|Version 3 (modified by simonpj, 6 years ago) (diff)|
Overview of modules in the new code generator
This page gives an overview of the new code generator, including discussion of:
The new Cmm data type
There is a new Cmm data type:
- compiler/cmm/ZipCfg.hs contains a generic zipper-based control-flow graph data type. It is generic in the sense that it's polymorphic in the type of middle nodes and last nodes of a block. (Middle nodes don't do control transfers; last nodes only do control transfers.) There are extensive notes at the start of the module.
The key types it defines are:
- Block identifiers: BlockId, BlockEnv, BlockSet
- Control-flow blocks: Block
- Control-flow graphs: Graph
- ZipDataFlow contains a generic framework for solving dataflow problems over ZipCfg. It allows you to define a new optimization simply by defining a lattice of dataflow facts (akin to a specialized logic) and then writing the dataflow-transfer functions found in compiler textbooks. Handing these functions to the dataflow engine produces a new optimization that is not only useful on its own, but that can easily be composed with other optimizations to create an integrated "superoptimization" that is strictly more powerful than any sequence of individual optimizations, no matter how many times they are re-run. The dataflow engine is based on (Lerner, Grove, and Chambers 2002); you can find a functional implementation of the dataflow engine presented in (Ramsey and Dias 2005).
- compiler/cmm/ZipCfgCmmRep.hs instantiates ZipCfg for Cmm, by defining types Middle and Last and using these to instantiate the polymorphic fields of ZipCfg. It also defines a bunch of smart constructor (mkJump, mkAssign, mkCmmIfThenElse etc) which make it easy to build CmmGraph.
- CmmExpr contains the data types for Cmm expressions, registers, and the like. Here is a fuller description of these types is at Commentary/Compiler/BackEndTypes. It does not depend on the dataflow framework at all.
Module structure of the new code generator
The new code generator has a fair number of modules, which can be split into three groups:
- basic datatypes and infrastructure
- analyses and transformations
- linking the pipeline
All the modules mentioned are in the cmm/ directory, unless otherwise indicated.
Basic datatypes and infrastructure
- CLabel (CLabel): All sorts of goo for making and manipulating labels.
- BlockId (BlockId, BlockEnv, BlockSet): The type of a basic-block id, along with sets and finite maps.
- CmmExpr (CmmType, LocalReg, GlobalReg, Area, CmmExpr): Lots of type definitions: for Cmm types (bit width, GC ptr, float, etc), registers, stack areas, and Cmm expressions.
- Cmm (GenCmm, CmmInfo, CmmInfoTable): More type definitions: the parameterized top-level Cmm type (GenCmm), along with the type definitions for info tables.
- ZipCfg (Graph, LGraph, Block): Describes a zipper-like representation for true basic-block control-flow graphs. A block has a single entry point, which is a always a label, followed by zero or mode 'middle nodes', each of which represents an uninterruptible single-entry, single-exit computation, then finally a 'last node', which may have zero or more successors. ZipCFG is polymorphic in the type of middle and last nodes.
- ZipCfgCmmRep (Middle, Last, CmmGraph) Types to instantiate ZipCfg for C--: middle and last nodes, and a bunch of abbreviations of types in ZipCfg and Cmm.
- MkZipCfg (AGraph, mkLabel, mkMiddle, mkBranch) Smart constructors for control-flow graphs (and the constructors have non-monadic types). Like ZipCfg, MkZipCfg is polymorphic in the types of middle and last nodes.
- MkZipCfgCmm (mkNop, mkAssign, mkStore, mkCall, ...) Smart constructors for creating middle and last nodes in control-flow graphs (and the constructors have non-monadic types).
- CmmInfo (cmmToRawCmm, mkBareInfoTable): Converts Cmm code to "raw" Cmm. What this means is: convert a CmmInfo data structure describing the info table for each CmmProc to a [CmmStatic]. mkBareInfoTable is the workhorse that produces the [CmmStatic]. It is also used to produce the info table required for safe foreign calls (a middle node).
- CmmCallConv (ArgumentFormat, assignArgumentsPos): Implements Cmm calling conventions: given arguments and a calling convention, this module decides where to put the arguments. (JD: Crufty. Lots of old code in here, needs cleanup.)
