[LLVMdev] RFC: llvm-convert.cpp Patch

[Duncan Sands]

Hi,

> > I don't get it Bill.  llvm should be making memcpy with the right
> > alignment regardless of how Bar is formed.  Changing the alignment of
> > bar papers over this problem, but we will still get improper alignment
> > for other cases.  Also, DECL_USER_ALIGN isn't appropriate to tweak
> > here... the user isn't playing around with __attribute__((aligned))
>
> the fundamental problem is that the DestLoc argument to Emit doesn't
> come with alignment (or volatility for that matter).  When emitting
> an aggregate into DestLoc only the alignment of the source is used,
> not that of DestLoc.  But that's wrong if DestLoc is less aligned
> than the source (either because the source is highly aligned, eg
> because the user marked it as being highly aligned, or because the
> destination is little aligned as in the testcase).  So why not pass
> the alignment and volatility in too?  Probably Chris doesn't want
> to add extra arguments to Emit, so instead we could define a
> "DestLoc descriptor" struct:
> 	struct DLD {
> 		Value *Ptr;
> 		unsigned Alignment;
> 		bool Volatile;
> 	}
> declare one on the stack and pass its address, something like this:
> 	DLD D = { Ptr, Align, isVolatile };
> 	Emit(exp, &D);

I just did this in a sudden fit of activity.  It was quite easy and
looks like it catches a ton of volatility and alignment bugs.  Maybe
it will even work and fix the original problem :)  I will let you
know later.

Ciao,

Duncan.