[llvm-dev] ORC JIT Weekly #2 -- COFF COMDAT Constants and Emulated TLS

[David Chisnall via llvm-dev]

Hi Lang,

Thank you very much for writing these.  I suspect that they are likely 
to be drowned in the noise of this list and hidden from a lot of the 
people that would find them valuable.

Would it be possible for them to be hosted somewhere permanently? 
Perhaps they could either be incorporated as a subsection of LLVM Weekly 
or put in a different location on the same server?

David

On 24/01/2020 03:27, Lang Hames via llvm-dev wrote:
> Hi All,
> 
> This week I've been focused on removing some of the blockers for people 
> transitioning from ORCv1 to ORCv2.
> 
> Issue #1 (http://llvm.org/PR40074, http://llvm.org/PR44337):
> 
> When LLVM codegens floating point constants for COFF we produce named 
> constant pool entries of the form __real@<bitval>. These are stored in 
> COFF COMDAT sections [1] which allow duplicate symbol definitions to be 
> discarded (roughly equivalent to the concept of a weak symbol). This 
> posed two problems: (1) ORC had no idea about the (rough) mapping 
> between COMDAT and Weak, and (2) did not support late introduction of 
> weak symbols via MaterializationResponsibility::defineMaterializing. 
> This caused use of the same floating point constant in multiple files to 
> result in duplicate definition errors for the constant pool entries.
> 
> Both issues have been addressed in 84217ad6611, and floating point 
> constants should now work in COFF. If you run in to any trouble with 
> them please file a bug and CC me.
> 
> Issue #2: Emulated TLS and name mangling (raised by Geoff Levner on 
> llvm-dev):
> 
> Use of emulated-tls affects the name mangling of LLVM IR symbols. For a 
> thread local symbol with name <Name>, when emulated-tls is enabled we 
> will produce at least one emulated-tls symbol: __emutls_v.<Name> (which 
> must then be linker-mangled to get the final symbol name). If that 
> symbol has a non-zero initializer a second symbol will also be produced: 
> __emutls_t.<Name>. ORCv2 requires up-front specification of the exact 
> set of /linker-level/ symbol names that a module produces, so that it 
> can dispatch and track compiles, however none of the existing mangling 
> code took emulated-tls into account. This has been fixed in ce2207abaf9 
> (with follow up fixes for the various examples I broke), and you should 
> now be able to use thread-locals in ORCv2.
> 
> Besides these two bugs, I have continued work on the big static 
> initializer patch. Unfortunately issue #2 interacts with it, so I have 
> to do some integration work before it is ready to land, but I'm hopeful 
> that it will be ready some time next week.
> 
> Finally, I have added a new example (bb7a5707ac0)to 
> llvm/examples/LLJITExamples: LLJITWithObjectLinkingLayerPlugin 
> demonstrating how to write a custom ObjectLinkingLayer plugin. 
> ObjectLinkingLayer plugins can be used to receive and respond to events 
> in the linking layer, and to inspect and modify JITLink LinkGraphs 
> during the linking process. ORC uses this internally to register EH 
> frames, but the system is likely to be of interest to anyone who wants 
> low level control/visibility into the JIT linker.
> 
> — Lang.
> 
> [1] 
> https://docs.microsoft.com/en-us/windows/win32/debug/pe-format#comdat-sections-object-only 
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>