Re: Aliasing problem? gcc-3.3 miscompiles mozilla-1.4

["Kevin B. Hendricks" <kevin dot hendricks at sympatico dot ca>]

Hi Franz,

FWIW:

The your original ii file and my little main added to the end when compiled 
with -O2 -fno-exceptions does in fact show the segfault on Jack's machine.

Just in case this helps,

Kevin




On Thursday 19 June 2003 06:07, Franz Sirl wrote:
> On Wednesday 18 June 2003 20:48, Franz Sirl wrote:
> > Hi,
> >
> > this little routine is miscompiled with -O2 with gcc-3.3 on ppc-linux:
> >
> > nsresult imgContainerGIF::DecodingComplete(void)
> > {
> >   mDoneDecoding = 1;
> >
> >
> >   PRUint32 numFrames = inlinedGetNumFrames();
> >   if (numFrames == 1) {
> >     nsCOMPtr<gfxIImageFrame> currentFrame;
> >     inlinedGetFrameAt(0, getter_AddRefs(currentFrame));
> >     currentFrame->SetMutable(0);
> >   }
> >   return 0;
> > }
> >
> > This is the assembly:
> >
> > _ZN15imgContainerGIF16DecodingCompleteEv:
> >         stwu %r1,-48(%r1)
> >         mflr %r0
> >         stw %r0,52(%r1)
> >         li %r0,1
> >         stw %r0,104(%r3)         #  <variable>.mDoneDecoding
> >         stw %r31,44(%r1)
> >         addi %r31,%r3,16         #  this,  this
> >         lwz %r0,16(%r31)         #  <variable>.mCount
> >         stw %r30,40(%r1)
> >         addi %r30,%r1,16         #  this
> >         cmpwi %cr0,%r0,1
> >         stw %r29,36(%r1)
> >         mr %r3,%r30      #  this
> >         beq- %cr0,.L202
> > .L188:
> >         lwz %r0,52(%r1)
> >         li %r3,0
> >         lwz %r29,36(%r1)
> >         lwz %r30,40(%r1)
> >         mtlr %r0
> >         lwz %r31,44(%r1)
> >         addi %r1,%r1,48
> >         blr
> > .L202:
> >         li %r0,0         #  rawPtr
> >         stw %r0,16(%r1)  #  <variable>.mRawPtr,  rawPtr
> >         bl _ZN13nsCOMPtr_base16begin_assignmentEv@plt
> >         mr %r29,%r3      #  <anonymous>
> >         li %r4,0
> >         mr %r3,%r31      #  this
> >         bl _ZN15nsSupportsArray9ElementAtEj@plt
> >         lwz %r9,16(%r1)  #  <variable>.mRawPtr,  <anonymous> XXXXX
> >         cmpwi %cr0,%r3,0         #  _elem
> >         mr %r0,%r3       #  _elem
> >         li %r4,0
> >         mr %r3,%r9       #  <anonymous>
> >         beq- %cr0,.L196
> >         stw %r0,0(%r29)  # * <anonymous>,  _elem XXXXX
> > .L196:
> >         lwz %r9,0(%r9)   #  <variable>._vptr.nsISupports
> >         lwz %r9,20(%r9)
> >         mtctr %r9
> >         bctrl
> >         mr %r3,%r30      #  this
> >         bl _ZN13nsCOMPtr_baseD2Ev@plt
> >         b .L188
> >
> >
> > The problem are the marked lines, they are swapped ( 0(%r29) and 16(%r1)
> > point to the same stack slot, r29 is the return value of
> > _ZN13nsCOMPtr_base16begin_assignmentEv). Compiling with -O1 or
> > -fno-strict-aliasing lets the code work.
> >
> > I've attached the preprocessed source of the miscompiled file.
>
> I just found out that in addition to -O2, -fno-exceptions is a prerequisite
> to show the bug.
>
> Franz.