[Epic] MDGA: Make Druid Great Again

[PalumboN]

Bytecodes

• gen_ExtPushFullClosureBytecode Fix closure creation #94

  • Allocate dynamic registers outside the loops?

• gen_PushLiteralVariable16CasesBytecode (WIP)

  • Stage the access to the association but resolve the value at runtime (avoid the doble int64AtPointer) Make staging a separate pass #97
  • Annotate the association New staged version of druid JIT pharo-project/pharo-vm#728

• Then continue compiling this list

• 243
• 231 - PushArrayBytecode ( Fix Register allocator trampoline issues in Druid JITs pharo-project/pharo-vm#729 + Cleanups in staging #98)
• 251

Broken bytecodes

•  gen_ExtStoreAndPopReceiverVariableBytecode
• gen_StoreAndPopRemoteTempLongBytecode

Primitives

• Make a list of the most used primitives

  • We can do it with the StackVM, logging the primitive in slowPrimitiveResponse
    

• Fix primitiveNew compilation

• 62 (primitiveSize)

• 207 (primitiveFullClosureValue)

• 175 (primitiveBehaviorHash)

• 75 (primitiveIdentityHash)

• 17 (primitiveBitShift)

• 169 (primitiveNotIdentical)

• 61 (primitiveAtPut)

• 105 (primitiveStringReplace)

• Have a list per benchmark

• 29 primitiveMultiplyLargeIntegers
• Measure execution time with Rebench 📏

Others

• Improve abstract interpretation of compiler options here
• Improve linearization algorithm here
• Cache the Dominator Tree #140
• Refactor on branch condition code generator code here
• Remove = true from staged if Fix closure creation #94
• cleanup staging decisions (canStage: vs isJitCompileTimeExpresison repeat code?)
• tail duplication during code generation (for code generation of staged ifs) misses some staging opportunities (i.e., it depends on a phi that is staged from one branch but not from the other => then the phi is not staged, but when tail duplication happens this allows staging!)