COMP 3103 · Progressive Lowering: From Ideas to Machines

Led by Lattnerian Compiler Simulacrum

From high-level to machine code through intermediate stages — and why staging matters for every complex translation.

1. Module 1

   What Is Progressive Lowering?

   Led by Lattnerian Compiler Simulacrum

   The question

   Swift source → SIL → LLVM IR → machine code. What does each level see that the others cannot? And why does no single representation see everything?

   Outcome

   The student can trace a function through four compilation stages and identify level-specific optimisations.

   Sub-units

   1. ○ 1.1 Four Stages Open this unit → ✓ ↻
   2. ○ 1.2 Level-Specific Optimisations Open this unit → ✓ ↻
   Begin the course → Print certificate of completion
2. Module 2

   MLIR: Generalising Progressive Lowering

   Led by Lattnerian Compiler Simulacrum

   The question

   MLIR has multiple intermediate representations — dialects — each optimised for its domain. TF dialect → Linalg → LLVM IR → machine code. What does the ML compilation problem require that one IR level cannot provide?

   Outcome

   The student can explain MLIR's dialect concept and trace an ML computation through multiple levels.

   Sub-units

   1. ○ 2.1 Trace an ML Computation Open this unit → ✓ ↻
   Open this module → Print certificate of completion
3. Module 3

   Progressive Lowering Outside Compilers

   Led by Lattnerian Compiler Simulacrum

   The question

   Legal drafting, business strategy, recipe development, teaching — all involve translating from a high-level, ambiguous description to a low-level, executable specification. Where are the natural intermediate stages, and what does each one make explicit?

   Outcome

   The student can identify progressive lowering in a non-software domain and explain each stage's contribution.

   Sub-units

   1. ○ 3.1 Non-Software Lowering Open this unit → ✓ ↻
   Open this module → Print certificate of completion
4. Module 4

   The Cost of Lowering

   Led by Lattnerian Compiler Simulacrum

   The question

   Every lowering step loses information. Optimised code is hard to debug because the optimiser has discarded source structure. What is the trade-off between optimisation and debuggability — and can it be managed?

   Outcome

   The student can explain the information-loss trade-off and the debug information problem.

   Sub-units

   1. ○ 4.1 The Debuggability Trade-off Open this unit → ✓ ↻
   Open this module → Print certificate of completion
5. Module 5

   The Right Stages

   Led by Lattnerian Compiler Simulacrum

   The question

   Too few stages: each does too much. Too many: overhead without benefit. The right stages are the natural levels of description — where an expert would pause and say "I need to decide X before proceeding to Y." How do you find them?

   Outcome

   The student can take a defended position on how to design a progressive lowering pipeline.

   Sub-units

   1. ○ 5.1 Final Essay: How Many Stages? Open this unit → ✓ ↻
   Open this module → Print certificate of completion