Get all symbols stored(?)in llvm::orc::ExecutionSession

Hey everyone,

is there a way to get the name of all symbols that are stored (“stored” is not the right term – is it?) in my current ExecutionSession?

I know by now that you can use “lookup” to get symbols, but this requires knowing those names.

Mhh… I guess that is my shortest so far question xD Thank you in advance for any help!

Kind greetings

Björn

Hi Bjoern,

There’s no way to do this currently. ORC assumes you know all the symbols, since you added the modules defining them.

For testing / debugging you can dump the modules to stderr using ExecutionSession::dump, but that’s about it.

Do you want the symbols for diagnostic purposes, or some other reason?

Regards,
Lang.

This issue is still not solved for me…

What can I improve on my question? :o

Adding more details or something? >_x

Kind greetings

Björn

Hi Bjoern,

Did you see my previous reply?

There’s no way to do this currently. ORC assumes you know all the symbols, since you added the modules defining them.
For testing / debugging you can dump the modules to stderr using ExecutionSession::dump, but that’s about it.
Do you want the symbols for diagnostic purposes, or some other reason?

Regards,
Lang.

Hey Lang,

Oh shoot… seems like I oversaw it… Sorry sorry sorry! (I was 5 weeks away but I thought I checked all mails correctly >o<)

I think ORCs assumption is correct, currently I do extract the symbols which I care about via the modules, but when writing the code I was like “why do I have to do it like that, when the ExecutionSession already knows all symbols?” – that is where my question came from.

However… I’m a bit worried about taking symbols from modules, because some of the symbols might change later when the code was compiled with the JIT. So I felt like that the ExecutionSession would been more reliable.

But if I reaaaally want such a function, I could basically have a look into the dump function… Cause I think my motivation is not convincing xD

Thank you Lang!

Kind greetings

Björn

Hi Bjoern,

I’m a bit worried about taking symbols from modules, because some of the symbols might change later when the code was compiled with the JIT

Symbol names shouldn’t change from the moment they’re added to the JIT (i.e. after being fully mangled). Consistency of symbol names is a requirement for ORC to work.

You could try hooking into the orc::Platform API for this: https://github.com/llvm/llvm-project/blob/ec29538af2e0886a65f479d6a533956a1c478132/llvm/include/llvm/ExecutionEngine/Orc/Core.h#L1062

The notifyAdding method will be called for each MaterializationUnit added to the JIT, which gives you an opportunity to iterate over the symbols that it provides and record them.

Regards,
Lang.

Hey Lang,

Symbol names shouldn’t change from the moment they’re added to the JIT (i.e. after being fully mangled). Consistency of symbol names is a requirement for ORC to work.

Right, what I meant was, that some of the symbols I take from the modules might disappear… I don’t remember the details anymore, but I think I encountered the symbol for an intrinsic version of memcpy but after the compilation it became a call to memcpy.

Does that explain what I mean?

You could try hooking into the orc::Platform API for this: https://github.com/llvm/llvm-project/blob/ec29538af2e0886a65f479d6a533956a1c478132/llvm/include/llvm/ExecutionEngine/Orc/Core.h#L1062

So I would use the function “setPlatform” of the ”ExecutionSession”? But wouldn’t I override something really important with that?

Thank you so far Lang :smiley:

Kind greetings

Björn

Hi Bjoern,

Symbol names shouldn’t change from the moment they’re added to the JIT (i.e. after being fully mangled). Consistency of symbol names is a requirement for ORC to work.
Right, what I meant was, that some of the symbols I take from the modules might disappear… I don’t remember the details anymore, but I think I encountered the symbol for an intrinsic version of memcpy but after the compilation it became a call to memcpy.
Does that explain what I mean?

Oh – You only want to look at definitions, not declarations. It’s totally fine if your code adds or removes uses of other symbols – that only changes the lookups that the JIT linker will need to perform at the end of the process. What’s really important is that the definitions do not change.

That said: Sometimes the compiler does add or remove definitions. E.g. COFF COMDATS for floating point constants (see http://lists.llvm.org/pipermail/llvm-dev/2020-January/138584.html). These cases are rare, but if you hit one please let me know: We probably need to teach ORC to recognize it.

So I would use the function “setPlatform” of the ”ExecutionSession”? But wouldn’t I override something really important with that?

It depends on whether you’re using LLJIT or a custom JIT class. ExecutionSession doesn’t set a Platform by default so you’re free to (ab)use the Platform when writing a custom JIT class. LLJIT does set up a Platform, but we could look at moving that into a header so that you could extend it with your own behavior.

– Lang.