Rewriter

Publications

• Jason Gross, Andres Erbsen, Miraya Poddar-Agrawal, Jade Philipoom, Adam Chlipala. Accelerating Verified-Compiler Development with a Verified Rewriting Engine. Proceedings of the Interactive Theorem Proving - Thirteenth International Conference (ITP'22). August 2022.
• Jason Gross, Andres Erbsen, Jade Philipoom, Rajashree Agrawal, and Adam Chlipala. Towards Towards a Scalable Proof Engine: A Performant Prototype Rewriting Primitive for Coq. Journal of Automated Reasoning, 68(3):19, Aug 2024.

Building

This repository requires Coq 8.19 or later, and requires that the version of OCaml used to build Coq be installed and accessible on the system.

Git submodules are used for some dependencies. If you did not clone with --recursive, run

git submodule update --init --recursive

To build:

make

Usage

The entry point is src/Rewriter/Util/plugins/RewriterBuild.v. (Better entry point coming in the future.) There are some examples in src/Rewriter/Rewriter/Examples.v.

The examples from the paper are in src/Rewriter/Rewriter/Examples/PerfTesting/. On the branch with performance data, the perf-*.txt files in the top-level directory contain selected performance data in two-column format, while the perf*.csv files contain all performance data in CSV format. The logs with raw timing data live in the subdirectories of src/Rewriter/Rewriter/Examples/PerfTesting/, and were made with the targets perf-SuperFast, perf-Fast, and perf-Medium targets. The text and csv files were generated with make perf-csv. (The targets perf-Slow and perf-VerySlow exist but take a very, very, very long time to run.)

License

This project is distributed under the terms of the MIT License, the Apache License (Version 2.0), and the BSD 1-Clause License; users may pick which license to apply.

See COPYRIGHT, LICENSE-MIT, LICENSE-APACHE, and LICENSE-BSD-1 for details.

Reading The Code

Actual Synthesis Pipeline

The entry point for clients of the PHOAS expressions we use is Language/API.v. Refer to comments in that file for an explanation of the interface; the following text describes how the expressions are generated, not how to interact with them.

The ordering of files (eliding *Proofs.v files) is:

Language/*.v
    ↑
Rewriter/*.v

Within each directory, the dependency graphs (again eliding *Proofs.v and related files) are:

Within Language/:

  PreCommon.v ←─────────── Pre.v
    ↑                        ↑
Language.v ←── IdentifiersBasicLibrary.v ←──── IdentifiersBasicGenerate.v
    ↑                        ↑
    ├────────────────┐       └────────────────────────────┐
UnderLets.v    IdentifiersLibrary.v ←──────────── IdentifiersGenerate.v
                     ↑                                       ↑
              IdentifiersLibraryProofs.v ←─── IdentifiersGenerateProofs.v

We will come back to the Rewriter/* files shortly.

The files contain:

• Rewriter/*.v: rewrite rules, rewriting.

• Inside Language/:

  • PreCommon.v: A few definitions which are used in writing out rewrite rules and the interpretations of PHOAS identifiers, e.g., ident.cast, ident.eagerly, etc.

  • Pre.v: A few definitions which are used in writing out rewrite rules, which are not used in the parameterized rewriter, only in reifying rewrite rules.

  • Language.v: Defines parts of the PHOAS basic infrastructure parameterized over base types and identifiers including:

    • PHOAS
    • reification
    • denotation/intepretation
    • utilities for inverting PHOAS exprs
    • default/dummy values of PHOAS exprs
    • default instantiation of generic PHOAS types
    • gallina reification of ground terms
    • Flat/indexed syntax trees, and conversions to and from PHOAS

    Defines the passes:

    • ToFlat
    • FromFlat
    • GeneralizeVar

  • IdentifiersBasicLibrary.v: Defines the package type holding basic identifier definitions.

  • IdentifiersBasicGenerate.v: Defines the tactics that generate all of the identifier-list-specific definitions used by the PHOAS machinery, in addition to defining the tactics that do reification based on the generated package.

  • UnderLets.v: the UnderLets monad, a pass that does substitution of var-like things, a pass that inserts let-binders in the next-to-last line of code, substituting away var-like things (this is used to ensure that when we output C code, aliasing the input and the output arrays doesn't cause issues). Defines the passes:

    • SubstVar
    • SubstVarLike
    • SubstVarOrIdent

  The following files in Language/ are used only by the rewriter:

  • IdentifiersLibrary.v: Some definitions about identifiers and pattern identifiers and raw identifiers. Some of these definitions take generated definitions as arguments. Also defines a package record to hold all of the generated definitions.

  • IdentifiersGenerate.v: Tactics to generate definitions about untyped and pattern versions of identifiers for the rewriter. Culminates in a tactic which inhabits the package type defined in IdentifiersLibrary.v

  • IdentifiersLibraryProofs.v: proofs about definitions in IdentifiersLibrary. Also defines a package to hold generated proofs that require destructing inductives not yet defined in this file.

