Home My Page Projects Code Snippets Project Openings SML/NJ
Summary Activity Forums Tracker Lists Tasks Docs Surveys News SCM Files

SCM Repository

[smlnj] View of /sml/trunk/compiler/DEVNOTES/Primops/primop2-note-1
ViewVC logotype

View of /sml/trunk/compiler/DEVNOTES/Primops/primop2-note-1

Parent Directory Parent Directory | Revision Log Revision Log


Revision 5026 - (download) (annotate)
Thu May 2 11:44:23 2019 UTC (2 months, 2 weeks ago) by jhr
File size: 25465 byte(s)
changed InLine to Inline
Primop2 Notes

Primop2 Branch Info
===================

The development branch name is:

  primop-branch-2

The tag for the root of primop-branch-2 is:

  dbm-20060615-base-primop-branch-2


======================================================================
The Program

Our goal is to clean-up and simplify the code of the SML/NJ Front End
(everything before the translate phase), and as far as possible turn it
into a self-contained SML front end "library" that could be used outside
the SML/NJ compiler. There are several related subgoals:

1. A major part of this effort aims to remove dependencies on FLINT
constructs and interfaces from the front end. This requires that we
understand what the FLINT-related values and computations are used for
so that we can reproduce that functionality in the translate phase,
ultimately producing the same FLINT intermediate language.

2. We want to avoid transformations of the abstract syntax before type
checking, since these transformations reduce the faithfulness of the
abstract syntax to the original source and lead to less useful and less
accurate type error messages. Abstract syntax transformations/normalizations
that have presumably been introduced to be helpful for FLINT should be
be moved into the translate phase.  All we need to do is make sure
that the abstract syntax captures and retains all static information that
the translate phase will need to produce the correct plambda code.

3. We want to simplify the FLINT code (and ultimately the object code)
by eliminating the real array => realarray representation
optimization.  We need to understand what parts of the translation is
specifically supporting this optimization, so that those parts can be
simplified.


======================================================================
Questions:

1. Where have wordTycon and wordTy gone?
In primop, they were defined in ElabData/types/core-basictypes.sml.
Now they are in Elaborator/types/basictypes.s??

1.1. What is the reason for the split of basic type building between
core-basictypes (in ElabData/types) and basictypes (in
Elaborator/types)?

2. What is the purpose of the ty lists in VARexp and CONexp
constructors of exp in ElabData/syntax/absyn.s??
Should be able to replace list of metavariables by (instantiated)
variable type. Or, since in most cases this is not relevant (it's
represented by [] currently), it should be a type option.  This type
would the the instantiated "variable type", i.e. based on a generic
instantiation of the typ field of the VALvar, etc. [We could just use
UNDEFty as the default, unneeded value instead of NONE, but using
ty option is cleaner and more explicit about what is going on.]

Files affected:
 ElabData/syntax/absyn.s??  [done - except for option]
 Elaborator/types/typecheck.sml [done - except for option]
 Elaborator/modules/sigmatch.sml [?]
 FLINT/trans/translate.sml [?]

3. Question in comment on generalizePat in Elaborator/types/typecheck.sml.

4. Why do CONpats (see CONpat case of patType in typecheck.sml) need
the "insts" value? Similarly for CONexps. What do they have to do with
primops? Where is this information used?  Is this related to the
special role of ref?

5. Catalogue of primops:
We need a complete catalog of all primops, with their initial types (primop
generic type), and the modules and types where they are ultimately defined
in the Basis modules.

6. Numbered primops:
Perhaps in the front end, primops can be respresented by a simple number
(defined in the basic primop module). The mapping between primop numbers
and their generic types could be implemented as a table or finite map.
Generic primops that are used at different types could be split into a set
of primop numbers, one for each final (variable) type. FLINT would be
responsible for recognizing that several primop numbers represent the same
generic underlying primop.  Then the primop numbers could be mapped to their
exact (variable) type in the Inline module. One problem here is that some
type constructors appearing in variable types are not defined (yet) in
the PrimTypes module -- presumably we could fix this by adding them as
additional primite types. Examples are probably array and vector tycons.
notes

7. InlInfo datatype ii/selectors and utility functions
Why is the module ElabData/Basics/II (which defines the datatype ii
and selectors) separate from module Semant/Basics/InlInfo (which
defines utility functions and on the datatype ii)?

