llvm dev - Jul 2008 - [LLVMdev] instcombine Question

I see instcombine doing something I'm not sure is right.

Incoming, I have this:

	%r29 = ptrtoint [71 x i64]* %"t$3" to i64		; <i64> [#uses=1]
	%r30 = inttoptr i64 %r29 to i64*		; <i64*> [#uses=1]
	store i64 72057594037927936, i64* %r30, align 8

Outgoing, I have this:

	%r30 = getelementptr [71 x i64]* %"t$3", i32 0, i32 0		; <i64*>
[#uses=1]
	store i64 72057594037927936, i64* %r30, align 16

I believe the alignment on the outgoing store is wrong, but my question
concerns the optimization of the ptrtoint / intoptr to a gep.  Is this 
optimization allowed to pun the pointer type like this or should there be
a bitcast to i64* in there somewhere?

                                                         -Dave

On Jul 16, 2008, at 1:32 PM, David Greene wrote:
> I see instcombine doing something I'm not sure is right.
>
> Incoming, I have this:
>
> 	%r29 = ptrtoint [71 x i64]* %"t$3" to i64		; <i64>
[#uses=1]
> 	%r30 = inttoptr i64 %r29 to i64*		; <i64*> [#uses=1]
> 	store i64 72057594037927936, i64* %r30, align 8
>
> Outgoing, I have this:
>
> 	%r30 = getelementptr [71 x i64]* %"t$3", i32 0, i32 0		;
<i64*>
> [#uses=1]
> 	store i64 72057594037927936, i64* %r30, align 16
>
> I believe the alignment on the outgoing store is wrong, but my  
> question
> concerns the optimization of the ptrtoint / intoptr to a gep.  Is this
> optimization allowed to pun the pointer type like this or should  
> there be
> a bitcast to i64* in there somewhere?
With what you've shown, the transformation is valid.  instcombine will  
increase the alignment when it knows it is safe.  Is %"t$3" derived  
from something provably 16-byte aligned?

-Chris

On Jul 16, 2008, at 1:32 PM, David Greene wrote:
> I see instcombine doing something I'm not sure is right.
>
> Incoming, I have this:
>
> 	%r29 = ptrtoint [71 x i64]* %"t$3" to i64		; <i64>
[#uses=1]
> 	%r30 = inttoptr i64 %r29 to i64*		; <i64*> [#uses=1]
> 	store i64 72057594037927936, i64* %r30, align 8
>
> Outgoing, I have this:
>
> 	%r30 = getelementptr [71 x i64]* %"t$3", i32 0, i32 0		;
<i64*>
> [#uses=1]
> 	store i64 72057594037927936, i64* %r30, align 16
>
> I believe the alignment on the outgoing store is wrong, but my  
> question
> concerns the optimization of the ptrtoint / intoptr to a gep.  Is this
> optimization allowed to pun the pointer type like this or should  
> there be
> a bitcast to i64* in there somewhere?
Implicit punning of pointer types in registers is not permitted. The  
above
getelementptr isn't doing that though. The first index of a GEP is the
one that just adds to the pointer value and doesn't change the type. The
second index goes into the pointee type -- the array -- so the result
is i64*.

Dan