GitLab

Some constraints were missing, some constraints were too coarse.
Also, switched to using [FixSolver] instead of a home-made system.

that is, before %inlining.

Only in internal form, for now -- does not exist in concrete syntax.
Not yet supported by the back-ends.

Updated the code back-end to produce the same positions as ocamlyacc and the table back-end, in principle.
This involves:
1- moving the [endp] field (when it exists) to offset 1 in every stack cell,
   so it can be found when reducing an epsilon production;
2- creating a sentinel cell with an [endp] field at offset 1;
3- changing the code for epsilon productions to read the [endp] field in the
   top stack cell.

This should fix an old bug where the user prelude could hide [Error].

Renamed to [listif] and [elementif].

  defined at different places in the generated file (in order to avoid
  name clashes, and for clarity).

['a lr1state -> 'a symbol] and maps a state (an abstract datum)
to (a code for) its incoming symbol. This allows inspecting the
semantic values stored in the stack.
Updated the demo [calc-incremental] to display the contents of
the stack at every token. It seems to work!

more general and correct (the old code could change a set of nonrecursive
type definitions into a set of recursive type definitions).

[IL.structure]. This makes the IL more reasonable and flexible. In
principle, the generated code is unchanged.

to insert comments a posteriori. Now done a priori where appropriate.

the type [terminal], which is a GADT. Produced only in [--table]
mode. The flags [--only-tokens] and [--external-tokens] are respected.

This removes the dependency of [Interface] on [PreFront].

Explicitly parameterize [TokenType] over a grammar.
This removes the dependency of [TokenType] on [PreFront].

by reading the positions from the lexbuf just in time. No speed
difference, but this should imply less frequent minor collections.

the environment record instead of updating it. This makes
the code back-end TWICE FASTER.

function is no longer required. This does not change the generated
code, but simplifies things slightly and reduces the generation time
(less work for the inliner).

to a Boolean field [env.error]. This yields slightly simpler
code and a small speed increase (maybe 5%).

in the special case where the language is empty or {epsilon} is now fixed.

The first token is now read during the first call to [run].
This exchanges the order of the first two messages in the trace,
but is actually more regular.
In the code back-end, this fixes a "bug" in the case where the grammar
accepts only the empty word. In the table back-end, the bug is not yet fixed.

A lot of cosmetic changes in order to avoid warnings (mostly unused variables).

git-svn-id: svn+ssh://scm.gforge.inria.fr/svnroot/menhir@320 0f8b5475-4b4e-0410-85a8-ee3154a6bfe7