[llvm-dev] C++ JIT Compiler with LLVM on Windows 10 - part 5

[Emmanuel Roche via llvm-dev]

Hello LLVM developers!

So, I'm continuing my journey with my toy C++ JIT compiler implementation, and I wrote another article on the issues/solutions I've been working on in the past few days, mainly:

- Precompiled header handling,

- Emulated TLS desactivation,

- Globals construction & destruction,

- C++ exceptions handling,

- Multi modules linking,

=> In case this could be helpful to anyone, you will find this article here: http://wiki.nervtech.org/doku.php?id=blog:2020:0425_jit_compiler_part5_improvements

And of course, if you have any questions for me, just let me know.

Meanwhile, happy hacking everyone ;-)!

Best regards,

Manu.

[David Blaikie via llvm-dev]

+Lang for LLVM Orc things

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

[Lang Hames via llvm-dev]

I was just reading the blog post -- very cool!

Regarding Globals construction & destruction: There definitely has been a lot of churn in that area. There will probably be more before  LLVM 11 is released, but I can see light at the end of the tunnel. I think the Right Way to run initializers in a JITDylib is to treat it as equivalent to a dlopen operation (with extra allowances for the fact that new initializers can be added to a JITDylib at runtime). This is what the LLJIT::initialize method is doing. Now we just need to generalize it to support out-of-process JITs. That will at least require a new remote-target abstraction, and an RPC implementation for testing in-tree.

Regarding emulated TLS deactivation: I've never looking into how/whether this works on Windows. If it doesn't make sense to have it on by default there we can change the default for Windows targets.

Regarding merging of multiple modules:

"But of course this would not work because as soon as I try to load the second script, I get a duplicate symbol error from LLVM (and this completely makes sense):

[ERROR]: LLVM error: Duplicate definition of symbol '??_7success@lest@@6B@'"

I think that anything that llvm-link can merge should, in theory, be safe to add to the JIT. This actually sounds like a bug. Are you able to share the full modules that you were merging?

Regards,

Lang.

[Emmanuel Roche via llvm-dev]

Hi Lang,

Thank you for your feedback on the blog post, please find below some additional inputs from my side on the comments you provided:

> Regarding emulated TLS deactivation: I've never looking into how/whether this works on Windows. If it doesn't make sense to have it on by default there we can change the default for Windows targets.

I'm really no expert, so you probably want to collect additional inputs/checks on this point, but yes, from my current perspective, it seems that it doesn't really work to try to use Emulated TLS in LLJIT on Windows, but one should keep in mind that this is when using the microsoft crt from Visual Studio (FYI, I'm using version 2017 for now, maybe it was different in previous versions too): I would expect things to be different if you are using another compiler as base (?) [I mean, if you change the default, this might then break previously working code that was based on mingw for instance... But not sure if this should be a concern or not for a not yet released version of LLVM ;-)]

> I think that anything that llvm-link can merge should, in theory, be safe to add to the JIT. This actually sounds like a bug. Are you able to share the full modules that you were merging?

Initially, the test I made on this point was using the Lest test framework as described in my post, but your feedback above got me thinking, and I could actually come up with a more simple setup without any C++ library dependency: this new test is simply based on a couple of C++ functions that are using exceptions. I used the attached lua script "exception3.lua" to generate the bitcode files and load the corresponding modules [of course I don't expect you to run lua scripts, but I thought that this could help understanding the code generation context anyway]. So with that script I generated the test_func1.bc/test_func2.bc/test_func1_and_func2.bc and the corresponding .ll file (generated from the .bc files using llvm-dis): all those files are attached to this email too.

=> So, I can load for instance the module corresponding to test_func1.bc in my main JITDylib, but then, if I try to load the test_func2 IR module in the same Dylib, I get the following error:

Duplicate definition of symbol '??_7exception@std@@6B@'

In all the .ll files you will find multiple references of that symbol... but unfortunately, my skills won't get me any further than that [=> ie. I have no idea what to think about the content of those files :-)].

