llvm-project/lldb/www/projects.html

505 lines
41 KiB
HTML
Executable File
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1" />
<link href="style.css" rel="stylesheet" type="text/css" />
<title>LLDB Projects</title>
</head>
<body>
<div class="www_title">
Project ideas for the <strong>LLDB</strong> Debugger
</div>
<div id="container">
<div id="content">
<!--#include virtual="sidebar.incl"-->
<div id="middle">
<div class="post">
<h1 class ="postheader">Projects</h1>
<div class="postcontent">
<p>The following is a mostly unordered set of the ideas for improvements
to the LLDB debugger. Some are fairly deep, some would require less
effort.
</p>
<ol>
<li>
Speed up type realization in lldb.  
<p>
The type of problem I'm addressing here is the situation where you are
debugging a large program (lldb built with debug clang/swift will do)
and you go to print a simple expression, and lldb goes away for 30 seconds.
When you sample it, it is always busily churning through all the CU's in the
world looking for something.  The problem isn't that looking for something in
particular is slow, but rather that we somehow turned an bounded search
(maybe a subtype of "std::string" into an unbounded search (all things with the
name of that subtype.)  Or didn't stop when we got a reasonable answer
proximate to the context of the search, but let the search leak out globally.
And quite likely there are other issues that I haven't guessed yet.
But if you end up churning though 3 or 4 Gig of debug info, that's going to be slow
no matter how well written your debug reader is...
</p>
<p>
My guess is the work will be more in the general symbol lookup than in the DWARF
parser in particular, but it may be a combination of both.
</p>
<p>
As a user debugging a largish program, this is the most obvious lameness of lldb.
</p>
</li>
<li>
Symbol name completion in the expression parser.
<p>
This is the other obvious lameness of lldb.  You can do:
</p>
<code><pre><tt>
(lldb) frame var foo.b<TAB>
</tt></pre></code>
<p>
and we will tell you it is "foo.bar". But you can't do that in the expression parser.
This will require collaboration with the clang/swift folks to get the right extension
points in the compiler. And whatever they are, lldb will need use them to tell the
compiler about what names are available. It will be important to avoid the pitfalls
of #1 where we wander into the entire DWARF world.
</p>
</li>
<li>
Make a high speed asynchronous communication channel to replace the gdb-remote protocol.
<p>
All lldb debugging nowadays is done by talking to a debug agent. We used the
gdb-remote protocol because that is universal, and good enough, and you have
to support it anyway since so many little devices & JTAG's and VM's etc
support it. But it is really old, not terribly high performance, and
can't really handle sending or receiving messages while the process is
supposedly running. It should have compression built in, remove the hand-built
checksums and rely on the robust communication protocols we always have nowadays,
allow for out-of-order requests/replies, allow for reconnecting to a temporarily
disconnected debug session, regularize all of the packet formatting into JSON or
BSON or whatever while including a way to do large binary transfers. It must be
possible to come up with something faster, and better tunable for the many
communications pathways we end up supporting.
</p>
</li>
<li>
Fix local variable lookup in the lldb expression parser.
<p>
The injection of local variables into the clang expression parser is
currently done incorrectly - it happens too late in the lookup. This results
in namespace variables & functions, same named types and ivars shadowing
locals when it should be the other way around. An attempt was made to fix
this by manually inserting all the visible local variables into wrapper function
in the expression text. This mostly gets the job done but that method
means you have to realize all the types
and locations of all local variables for even the simplest of expressions, and
when run on large programs (e.g. lldb) it would cause unacceptable delays. And
it was very fragile since an error in realizing any of the locals would cause
all expressions run in that context to fail. We need to fix this by adjusting
the points where name lookup calls out to lldb in clang.
</p>
</li>
<li>
Fix the event handling/process control machinery to support calling SB & Commands
everywhere, and to support non-stop debugging
<p>
There is a fairly ad-hoc system to handle when it is safe to run SB API's and
command line commands. This is actually a bit of a tricky problem, since we
allow access to the command line and SB API from some funky places in lldb.
The Operating System plugins are the most obvious instance, since they get
run right after lldb is told by debugserver that the process has stopped, but
before it has finished collating the information from the stop for presentation
to the higher levels. But breakpoint callbacks have some of the same problems,
and other things like the scripted stepping operations and any fancier extension
points we want to add to the debugger are going to be hard to implement robustly
till we work on a finer-grained and more explicit control over who gets to control
the process state.
<p>
We also won't have any chance of supporting non-stop debugging - which is a useful
mode for programs that have a lot of high-priority or real-time worker threads -
until we get this sorted out.
