Aug 7, 2019, 5:08:58 PM (5 years ago)
Michael Brooks <mlbrooks@…>
ADT, arm-eh, ast-experimental, enum, forall-pointer-decay, jacob/cs343-translation, master, new-ast, new-ast-unique-expr, pthread-emulation, qualifiedEnum
ba417e2, c570806

Two fixes of MemberExpr? handling on new-AST.

Each bug on its own, and also both bugs together, produce the same symptom: Expressions that use function-pointer typed members of structs, where the function has a generically-typed argument, were lacking casts (vs old-AST) in the output of:

./driver/cfa-cpp --prelude-dir ./libcfa/x64-debug/prelude -tm ../cfa-cc/libcfa/prelude/bootloader.cf bootloader.c

First bug: a conversion inaccuracy, which is a regression of convert-convert-post-resolve, introduced at Jun 28 rev 55b647. The constructor of a MemberExpr?, that Convert old-to-new uses, does type-variable substitution. This is intended behaviour when it occurs during resolve, but is incorrect when done twice. In absence of this fix, the old-to-new converter runs this substitution a second time, both in: oldresolve-old2new-new2old and old2new-newresolve-new2old. The fix is offering/using a no-op constructor to/at the converter.

Second bug: a mutation of a transitively-multiply-referenced object. The type-variable substitution just mentioned, which should be happening to the result type of the MemberExpr? and not to the top-level struct-member declaration, actually happens as:

  • old ast: as should; all type-value intentions are clones
  • new ast pre this fix: incorrectly to both types; ref-counted sharing means theExpr.result == theDecl.type, substitution happens on sub-object (see details in code comment)
  • new ast with this fix: as should; deep copy of the type is explicitly forced upfront

All bootloader output differences (new-AST's resolver vs old-AST's resolver) that remain after this fix are: four constructor-destructor calls in builtin function bodies, and a missing copy-constructor call on invoke-main. These differences were present ahead of this change.

This change has been tested for a clean convert-convert difference (vs no new-AST at all), for convert-convert at: pre-resolve, post-resolve, post-parse. In the case of cvcv post-resolve, this clean difference is being restored from the first-bug regression.

This change has been tested for, and causes a small-to-questionable regression of, cfa-cpp speed of compiling bootloader. Old AST = 6.04 s, New AST before this change = 4.64 s, New ast after this change 4.76 s. All numbers are mean of 5 samples with sample SD ~= 0.3 sec. New-AST improvement before this fix = 23% = 4.2 SDs. New-AST improvement after this fix = 21% = 3.9 SDs.

1 edited


  • src/AST/Convert.cpp

    r2890212 rae265b55  
    21242124                                old->location,
    21252125                                GET_ACCEPT_1(member, DeclWithType),
    2126                                 GET_ACCEPT_1(aggregate, Expr)
     2126                                GET_ACCEPT_1(aggregate, Expr),
     2127                                ast::MemberExpr::NoOpConstructionChosen
    21272128                        )
    21282129                );
Note: See TracChangeset for help on using the changeset viewer.