[LLVMdev] Crash with optimization for size

Somorjai, Akos ASomorjai at graphisoft.com
Fri Oct 21 00:30:47 PDT 2011


Thanks, Bob!

I guess we should be expecting a 4.2.1 update after clang 3.0 has been released, shouldn't we?

Best,

Akos

From: Bob Wilson <bob.wilson at apple.com<mailto:bob.wilson at apple.com>>
Date: Thu, 20 Oct 2011 08:46:38 -0700
To: Ákos Somorjai <asomorjai at graphisoft.com<mailto:asomorjai at graphisoft.com>>
Cc: "llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu>" <llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu>>
Subject: Re: [LLVMdev] Crash with optimization for size

This is  http://llvm.org/pr10514
Unfortunately the fix did not make it into that version of clang.

On Oct 20, 2011, at 7:47 AM, Somorjai, Akos wrote:

Here's a code generated with -Os on darwin/x86_64 with clang from the Xcode 4.2 GM toolset on Mac OSX 10.7.2 (Apple clang version 3.0 (tags/Apple/clang-211.10.1) (based on LLVM 3.0svn), Target: x86_64-apple-darwin11.2.0)

0x000000010277d281  <+2102>  lea    0x1d43bd0(%rip),%rax        # 0x1044c0e58 <gFloorPlanCutData>
0x000000010277d288  <+2109>  movaps 0x80(%rax),%xmm0
0x000000010277d28f  <+2116>  movaps %xmm0,-0x40(%rbp)


The documentation for 'movaps' states: "When the source or destination operand is a memory location, it must be aligned on a 16-byte boundary." The problem is that the address of the global variable loaded into AX is not necessarily aligned; as it isn't in our case, which leads to the crash.

Here's is the code; it is loading the address of a global variable into AX, and tries to copy the data member, coming from the following class:

MD5::FingerPrintcutDataCheckSum = ::gFloorPlanCutData.GetCutDataCheckSum ();
memcpy (elemData->cutDataCheckSum, cutDataCheckSum.data, sizeof (elemData->cutDataCheckSum));


namespace MD5 {
class GSROOT_DLL_EXPORT FingerPrint {
public:
FingerPrint ();
virtual ~FingerPrint () {};
FingerPrint (unsignedchar fp[16]);

void                Set (unsignedchar fp[16]);
virtualGSErrCode Write (GS::OChannel& oc) const;
virtualGSErrCode Read (GS::IChannel& ic);
virtualGSErrCode GetAsString (char* md5String) const;
boolIsEmpty (void) const;
voidClear (void);
bool                operator == (const FingerPrint&rhs) const;
bool                operator != (const FingerPrint&rhs) const;
FingerPrint&operator= (const FingerPrint& source);
public:
unsignedchar       data[16];
};
}

The type of gFloorPlanCutData is

class   FloorPlanCutData: public GS::Object,
  public GS::StateEventSource {

DECLARE_CLASS_INFO

private:
double  currCutLevel;
double  topCutLevel;
double  bottomCutLevel;
short   topCutBaseStoryRelativeIndex;
short   bottomCutBaseStoryRelativeIndex;

double  fixLevel2Absolute0;

Int32lock;

// temporary data dependent from the current story
doublecurrCutLevelCalculated;
doubletopCutLevelCalculated;
double  currentStoryLevel;
double  topBase2Current;
double  bottomBase2Current;
short   currentFloorIndex;
boolrelativeCalculated;
MD5::FingerPrintcutDataCheckSum;

...
}

The optimization kicks in because 'data' is 16-bytes; but there's no alignment check…

This seems to be a code generation error, but I couldn't reproduce it in a small application. Is this a known bug, or shall I adjust my build settings?

Thanks,

Ákos Somorjai

Developer Support Manager

GRAPHISOFT | Graphisoft Park 1. Budapest 1031 Hungary | +36 1 437-3000 | asomorjai at graphisoft.com<mailto:asomorjai at graphisoft.com>

_______________________________________________
LLVM Developers mailing list
LLVMdev at cs.uiuc.edu<mailto:LLVMdev at cs.uiuc.edu>         http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20111021/61d34246/attachment.html>


More information about the llvm-dev mailing list