Adopt ReversePropagation 0.8 SSAFunction + RuntimeGeneratedFunctions API#228
Open
Adopt ReversePropagation 0.8 SSAFunction + RuntimeGeneratedFunctions API#228
Conversation
Contractor now lowers its expression to an SSAFunction once via cse_equations and passes that SSA to forward_backward_contractor. The SSA is stored on the contractor so it can be reused across multiple constraints over the same expression. Separator shares the SSA between its forward evaluator and its HC4Revise contractor, so each constructor does the symbolic CSE pass only once. make_function switches from eval to @RuntimeGeneratedFunction, matching RP's own shift: generated functions can now be built and invoked within the same dynamic extent without hitting Julia world-age errors, with much faster build time. Bumps compat to ReversePropagation = "0.8" and adds RuntimeGeneratedFunctions as a direct dep. ICP package version bumped to 0.16.0. Drive-by fix to a garbled .gitignore entry. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
IntervalArithmetic v1 taints any interval that touches a plain (non- ExactReal) Real with the "not guaranteed" (NG) flag. ICP's generated code is full of those — the literals from the user's expression (the 1 in x^2+y^2==1) and the integer exponents CSE inserts (the 2 in x^2) — so C.f([3..4, 5..6]) came back as [33.0, 51.0]_com_NG. Wrapping the literals inside the symbolic expression doesn't survive: Symbolics normalises ExactReal back to a plain number during its own simplification. The fix has to happen on the generated Expr. Add exactify(ex), which walks an Expr and wraps every non-Bool Real literal in IntervalArithmetic.exact. Thread an exact::Bool = false keyword through Contractor, Separator, and separator/constraint: - make_function applies exactify to the build_function output before @RuntimeGeneratedFunction when exact=true. - build_fb_contractor replicates RP's outer wrapper locally so it can exactify the body before RGF. When exact=false it forwards to forward_backward_contractor unchanged (no overhead in the default). Default behaviour is unchanged; opting in gives julia> constraint(x^2+y^2==1, vars; exact=true).f([3..4, 5..6]) [33.0, 51.0]_com Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
IEEE 1788 weakens intersect_interval(a, b) to the trv decoration because, in general, the result may not be a subset of the range of any function. Every reverse op in IntervalContractors is built on that same intersect, so both the outer intersect_interval(_value, _constraint) that RP emits and every narrowing step in the reverse sweep drag trv through the contractor output. In HC4Revise, however, each intersection/reverse-op is narrowing *within* an already-evaluated function range, so min(decoration(inputs)...) is the valid decoration — not trv. Add preserve_decorations(ex) in src/utils.jl: a walker that rewrites the body returned by ReversePropagation.forward_backward_expr so that every IntervalArithmetic.intersect_interval call and every ReversePropagation._rev_* reverse-op call has its output re-decorated with min(decoration(inputs)...). Because Symbolics' toexpr emits the call head as the actual function value (not Expr(:., Module, :name)), the matchers compare ex.args[1] === IntervalArithmetic.intersect_interval and parentmodule(f) === ReversePropagation && startswith(nameof(f), "_rev_") directly. _redecorate(x, d) unconditionally replaces the decoration (not min with the op's own output decoration, which is precisely the weakened trv we're overriding). Empty intervals stay trv (empty semantics). build_fb_contractor now runs preserve_decorations alongside exactify under exact=true. With both on: julia> constraint(x^2+y^2==1, vars; exact=true).contractor((-10..10, -10..10)) [-1.0, 1.0]_com² # was [-1.0, 1.0]_trv_NG² (The outer of the Separator triple can still end up _trv because Separator computes it via boundary ⊔ corner and hull intrinsically weakens under IEEE 1788 — intended semantics.) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Contractornow lowers its expression to aReversePropagation.SSAFunctiononce viacse_equationsand passes that SSA intoforward_backward_contractor(ssa, vars); the SSA is stored on the contractor so it can be reused across multiple contractors/constraints over the same expression.Separator(ex, vars, constraint)shares that SSA between its forward evaluator and its HC4Revise contractor, so the symbolic CSE pass runs once per separator instead of twice.make_functionswitches fromevalto@RuntimeGeneratedFunction, matching RP's own shift — generated functions can now be built and invoked within the same dynamic extent without Julia world-age errors, with much faster build time.[compat] ReversePropagation = "0.8", addsRuntimeGeneratedFunctions = "0.5"as a direct dep, and bumps package version to0.16.0..gitignoreentry.Blocker
This PR depends on ReversePropagation 0.8.0, which is not registered yet — it lands with the SSAFunction / RGF refactor currently split across the open PRs in that repo. Until it is released, CI here will fail to resolve; local reviewers can
]devthe local RP checkout (as this was developed and tested against).Test plan
Pkg.test()passes locally (Contractors 4/4, Separators 4/4, pave 7/7, Paving a 3D torus 1/1) against a dev'd RP 0.8.0🤖 Generated with Claude Code