Skip to content

IR_TODOS.md

Latest commit

 

History

History
57 lines (49 loc) · 3.24 KB

IR_TODOS.md

File metadata and controls

57 lines (49 loc) · 3.24 KB

IR Implementation Progress

Plan: Structural IR Entities (robust-plotting-rocket.md)

Phase 1: FunctionKind + sourceDeclEntityId rename — COMPLETE

Phase 2: IRStructure entity + scope tracking — COMPLETE

  • StructureKind enum (18 kinds), IRStructure entity, IRStructureId
  • Structure capnp schema, entity providers, pack/unpack
  • Structure generation (PushStructure/PopStructure) for all control flow
  • FUNCTION_SCOPE, nested SCOPE for CompoundStmt, implicit SCOPE for for-init
  • ENTER_SCOPE/EXIT_SCOPE opcodes on all paths (including break/continue/return/goto)
  • Block parentStructureId, scope object association via AssociateObjectWithScope
  • IRFunction::body_scope(), IRBlock::parent_structure() accessors
  • SWITCH_CASE structures for each case/default
  • IRSwitchCase → IRStructure(SWITCH_CASE) full migration (deferred)
  • Full Python bindings (stub exists, requires bootstrap regen)

Phase 3: Control flow region structures — MERGED INTO PHASE 2

Phase 4: Global initializer functions — COMPLETE

  • GenerateGlobalInit creates synthetic FunctionIR for globals with initializers
  • FunctionKind::GLOBAL_INITIALIZER, sourceDeclEntityId = VarDecl
  • Entry block with ADDRESS_OF → EmitRValue(init) → STORE → RET
  • VarDecl maps to its GLOBAL_INITIALIZER IRFunction

Known Extensions — PARTIALLY COMPLETE

  • MEMSET (opcode 68): dest, byte_value, size — lowered from memset/builtin calls
  • MEMCPY (opcode 69): dest, src, size — lowered from memcpy/memmove/builtin calls
  • MemsetInst, MemcpyInst instruction class wrappers
  • VAR_INIT block kind + structure kind (for variable initialization regions)

Known Issues / Lies Remaining

  1. IRSwitchCase still separate entity type (coexists with SWITCH_CASE structures)
  2. string_bytes() missing on IRObject for string literals
  3. Python bindings are stub only for IRStructure
  4. Global initializer quality depends on EmitRValue handling of all init expressions
  5. Goto/Duff's device scope compensation: goto that jumps into the middle of a scope bypasses the normal ENTER_SCOPE path. An interpreter following the goto would not see the scope entry for variables in that scope. Similarly, Duff's device-style switch cases can interleave with loop bodies, creating scope entry paths that don't go through ENTER_SCOPE. We need "compensation blocks" — synthetic blocks inserted on goto/case edges that emit the ENTER_SCOPE instructions for any scopes being entered. This is tricky because the label/case might also be reachable from normal control flow (which already has the ENTER_SCOPE), so we can't just add ENTER_SCOPE at the label — we need it on the specific edge. Always using compensation blocks (even for the normal case) would be the simplest correct solution.

Decisions Made

  • Phase 3 merged into Phase 2
  • SWITCH_CASE structures coexist with IRSwitchCase for backward compat
  • ENTER_SCOPE/EXIT_SCOPE carry IRStructureId extra in entity pool
  • StructureKind embedded in IRStructureId (18 sub_kind offsets)
  • Global initializer uses same EmitRValue path as function body codegen
  • MEMSET/MEMCPY lowered from memset/memcpy/memmove and all _builtin variants
  • Goto conservatively exits all scopes to FUNCTION_SCOPE