[PATCH] Branch folding causes different code generation at "-O2 -g" and "-O2"

[Katya Romanova]

  Hi Chandler,

  I have posted my previous comment and didn't hear back from you for a while. Sorry to ping you again. I think I have shortened the metadata as much as I could. Out of 91 metadata entries that I had in the original patch, I have only 7 metadata entries now.

  I wasn't able to shorten the testcase more by replacing
  !<n> = metadata !{blah, blah, blah, ..., blah} ;  with
  !<n> = metadata !{null} ;
  because either the test stops to reproduce the original bug or it asserts. If you meant to shorten metadata differently, please give me a specific example of what you had in mind or point me to the testcase that is doing a similar thing.

  Thanks!
  Katya.

  P.S. Out of curiosity, I checked how many metadata entries other debug-info related tests in llvm/etst/CodeGen/X86 have. This amount ranged from 12 to 72 entries. Is there a specific reason why having only 7 metadata entries in my test is not acceptable?

http://llvm-reviews.chandlerc.com/D2970