Functions in InlInfo appear to be used only in the TopLevel profiling
hooks. The remaining use is in the ElabData/types/TypesUtil.isValue
function. Unfortunately, there is some funny business going on in that
the functions in InlInfo require the PrimOp module which is not
available in ElabData/types/TypesUtil.

We've merged these into a single module InlInfo. This depends on PrimOp.
This caused the sigmatch functor to be moved to Semant/modules.


8. Dynamic access and inline info.  Are these mutually exclusive (i.e.
something has a dynamic value accessed via the access info, or it is
a primop accessed via the inline info), or to they both work in parallel
(i.e. a primop also has a function closure representation stored in
a slot in the dynamic vector for its home structure)?

9. Structure/substructure primop IDs and signature slot information
are split among the structure and signature representations
respectively. Are substructure primop IDs (and other substructure
components such as realization, access, signature) available directly
in the static environment?

No. Although the root structure strrec's are entered into the static
environment, all substructure information can only be found within the
strrec of the root structure.


======================================================================
Old Primop Notes (from old primop branch, mostly obsolete)

These notes are my attempt to figure out what appears to be a very kludgy
treatement of how primops interact with the type checker.  I want to define
what the basic problem is, understand what the current code is doing to deal
with the problem, and then either eliminate the problem or find a more elegant
solution.

1. PrimOp dependency in type checker

The isValue function in TypesUtil needs to check for the CAST primop, which
it does using InlInfo.pureInfo (badly named function).  Thus

	TypeCheck --> TypesUtil --> InlInfo --> PrimOp

Before, this was factored out by making TypeCheck into a functor TypeCheckFn
which was then applied (in Semant/types/typecheck.sml) to InlInfo.pureInfo.

The other dependency on ii2ty is eliminated by having VARexp carry the
whole instantiated type rather than the list of generic meta-type-variables
(with their instantiations).


2. Role of boundtvs

"boundtvs" is a field of the VB (variable binding) record, containing
a list of tyvars.

Where is boundtvs set?

Most VB constructions throughout the front end (e.g. all in ElabCore)
set boundtvs to [] or pass a value of boundtvs through
(Elaborator/elaborate/elabutil.sml).  There are 3 places where it is
potentially set to a non-nil list:

(1) TypeCheck/decType0 (Elaborator/types/typecheck.sml).
In the VALdec case, variable bindings are constructed where boundtvs
is the set of meta tyvars that are generalized by generalizePat
applied to the pattern.

[From the way generalizePat is written, there can be only one variable
in the pattern whose type is polymorphic (i.e. has tyvars generalized).]


(2) SigMatch/matchElems (Elaborator/modules/sigmatch.sml).
In the cases for VALspec/VALspec, and VALspec/CONspec. Here the value for
boundtvs (btvs) is computed by SigMatch/matchTypes, which in turn calls
SigMatch/eqvTnspTy after checking that the spec type and actual type are
compatible.

eqvTnspTy (mnemonic for what?) creates a generic instantiation of the
actual type (actInst, with generic meta-tyvars actInstTvs), a separate
generic instantiation (specInst, specInstTvs) of the specty type,
which may be less polymorphic than actualty, and then unifies actInst
and specInst.  This unification may cause tyvars in actInstTvs to be
instantiated.  Then specInstTvs is transformed by TU.tyvarType to
produce the second result, boundTvs.

