Backend Internals - smarter/dotty GitHub Wiki

The code for the backend is split up by functionality and assembled in the objet GenBCode.

object GenBCode    --- [defines] -->        PlainClassBuilder       GenBCode also defines class BCodePhase, the compiler phase
     |                                              |
 [extends]                                      [extends]
     |                                              |
BCodeSyncAndTry    ---------------->        SyncAndTryBuilder
     |                                              |
BCodeBodyBuilder   ---------------->        PlainBodyBuilder
     |                                              |
BCodeSkelBuilder   ---------------->        PlainSkelBuilder
     |                                       /      |       \
 BCodeHelpers      ---------------->  BCClassGen BCAnnotGen ...  (more components)
     |      \       \
     |       \       \------------->  helper methods
     |        \       \------------>  JMirrorBuilder, JBeanInfoBuilder (uses some components, e.g. BCInnerClassGen)
     |         \
     |   BytecodeWriters  --------->        methods and classes to write byte code files
     |
 BCodeTypes        ---------------->        maps and fields for common BTypes, class Tracked, methods to collect information on classes, tests for BTypes (conforms), ...
     |
BCodeIdiomatic     ---------------->        utilities for code generation, e.g. genPrimitiveArithmetic
     |
  BCodeGlue        ---------------->        BType class, predefined BTypes

Data Flow

Compiler creates a BCodePhase, calls runOn(compilationUnits).

  • initializes fields of GenBCode defined in BCodeTypes (BType maps, common BTypes like StringReference)
  • initialize primitives map defined in scalaPrimitives (maps primitive members, like int.+, to bytecode instructions)
  • creates BytecodeWriter, JMirrorBuilder and JBeanInfoBuilder instances (on each compiler run)
  • buildAndSendToDisk(units): uses work queues, see below.
    • BCodePhase.addToQ1 adds class trees to q1
    • Worker1.visit creates ASM ClassNodes, adds to q2. It creates one PlainClassBuilder for each compilation unit.
    • Worker2.addToQ3 adds byte arrays (one for each class) to q3
    • BCodePhase.drainQ3 writes byte arrays to disk

Architecture

The architecture of GenBCode is the same as in Scalac. It can be partitioned into weakly coupled components (called "subsystems" below):

(a) The queue subsystem

Queues mediate between processors, queues don't know what each processor does.

The first queue contains AST trees for compilation units, the second queue contains ASM ClassNodes, and finally the third queue contains byte arrays, ready for serialization to disk.

Currently the queue subsystem is all sequential, but as can be seen in http://magarciaepfl.github.io/scala/ the above design enables overlapping (a.1) building of ClassNodes, (a.2) intra-method optimizations, and (a.3) serialization to disk.

This subsystem is described in detail in GenBCode.scala

(b) Bytecode-level types, BType

The previous bytecode emitter goes to great lengths to reason about bytecode-level types in terms of Symbols.

GenBCode uses BType as a more direct representation. A BType is immutable, and a value class (once the rest of GenBCode is merged from http://magarciaepfl.github.io/scala/ ).

Whether value class or not, its API is the same. That API doesn't reach into the type checker. Instead, each method on a BType answers a question that can be answered based on the BType itself. Sounds too simple to be good? It's a good building block, that's what it is.

The internal representation of a BType is based on what the JVM uses: internal names (eg Ljava/lang/String; ) and method descriptors; as defined in the JVM spec (that's why they aren't documented in GenBCode, just read the spec).

All things BType can be found in BCodeGlue.scala

(c) Utilities offering a more "high-level" API to bytecode emission

Bytecode can be emitted one opcode at a time, but there are recurring patterns that call for a simpler API.

For example, when emitting a load-constant, a dedicated instruction exists for emitting load-zero. Similarly, emitting a switch can be done according to one of two strategies.

All these utilities are encapsulated in file BCodeIdiomatic.scala. They know nothing about the type checker (because, just between us, they don't need to).

(d) Mapping between type-checker types and BTypes

So that (c) can remain oblivious to what AST trees contain, some bookkeepers are needed:

  • Tracked: for a bytecode class (BType), its superclass, directly declared interfaces, and inner classes.

To understand how it's built, see:

final def exemplar(csym0: Symbol): Tracked = { ... }

Details in BCodeTypes.scala

(e) More "high-level" utilities for bytecode emission

In the spirit of BCodeIdiomatic, utilities are added in BCodeHelpers for emitting:

  • bean info class
  • mirror class and their forwarders
  • android-specific creator classes
  • annotations

(f) Building an ASM ClassNode given an AST TypeDef

It's done by PlainClassBuilder