute((packed)) and program behavior will not change. But warning is generated only in the first case.

[DmitryPolukhin]

It seems that check for type alignment <= 8 was there practically forever http://llvm.org/viewvc/llvm-project/cfe/trunk/Sema/SemaDecl.cpp?r1=47197&r2=47196&pathrev=47197 and there is no good explanation why it was implemented. Subsequent changes only add more condition to exclude cases when it shouldn't be reported.

[hfinkel]

Please upload this patch with full context, see: http://llvm.org/docs/Phabricator.html#requesting-a-review-via-the-web-interface

Should we add warning about changes in layout for packed bit-fileds of char type? GCC has it "note: offset of packed bit-field ‘b’ has changed in GCC 4.4" will add if needed.

Warning about this is a good idea. We did something similar when we (re-)introduced sync_fetch_and_nand, see:

def warn_sync_fetch_and_nand_semantics_change : Warning<
  "the semantics of this intrinsic changed with GCC "
  "version 4.4 - the newer semantics are provided here">,
  InGroup<DiagGroup<"sync-fetch-and-nand-semantics-changed">>;

[DmitryPolukhin]

Added warning about semantic change + uploaded context. PTAL

[hfinkel]

include/clang/Basic/DiagnosticSemaKinds.td

2783 ↗

(On Diff #41131)

Calling this "a semantic" reads oddly to me. This sounds better to me:

def note_attribute_packed_for_bitfield_offset_changed : Warning<
  "the offset assigned to packed bit-field member %0 has changed with GCC version 4.4 - "
  "the newer offset is used here">,
  InGroup<DiagGroup<"attribute-packed-bitfield-offset-changed">>;

lib/Sema/SemaDeclAttr.cpp

1040 ↗

(On Diff #41131)

This comment is now out of date?

[DmitryPolukhin]

Changed note text message + fixed outdated comment.

[DmitryPolukhin]

This is another GCC ABI compatibility issue. If there is no more comments, could someone please approve this CL?

[aaron.ballman]

lib/Sema/SemaDeclAttr.cpp

1046 ↗

(On Diff #41222)

No need to call getDeclName(), the diagnostic can accept a NamedDecl directly (which properly quotes the name in the diagnostic).

[DmitryPolukhin]

Don't call getDeclName() that it is not required. PTAL

[rjmccall]

include/clang/Basic/DiagnosticSemaKinds.td

2790 ↗

(On Diff #41510)

No, this diagnostic shouldn't make such an affirmative statement about the offset changing. I would instead suggest:

warning: 'packed' attribute was ignored on bit-fields with alignment 1 in older versions of GCC and Clang

[DmitryPolukhin]

Warning text updated, PTAL.

[aaron.ballman]

include/clang/Basic/DiagnosticSemaKinds.td

2788 ↗

(On Diff #41597)

Would this read better as "on bit-fields with single-byte alignment" instead of "on bit-fields with alignment 1"?

test/Sema/struct-packed-align.c

155 ↗

(On Diff #41597)

"ignores uses" -> "uses"

[DmitryPolukhin]

Fixed warning and comment.

[rjmccall]

LGTM, thanks!

[aaron.ballman]

LGTM! Thank you!

cfe/trunk/include/clang/Basic/DiagnosticSemaKinds.td

 // for their workflow.
 def warn_int_to_pointer_cast : Warning<
   "cast to %1 from smaller integer type %0">,
   InGroup<IntToPointerCast>;
 def warn_int_to_void_pointer_cast : Warning<
   "cast to %1 from smaller integer type %0">,
   InGroup<IntToVoidPointerCast>;
 
-def warn_attribute_ignored_for_field_of_type : Warning<
-  "%0 attribute ignored for field of type %1">,
-  InGroup<IgnoredAttributes>;
+def warn_attribute_packed_for_bitfield : Warning<
+  "'packed' attribute was ignored on bit-fields with single-byte alignment "
+  "in older versions of GCC and Clang">,
+  InGroup<DiagGroup<"attribute-packed-for-bitfield">>;
 def warn_transparent_union_attribute_field_size_align : Warning<
   "%select{alignment|size}0 of field %1 (%2 bits) does not match the "
   "%select{alignment|size}0 of the first field in transparent union; "
   "transparent_union attribute ignored">,
   InGroup<IgnoredAttributes>;
 def note_transparent_union_first_field_size_align : Note<
   "%select{alignment|size}0 of first field is %1 bits">;
 def warn_transparent_union_attribute_not_definition : Warning<