TU.tyvarType turns type expressions that represent meta type variables
into the tyvars that they contain, stripping off VARtys and following
instantiations.  Note that instantiatePoly returns a list of type
_expressions_ (VARty's) as its second result.

    fun eqvTnspTy (specty: ty, actualty: ty) : (ty * tyvar list) =
      let val actualty = TU.prune actualty
	  val (actInst, actInstTvs) = TU.instantiatePoly actualty
	  val (specInst, specInstTvs) = TU.instantiatePoly specty
	  val _ = ((Unify.unifyTy(actInst, specInst))
		   handle _ => bug "unexpected types in eqvTnspTy")
	  val boundTvs = map TU.tyvarType specInstTvs
       in (actInst, boundTvs)
      end

The unification will cause some tyvars in actInstTvs to be
instantiated to non-type-variable type terms (subterms of specInst),
where specty is more specific than actualty). Where corresponding
subterms are both type variables, the specty tyvar (a member of
specInstTvs) will be instantiated to the corresponding actualty type
variable (a member of actInstTvs) because of the right-to-left
instantiation bias in Unify/unifyTyvars [Elaborator/types/unify.sml].

So the resultant type term actInst may contain (uninstantiated) tyvars
in specInstTvs as well as some (uninstantiated) tyvars in actInstTvs.
For the latter tyvars, the matching tyvar in specInstTvs will have
been instantiated to it.  So the specInstTvs will contain all the
tyvars occuring in actInst, some of which will be uninstantiated,
while others will be instantiated to tyvars in actInstTvs.  Applying
TU.tyvarType will strip off the instantiations, so that the resulting
boundtvs will contain a mixture of the tyvars from actInstTvs and
specInstTvs.

Example:

actualty  = 'a * 'b -> 'b -> 'a
specty    = 'c list * 'c -> 'c -> 'c list
actInst   = 'X * 'Y -> 'Y -> 'X
actInstTvs = ['X, 'Y]
specInst  = 'Z list * 'Z -> 'Z -> 'Z list
specInstTvs = ['Z]
-------------------
after unification:
-------------------
actInst   = {{'Y/'Z} list/'X} * 'Y -> {{'Y/'Z} list/'X}
boundtvs  = [tyvar('Y)]

where {ty/'X} represents the meta type variable 'X instantiated
to type ty.

If unifyTyvars had a left-to-right instantiation bias instead of a
right-to-left bias, the unification would instantiate all the tyvars
in actInstTvs and none of the tyvars in specInstTvs, so the resultant
type actInst will be equivalent to specInst -- note that in this
special case unification will not have to adjust any type variable
attributes, since all depths will be infinity and the consistency of
equality properties will already have been guaranteed by the previous
call of TU.compareTypes in matchTypes.

In this case, eqvTnspTy could just return specInst and specInstTvs
without bothering to perform the useless unification.  It is not clear
why this would not suffice anyway, i.e. eqvTnspTy could simply be
replaced by TU.instantiatePoly applied to specty, with tyvarType then
mapped over the resulting specInstTvs. The simplified version of
eqvTnspTy would be

    fun eqvTnspTy (specty: ty) : (ty * tyvar list) =
      let val (specInst, specInstTvs) = TU.instantiatePoly specty
       in (specInst, map TU.tyvarType specInstTvs)
      end

and the actualty argument would not be needed.

Maybe the interaction is more complicated if, for instance, specty
contains type abbreviations that are not used in the corresponding
places in actualty?  Unification would then cause these type
abbreviations to be expanded. But how could this matter?


(3) SigMatch/packElems in the VALspec case.
Here the value btvs for boundtvs is calculated by SigMatch/absEqvTy.
absEqvTy is similar to eqvTnspTy, in that it generically instantiations
the actual and spec types and then unifies them.

fun absEqvTy (specty: ty, actualty: ty) : (ty * tyvar list * ty * bool) =
  let val actualty = TU.prune actualty
      val specty = TU.prune specty
      val (actInst, actInstTvs) = TU.instantiatePoly actualty
      val (specInst, specInstTvs) = TU.instantiatePoly specty
      val _ = ListPair.app Unify.unifyTy (actInstTvs, specInstTvs)

      val res = (Unify.unifyTy(actInst, specInst); true) handle _ => false
      (* ???? what is this doing? *)

      val boundTvs = map TU.tyvarType actInstTvs
      (* should I use specInstTvs here instead, why actInstTvs ? *)

   in (actInst, boundtvs, specInst, res)
  end

This function is defined in a very obscure way, and returns both the
instantiated actualTy (actInst) and instantiated specTy (specInst),
after attempting to unify them.  But first, _some_ of the type
variables of actInstTvs and specInstTvs are identified by the
ListPair.app unifyTy line (note that these two lists may have
different lengths, and it is not clear why identifying initial
segments of these two lists of type variables makes sense.