- CmmTx (Tx, TxRes): A simple monad for tracking when a transformation has occurred (something has changed). Used by the dataflow analysis to keep track of when the graph is rewritten.
- OptimizationFuel (OptimizationFuel, FuelMonad, maybeRewriteWithFuel) We can use a measure of "fuel" to limit the number of rewrites performed by a transformation. This module defines a monad for tracking (and limiting) fuel use. (JD: Largely untested.)
- DFMonad (DataflowLattice, DataflowAnalysis, runDFM): Defines the type of a dataflow lattice and an analysis. Defines the monad used by the dataflow framework. The monad keeps track of dataflow facts, along with fuel, and it can provide unique id's. All in support of the dataflow module.
- ZipDataflow (ForwardTransfers, BackwardTransfers, ForwardRewrites, BackwardRewrites, zdfSolveFrom, zdfRewriteFrom, etc) This module implements the Lerner/Grove/Chambers dataflow analysis frameword. Given the definitions of a lattice and dataflow transfer/rewrite functions, this module provides all the work of running the dataflow analysis and transformation. A number of the phases of the back end rely on this code, and hopefully more optimizations will target it in the future.
And a few basic utilities:
- CmmZipUtil: (JD: Unused, I believe, but probably should be used in a few places.) A few utility functions for manipulating a zipcfg.
- PprC: Prettyprinting to generate C code.
- PprCmm: Prettyprinting the C-- code.
- PprCmmZ: (JD: Unused, I believe.) Prettyprinting functions related to ZipCfg and ZipCfgCmm.
Analyses and transformations
- CmmLint (cmmLint, cmmLintTop): Some sanity checking on the old Cmm graphs. Not sure how effective this is.
- CmmLiveZ (CmmLive, livelattice, cmmLivenessZ): Liveness analysis for registers (uses dataflow framework).
- CmmProcPointZ (ProcPointSet, callProcPoints, minimalProcPointSet, procPointAnalysis, splitAtProcPoints) A proc point is a block in a control-flow graph that must be the entry point of a new procedure when we generate C code. For example, successors of calls and joinpoints that follow calls are procpoints. This module provides the analyses to find procpoints, as well as the transformation to split the procedure into pieces. The procpoint analysis doesn't use the dataflow framework, but it really should - dominators are the way forward.
- CmmSpillReload (DualLive, dualLiveLattice, dualLiveness, dualLivenessWithInsertion, insertLateReloads, removeDeadAssignmentsAndReloads): Inserts spills and reloads to establish the invariant that at a safe call, there are no live variables in registers.
- CmmCommonBlockElimZ (elimCommonBlocks): Find blocks in the CFG that are identical; merge them.
- CmmContFlowOpt (branchChainElimZ, removeUnreachableBlocksZ, runCmmOpts): Branch-chain elimination and elimination of unreachable code.
- CmmStackLayout (SlotEnv, liveSlotAnal, manifestSP, stubSlotsOnDeath): The live-slot analysis discovers which stack slots are live at each basic block. We use the results for two purposes: stack layout (manifestSP) and info tables (in CmmBuildInfoTables). The function `stubSlotsOnDeath' is used as a debugging pass: it stubs each stack slot when it dies, hopefully causing bad programs to fail faster.
- CmmBuildInfoTables (CAFEnv, cafAnal, lowerSafeForeignCalls, setInfoTableSRT, setInfoTableStackMap): This module is responsible for building info tables. Specifically, it builds the maps of live variables (stack maps) and SRTs. It also has code to lower safe foreign calls into a sequence that makes them safe (but suspending and resuming threads very carefully). (JD: The latter function probably shouldn't be here.)
Linking the pipeline
- CmmCvt: Converts between Cmm and ZipCfgCmm representations. (JD: The Zip -> Cmm path definitely works; haven't tried the other in a long time -- there's no reason to use it with the new Stg -> Cmm path).
- CmmCPSZ: Links the phases of the back end in sequence, along with some possible debugging output.
- CmmCPSGen, CmmCPS (Michael Adams), CmmBrokenBlock, CmmLive, CmmPprCmmZ, StackColor, StackPlacements