Code

Design Overview

• emulates linking and symbol resolution resolution
• allows ORC to jit arbitary LLVM IR

ExecutionEngine

https://llvm.org/docs/ORCv2.html#how-to-add-process-and-library-symbols-to-jitdylibs

Experiments

Running Program as a REPL

Design:

• Aim to execute code arbitrarily in running process, with access to memory, etc.
• Shared API for memory

Other thoughts:

• We could probably

Compile to LLVM IR with clang

1. Code /Users/miguelmartin/repos/langdev/repl-tests

   • main program
   • shared api
     • shared api to hold state defns
   • repl api
     • given a string => execute it in the main process

2. Commands to Execute

Tutorial

https://llvm.org/docs/ORCv2.html

Chapter 1

https://llvm.org/docs/tutorial/BuildingAJIT1.html

Bare bones JIT API has just two functions:

1. addModule(Unique<Module>)
   • makes the given IR module available for execution
2. Expected<JITEvaluatedSymbol> lookup()
   • searches for pointers to symbols

KaleidoscopeJIT class has the following members:

1. ExecutionSession (ES)
   • context for running JIT'd code, includes a:
     • string pool, global mutex and error reporting
2. RTDyldObjectLinkingLayer (ObjectLayer)
   • Can be used to add object files to the JIT
   • Not used directly
3. IRCompileLayer (CompileLayer)
   • Add LLVM modules to the JIT
4. DataLayout (DL)
5. MangleAndInterner (Mangle)
   • symbol and mangle layout
6. ThreadSafeContext (Ctx)
   1. Used when building IR files

JITTargetMachineBuilder, used by DataLayout & IRCompileLayer

Talks

2018 Meeting

Implementing LLJit

Construct:

• ExecutionSession, contains:
  • String pool
  • Session mutex
  • Error reporting
• RTDylibObjectLinkingLayer
  • adds object files to JITDylib
• ConcurrentIRCompiler
  • Compiles LLVM IR to asm
• For lazy:
  • ComplieOnDemandLayer
• For multi-threaded:
  • ES.setDispatchMaterialization
    • <insert code to execute on a thread pool / seperate thread>

Layers wrap compilers MaterializationUnits wrap program representations MaterializationResponsibility tracks compiler responsibilities

2021 ORCv2

JITLink

• LinkGraph
  1. object file
  2. LinkGraph builder
  3. LinkGraph
  4. JITLinker
  5. Linking Context
     1. Pass pipeline modification
     2. Add passes, any callable of Error(LinkGraph& G)
     3. Phases:
        • pre-prune
          • mark elements alive before dead-stripping
        • post-prune
          • add/remove content (GOT, PLT entries). Global Offset Table entries, Procedure Linkage Table entries
        • post-alloc
          • respond to assinged addressed for defined symbols
          • use-cases: internal book-keeping
        • pre-fixup
          • access final addresses, e.g. for instruction optimization
        • post-fixup
          • react to linked content after fixups have been applied

Error interpoFunctions(LinkGraph& G) {
  for(auto *B: G.blocks()) {
	for(auto& E : B->edges()) {
	  E.setTarget(getStubeFor(E.getTarget()));
	}
  }
  return Error::success();
}

Can do many things:

• Instruction stream optimization
• Redirect calls and branches
• Insert instrumentation