Record include(mapexpr) functions

[timholy]

Package cache headers store information needed to determine validity and a cache of source-text. The latter gets consumed by Revise, as a record of what the code was before edits were made. This allows Revise to skip defensively parsing every loaded package. However, there has long been a gap in this system: include(mapexpr, …) calls, which can modify the "effective" source text of the included files. Currently, Revise cannot even determine whether a package has used a non-identity mapexpr function without defensively parsing every included file.

Add Base.include_mapexprs(mod), returning the non-identity mapexpr functions passed to include(mapexpr, …) calls while loading the package rooted at mod, keyed by (including_module, absolute_path), or nothing when there were none.

These functions are recorded into a hidden const binding in the package's own root module as the includes run, so they serialize into the package image and remain available after the package is loaded from its cache.

This lets revision tools (e.g. Revise) re-apply the original transform when an include(mapexpr, …)-ed file is edited. Recording the actual function object is necessary because a mapexpr that captures runtime state cannot be reconstructed by re-parsing the source; the precompile environment is closed and deterministic, so the object is serializable.

[timholy]

On my bike ride home I asked myself, "if I wanted to break this, how would I do it?" And unfortunately there is a clear path: piracy. Suppose PkgA has a mapexpr that relies on some Base function foo, with the output Exprs depending on values returned by foo. PkgA does not depend on package Pirate. Pirate, being a bad citizen, redefines Base.foo causing it to return different values for some types that PkgA will pass into foo. Now a user does using Pirate, PkgA. Presto, Revise has the wrong idea about what code actually got serialized in the pkgimage.

This is tough because there is no monotonic world-age in which it works; you'd have to insist that mapexpr be precompiled and somehow call the invalidated version that loaded from the cache file. I can see one viable way to prevent that, however: require that mapexprs be @opaque. Thoughts?