1) The bootstrap problem...

The current FriCAS compiler compiles each constructor (category, domain
or package) separately.  In order to perform type checking FriCAS
needs type information from other constructors.  To collect
type information about a constructor FriCAS must compile it.
This creates a bootstrap problem: in order to compile a constructor
one must first compile many other constructors.

In principle the bootstrap problem could be avoided by organizing algebra
into layers, so that layer 0 is independent of other layers and
each higher layer depends only on lower layers.  However, the current
FriCAS algebra contains a lot of cyclic dependencies: for example, a
substantial part of algebra depends on 'Integer', but 'Integer' has
a rich structure and depends on many categories which in turn depend
on 'Integer'.

2) ... and its solution

In the past (in Axiom era) significant part of bootstrap problem
was avoided simply by using type information collected in the earlier
and stored in databases 'interp.daase' and 'category.daase'.
This solution meant that any attempt to change the signature of an
exported operation was likely to cause build failure.

FriCAS bootstrap procedure uses different approach.  It proceeds
in several phases.  First phases of bootstrap collect database
information.  To save time first phases work in a single
Lisp image and keep database information only in memory.
Then we dump databases and use them for main compilation.

Using the type information stored in databases solves only part of
the bootstrap problem: internally, FriCAS types are represented by
executable code and sometimes FriCAS needs to load actual compiled
code.  FriCAS needs to load categories given as arguments to 'Join',
because 'Join' is implemented as a runtime operation.  Also
FriCAS needs to load a category given as an argument to a constructor
if this category itself has arguments.

During normal compilation FriCAS tries to optimize (i.e., expand inline)
calls to operations from core domains, which are listed in the
'$optimizableConstructorNames' variable.  To do this, FriCAS has
to load these domains.  In the Axiom period this optimization required
keeping compiled Lisp corresponding to a collection of categories and
domains sufficient to cut cycles involving core domains.
Currently FriCAS has a special flag '$bootstrapDomains' to
disable this optimization.

Even after using the '$bootstrapDomains' flag there are still
cyclic load dependencies caused by constructor arguments, but
those problem are resolved using the '$bootStrapMode' flag
which causes relaxed type checking and skips executable parts
of categories.

First three stages of bootstrap are done by 'boo_db.input'
input file.  In first stage FriCAS parses all of algebra
and saves resulting parse tree (using special '$SaveParseOnly'
flag to delay processing).  Next, the 'processGlobals' function
from 'src/interp/ncomp.boot' is run.  This function examines
(via 'processGlobals1') all constructor definitions and
computes basic database information about domains
and categories.  Next, 'processGlobals' compiles all
categories.  Normal category compilation immediately loads
freshly compiled category which causes load dependencies.
'processGlobals' compiles categories without loading them
which avoids most of dependencies (in particular avoids
dependencies due to arguments of 'Join').  'processGlobals'
avoids dependencies due to types of category arguments by
first compiling categories which have simple argument types
and only later categories with complicated argument types.
In this way complicated argument types can run  code of
categories which were compiled earlier.

When 'processGlobals' finishes the running FriCAS image has compiled
code of all categories and enough database information to
do '$bootStrapMode' compilation of all domains.  Note that
'processGlobals' in doing only in-memory work, nothing is
stored on disc.

The second bootstrap stage is '$bootStrapMode' compilation of
all constructors.

In the third stage we compile all categories and core domains
(the one which are subject to optimization, so must be
available for loading) using '$bootstrapDomains' flag.

At the end of stage three we dump databases.

In the next stage we compile categories and core domains in
normal mode (this is done by the 'oboo3.input' input file
generated from the makefile).

The first four stages are purely sequential.  After that
we a ready to compile the rest of algebra in arbitrary
order.  In fact, since all constructors which we need to
load were compiled earlier and all type information
needed for compilation is in databases there are no
dependencies.  So, the last bootstrap stage goes in
parallel.

At the end of last stage we rebuild databases to include
information not needed for compilation (basically info for
FriCAS browser) which were absent in earlier databases.

Note that after bootstrap running 'make' recompiles modified
files using old database information and old categories and
core domains ignoring any dependencies between modified files.
This means that 'make' may fail or produce wrong result
due to ignored dependencies.  However, changing only
implementation part, without change to list of exported
functions should work fine.  Also, adding new exported functions
that nobody uses is fine too.  Removing exported function and
all its users should work, but if somebody forgets about some
remaining use the compiler will not detect an error and the
failure will happen only later (at runtime or during next
recompilation).  Change that adds exported function and
tries to use it may fail.  Since after compilation 'make'
rebuilds databases such change may be split into two, one
which only adds exported function, and another which uses
new function.  Still, in case of more complicated changes
it is probably simpler just to re-run bootstrap.
