Mimosa compilation phases
Mimosa is primarily based on the Cedar 6.0 & Cedar 7.0 releases of the Cedar/Mesa compiler. Its pass structure, algorithms, and internal structure are inherited, rather than created, which leads to some confusion at times. There has also been substantial revision, which creates a different variety of confusion.
Pass 1 - Scanning & Parsing
Definitions: MimP1
Exports: CompilerUtil, ScannerExtras
Implementations: MimScanner Pass1 MimParserImpl Pass1T Pass2
The scanner is an ad hoc program that accepts an input stream and produces a stream of tokens. The implementation is in MimScanner.mesa. The scanner is meant to be compatible with the scanner accessible through the Cedar system (see IOScanImpl.mesa). Unfortunately, the scanners are not identical, and can give different results, although this is regarded as erroneous.
The parser is a table-driven LALR(1) parser with tables produced by a separate program (PGS - Parser Generation System). Essentially the parser accepts a stream of tokens and produces an abstract syntax tree. The productions and associated actions used by PGS to produce the tables are contained in the file Pass1T.pgs. PGS also produces the file Pass1T.mesa, which is a renumbered version of Pass1T.pgs. The mesa code in Pass1T.pgs (copied to Pass1T.mesa) specifies the actions at each production, and specialized comments specify the productions.
There is a current limitation of 256 (maybe 255) productions imposed by PGS. There is some indication that the PGS code was meant to handle more, but it has not been made to work. [RRA]
Pass 2 - Semantic table building
Definitions: none
Exports: CompilerUtil
Implementations: Pass2
This pass (implemented by Pass2) scans the tree, producing some simplifications, and constructs semantic entries for all type constructors, body table and context entries for all observed bodies, and other housekeeping.
Pass 3 - Type checking
Definitions: MimP3 MimP3S Pass3Attributes Pass3
Exports: CompilerUtil, MimosaCopier
Implementations: MimosaCopierImpl, Pass3AttributesImpl Pass3B Pass3D Pass3I Pass3M Pass3S Pass3V Pass3Xa Pass3Xb Pass3Xc Pass3P
Pass3 performs most of the work in the compiler not related to actual code generation: type checking, definitions reading, and symbol copying. Much of the work comes from making sense out of generic operations like generic arithmetic.
Type checking is carried out up to but not including range checking, so subranges are considered to be equivalent to base types.
MimosaCopierImpl - symbol copying
MimosaCopierImpl copies symbols (and trees) from symbols files into the current symbol table within the compiler.
Pass3B - DIRECTORY processing
Pass3D - declaration processing
Pass3I - identifier processing
Pass3I performs most declaration and scope processing for identifiers. Pass3I also performs RC map completion and REF literal processing.
Pass3M - miscellaneous processing
Pass3M handles type checking for a variety of statements and expressions. Pass3M.MiscStmt handles the following nodes: signal, error, start, join, wait, xerror, resume, reject, continue, retry, restart, stop, notify, broadcast, free, and enable. Pass3M.MiscStmt even handles some obsolete forms (dst, lst, lste, lstf). Pass3M.New handles new (New), cons (Cons) and listcons (ListCons) nodes (all called by Pass3XB). Unusual control transfers are handled by Pass3M.MiscXfer (signalx, errorx, fork, joinx, create, startx). Catch phrases are handled by Pass3M.CatchPhrase.
Pass3S - statement processing
general statement processing
Pass3V - variant processing
variant record processing
Pass3Xa, Pass3Xb, Pass3Xc - expression processing
general expression processing
Pass3P -
INLINE expansion
Pass3P performs most of the inline expansion, transforming calls on inline procedures into
subst nodes (for calls that are statements) and
substx nodes (for calls that are expressions). In many ways this module is a completely separate pass.
There is a funny problem with inline expansion due to the fact that trees are saved at the end of Pass4 processing, while they are expanded before Pass4 processing. Someday this crock should be fixed! [RRA]
Pass 4 - Range checking & constant folding
Pass 4 performs the range checking part of type checking, which requires that constant folding be performed as well. During Pass 4 the sizes of all types and variables are determined, as well as the layout of records (the bit positions of all fields).
The constant folding is still a bit of a mess since there are still implicit assumptions about the big-endian nature of the target, yet we eventually want to target little-endian machines. For now most constant-fodling is disabled for record and array constructors. This problem is partly relieved by constant expression collection and evaluation during global frame initialization. However, this area still needs work. [RRA]
Pass4B - Binary file processing
Pass4B deals with Pass 4 initialization and with object files. Pass4B assigns links to imported items in AssignImports. Pass4B determines exported type mismatches in ExportId.
Pass4L - Layout
Pass4L lays out records, frames, interfaces, and so forth. Ideally all decisions would be determined by parameters, but this still needs some work.
Pass4L has facilities (currently not used) for performing naive local variable allocation using static frequency. Unfortunately, on some machines (like Dragon), variables have penalties for being placed in arbitrary locations, and register assignment is better left to machine dependent modules anyway.
Pass4L puts out the LinkFrag part of the mob file (indicating which items are imported).
Pass 5 - Intermediate code generation
Definitions: MimCode MimGenRC MimP5 MimP5S MimP5U
Exports: ??
Implementations: MimCalls MimP5UImpl MimCons MimExpr MimFlow MimSelect MimStmt MimStore MimDriver
??
MimP5UImpl - utility routines
MimCalls - generation of procedure calls
MimCons - constructors
MimExpr - expressions
general expression generation: binary & unary arithmetic
MimFlow - flow of control
relational expressions, IF, labels, EXIT, LOOP, GO TO, arithmetic conversions
MimSelect - selectors
SELECT, WITH ... SELECT
ISTYPE, NARROW
type & tag predicates
MimStmt - statements
general statements
MimStore - storing of values
assignment, extraction, indexing of arrays
MimDriver - module level generation
modules, procedures, returns, monitors locks, proc descriptors
Int Code - Intermediate code transformations
Definitions: ??
Exports: IntCodeTarget, IntCodeTwig, IntCodeUtils, ParseIntCode, TargetConversions
Implementations: IntCodeTargetImpl, IntCodeTwigImpl, IntCodeUtilsImpl, ParseIntCodeImpl, TargetConversionsImpl
IntCodeTwigImpl - intermediate code transformations
IntCodeTwigImpl handles transformations on the intermediate code. Most of these transformations have to do with the procedure model. It is this module that handles argument record and return record limits, nested procedures, and many others. See IntCodeTwigDoc for more information.
Utilities
Definitions: ??
Exports: ??
Implementations: SymbolCacheImpl, SymbolOpsImpl, TreeOpsImpl, LiteralOpsImpl, SymLiteralOpsImpl, TypesImpl, MimFilesImpl, BcdOutput, MesaTab, ConstArithImpl, MimSysOpsImpl
Control & Data
Definitions: ??
Exports: ??
Implementations: DebugTab, ErrorTab, FileParmsImpl, MimCommandUtilImpl, MimData, MimosaDebug, MimosaInterface, MimosaLogImpl, MimosaSequencer
Other
Definitions: MimosaType2CType
Exports: MimosaType2CType
Implementations: MimosaType2CTypeImpl
MimosaType2CTypeImpl - C type hacking
MimosaType2CTypeImpl turns a Mimosa type description into a C type description. This translation is only partial, and has not been verified.