</p>
</li>
<li>
Finish the language abstraction and remove all the unnecessary C++/clang-based API's
<p>
An important part of making lldb a more useful "debugger toolkit" as opposed to a
C/C++/ObjC/Swift debugger is to have a clean abstraction for language support.
We did most, but not all, of the physical separation.  We need to finish that.
And then by force of necessity the API's really look like the interface to a C++
type system with a few swift bits added on.  How you would go about adding a new
language is unclear and much more trouble than it is worth at present. But if
we made this nice, we could add a lot of value to other language projects.
</p>
</li>
<li>
Add some syntax to generate data formatters from type definitions
<p>
Uses of the data formatters fall into two types. There are data formatters for
types where the structure elements pretty much tell you how to present the data,
you just need a little expression language to express how to turn them into what
the user expects to see. Then there are the ones (like pretty much all our
Foundation/AppKit/UIKit formatters) that use deep magic to figure out how the type
is actually laid out. The latter are pretty much always going to have to be done by hand.
</p>
<p>
But for the ones where the information is expressed in the fields, it would be great
to have a way to express the instructions to produce summaries and children in some
form you could embed next to the types and have the compiler produce a byte code
form of the instructions and then make that available to lldb along with the library.
This isn't as simple as having clang run over the headers and produce something from the
types directly. After all, clang has no way of knowing that the interesting thing about
a std::vector is the elements that you get by calling size (for the summary) and [] for
the elements. But it shouldn't be hard to come up with a generic markup to express this.
</p>
</li>
<li>
Allow the expression parser to access dynamic type/data formatter information
<p>
This seems like a smaller one. The symptom is your object is Foo child of Bar, and in
the Locals view you see all the fields of Foo, but because the static type of the object
is Bar, you can't see any of the fields of Foo. But if you could get this working,
you could hijack the mechanism to make the results of the value object summaries/synthetic
children available to expressions. And if you can do that, you could add other properties
to an object externally (through Python or some other extension point) and then have
these also available in the expression parser. You could use this to express invariants
for data structures, or other more advanced uses of types in the debugger.
</p>
<p>
Another version of this is to allow access to synthetic children in the expression
parser. Otherwise you end up in situations like:
</p>
<code><pre><tt>
(lldb) print return_a_foo()
(SomeVectorLikeType) $1 = {
[0] = 0
[1] = 1
[2] = 2
[3] = 3
[4] = 4
}
</tt></pre></code>
<p>
That's good but:
</p>
<code><pre><tt>
(lldb) print return_a_foo()[2]
</tt></pre></code>
<p>
fails because the expression parser doesn't know anything about the array-like nature of
SomeVectorLikeType that it gets from the synthetic children.
</p>
</li>
<li>
Recover thread information lazily
<p>
LLDB stores all the user intentions for a thread in the ThreadPlans stored in
the Thread class. That allows us to reliably implement a very natural model for
users moving through a debug session. For example, if step-over stops at a breakpoint
in an function in a younger region of the stack, continue will complete the step-over
rather than having to manually step out. But that means that it is important that
the Thread objects live as long as the Threads they represent. For programs with many
threads, but only one that you are debugging, that makes stepping less efficient,
since now you have to fetch the thread list on every step or stepping doesn't work
correctly. This is especially an issue when the threads are provided by an
Operating System plugin, where it may take non-trivial work to reconstruct the
thread list. It would be better to fetch threads lazily but keep "unseen" threads in
a holding area, and only retire them when we know we've fetched the whole thread list
and ensured they are no longer alive.
</p>
</li>
<li>
Add an extension point in the breakpoint search machinery.
<p>
This would allow highly customizable, algorithmic breakpoint types, like "break on every
use of some particular instruction, or instruction pattern, etc."
</p>
</li>
<li>
Make Python-backed commands first class citizens
<p>
As it stands, Python commands have no way to advertise their options. They are
required to parse their arguments by hand. That leads to inconsistency, and more
importantly means they can't take advantage of auto-generated help and command
completion. This leaves python-backed commands feeling worse than built-in ones.
</p>
<p>
As part of this job, it would also be great to hook automatically hook the "type" of an option value
or argument (e.g. eArgTypeShlibName) to sensible default completers. You need to be able to
over-ride this in more complicated scenarios (like in "break set" where the presence of
a "-s" option limits the search for completion of a "-n" option.) But in common cases it is
unnecessary busy-work to have to supply the completer AND the type. If this worked, then
it would be easier for Python commands to also get correct completers.
