We're currently developing a PL/I compiler producing LLVM IR, and in particular debugging information which would eventually allow us to leverage LLDB as a PL/I debugger.
PL/I heavily relies on descriptors of various flavours to describe data at runtime, a principle which is somewhat akin to dope vectors in Fortran. The DWARF specification gives an example of how to describe such a descriptor, and it basically relies on the DW_OP_push_object_address operator within DWARF expressions, and we believe that it would be a promising way to implement what we need.
Currently, the LLVM IR does not support this operator within DIExpressions, and we would thus be interested in adding it (with corresponding support in LLDB further down the line).
Would a patch adding the operator into the DIExpression language, but only supporting DWARF generation at this time and not e.g. CodeView, be accepted in principle? What else would the patch have to provide?
I'm not at all familiar with CodeView, so I can't even be sure the operator would map onto it, and I assume that we want the DIExpression subset to be more or less target independent?
Thanks for any pointers,
OpenVMS uses descriptors for many languages (Ada, Fortran, Pascal, PL/I,
BASIC, etc.) on our Itanium platform (uses DWARF but our own code
generator). For descriptors, we produced compiler-generated thunks that
know how to extract/decode the various descriptor formats and use thunks
and DW_OP_CALL. For languages like Pascal or PL/I, such thunk routines
sometimes even have to perform uplevel references to compute the
requested info (length, bounds, address, etc.). Doing those as
DIExpressions can get complicated (I tried that way back in the VAX/VMS
days with our proprietary debug info but it had exactly the same
concepts as DIExpressions and thunks). I would often end up with debug
stack machine routines that were 100s of operations.
We also created a descriptor tag (DW_TAG_descriptor) to tell our
debugger (not LLDB) about the "by descriptor" usage. However, you then
get into the business of having to document the descriptor formats
(which we do on OpenVMS)
With all that said, I think your use of push_object_address is
reasonable and avoids pushing descriptor knowledge into LLDB.
We're about to start doing extending of our own for COBOL, BASIC, etc.