[llvm-dev] Debug info for Cuda

[Robinson, Paul via llvm-dev]

> Hi everybody,
> As you know, Cuda/NVPTX target has very limited support of the debug
> info in Clang/LLVM. Currently, LLVM supports only emission of the line
> numbers debug info.
> This is caused by limitations of the Cuda/NVPTX codegen. Clang/LLVM
> translates the source code to LLVM IR, which is then lowered to PTX
> (parallel thread execution) intermediate file. This PTX file represents
> special kind of the assembler code in text format, which contains the
> code itself + (possibly) debug info. Then this PTX file is compiled by
> ptxas tool into the CUDA binary representation.
>
> Debug info representation in PTX file.
> ========================
> According to PTX Writer's Guide to Interoperability, Debug information
> (http://docs.nvidia.com/cuda/ptx-writers-guide-to-interoperability/index.html#debug-information)
> , debug information must be encoded in DWARF (Debug With Arbitrary
> Record Format). The responsibility for generating debug information is
> split between the PTX producer and the PTX-to-SASS backend. The PTX
> producer is responsible for emitting binary DWARF into the PTX file,
> using the .section and .b8-.b16-.b32-and-.b64 directives in PTX. This
> should contain the .debug_info and .debug_abbrev sections, and possibly
> optional sections .debug_pubnames and .debug_aranges. These sections
> are standard DWARF2 sections that refer to labels and registers in the
> PTX.
>
> The PTX-to-SASS backend is responsible for generating the .debug_line
> section from the .file and .loc directives in the PTX file. This
> section maps source lines to SASS addresses. The PTX-to-SASS backend
> also generates the .debug_frame section.

All this sounds like the standard division of responsibilities between
an LLVM code generator and the assembler.

> LLVM is able to emit debug info in DWARF. But ptxas compiler has some
> limitations, that make it hard to adapt LLVM for correct emission of
> the debug info in PTX files.
>
> Limitations/features of the PTX format/ptxas compiler.
> ==================================
> a) Supports DWARF-2 only.

IIRC, Darwin had a similar restriction until recently.

> b) Labels are allowed only in code section (only in functions).

If you have static/global variables, I guess their locations would
have to be described using a section+offset expression?  Normally
we emit a location attribute that is just a reference to a label
for the variable.

> c) Does not support label arithmetic in DWARF sections.
>     “.b32 L1 – L2” as the size of the section is not allowed, so the
> sections sizes should be calculated explicitly.

MachO has a similar restriction, this should not be a problem if you
can do something like:
    L3 = L1 - L2
    .b32 L3

> d) Debug info must point to the sections, not to labels inside these
> sections.
>     “.b32 .debug_abbrevs”

Offhand for DWARF-2 I can't think of a reference that couldn't be done
this way.

> e) Sections itself must be enclosed into braces
>     “.section .debug_info {…}”
> f) Frame info is non-register based
>     Based on function local “__local_depot” array, that represents the
> stack frame.
> g) All variables must have non-standard DW_AT_address_class attribute
> so the debuger had the info about address class of the variable -
> global or local. DWARF standard does support this attribute, but it can
> be appiled to pointer/reference types only, not variables.

For variables it would be more usual to use DW_AT_segment for this.
But that's an agreement that the compiler and debugger need to reach.

> h) The first label in the function must follow the debug location macro.
> In LLVM, it is followed by the debug location macro.

I am not 100% sure what you mean by this, but I think it has to do with
the fact that LLVM attaches locations to instructions, not labels.  It
might or might not be easy to work around this; there might be an
unfortunate interaction with how emitting line-0 records works.

> i) .debug_frame section is emitted by txas compiler.
>     DW_AT_frame_base must be set to dwarf::DW_FORM_data1
> dwarf::DW_OP_call_frame_cfa value.

I doubt that's a problem.

> j) Strings cannot be referenced by the labels, instead they must be
> inlined in the sections in form of array of chars.

LLVM used to do inline strings, but switched to the .debug_str section
quite a while ago.  On the other hand, I spent a little time maybe a
year ago looking into whether we could emit short strings inline as a
space-saving measure, and decided it was feasible.  (I didn't do it 
because the space savings was really trivial.)  So I think doing this
would not be terribly hard.

> Some changes in LLVM are required to support all these
> limitation/features in the output PTX files.
> Required changes in LLVM.
> ==================
> •include/llvm/CodeGen/AsmPrinter.h.
>     •Add “virtual MCSymbol *getFunctionFrameSymbol(const
> MachineFunction *MF) const” for non-register-based frame info.
>     •Override “NVPTXMCAsmPrinter.cpp” to return the name of the
> “__local_depot” frame storage.
> •Add ”cuda-gdb” specific tuning.

Note that our design philosophy for "tuning" is that a tuning option
unpacks into other separate flags.  Not a problem, just an observation.

>     •Inlined strings must be used in sections, not string references.
>     •Label arithmetic is replaced by the absolute section size
> evaluation.

This one isn't a debug-info tuning decision, it's how your assembler works
and so is a target decision.

>     •Use “AsmPrinter::doInitialization()” instead of NVPTX-custom manual
> initialization.
>    •Local variables address emitted as “__local_depot” + <var offset>.
> •Add NVPTX specific “NVPTXMCAsmStreamer” class.
>     •Requires moving to includes of “MCAsmStreamer” class declaration.
>     •Overrides emission of the labels (names of the section are emitted
> instead).
>     •Overrides emission of the sections (emit braces)
>     •Overrides string emission (as sequence of bytes, not as strings)
>     •Overrides emission of files/locations debug info
> Required changes in Clang.
> =================
> •Add option “-gcuda-gdb” to driver.
>     •Emit cuda-gdb compatible debug info (DWARF-2 by default + CudaGDB
> tuning).
> •Add options “-g --dont-merge-basicblocks --return-at-end” to “ptxas”
> call.
>     •ptxas is able to translate debug information only if -O0
> optimization level is used. It means, that we can use optimization
> level in LLVM > O0, but still have to use O0 when calling ptxas
> compiler.
>
> This approach was implemented in https://github.com/clang-ykt to support
> debug info emission for NVPTX target when generating code for OpenMP
> offloading constructs. You can try to use it.

I haven't looked at your code but all the things you describe seem
reasonably feasible.  Certainly the details of what you want to do
to the emitted DWARF are fine; I am less sure about the assembler
details, but if you have a worked example that makes it likely that
part is okay as well.
--paulr