[PATCH] Fixing a heisenbug where the memory dependence analysis behaves differently with and without -g

Yunzhong Gao Yunzhong_Gao at playstation.sony.com
Tue Nov 12 16:33:19 PST 2013

  From: David Blaikie [mailto:dblaikie at gmail.com]
  Sent: Tuesday, November 12, 2013 2:26 PM

  > Are there existing test cases that ensure the threshold is respected? If you massively increase or
  > decrease the threshold, do any tests fail? If not, then it seems we're missing test coverage. We
  > should have some tests that check that, and then we could have another variant that has a trash
  > llvm.dbg.value intrinsic designed to push the block over the limit and ensure that we get the same
  > behavior as if the intrinsic is not present.
  > - David

  No existing test fails if I increase the threshold from 100 to 5000 or decrease it to 15, so it seems
  that we are missing test cases. Err I do not quite have an idea how to test the threshold directly.
  Who should I ask? The CODE_OWNERS.txt does not say who owns the analysis passes. I am
  going to cc Owen, since he seemed to have added this pass in the beginning, and Bill, since he
  changed the threshold most recently. Sorry for the spam.


More information about the llvm-commits mailing list