Add Ir Instruction Agent Skill

Adding a New IR Instruction to Hermes

When adding a new IR instruction, you must touch a specific set of files. This

skill describes each file, the pattern to follow, and important conventions.

Checklist of Files to Modify

• doc/IR.md — Documentation (the only place for doc-comments)
• include/hermes/IR/Instrs.def — Instruction registration
• include/hermes/IR/Instrs.h — Class definition (NO doc-comments here)
• include/hermes/IR/IRBuilder.h — Builder declaration
• lib/IR/IRBuilder.cpp — Builder implementation
• lib/IR/IRVerifier.cpp — Verification logic
• lib/Optimizer/Scalar/TypeInference.cpp — Type inference stub
• lib/BCGen/HBC/ISel.cpp — HBC instruction selection (stub or implementation)
• lib/BCGen/SH/SH.cpp — Static Hermes codegen (stub or implementation)

10. lib/BCGen/facebook/Mins/Mins.cpp — Mins codegen (stub or implementation)

11. Tests — At minimum, update or add tests in test/

If the instruction needs lowering (i.e., it does not map directly to a bytecode

opcode), you also need:

12. include/hermes/BCGen/Lowering.h — Lowering pass declaration

13. lib/BCGen/Lowering.cpp — Lowering pass implementation

14. lib/BCGen/HBC/LoweringPipelines.cpp — Register pass in HBC pipeline

15. lib/BCGen/SH/SH.cpp (in lowerModuleIR) — Register pass in SH pipeline

Step-by-Step Guide

1. Document in doc/IR.md

This is the ONLY place to put documentation for the instruction. Do NOT add

doc-comments to Instrs.h.

Add a markdown table entry in the appropriate section:

### MyNewInst

MyNewInst | _
--- | --- |
Description | Brief description of what the instruction does.
Example |   `MyNewInst %arg1, %arg2 : type`
Arguments | *%arg1* is ... *%arg2* is ...
Semantics | Describe the semantics, referencing the spec where appropriate.
Effects | Describe side effects (e.g., "May read and write memory.", "May read memory and throw.", "Does not read or write memory.").

2. Register in Instrs.def

Add a DEF_VALUE entry. Place it near related instructions:

DEF_VALUE(MyNewInst, Instruction)

If it's a subclass of another instruction, use the parent as the second argument.

If it's a terminator, use TERMINATOR instead of DEF_VALUE.

3. Define the class in Instrs.h

Do NOT add doc-comments to this file. Documentation belongs in doc/IR.md.

Follow this exact pattern:

class MyNewInst : public Instruction {
  MyNewInst(const MyNewInst &) = delete;
  void operator=(const MyNewInst &) = delete;

 public:
  enum { Arg1Idx, Arg2Idx };
  explicit MyNewInst(Value *arg1, Value *arg2)
      : Instruction(ValueKind::MyNewInstKind) {
    // Optional assertions on operand types:
    // assert(arg2->getType().isSomeType() && "message");

    // Set the result type:
    setType(Type::createNoType());  // for instructions with no output
    // or: setType(Type::createFoo()); for typed instructions

    pushOperand(arg1);
    pushOperand(arg2);
  }

  explicit MyNewInst(
      const MyNewInst *src,
      llvh::ArrayRef<Value *> operands)
      : Instruction(src, operands) {}

  Value *getArg1() const {
    return getOperand(Arg1Idx);
  }
  Value *getArg2() const {
    return getOperand(Arg2Idx);
  }

  static bool hasOutput() {
    return false;  // true if the instruction produces a value
  }
  static bool isTyped() {
    return false;  // true if the output type is meaningful
  }

  SideEffect getSideEffectImpl() const {
    // Compose side effects. Common patterns:
    //   return {};                                          // pure
    //   return SideEffect{}.setReadHeap();                  // reads memory
    //   return SideEffect{}.setReadHeap().setWriteHeap();   // reads+writes
    //   return SideEffect{}.setThrow().setReadHeap();       // may throw + read
    return SideEffect{}.setThrow().setReadHeap();
  }

  static bool classof(const Value *V) {
    ValueKind kind = V->getKind();
    return kind == ValueKind::MyNewInstKind;
  }
};

4. Add IRBuilder declaration in IRBuilder.h

MyNewInst *createMyNewInst(Value *arg1, Value *arg2);

