|Titan Architecture Internals: On how the compiler and the Designer are optimized differently [message #1750053]
||Thu, 15 December 2016 14:16
| Kristof Szabados
Registered: July 2009
When looking at the compiler and the Designer from the outside at first they might seem to be doing the same thing: reading and checking TTCN-3 files.|
- The core compiler reads TTCN-3 and ASN.1 files, checks them for syntactic and semantic errors, and finally generates C/C++ code.
- The Designer reads TTCN-3 and ASN.1 files, checks them for syntactic and semantic error, and offers some IDE functionalities.
Although they seem to do the same things in most of their source code too ... they are actually working in very different constraints, and optimized for different use cases:
- is usually used rarely and in the background, as part of a CI system, maybe doing nightly builds and builds after each commit.
- is viewed by users as part of the build process, which starts with TTCN-3 and ASN.1 files and finishes with the creation of the executable tests ... also involving GCC/CLANG. In a full build the compiler might take up as little as only 1%-5% of the build process.
- as such optimization of the compiler is done to minimize the whole build time (the compiler could take a few more seconds, if it can save minutes on the C/C++ compiler)
- when invoked all code is generated (even if just into memory at first).
- the memory consumption is usually not a problem ... as GCC/CLANG will require much more from the same system to finish the build.
- when there is an error in the code the build stops and the user needs to correct it before the code can compile.
When the compiler finds and unparsable piece of code like "variable_name." it can report and error and stop processing.
- users don't expect interaction.
- extending the generated code can be done will scripts and modified makefiles if needed.
- up close and personal, this is the tool used to write TTCN-3, ASN.1 code.
- after each keystroke the user can expect to have some syntactic and semantic checks done to ensure that the code just written is ok.
(without optimization this would mean 6-10 semantic check within a second)
- As this is the editing interface interactivity is expected.
Usually expecting below 0.1 second response time ... whatever the operation was.
- the TTCN-3 code is most of the time erroneous (sometimes on several levels) as it is being written.
But the tool must not stop working.
- In fact the Designer is expected to work best and offer helpful support when the code has the most problems.
When the user presses ctrl+space after an unparsable piece of code like "variable_name." Designer is expected to list all possible completion methods that could make the code correct.
- advanced interaction, like outline views, module structure graph visualization, code smell checking, refactoring are expected to operate ... also interactively.
For these reasons, even though most of the code seems to be similar, they follow very different ways of workings:
- reads all TTCN-3 and ASN.1 files it received as parameters.
- does the semantic checking.
- if no error is found, generates the C/C++ code.
- exits leaving no information behind.
The Designer works in loops:
In the first loop:
- reads all TTCN-3 and ASN.1 files that are in the project (or in related projects)
- does the syntactic and semantic analysis.
- keeps the results in memory to speed up further operations.
In subsequent loops (after something was edited):
- using the information present in memory tries to calculate the minimum amount of data that needs to be re-processed
- does the syntactic and semantic checking using the already existing information, to minimize work.
- keeps the results in memory to speed up further operations.
If some data is needed by the user or a tool:
- the information is always available in memory to be used.
This however complicates things for the Designer to meet the user's harder needs:
- while the compiler is always starting from zero, the Designer has to keep up the consistency.
In the compiler this means that a single boolean "checked" variable per AST node is enough to know if the was already checked during the semantic processing, or it still needs to be checked (in the semantic graph the order of processing is not linear)
In the Designer the AST nodes have to have a timestamp ("lastTimeChecked") showing in which semantic checking cycle that node was last checked. If it was already checked in an earlier loop and the change triggering the current loop cannot change its semantic properties ... we don't need to re-analyze it.
This timestamp is created at the beginning of each semantic check loop ... and passed down from the Module level to each checked AST node via their "check" function (that does the semantic checking).
Please note, that at any point in time the semantic graph can have coexisting regions that were last checked in very different points in time.
- To read the files as fast as possible the syntactic checking of files is happening in parallel, before the semantic checking starts (in org.eclipse.titan.designer.parsers.ProjectSourceSyntacticAnalyzer the internalDoAnalyzeSyntactically function).
As many files are processed at the same as many cores there are in the user's machine (if not configured otherwise).
- In the compiler before the semantic check we know that all nodes need to be checked ... in the Designer the nodes in the semantic graph that might be affected by a change have to be discovered (in org.eclipse.titan.designer.AST.brokenpartsanalyzers.BrokenPartsViaReferences).
/* using data from the previous semantic check iteration ... that might have changed just now */
- while the compiler can run in its own process (as no interactivity is needed) ... the Designer (and all functions accessing its data) needs to be prepared to work as a distributed system.
The user (or system) can trigger changes in many files and many line at the same time.
The user could also try to invoke some functionality while the syntactic and semantic checking of the last changes are still running.
... invoke operations on incorrect code ... and in situations where they cannot realistically work (try renaming a word doc to a TTCN-3 file, open it and ask for code completion).
For this reason operations reaching the semantic data are run in WorkspaceJob -s to provide protection from concurrent access.
Powered by FUDForum
. Page generated in 0.01294 seconds