llvm-debuginfo-analyzer

These are the notes collected during the development, review and test. They describe limitations, known issues and future work.

Remove the use of macros in

LVReader.h

that describe the

bumpallocators

.

D137933

Use a standard (or LLVM)

with

typeinfo

(would need a specialization to expose equality and hasher) for the allocators and the creation functions could be a function template.

Use a lit test instead of a unit test for the logical readers.

D125783

As the

library is sufficiently exposed via the

llvm-debuginfo-analyzer

tool, follow the LLVM general approach and use

lit

tests to validate the logical readers.

Convert the

:

llvm-project/llvm/unittests/DebugInfo/LogicalView/CodeViewReaderTest.cpp
llvm-project/llvm/unittests/DebugInfo/LogicalView/DWARFReaderTest.cpp

into

tests:

llvm-project/llvm/test/DebugInfo/LogicalView/CodeViewReader.test
llvm-project/llvm/test/DebugInfo/LogicalView/DWARFReader.test

Eliminate calls to

getInputFileDirectory()

in the

unittests

.

D125783

Rewrite the unittests

and

CodeViewReaderTest

to eliminate the call:

  getInputFileDirectory()

as use of that call is discouraged.

Fix mismatch between

%d/%x

format strings and

uint64_t

type.

D137400 / 58758

Incorrect printing of

on

32-bit

platforms. Add the

PRIx64

specifier to the printing code (

format()

).

Remove

LVScope::Children

container.

D137933

Use a chaining iterator over the other containers rather than keep a separate container

that mirrors their contents.

Use

TableGen

for command line options.

D125777

The current trend is to use

for command-line options in tools. Change command line options to use

tablegen

as many other LLVM tools.

LVDoubleMap

to return

optional<ValueType>

instead of

null pointer

.

D125783

The more idiomatic LLVM way to handle this would be to have

return

Optional<ValueType>

.

Pass references instead of pointers (Comparison functions).

D125782

In the comparison functions, pass references instead of pointers (when pointers cannot be null).

Use

StringMap

where possible.

D125783

LLVM has a

class that is advertised as more efficient than

std::map<std::string, ValueType>

. Mainly it does fewer allocations because the key is not a

std::string

.

Replace the use of

with

StringMap

. One specific case is the

LVSymbolNames

definitions.

Calculate unique offset for CodeView elements.

In order to have the same logical functionality as the DWARF reader, such as:

• find scopes contribution to debug info
• sort by its physical location

The logical elements must have an unique offset (similar like the DWARF

offset).

Move

initializeFileAndStringTables

to the CodeView Library.

There is some code in the CodeView reader that was extracted/adapted from

that can be moved to the CodeView library.

We had a similar case with code shared with

that was moved to the PDB library: D122226

Move

getSymbolKindName

and

formatRegisterId

to the CodeView Library.

There is some code in the CodeView reader that was extracted/adapted from

that can be used.

Use of

std::unordered_set

instead of

std::set

.

D125784

Replace the

usage for

DeducedScopes

,

UnresolvedScopes

and

IdentifiedNamespaces

with

std::unordered_set

and get the benefit of the O(1) while inserting/searching, as the order is not important.

Optimize

LVNamespaceDeduction::find

funtion.

D125784

Optimize the

method to use the proposed code:

  LVStringRefs::iterator Iter = std::find_if(Components.begin(), Components.end(),
    [](StringRef Name) {
        return IdentifiedNamespaces.find(Name) == IdentifiedNamespaces.end();
    });
  LVStringRefs::size_type FirstNonNamespace = std::distance(Components.begin(), Iter);

Move all the printing support to a common module.

Factor out printing functionality from the logical elements into a common module.

Refactor

LVBinaryReader::processLines

.

D125783 / D137156

During the traversal of the debug information sections, we created the logical lines representing the disassembled instructions from the text section and the logical lines representing the line records from the debug line section. Using the ranges associated with the logical scopes, we will allocate those logical lines to their logical scopes.

Consider the case when any of those lines become orphans, causing incorrect scope parent for disassembly or line records.

Add support for

-ffunction-sections

.

D125783

Only linked executables are handled. It does not support relocatable files compiled with

.

Add support for DWARF v5

.debug_names

section / CodeView public symbols stream.

D125783

The DWARF and CodeView readers use the public names information to create the instructions (

). Instead of relying on DWARF section names (

.debug_pubnames

,

.debug_names

) and CodeView public symbol stream (

S_PUB32

), the readers should collect the needed information while processing the debug information.

If the object file supports the above section names and stream, use them to create the public names.

Add support for some extra DWARF locations.

The following DWARF debug location operands are not supported:

• DW_OP_const_type
• DW_OP_entry_value
• DW_OP_implicit_value

Add support for additional binary formats.

• Extended COFF (
  XCOFF
  )

Add support for

JSON

or

YAML

The logical view uses its own and non-standard free form text when displaying information on logical elements.