But anyway, if you notice anything interesting yourself, or you have an idea you think I should try, or you need more info on anything here please let me know.

PS: I'm thinking maybe the command line arguments I'm providing to the compilerInvocation might also be important here so, just in case: here is the list of args I'm using currently:

{    "-triple=x86_64-pc-windows-msvc19.16.27030",
    "-x", "c++",
    "-mrelax-all",
    "-mincremental-linker-compatible",
    "-disable-free",
    "-discard-value-names",
    "-mrelocation-model", "pic",
    "-pic-level", "2",
    "-mthread-model", "posix",
    "-mframe-pointer=none",
    "-relaxed-aliasing",
    "-fmath-errno",
    "-fno-rounding-math",
    "-mconstructor-aliases",
    "-munwind-tables",
    "-target-cpu", "x86-64",
    "-mllvm",
    "-x86-asm-syntax=intel",
    "-stack-protector","2",
    "-fcxx-exceptions",
    "-fexceptions",
    "-fexternc-nounwind",
    "-fms-volatile",
    "-fdiagnostics-format", "msvc",
    "-dwarf-column-info",
    "-Wall",
    "-std=c++17",
    "-fdeprecated-macro",
    "-ferror-limit","19",
    "-fmessage-length=174",
    "-fms-extensions",
    "-fms-compatibility",
    "-fms-compatibility-version=19.16.27030",
    "-fdelayed-template-parsing",
    "-fcolor-diagnostics",
    "-fobjc-runtime=gcc",
    "-faddrsig",
  }

Regards,

Manu.

[Lang Hames via llvm-dev]

Hi Emmanuel,

Thank you very much for these! I will check them out today and see what's going on.

Regards,

Lang. 

[Lang Hames via llvm-dev]

Hi Emmanuel,

I see the problem:

$"??_7exception@std@@6B@" = comdat largest

The JIT knows how to honor linkages, but it does not know about comdats yet except in limited circumstances, so it sees these as conflicting definitions.

Adding support for comdats will be non-trivial as their selection rules are more complex than the selection rules for linkage types. For now I would stick to using llvm-link. I will file a bug tomorrow to add COMDAT support to ORC. I won't have time to work on this feature any time soon, but I will try to provide a sketch of the solution in the bug report in case you or anyone else is interested in taking up the challenge. :)

Cheers,

Lang.

[Emmanuel Roche via llvm-dev]

Hello Lang,

Good work! That was very quick ;-) 

> I will file a bug tomorrow to add COMDAT support to ORC. I won't have time to work on this feature any time soon, but I will try to provide a sketch of the solution in the bug report in case you or anyone else is interested in taking up the challenge. :)

=> I suspect this could be an issue that could get in my way more significantly than I initially thought, so yes, if you could provide some details in the bug report that would be great indeed, and I might eventually find enough motivation and energy to try to handle it myself.

=> Please let me know when you have a link to that new bug report available, and meanwhile thank you very much for your time and efforts on this issue!

Have a great day!

Cheers,

Emmanuel.

[Lang Hames via llvm-dev]

Hi Emmanuel,

I have filed http://llvm.org/PR45822 to track any COMDAT support work.

For the next few weeks at least I will be focused on removable code and other features. As I've noted in the bug report it's probably not worth trying to implement this until removable code lands as there will be a lot of churn in Core.cpp until then. If you're interested in working on this though I would recommend starting by taking a look at how JITDylib::defineImpl currently uses SymbolFlagsMap to decide which definitions to keep/discard. That's the logic we will need to generalize to support COMDATs.  

Regards,

Lang.

[Emmanuel Roche via llvm-dev]

Hi Lang,

Well received.

=> When I get some time to work on this now I'll try to read the documentations pages you reported, and the corresponding code in JITDylib: just to build an understanding what this is all about first. Then I'll see if I get an idea on what to do about it, and I'll keep you informed if I have anything new to add on this problem ;-)

Meanwhile, good luck for the removable code support implementation !

Cheers,

Emmanuel.