Lazarus Debugger Implementation
English (en) │
This page provides an overview of how the debugger is implemented in the Lazarus IDE.
The debugger comprises the following parts:
- The controls (run, step, stop buttons) and dialogs (watches, stack) in the IDE.
- A set of base-classes defining the communication between the front-end and each back-end
- Any code to control the debugged app, and retrieve/modify data from/in that app. The back-end may itself be split into an IDE back-end and an external back-end (e.g. gdb).
The code for the debugger frontend belongs currently directly to the IDE (to the project ide/lazarus.lpi). It is not a package of it's own.
Some of this code may one day be moved into a package. And maybe even introduce a package DebuggerFrontendIntf, to allow alternative 3rd party front-ends.
The code currently is located in
- folder /debugger
- folder /ide, files: debugmanager.pas and basedebugmanager.pas
- folders ide/frames and debugger/frames for configuration settings.
BaseDebugManager is currently based on components\debuggerintf\dbgintfdebuggerbase.pp
TDebuggerIntf - main class
TDebuggerIntf provides the communication between front ad back-end.
Note: Currently State contains some value internal to the backend. (dsIdle/dsInit). Those may vary from back-end to back-end
There are 2 pause states.
- The proper pause (debugged app is paused) = dsPause.
- And dsInternalPause used for example for breakpoints that will automatically continue to run the debuggee.
See documentation in source code.
The front-end can send asynchronous commands to the back-end. A result - if any - will be provided via an event.
Commands are (some entries of the list are outdated)
TDBGCommand = ( dcRun, dcPause, dcStop, dcStepOver, dcStepInto, dcStepOut, dcStepTo, dcRunTo, dcJumpto, dcAttach, dcDetach, dcBreak, dcWatch, dcLocal, dcEvaluate, dcModify, dcEnvironment, dcSetStackFrame, dcDisassemble, dcStepOverInstr, dcStepIntoInstr, dcSendConsoleInput //, dcSendSignal );
The back-end does not need to implement all commands. (see procedure SupportedCommands) The availability of commands also depends on the State
There are helpers like
They will take additional care of calling Init if needed.
TDebuggerDataMonitor and TDebuggerDataSupplier - stack, watches, ....
The debugger can supply data like Threads, Stack, Locals, Watches... Any of this data is provided by a ___DataSupplier. On the IDE site to each Supplier, one ___DataMonitor is attached. The Supplier/Monitor concept was introduced to allow both sides to have their own, yet synchronized list of items for the list-entries.
For an example: Watches.
The IDE creates a Watch as "TIdeWatch". The Ide can stick any data to a watch that the IDE may require. For example watch-properties (display, enabled, ...) or streaming.
The back-end may need different information, such as internal details retrieved from the external-backend (gdb). So the back-end uses "TGDBMIWatches = class(TWatchesSupplier)". And other back-ends have their own sub-classes.
The supplier/monitor concept keeps the two lists synchronized. For Watches, the IDE side (the monitor) creates and deletes the entries. For Stack and locals, it is the back-end that creates/deletes entries.
When the IDE needs a value then the request is forwarded to the data-supplier. A value in this context can be anything read from any of the involved objects. The count of thread/stackframes is a value, the frame itself is a value (which is an object with multiple data fields). And the evaluated result of a watch is a value too.
All data is retrieved event driven. Asking for Stackcount, or watch-result will not return the value immediately. It will trigger its evaluation, and the data-supplier will sent an update event when the value (any value) becomes available.
For this to work, each value has a validity "TDebuggerDataState".
- AWatch.Values[ThreadId, Stackframe].Value and AWatch.Values[ThreadId, Stackframe].Validity
- In other cases validity is not exposed. Count may return 0, if it is not yet valid. (Maybe a TODO, to expose the validity? Just maybe.)
There is no rule how many update-events may be triggered before some value becomes valid. Any code listening to events from the monitor simply has to start over when it gets an event. Events may occur nested, code must protect itself again re-entrance in such a case.
The validity is internally also used to avoid duplicate triggers. The IDE can ask for a value as often as it wants. When it asks for the first time the validity changes to "requested", preventing further triggers to the back-end. Validity is usually reset, when the debugger leaves the dsPause state. (dsPause is a state related to the debuggers execution of the debuggee).
- "Command Queue"
This is an item most/all backends have in common. Commands are received via RequestCommand and may take any amount of time to execute. Other work for the backend comes from the DataMonitor/Suppliers. And Backends may also schedule work of there own (for example retrieving the current location when an app enters pause). All this work is usually serialized in the backend (though backends are free to process multiple items paralel). For this backends have (at least) one CommandQueue. Usually this queue supports some sort of priority, so items in response to a user action (e.g. evaluating an expression for source editor hint) can be executed first. This means the user does not need to wait unnecessary long for this.
GDB / GDBMI based
Location: components/lazdebuggers/LazDebuggerLldb components/lazdebuggers/LazDebuggerFpLldb