cfe/trunk/lib/Sema/SemaDeclAttr.cpp

static void handleExtVectorTypeAttr(Sema &S, Scope *scope, Decl *D,
   S.ExtVectorDecls.push_back(cast<TypedefNameDecl>(D));
 }
 
 static void handlePackedAttr(Sema &S, Decl *D, const AttributeList &Attr) {
   if (TagDecl *TD = dyn_cast<TagDecl>(D))
     TD->addAttr(::new (S.Context) PackedAttr(Attr.getRange(), S.Context,
                                         Attr.getAttributeSpellingListIndex()));
   else if (FieldDecl *FD = dyn_cast<FieldDecl>(D)) {
-    // If the alignment is less than or equal to 8 bits, the packed attribute
-    // has no effect.
+    // Report warning about changed offset in the newer compiler versions.
     if (!FD->getType()->isDependentType() &&
-        !FD->getType()->isIncompleteType() &&
+        !FD->getType()->isIncompleteType() && FD->isBitField() &&
         S.Context.getTypeAlign(FD->getType()) <= 8)
-      S.Diag(Attr.getLoc(), diag::warn_attribute_ignored_for_field_of_type)
-        << Attr.getName() << FD->getType();
-    else
-      FD->addAttr(::new (S.Context)
-                  PackedAttr(Attr.getRange(), S.Context,
-                             Attr.getAttributeSpellingListIndex()));
+      S.Diag(Attr.getLoc(), diag::warn_attribute_packed_for_bitfield);
+
+    FD->addAttr(::new (S.Context) PackedAttr(
+        Attr.getRange(), S.Context, Attr.getAttributeSpellingListIndex()));
   } else
     S.Diag(Attr.getLoc(), diag::warn_attribute_ignored) << Attr.getName();
 }
 
 static bool checkIBOutletCommon(Sema &S, Decl *D, const AttributeList &Attr) {
   // The IBOutlet/IBOutletCollection attributes only apply to instance
   // variables or properties of Objective-C classes.  The outlet must also
   // have an object reference type.

cfe/trunk/test/Sema/struct-packed-align.c

 // RUN: %clang_cc1 %s -fsyntax-only -verify
-// expected-no-diagnostics
+// RUN: %clang_cc1 %s -fsyntax-only -triple=x86_64-windows-coff -verify
 
 // Packed structs.
 struct s {
     char a;
     int b  __attribute__((packed));
     char c;
     int d;
 };

 #if defined(_WIN32) && !defined(__declspec) // _MSC_VER is unavailable in cc1.
 // Alignment doesn't affect packing in MS mode.
 extern int n1[sizeof(struct nS) == 16 ? 1 : -1];
 extern int n2[__alignof(struct nS) == 8 ? 1 : -1];
 #else
 extern int n1[sizeof(struct nS) == 9 ? 1 : -1];
 extern int n2[__alignof(struct nS) == 1 ? 1 : -1];
 #endif
+
+// Packed attribute shouldn't be ignored for bit-field of char types.
+// Note from GCC reference manual: The 4.1, 4.2 and 4.3 series of GCC ignore
+// the packed attribute on bit-fields of type char. This has been fixed in
+// GCC 4.4 but the change can lead to differences in the structure layout.
+// See the documentation of -Wpacked-bitfield-compat for more information.
+struct packed_chars {
+  char a:4;
+  char b:8 __attribute__ ((packed));
+  // expected-warning@-1 {{'packed' attribute was ignored on bit-fields with single-byte alignment in older versions of GCC and Clang}}
+  char c:4;
+};
+
+#if defined(_WIN32)
+// On Windows clang uses MSVC compatible layout in this case.
+extern int o1[sizeof(struct packed_chars) == 3 ? 1 : -1];
+extern int o2[__alignof(struct packed_chars) == 1 ? 1 : -1];
+#else
+extern int o1[sizeof(struct packed_chars) == 2 ? 1 : -1];
+extern int o2[__alignof(struct packed_chars) == 1 ? 1 : -1];
+#endif