</p>
</li>
<li>
Reimplement the command interpreter commands using the SB API
<p>
Currently, all the CommandObject::DoExecute methods are implemented
using the lldb_private API's. That generally means that there's code
that gets duplicated between the CommandObject and the SB API that does
roughly the same thing. We would reduce this code duplication, present a
single coherent face to the users of lldb, and keep
ourselves more honest about what we need in the SB API's if we implemented
the CommandObjects::DoExecute methods using the SB API's.
</p>
<p>
BTW, it is only the way it was much easier to develop lldb if it had a functioning
command-line early on. So we did that first, and developed the SB API's when lldb
was more mature. There's no good technical reason to have the commands use the
lldb_private API's.
</p>
</li>
<li>
Documentation and better examples
<p>
We need to put the lldb syntax docs in the tutorial somewhere that is more easily accessible.
On suggestion is to add non-command based help to the help system, and then have a "help lldb"
or "help syntax" type command with this info. Be nice if the non-command based help could
be hierarchical so you could make topics.
</p>
<p>
There's a fair bit of docs about the SB API's, but it is spotty. Some classes are
well documented in the Python "help (lldb.SBWhatever)" and some are not.
</p>
<p>
We need more conceptual docs. And we need more examples. And we could provide a
clean pluggable example for using LLDB standalone from Python. The process_events.py
is a start of this, but it just handles process events, and it is really a quick sketch
not a polished expandable proto-tool.
</p>
</li>
<li>
Make a more accessible plugin architecture for lldb.
<p>
Right now, you can only use the Python or SB API's to extend an extant lldb. You can't
implement any of the actual lldb Plugins as plugins. That means anybody that wants to
add new Object file/Process/Language etc support has to build and distribute their own
lldb. This is tricky because the API's the plugins use are currently not stable (and
recently have been changing quite a lot.) We would have to define a subset of lldb_private
that you could use, and some way of telling whether the plugins were compatible with
the lldb. But long-term, making this sort of extension possible will make lldb more
appealing for research and 3rd party uses.
</p>
</li>
<li>
Use instruction emulation to avoid the overhead of swapping trap and instruction for breakpoints
<p>
At present, breakpoints are implemented by inserting a trap instruction, then when the trap
is hit, replace the trap with the actual instruction and single step. Then swap back and
continue. This causes problems for read only text, and also means that no-stop debugging
ust either stop all threads briefly to handle this two-step or risk missing some
breakpoint hits. If you emulated the instruction and wrote back the results, you wouldn't
have these problems, and it would also save a stop per breakpoint hit. Since we use breakpoints
to implement stepping, this savings could be significant on slow connections.
</p>
</li>
<li>
Use the JIT to speed up conditional breakpoint evaluation
<p>
We already JIT and cache the conditional expressions for breakpoints for the C family of
languages, so we aren't re-compiling every time you hit the breakpoint. And if we couldn't
IR interpret the expression, we leave the JIT'ed code in place for reuse. But it would
be even better if we could also insert the "stop or not" decision into the code at the
breakpoint, so you would only actually stop the process when the condition was true.
Greg's idea was that if you had a conditional breakpoint set when you started the
debug session, Xcode could rebuild and insert enough no-ops that we could instrument
the breakpoint site and call the conditional expression, and only trap if the conditional was true.
</p>
</li>
<li>
Broaden the idea in "target stop-hook" to cover more events in the debugger
<p>
Shared library loads, command execution, User directed memory/register reads and writes
are all places where you would reasonably want to hook into the debugger.
</p>
</li>
<li>
Mock classes for testing
<p>
We need "ProcessMock" and "ObjectFileMock" and the like. These would be real
plugin implementations for their underlying lldb classes, with the addition
that you can prime them from some sort of text based input files. For classes
that manage changes over time (like process) you would need to program the state
at StopPoint 0, StopPoint 1, etc. These could then be used for testing reactions
to complex threading
problems & the like, and also for simulating hard-to-test environments (like bare
board debugging).
</p>
</li>
<li>
A Bug-Trapper infrastructure
<p>
We very often have bugs that can't be reproduced locally. So having a bug-report-trapper
that can gather enough information from the
surroundings of a bug so that we can replay the session locally would be a big help
tracking down issues in this situation. This is tricky because you
can't necessarily require folks to leak information about their code in order to
file bug reports. So not only will you have to figure out what state to gather,
you're also going to have to anonymize it somehow. But we very often have bugs
from people that can't reduce the problem to a simple test case and can't give us
our code, and we often just can't help them as things stand now. Note that adding
the ProcessMock would be a good first stage towards this, since you could make a ProcessMock
creator/serializer from the current lldb state.
