[PATCH] D28390: [DWARF] LICM should null out the debug loc of hoisted loop invariant instructions

[Vedant Kumar via Phabricator via llvm-commits]

vsk added a comment.
Herald added a subscriber: asbirlea.
Herald added a project: LLVM.

As an alternative to removing the debug location, what about setting a line 0 location (with appropriate scope/inlinedAt info) to the hoisted instruction? It seems like that allows debuggers to give more helpful/specific backtraces. It also doesn't affect stepping (at least not in lldb, which collapses line 0 ranges). Example:

  int load(int *p) { return *p; }
  
  int licm(int seed, int n, int *p /* Points to garbage memory. */) {
    int hash = seed;
    for (int i = 0; i < n; ++i)
      hash ^= hash + (seed >> i) + load(p); // <- Crash occurs here.
    return hash;
  }

With no location (current behavior, the crash looks like it occurs somewhere in 'main'):

  (lldb) bt
  * thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0xdead)                                                          
    * frame #0: 0x0000000100000f73 licm`main at licm.cc:0:3 [opt]
      frame #1: 0x0000000100000f6c licm`main(argc=<unavailable>, argv=<unavailable>) at licm.cc:18 [opt] 

With line 0 location:

    * frame #0: 0x0000000100000f73 licm`main [inlined] load(p=<unavailable>) at licm.cc:0:3 [opt]
  ...
  (lldb) up
  frame #1: 0x0000000100000f73 licm`main at licm.cc:9 [opt]
     6    int licm(int seed, int n, int *p /* Points to garbage memory. */) {
     7      int hash = seed;
     8      for (int i = 0; i < n; ++i)
  -> 9        hash ^= hash + (seed >> i) + load(p); // <- Crash occurs here.

@danielcdh @wolfgangp -- Would switching to line 0 locations for hoisted instructions have an adverse effect on Sample PGO?


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D28390/new/

https://reviews.llvm.org/D28390