[LLVMdev] Use list preservation when using Instruction::clone

[Tyler Denniston]

Ok, it took me a little while to isolate this into a test case, but I
have one, in the attached tarball. Please run the attached pass on the
attached bitcode (LLVM 3.4).

The bug I'm seeing is summarized as follows:

1. Save a list of all of the function F's instructions.
2. Create a clone of the function F.
3. JIT and execute the cloned function.
4. Assert that the body of F has not changed.

The bug is the assertion in (4) failing. If you run the test case, you
should see a crash with this message:

Assertion failed: (false), function validateFunctionBody, file  
Test.cpp, line 50.

The reason I started this thread asking about use lists is I think
this has to do with instructions in the cloned function listing
instructions in the original function as Uses, and thus when the JIT
erases some instructions in the cloned function, the originals are
deleted as dead code. That's just a hypothesis, anyway, I'm not sure
it's correct.

I don't think this seems like intended behavior, but perhaps with this
additional information about the behavior I'm seeing, someone can shed
some light on this for me.

Tyler

Quoting Jingyue Wu <jingyue at google.com>:

> Hi Tyler,
>
> Doesn't look right. Would you mind attaching the test?
>
> Btw, instead of creating an empty function and CloneFunctionInto, can you
> just CloneFunction?
>
> Jingyue
>
> On Thu Oct 02 2014 at 2:47:51 PM Tyler Denniston <tyler at csail.mit.edu>
> wrote:
>
>> I'm trying to create a clone of a function using Function::Create()
>> and CloneFunctionInto. However, I'm running into an issue. I believe
>> that the instructions in the function clone still have Use edges to
>> values in the original function. This is a problem for my purposes.
>>
>> For example, consider an original function F. I create a new function
>> G belonging to the same module and call CloneFunctionInto(G, F) to
>> copy over the function body. Now, consider a phi node p in F, and an
>> instruction I in F which uses p. The problem is, the cloned phi node
>> p' in G still lists I as a user.
>>
>> Is this intended behavior? And if so, is there a way I can remove
>> these Use edges in my cloned function?
>>
>> Thanks,
>>
>> Tyler
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: bug.tar.gz
Type: application/x-gzip
Size: 6656 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20141002/12ba25e4/attachment.bin>