[DEBUG_INFO][NVPTX] Generate DW_AT_address_class to get the values in debugger.
authorAlexey Bataev <a.bataev@hotmail.com>
Tue, 5 Feb 2019 19:33:47 +0000 (19:33 +0000)
committerAlexey Bataev <a.bataev@hotmail.com>
Tue, 5 Feb 2019 19:33:47 +0000 (19:33 +0000)
commitf3a9150324c4fe88b7f9e28be18e0f989dff8af7
treedc656043836629adf1ad8f924c64d810c9642314
parenta3f9b71c09613fa51df44a6253afc52e5fff2947
[DEBUG_INFO][NVPTX] Generate DW_AT_address_class to get the values in debugger.

Summary:
According to
https://docs.nvidia.com/cuda/archive/10.0/ptx-writers-guide-to-interoperability/index.html#cuda-specific-dwarf,
the compiler should emit the DW_AT_address_class attribute for all
variable and parameter. It means, that DW_AT_address_class attribute
should be used in the non-standard way to support compatibility with the
cuda-gdb debugger.
Clang is able to generate the information about the variable address
class. This information is emitted as the expression sequence
`DW_OP_constu <DWARF Address Space> DW_OP_swap DW_OP_xderef`. The patch
tries to find all such expressions and transform them into
`DW_AT_address_class <DWARF Address Space>` if target is NVPTX and the debugger is gdb.
If the expression is not found, then default values are used. For the
local variables <DWARF Address Space> is set to ADDR_local_space(6), for
the globals <DWARF Address Space> is set to ADDR_global_space(5). The
values are taken from the table in the same section 5.2. CUDA-Specific
DWARF Definitions.

Reviewers: echristo, probinson

Subscribers: jholewinski, aprantl, llvm-commits

Differential Revision: https://reviews.llvm.org/D57157

llvm-svn: 353203
llvm/include/llvm/IR/DebugInfoMetadata.h
llvm/lib/CodeGen/AsmPrinter/DwarfCompileUnit.cpp
llvm/lib/IR/DebugInfoMetadata.cpp
llvm/test/DebugInfo/NVPTX/dbg-declare-alloca.ll
llvm/test/DebugInfo/NVPTX/debug-addr-class.ll [new file with mode: 0644]