If the unification is successful, then actInst and specInst are
essentially the same type.  If the unification fails, then we have two
inequivalent, but partially unified type expressions.  What is going
on here?


3. VARexp type checking hack for primops.

Here is the code for the VARexp case in expType (Elaborator/types/typecheck.sml):

     case exp
      of VARexp(r as ref(VALvar{typ, info, ...}), _) =>
	 (case ii2ty info of
	      SOME st =>
              let val (sty, insts) = instantiatePoly(st)
		  val (nty, _) = instantiatePoly(!typ)
              in
		  unifyTy(sty, nty) handle _ => ();  (* ??? *)
		  (VARexp(r, insts), sty)
              end
	    | NONE =>
	      let val (ty, insts) = instantiatePoly(!typ)
	      in (VARexp(r, insts), ty)
	      end)

This uses ii2ty (passed as a parameter to the TypeCheckingFn, but uniquely
defined to be InstantiateParam.ii2ty in Semant/types/typecheck.sml) to extract
a type from the info field, if possible (i.e. when the VALvar is associated
with a primop).

The reason for this use of unification seems to be that there are some primops
(e.g. length) that are used for several different operations (therefore
different VALvars), and the VALvar types are speciallized instances of a
more polymorphic type associtated with the primop (in the info field).


What is the ultimate consumer of the boundtv field?


Relevant Files and Modules
--------------------------
Primitive Types

  CorePrimTycNum (ElabData/basics/core-ptnum.sml)
      core prim-type numbers (essentially the same as Elaborator/basics/ptnum.sml
      except only language-standard types are included)
      see: Elaborator/basics/ptnum.sml
      defs: CORE_PRIM_TYC_NUM, CorePrimTycNum

  PrimTycNum (Elaborator/basics/ptnum.sml)
      prim type numbers, augmenting ElabData/basics/core-ptnum.sml with
      implementation dependent types
      see: ElabData/basics/core-ptnum.sml
      defs: PRIM_TYC_NUM, PrimTycNum: PRIM_TYC_NUM

  CoreBasicTypes (ElabData/types/core-basictypes.sml)
      building the primitive types and associated values (containing only
      implementation independent, language-standard basic types?)
      see: Elaborator/types/basictypes.sml
      defs: CoreBasicTypes

  BasicTypes     (Elaborator/types/basictypes.sig,sml)
      define basic (built-in) types; most are just defined in terms of
      types from CoreBasicTypes [MOVE to ElabData?]
      see: ElabData/types/core-basictypes.sml
      defs: BasicTypes

  PrimEnv: PRIM_ENV (Semant/statenv/prim.sml)
      define static env primEnv, containing module bindings for
      primitive types (PrimTypes), primops (Inline), and unrolled
      lists (UnrolledList)
      defs: PRIM_ENV, PrimEnv: PRIM_ENV


Primops

  PrimOp  (FLINT/kernel/primop.sml)
      the types (primop, etc.) used to represent primops and their
      attributes. There are no type specs in these primop representations.
      PrimOp imports CTypes, which are defined in MLRISC! (MLRISC/c-calls/ctype.sml)

  PrimEnv (Semant/statenv/prim.sml)
      builds Inline structure, containing bindings to primops
      this is where all primops are actually created and initially bound
      to identifiers. They will be rebound later in InLineT (below), and
      then again in the Basis module definitions. What role does the
      intermediate InLineT play?

      Matthias revised this to assign accurate (final?) types to the
      primop variables. Do these change in InLineT or in the Basis modules?