  • IdentifiersGenerateProofs.v: tactics to prove lemmas to inhabit the package defined in IdentifiersLibraryProofs.v

The files defined in Rewriter/ are split up into the following dependency graph (including some files from Language/ at the top):

IdentifiersLibrary.v ←───────────────────────── IdentifiersGenerate.v
    ↑ ↑                                                   ↑
    │ └──────────────── IdentifiersLibraryProofs.v ←──────┴─ IdentifiersGenerateProofs.v
    │                                     ↑
    │                                     │
    │                                     │
    │                                     │
    │                                     │
Rewriter.v ←────────────────────── ProofsCommon.v ←──────────────────── ProofsCommonTactics.v
    ↑                                 ↗        ↖                                ↑
Reify.v ←──────────────┐           Wf.v   InterpProofs.v                        │
                       │              ↖        ↗                                │
                       └──────────── AllTactics.v ──────────────────────────────┘

• Rewriter.v: Defines the rewriter machinery. In particular, all of the rewriter definitions that have non-rewrite-rule-specific proofs about them are found in this file.

• RewrierReify.v: Defines reification of rewrite rules, continuing on from Rewriter.v, and culminates in the tactic RewriteRules.Tactic.Build_RewriterT and the tactic notation make_Rewriter which define a package of type RewriteRules.GoalType.RewriterT. The Build_* tactic returns a constr, while the make_* tactic notation refines that constr in the goal. Both tactics take two arguments: first a boolean include_interp which specifies whether (true) or not (false) to prefix the list of rewrite rules with the reduction-to-literal rewrite rules; and second a list of bool * Prop which is the list of rewrite rule types to reify, each paired with a boolean saying whether or not to try rewriting again in the output of the replacement for that rule.

• ProofsCommon.v: Defines the notion of interp-goodness and wf-goodness for rewrite rules, defines tactics to prove these notions, and contains a semi-arbitrary collection of proofs and definitions that are mostly shared between the wf proofs and the interp proofs. Importantly, this file defines everything needed to state and prove that specific rewrite rules are correct. Additionally defines a package RewriteRules.GoalType.VerifiedRewriter which describes the type of the overall specialized rewriter together with its Wf and Interp proofs. (This package should perhaps move to another file?)

• ProofsCommonTactics.v: Defines the actual tactics used to prove that specific rewrite rules are correct, and to inhabit the packages defined in ProofsCommon.v.

• Wf.v: Proves wf-preservation of the generic rewriter, taking in wf-goodness of rewrite rules as a hypothesis.

• InterpProofs.v: Proves interp-correctness of the generic rewriter, taking in interp-goodness of rewrite rules as a hypothesis.

• AllTactics.v: Defines the tactic RewriteRules.Tactic.make_rewriter (and a similar tactic notation) which build the entire VerifiedRewriter. They take in the include_interp as in Rewriter.v tactics, as well as an hlist of proofs of rewrite rules indexed over a list (bool * Prop) of rewrite rule types. This is the primary interface for defining a rewriter from a list of rewrite rules.

Proofs files: For Language.v, there is a semi-arbitrary split between two files Language.Inversion and Language.Wf.

• Inversion.v:

  • classifies equality of type codes and exprs
  • type codes have decidable equality
  • correctness of the various type-transport definitions
  • correctness lemmas for the various expr.invert_* definitions
  • correctness lemmas for the various reify_* definitions in Language.v
  • inversion_type, which inverts equality of type codes
  • type_beq_to_eq, which converts boolean equality of types to Leibniz equality
  • rewrite_type_transport_correct, which rewrites with the correctness lemmas of the various type-transport definitions
  • type.invert_one e which does case analysis on any inductive type indexed over type codes, in a way that preserves information about the type of e, and generally works even when the goal is dependently typed over e and/or its type
  • ident.invert, which does case-anaylsis on idents whose type has structure (i.e., is not a var)
  • ident.invert_match, which does case-analysis on idents appearing as the discriminee of a match in the goal or in any hypothesis
  • expr.invert, which does case-anaylsis on exprs whose type has structure (i.e., is not a var)
  • expr.invert_match, which does case-analysis on exprs appearing as the discriminee of a match in the goal or in any hypothesis
  • expr.invert_subst, which does case-analysis on exprs which show up in hypotheses of the form expr.invert_* _ = Some _
  • expr.inversion_expr, which inverts equalities of exprs

• Wf.v: Depends on Inversion.v Defines:

  • expr.wf, expr.Wf, expr.wf3, expr.wf4, expr.Wf3, expr.Wf4
  • GeneralizeVar.Flat.wf
  • expr.inversion_wf (and variants), which invert wf hypotheses
  • expr.wf_t (and variants wf_unsafe_t and wf_safe_t) which make progress on wf goals; wf_safe_t should never turn a provable goal into an unprovable one, while wf_unsafe_t might.
  • expr.interp_t (and variants), which should make progress on equivalence-of-interp hypotheses and goals, but is not used much (mainly because I forgot I had defined it)
  • prove_Wf, which proves wf goals on concrete syntax trees in a more optimized way than repeat constructor Proves:
  • funext → (type.eqv ↔ Logic.eq) (eqv_iff_eq_of_funext)
  • type.related and type.eqv are PERs
  • Proper instances for type.app_curried, type.and_for_each_lhs_of_arrow
  • type.is_not_higher_order → Reflexive (type.and_for_each_lhs_of_arrow type.eqv)
  • iff between type.related{,_hetero} and related of type.app_curried
  • various properties of type.and{,b_bool}for_each_lhs_of_arrow
  • various properties of type.eqv and ident.{gen_,}interp
  • various properties of ident.cast
  • various properties of expr.wf (particularly of things defined in Language.v)
  • interp and wf proofs for the passes to/from Flat

• UnderLetsProofs.v: wf and interp lemmas for the various passes defined in UnderLets.v

Extended Walkthrough

There's an extended description of much of the codebase in rewriting-extended.md.