clang breaks strict-aliasing rules according to g++4.4

[edwintorok]

Bugzilla Link	4061
Resolution	WONTFIX
Resolved on	Jun 13, 2012 13:24
Version	unspecified
OS	Linux
CC	@akyrtzi,@lattner,@DougGregor,@tlattner

Extended Description

/home/edwin/llvm-svn/llvm/include/llvm/ADT/APInt.h: In member function ‘bool clang::Expr::isIntegerConstantExpr(llvm::APSInt&, clang::ASTContext&, clang::SourceLocation*, bool) const’:
/home/edwin/llvm-svn/llvm/include/llvm/ADT/APInt.h:99: warning: dereferencing pointer ‘’ does break strict-aliasing rules
/home/edwin/llvm-svn/llvm/include/llvm/ADT/APInt.h:548: warning: dereferencing pointer ‘’ does break strict-aliasing rules
/home/edwin/llvm-svn/llvm/include/llvm/ADT/APInt.h:549: warning: dereferencing pointer ‘’ does break strict-aliasing rules
/home/edwin/llvm-svn/llvm/include/llvm/ADT/APSInt.h:38: warning: dereferencing pointer ‘’ does break strict-aliasing rules
Expr.cpp:1334: note: initialized from here

The location info for the warning is not accurate (it doesn't show the very line causing the violation), but the aliasing violation is real:
return (APSInt)(void*)Data;

See URL for a possible solution.

[llvmbot]

A few more (there are tons of warnings, most of them duplicates; I just
grabbed two randomly).

clang/lib/Parse/DeclSpec.cpp:14:
clang/lib/Parse/../../include/clang/Parse/DeclSpec.h: In member function ‘clang::CXXScopeSpec& clang::DeclaratorChunk::MemberPointerTypeInfo::Scope()’:
clang/lib/Parse/../../include/clang/Parse/DeclSpec.h:605: warning: dereferencing type-punned pointer will break strict-aliasing rules

clang/lib/Parse/../../include/clang/Parse/DeclSpec.h: In member function ‘void clang::Parser::ParseDeclaratorInternal(clang::Declarator&, void (clang::Parser::*)(clang::Declarator&))’:
clang/lib/Parse/../../include/clang/Parse/DeclSpec.h:724: warning: dereferencing pointer ‘’ does break strict-aliasing rules
clang/lib/Parse/../../include/clang/Parse/DeclSpec.h:724: note: initialized from here

LLVM itself is clean.

[llvmbot]

You need to build with optimization (--enable-optimized) to see these warnings.

[DougGregor]

More issues:

type-punning will break strict-aliasing rules" warnings at clang/AST/Decl.h:308: return reinterpret_cast<Stmt **>(&Init);

[llvmbot]

there's also

/tmp/llvm/tools/clang/lib/Index/../../include/clang/AST/Decl.h: In member function 'clang::Stmt** clang::VarDecl::getInitAddress()':
/tmp/llvm/tools/clang/lib/Index/../../include/clang/AST/Decl.h:371: warning: dereferencing type-punned pointer will break strict-aliasing rules

[llvmbot]

This is impacting 2.6 as well. I get metric buttloads of these Decl.h:366 warnings while building clang 2.6 edition -- but with g++ 4.2.

[DougGregor]

Should be fixed here:

http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20090831/020877.html

[llvmbot]

Reopening because gcc-4.4 still reports a gazillion issues. Here are the
first two:

clang/lib/Basic/../../include/clang/Basic/SourceLocation.h: In member function ‘bool clang::SourceManager::isBeforeInTranslationUnit(clang::SourceLocation, clang::SourceLocation) const’:
clang/lib/Basic/../../include/clang/Basic/SourceLocation.h:43: warning: dereferencing pointer ‘’ does break strict-aliasing rules
/usr/include/c++/4.4/bits/stl_tree.h:175: note: initialized from here
...
clang/lib/Parse/../../include/clang/Parse/DeclSpec.h: In member function ‘clang::CXXScopeSpec& clang::DeclaratorChunk::MemberPointerTypeInfo::Scope()’:
clang/lib/Parse/../../include/clang/Parse/DeclSpec.h:637: warning: dereferencing type-punned pointer will break strict-aliasing rules

[lattner]

Is this on mainline or 2.6?

[llvmbot]

Both mainline and 2.6.

[DougGregor]

Yet another fix here:

http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20090907/021109.html

However, GCC 4.4 still emits a huge number of warnings, and I'm fairly certain that they are wrong: we are only accessing this data through a char and through the dynamic type of the object, both of which are allowed by C++ 3.10p15. For GCC 4.4., we should consider adding -fno-strict-aliasing.

[tlattner]

Is anyone running a buildbot with gcc 4.4? If not, I do not think this should be a showstopper for 2.6. Any objections to removing this as a blocking bug, speak now.

[lattner]

I think that makes sense, a few extra warnings are not the end of hte world for 2.6, we can fix the rest on mainline.

[llvmbot]

It would be nice to knock these out though, it looks kind of bad...

[tlattner]

Nice to fix, but this should not be a showstopper for 2.6. We have to draw the line somewhere, and I'd much rather fix the bugs in 2.6 than the warnings. If this gets fixed before 2.6, then great, but not going to block it.

[llvmbot]

Those about DeclaratorChunk::MemberPointerTypeInfo::Scope() are definitely false positives. Not only does C++ allow char arrays to alias with everything, we're never accessing this memory location in any way except through a CXXScopeSpec pointer, except for the struct's "constructor" DeclaratorChunk::getMemberPointer, which constructs the CXXScopeSpec into the char array.