  InLineT  (../system/smlnj/init/built-in.sml)


Inline Info

  II  (ElabData/basics/ii.sml)
      FLINT related information for inlining (why is this in elaborator?)
      datatype ii used for "inline info".  Actual atomic inline info
      represented here by type exn (a hack to delay referencing the actual
      inline info representation defined in
      defs: II

  InlInfo  (Semant/basics/inlinfo.sig,sml)
    inlinfo.sig/sml
      functions for creating and examining inlining info, using an exception
      E to package primop-type pairs into the II.ii datatype.
      defs: INL_INFO, InlInfo: INL_INFO

  [II has been folded into InlInfo, and the code of InlInfo has been
   simplified, eliminating the exn hack. Interface function names have
   bin changed (e.g. pureInfo => isPrimCast). ii2ty has been added with
   new name primopTy (moved from InstantiateParam).]


Semant

  ElabData and Elaborate CM groups cannot refernce FLINT and MLRISC stuff,
  which are part of the Core group.  Therefore places where the front end
  depend on FLINT (e.g. PrimOp, PLambdaType (in InstantiateParam)) have to
  be in Semant, which is part of Core group.

  This means that many of the elaboration modules are functorized, ultimately
  on PrimOp related values (former TypeCheckFn) or on PLambdaType stuff
  (InstantiateFn in Elaborate requires parameter InstantiateParam from Semant).


======================================================================

Example for VARexp typing:

	val x = Array.length

let

   v = the VALvar for Array.length.
   st = ii2ty v.info = 'a -> int  (the generic primop type)
   sty = X -> int  (generic instantiation of st, X a fresh type metavariable)
   insts = [X]

   v.typ = 'a array -> int
   nty = Y array -> int  (generic inst. of v.typ, Y a fresh type metavar)

   unify (sty, nty) ==>  X |-> Y array

   After type generalization of the val binding:
   boundtvs = [Y]  (in VB)
   insts = [X]     (in VARexp)

These last don't agree (different metavariables), so in mkVBS in
Translate (FLINT/trans/translate.sml), the first VB case will take the
else branch and apply mkPE.

It appears that the boundtvs is used in mkPE (Translate) to provide type
parameters to primop invocations.  [Why boundtvs and not insts?]


======================================================================

[FLINT] Questions:

The primop-branch-2 code contains a number of questions in comments
marked by the string "[KM ???]".  You can find them by grepping the
source for that string.

The following questions are in more or less random order, added as
they came up in our reading of the code.  Most but not all relate to
the interaction of FLINT and the front end, or to FLINT internals.


1. What is the relevance of the comment in system/smlnj/init/dummy.sml
to the code in that file?


2. What is the meaning of the comment after the PrimTypes
(re)declaration at the top of system/smlnj/init/built-in.sml?


3. The Single Generalization Conjecture (SGC): Each type metavariable
that is not instantiated is generalized at a single variable binding.
Consequence: the metavariable can be updated with a deBruijn index for
the corresponding generalized polymorphic variable.  [The code for
mkPE in translate.sml seems to imply this is not the case.]

If the SGC is true, then the "restore" function in Translate.mkPE is
unnecessary, as the metavariables in question will only be used once.

The SGC has been tested experimentally by removing the call of restore
in mkPE, and bootstrapping the compiler.  The runtime check in mkPE
that tests for previously used generalized metavariables was not
triggered.


4. More on Single Generalization Conjecture

Here is an expression containing two mutually recursive functions that
have the same return type, which will be polymorphic.  It seems likely
that the return type will be represented by the same type metavariable
in both parts of the mutually recursive definition, and that this
metavariable is generalized twice, once for the type of f and once for
g.  This needs to be verified by a careful look at the type checking
process though.

let fun f () = g ()
    and g () = f ()
in ()
end;

Another variant would that would show the types of f and g would be:

fun h () =
let fun f () = g ()
    and g () = f ()
in (f,g)
end;

val h = fn : unit -> (unit -> 'a) * (unit -> 'b)

Here it is clear that f and g have independent polymorphic types,
indicating two independent generalizations.


5. Two lexp types in FLINT
There are two types named lexp: PLambda.lexp (in FLINT/plambda/plambda.sml)
and FLINT.lexp (in FLINT/flint/flint.sml).  The two lexp's have quite different
definitions. What are the roles of these two lexp types?


6. mkAccInfo in Translate (FLINT/trans/translate.sml) ignores the
second argument (formerly info, now prim).  Why is this?  Doesn't this
mean that primops get lost?