</p>
</li>
<li>
Expression parser needs syntax for "{symbol,type} A in CU B.cpp" etc.
<p>
Sometimes you need to specify non-visible or ambiguous types to the expression parser.
We were planning to do $b_dot_cpp$A or something like. You might want to specify a
static in a function, in a source file, or in a shared library. So the syntax should
support all these.
</p>
</li>
<li>
Add a "testButDontAbort" style test to the UnitTest framework.
<p>
The way we use unittest now (maybe this is the only way it can work, I don't know)
you can't report a real failure and continue with the test. That is appropriate
in some cases: if I'm supposed to hit breakpoint A before I evaluate an expression,
and don't hit breakpoint A, the test should fail. But it means that if I want to
test five different expressions, I can either do it in one test, which is good
because it means I only have to fire up one process, attach to it, and get it to
a certain point. But it also means if the first test fails, the other four don't
even get run. So though at first we wrote a bunch of test like this, as time went
on we switched more to writing "one at a time" tests because they were more robust
against a single failure. That makes the test suite run much more slowly. It
would be great to add a "test_but_dont_abort" variant of the tests, then we could
gang tests that all drive to the same place and do similar things. As an added
benefit, it would allow us to be more thorough in writing tests, since each
test would have lower costs.
</p>
</li>
<li>
Convert the dotest style tests to use lldbutil.run_to_source_breakpoint.
<p>
run_to_source_breakpoint & run_to_name_breakpoint provide a compact API that
does in one line what the first 10 or 20 lines of most of the old tests now do by
hand. Using these functions makes tests much more readable, and by centralizing
common functionality will make maintaining the testsuites easier in the future.
This is more of a finger exercise, and perhaps best implemented by a rule like:
&quotIf you touch a test case, and it isn't using run_to_source_breakpoint, please
make it do so&quot.
</p>
</li>
<li>
Unify Watchpoint's & Breakpoints.
<p>
Option handling isn't shared, and more importantly the PerformAction's have a lot
of duplicated common code, most of which works less well on the Watchpoint side.
</p>
</li>
<li>
Reverse debugging.
<p>
This is kind of a holy grail, it's hard to support for complex apps (many
threads, shared memory, etc.) But it would be SO nice to have...
</p>
</li>
<li>
Non-stop debugging.
<p>
By this I mean allowing some threads in the target program to run while stopping
other threads. This is supported in name in lldb at present, but lldb makes the
assumption "If I get a stop, I won't get another stop unless I actually run the
program." in a bunch of places so getting it to work reliably will be some a
good bit of work. And figuring out how to present this in the UI will also be tricky.
</p>
</li>
<li>
Fix and continue.
<p>
We did this in gdb without a real JIT. The implementation shouldn't be that hard,
especially if you can build the executable for fix and continue. The tricky part is
how to verify that the user can only do the kinds of fixes that are safe to do.
No changing object sizes is easy to detect, but there were many more subtle changes
(function you are fixing is on the stack...) that take more work to prevent.
And then you have to explain these conditions the user in some helpful way.
</p>
</li>
<li>
Unified IR interpreter.
<p>
Currently IRInterpreter implements a portion of the LLVM IR, but it doesn't handle
vector data types and there are plenty of instructions it also doesn't support.
Conversely, lli supports most of LLVM's IR but it doesn't handle remote memory and
its function calling support is very rudimentary. It would be useful to unify these
and make the IR interpreter -- both for LLVM and LLDB -- better. An alternate strategy
would be simply to JIT into the current process but have callbacks for non-stack memory
access.
</p>
</li>
<li>
Teach lldb to predict exception propagation at the throw site
<p>
There are a bunch of places in lldb where we need to know at the point where an
exception is thrown, what frame will catch the exception.
</p>
<p>
For instance, if an expression throws an exception, we need to know whether the
exception will be caught in the course of the expression evaluation. If so it
would be safe to let the expression continue. But since we would destroy the
state of the thread if we let the exception escape the expression, we currently
stop the expression evaluation if we see a throw. If we knew where it would be
caught we could distinguish these two cases.
</p>
<p>
Similarly, when you step over a call that throws, you want to stop at the throw
point if you know the exception will unwind past the frame you were stepping in,
but it would annoying to have the step abort every time an exception was thrown.
If we could predict the catching frame, we could do this right.
</p>
<p>
And of course, this would be a useful piece of information to display when stopped
at a throw point.
</p>
</li>
</ol>
</div>
<div class="postfooter"></div>
</div>
</div>
</div>
</div>
</body>
</html>