[llvm-dev] DEBUG INFO: improve handling of DBG_VALUEs and DebugLocs in CodeGen

[Jonas Paulsson via llvm-dev]

Hi,                                   (Resent with proper subject line)

I recently found myself in trouble because the crash I had disappeared
with -g, so I could not debug the program. This happened because the
optimizer did not remember to consider DBG_VALUEs instruction so it
changed its behavior, and the bug went hiding.

I then started discussing this onhttps://reviews.llvm.org/D45878,  and
since this is something that should be handled by all different machine
passes, the immediate question is now what utility functions should be
made available for common use.

1. A pass such as MachineSink.cpp must first of all remember to consider
DBG_VALUEs when iterating over instructions, so that it does not e.g.
"break if next instruction is not X" and therefore change its results
with -g.

2. At a second priority,

  a) it should handle DBG_VALUEs when moving / erasing MachineInstrs.

b) it should " // Merge or erase debug location to ensure
consistent stepping in profilers and debuggers." (MachineSink.cpp).

Well, this is in fact the first point here: I think this should be
documented somewhere by a comment so that it is clear exactly how these
things should be handled. My personal understanding right now is that 2
is not vital but should be done on a best-effort basis. 1 however is
really bad if it happens (which it currently does).

We are discussing where to place these utility functions, and if
splice() should optionally be made to do this also. It seems that
collectDebugValues() in MachineSink.cpp is a good starting point, but we
probably want to do an even better search than just looking at the first
def operand.

It would be nice to get some general feedback from the community at this
point so we know which direction to take...

/Jonas