[LLVMdev] Reminder: LLVM 3.0 Branching Friday!

[Andrew Trick]

On Oct 14, 2011, at 6:31 AM, Duncan Sands wrote:
>> • a miscompile in one of the llvm-gcc testers:
>> 
>> http://lab.llvm.org:8011/builders/llvm-x86_64-linux-checks/builds/55
>> 
>> Comparing stages 2 and 3
>> warning: ./cc1plus-checksum.o differs
>> Bootstrap comparison failure!
>> ./cp/pt.o differs
> 
> I bisected this to commit 139841 by Andrew Trick.  He said he would look at it.

I was able to reproduce the problem, but only during a parallel make of the entire llvm-gcc tree and not consistently. I have not been able to reproduce yet with a single build step, which is making this quite painful.

I did carefully inspect all of the differences between the .o files and proved to myself that they are functionally equivalent. It appears to be random codegen behavior. But I don't understand yet why self-hosting the front end could affect it.

Duncan, if you provide the differing .o files I can take a look and see if I notice a pattern. Better yet, provide an address-stripped asm diff in case my binutils are inadequate. We can move this discussion to PR11133.

You mentioned that you need a loaded system to reproduce it. I agree, but can't understand the relationship between system load and codegen. Addresses are randomized either way.

-Andy