7. What is the point of PACKexp?  In mkExp in Translate (FLINT/trans/translate.sml)
it is just stripped off.


8. Complex abstract syntax for a simple structure declaration.

A simple declaration like

   structure X = struct val x = Array.length end

generates a surprising complex abstract syntax representation, looking
something like (using some shorthand notations) ...

   STRB{name = "X",
	str = bindStr,
	def = LETstr (
		STRdec [
		  STRB{name = "<tempStr>",
		       str = resStr
		       def = LETstr (
			       VB{pat = vv_x, exp = VARexp(vv_len)), ...}
			       STRstr([VALbind(vv_x)])}],
		VARstr(resStr)
	      )
       }

where resStr is the structure record returned for the rhs strexp by
elabStr, and bindStr is the same as resStr except the dacc field is
replaced by a new lvar, vv_x is the VALvar for the bound variable x,
and vv_len is the VALvar for Array.length.

This strips down to a skeleton like:

    let lv_b =             (* lv_b = lvar for bindStr *)
        let lv_r = ...     (* lv_r = lvar for resStr *)
         in let <component var decls>
	     in [component vars]
        in lv_r

What is the purpose of the outer layer of let binding, and the shift
in lvars from lv_r to lv_b?


9. In Absyn, why do CONexp's also have a tyvar list parameter (like VARexp?).
Is this only relevant to a few special constructors (maybe ref?), or is it
also used for ordinary user defined constructors?  Similarly for CONpat.

----------------------------------------------------------------------

10. In ltykernel.sml (lty.sml), the datatype rflag only has one variant,
and therefore doesn't seem to convey any useful information.  Was this
introduced to make some useful distinction in the future, but this
distinction was never used or implemented?  Can we get rid of this
useless flag?

11. In ltyextern.sml, the subkinding relation is actually an equality
relation because there are both rules:

BOX <: MONO
MONO <: BOX

Is there is reason why we are not just using equality for now? Do we
really need a subkinding relation?  Or was this complication introduced
in anticipation of some more refined analysis that never got implemented?

12. The kind checker in LtyExternal (tkChkGen) is incomplete, in that
it does not and cannot properly check closures (Env terms). When kind
checking an Env(tyc,i,j,tenv) expr, we must type check the tenv as well
as the body, and we need to extract kind environment info from the tenv
to use when type checking the body. This means we need to add kind
specifications to each binder in tenv, so we can translate tenv into
(a prefix of) a kind environment. We have modified tycEnv to include
the kind specifications at each binding level, and we have modified the
kind checker to do a complete check of Env tycs.  The kind checker is
now defined in kernel/lty.sml (structure Lty) and re-exported from
LtyExtern.

13. In translate.sml, mkVar is calling mkAccInfo, which just drops the
 prim!!!

  fun mkVar (v as V.VALvar{access, prim, typ, path}, d) =
	mkAccInfo(access, prim, fn () => toLty d (!typ), getNameOp path)
    | mkVar _ = bug "unexpected vars in mkVar"




======================================================================
Invariants for tyc and lty equivalence (and type normalization)

* TC_IND tycs will never have normal flag true, similarly for LT_IND ltys
and LT_TYC(TC_IND _) ltys.  Conversely, if the normal flag is true, the
body should not be an IND form.

* tycs returned by tc_lzrd, tc_whnm, and tc_norm with not be of the TC_IND form.
Similarly, ltys returned by lt_lzrd, lt_whnm, and lt_norm will not be of the
LT_IND or LT_TYC(TC_IND _) forms.

* AUX_REG(true_,_) is computed only by the hashing constructors (*_injX),
which detect normal forms when calling tc_mk and lt_mk to compute the
aux value for a type term being hashed.

* upd_tyc and upd_lty only operate on non-normal first args
(AUX_REG(false,_,_) or NO_AUX), and their result (the updated
first argument), retain this property of not having a true normal flag.

* tc_eq and lt_eq (pointer equality for tycs and ltys) will only be applied
to arguments with normalization flag set true.

root@smlnj-gforge.cs.uchicago.edu
ViewVC Help
Powered by ViewVC 1.0.0