5. Add IRBuilder implementation in IRBuilder.cpp

MyNewInst *IRBuilder::createMyNewInst(Value *arg1, Value *arg2) {
  auto *inst = new MyNewInst(arg1, arg2);
  insert(inst);
  return inst;
}

6. Add verification in IRVerifier.cpp

Add a visit method that checks invariants:

bool Verifier::visitMyNewInst(const MyNewInst &Inst) {
  AssertIWithMsg(
      Inst,
      Inst.getArg2()->getType().isSomeType(),
      "MyNewInst::Arg2 must be of SomeType");
  return true;
}

7. Add type inference in TypeInference.cpp

Add an infer method. For instructions without output, return createNoType():

Type inferMyNewInst(MyNewInst *inst) {
  return Type::createNoType();
}

8. Add code generation stubs

If the instruction is lowered before codegen, add fatal stubs. Otherwise,

implement the actual code generation.

HBC ISel (lib/BCGen/HBC/ISel.cpp):

void HBCISel::generateMyNewInst(MyNewInst *Inst, BasicBlock *next) {
  hermes_fatal("MyNewInst should have been lowered.");
}

SH (lib/BCGen/SH/SH.cpp):

void generateMyNewInst(MyNewInst &inst) {
  hermes_fatal("MyNewInst should have been lowered");
}

Mins (lib/BCGen/facebook/Mins/Mins.cpp):

Unless asked to, do not implement Mins codegen for new instructions.

Leave it as a stub.

void generateMyNewInst(MyNewInst &inst) {
  unimplemented(inst);
}

9. (If needed) Add a lowering pass

Declare in Lowering.h:

/// Brief description of what the lowering does.
Pass *createLowerMyNewInst();

Implement in Lowering.cpp:

Pass *hermes::createLowerMyNewInst() {
  class ThisPass : public FunctionPass {
   public:
    explicit ThisPass() : FunctionPass("LowerMyNewInst") {}

    bool runOnFunction(Function *F) override {
      IRBuilder builder{F};
      bool changed = false;

      // Collect instructions first to avoid iterator invalidation.
      llvh::SmallVector<MyNewInst *, 4> insts;
      for (auto &BB : *F) {
        for (auto &I : BB) {
          if (auto *MNI = llvh::dyn_cast<MyNewInst>(&I))
            insts.push_back(MNI);
        }
      }

      for (auto *MNI : insts) {
        // Replace MNI with lowered IR...
        MNI->eraseFromParent();
        changed = true;
      }

      return changed;
    }
  };
  return new ThisPass();
}

Register in pipelines:

In lib/BCGen/HBC/LoweringPipelines.cpp and in

lib/BCGen/SH/SH.cpp (lowerModuleIR), add PM.addPass(createLowerMyNewInst());

at the appropriate point (before any pass that would need to process instructions

introduced by the lowering).

10. Add or update tests

Add lit tests in the appropriate test/ subdirectory. Use %FileCheck or

%FileCheckOrRegen to verify the IR output. If existing tests cover the

feature, update their expected output to reflect the new instruction.

Which subdirectory to use depends on how the instruction interacts with the

compiler pipeline — not every instruction needs tests in every directory:

• test/IRGen/ — When the instruction is generated directly from JavaScript

source. Tests here verify that IRGen produces the expected IR.

• test/Optimizer/ — When the instruction affects or interacts with

optimization passes.

• test/BCGen/ — Not used as often, but if this IR instruction is used in

combination with new bytecode instructions, then place bytecode gen tests here.

Key Conventions

• No doc-comments in Instrs.h. All documentation goes in doc/IR.md.

The class in Instrs.h should have no /// or /** */ comments describing

what the instruction does. Brief inline comments explaining non-obvious

implementation details (like side effects) are fine.

• Placement matters. Place the new instruction near related instructions

in every file (e.g., private field instructions are grouped together).

• Consistent naming. The instruction name (e.g., FooBarInst) must be

consistent across all files: Instrs.def, Instrs.h, IRBuilder.h/cpp,

IRVerifier.cpp, TypeInference.cpp, and all codegen files.

• The ValueKind is derived automatically. When you add

DEF_VALUE(MyNewInst, Instruction) to Instrs.def, the enum value

ValueKind::MyNewInstKind is generated automatically.