The Glorious Driver (GD) is the part of GHC that orchestrates the interaction of all the other pieces that make up GHC. It supersedes the Evil Driver (ED), which was a Perl script that served the same purpose and was in use until version 4.08.1 of GHC. Simon Marlow eventually slayed the ED and instated the GD. The GD is usually called the Compilation Manager these days.
The GD has been substantially extended for GHCi, i.e., the interactive
variant of GHC that integrates the compiler with a (meta-circular)
interpreter since version 5.00. Most of the driver is located in the
directory
fptools/ghc/compiler/main/.
GHC's many flavours of command line options make the code interpreting
them rather involved. The following provides a brief overview of the
processing of these options. Since the addition of the interactive
front-end to GHC, there are two kinds of options: static
options and dynamic options. The former can only be set
when the system is invoked, whereas the latter can be altered in the
course of an interactive session. A brief explanation on the difference
between these options and related matters is at the start of the module
CmdLineOpts.
The same module defines the enumeration DynFlag, which
contains all dynamic flags. Moreover, there is the labelled record
DynFlags that collects all the flag-related information
that is passed by the compilation manager to the compiler proper,
hsc, whenever a compilation is triggered. If you like to
find out whether an option is static, use the predicate
isStaticHscFlag in the same module.
The second module that contains a lot of code related to the management
of flags is DriverFlags.hs.
In particular, the module contains two association lists that map the
textual representation of the various flags to a data structure that
tells the driver how to parse the flag (e.g., whether it has any
arguments) and provides its internal representation. All static flags
are contained in static_flags. A whole range of
-f flags can be negated by adding a -f-no-
prefix. These flags are contained in the association list
fFlags.
The driver uses a nasty hack based on IORefs that permits
the rest of the compiler to access static flags as CAFs; i.e., there is
a family of toplevel variable definitions in
CmdLineOpts,
below the literate section heading Static options, each of which
contains the value of one static option. This is essentially realised
via global variables (in the sense of C-style, updatable, global
variables) defined via an evil pre-processor macro named
GLOBAL_VAR, which is defined in a particularly ugly corner
of GHC, namely the C header file
HsVersions.h.
Inside the Haskell compiler proper (hsc), a whole series of
stages (``passes'') are executed in order to transform your Haskell program
into C or native code. This process is orchestrated by
main/HscMain.hscMain and its relative
hscReComp. The latter directly invokes, in order,
the parser, the renamer, the typechecker, the desugarer, the
simplifier (Core2Core), the CoreTidy pass, the CorePrep pass,
conversion to STG (CoreToStg), the interface generator
(MkFinalIface), the code generator, and code output. The
simplifier is the most complex of these, and is made up of many
sub-passes. These are controlled by buildCoreToDo,
as described below.
GHC has a large variety of optimisations at its disposal, many of which
have subtle interdependencies. The overall plan for program
optimisation is fixed in DriverState.hs.
First of all, there is the variable hsc_minusNoO_flags that
determines the -f options that you get without
-O (aka optimisation level 0) as well as
hsc_minusO_flags and hsc_minusO2_flags for
-O and -O2.
However, most of the strategic decisions about optimisations on the
intermediate language Core are encoded in the value produced by
buildCoreToDo, which is a list with elements of type
CoreToDo. Each element of this list specifies one step in
the sequence of core optimisations executed by the Mighty Simplifier. The type
CoreToDo is defined in CmdLineOpts.lhs.
The actual execution of the optimisation plan produced by
buildCoreToDo is performed by SimpleCore.doCorePasses.
Core optimisation plans consist of a number of simplification phases
(currently, three for optimisation levels of 1 or higher) with
decreasing phase numbers (the lowest, corresponding to the last phase,
namely 0). Before and after these phases, optimisations such as
specialisation, let floating, worker/wrapper, and so on are executed.
The sequence of phases is such that the synergistic effect of the phases
is maximised -- however, this is a fairly fragile arrangement.
There is a similar construction for optimisations on STG level stored in
the variable buildStgToDo :: [StgToDo]. However, this is a
lot less complex than the arrangement for Core optimisations.
RTS and libHSstd
Since the RTS and HSstd refer to each other, there is a Cunning
Hack to avoid putting them each on the command-line twice or
thrice (aside: try asking for `plaice and chips thrice' in a
fish and chip shop; bet you only get two lots). The hack involves
adding
the symbols that the RTS needs from libHSstd, such as
PrelWeak_runFinalizzerBatch_closure and
__stginit_Prelude, to the link line with the
-u flag. The standard library appears before the
RTS on the link line, and these options cause the corresponding
symbols to be picked up even so the linked might not have seen them
being used as the RTS appears later on the link line. As a result,
when the RTS is also scanned, these symbols are already resolved. This
avoids the linker having to read the standard library and RTS
multiple times.
This does, however, leads to a complication. Normal Haskell
programs do not have a main() function, so this is
supplied by the RTS (in the file
Main.c).
It calls startupHaskell, which
itself calls __stginit_PrelMain, which is therefore,
since it occurs in the standard library, one of the symbols
passed to the linker using the -u option. This is fine
for standalone Haskell programs, but as soon as the Haskell code is only
used as part of a program implemented in a foreign language, the
main() function of that foreign language should be used
instead of that of the Haskell runtime. In this case, the previously
described arrangement unfortunately fails as
__stginit_PrelMain had better not be linked in,
because it tries to call __stginit_Main, which won't
exist. In other words, the RTS's main() refers to
__stginit_PrelMain which in turn refers to
__stginit_Main. Although the RTS's main()
might not be linked in if the program provides its own, the driver
will normally force __stginit_PrelMain to be linked in anyway,
using -u, because it's a back-reference from the
RTS to HSstd. This case is coped with by the -no-hs-main
flag, which suppresses passing the corresonding -u option
to the linker -- although in some versions of the compiler (e.g., 5.00.2)
it didn't work. In addition, the driver generally places the C program
providing the main() that we want to use before the RTS
on the link line. Therefore, the RTS's main is never used and
without the -u the label __stginit_PrelMain
will not be linked.
Last modified: Tue Feb 19 11:09:00 UTC 2002