------------------------------------------------------------------------------ -- -- -- GNAT COMPILER COMPONENTS -- -- -- -- E X P _ U N S T -- -- -- -- S p e c -- -- -- -- Copyright (C) 2014-2017, Free Software Foundation, Inc. -- -- -- -- GNAT is free software; you can redistribute it and/or modify it under -- -- terms of the GNU General Public License as published by the Free Soft- -- -- ware Foundation; either version 3, or (at your option) any later ver- -- -- sion. GNAT is distributed in the hope that it will be useful, but WITH- -- -- OUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY -- -- or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License -- -- for more details. You should have received a copy of the GNU General -- -- Public License distributed with GNAT; see file COPYING3. If not, go to -- -- http://www.gnu.org/licenses for a complete copy of the license. -- -- -- -- GNAT was originally developed by the GNAT team at New York University. -- -- Extensive contributions were provided by Ada Core Technologies Inc. -- -- -- ------------------------------------------------------------------------------ -- Expand routines for unnesting subprograms with Table; with Types; use Types; package Exp_Unst is -- ----------------- -- -- The Problem -- -- ----------------- -- Normally, nested subprograms in the source result in corresponding -- nested subprograms in the resulting tree. We then expect the back end -- to handle such nested subprograms, including all cases of uplevel -- references. For example, the GCC back end can do this relatively easily -- since GNU C (as an extension) allows nested functions with uplevel -- references, and implements an appropriate static chain approach to -- dealing with such uplevel references. -- However, we also want to be able to interface with back ends that do -- not easily handle such uplevel references. One example is the back end -- that translates the tree into standard C source code. In the future, -- other back ends might need the same capability (e.g. a back end that -- generated LLVM intermediate code). -- We could imagine simply handling such references in the appropriate -- back end. For example the back end that generates C could recognize -- nested subprograms and rig up some way of translating them, e.g. by -- making a static-link source level visible. -- Rather than take that approach, we prefer to do a semantics-preserving -- transformation on the GNAT tree, that eliminates the problem before we -- hand the tree over to the back end. There are two reasons for preferring -- this approach: -- First: the work needs only to be done once for all affected back ends -- and we can remain within the semantics of the tree. The front end is -- full of tree transformations, so we have all the infrastructure for -- doing transformations of this type. -- Second: given that the transformation will be semantics-preserving, -- we can still used the standard GCC back end to build code from it. -- This means we can easily run our full test suite to verify that the -- transformations are indeed semantics preserving. It is a lot more -- work to thoroughly test the output of specialized back ends. -- Looking at the problem, we have three situations to deal with. Note -- that in these examples, we use all lower case, since that is the way -- the internal tree is cased. -- First, cases where there are no uplevel references, for example -- procedure case1 is -- function max (m, n : Integer) return integer is -- begin -- return integer'max (m, n); -- end max; -- ... -- end case1; -- Second, cases where there are explicit uplevel references. -- procedure case2 (b : integer) is -- procedure Inner (bb : integer); -- -- procedure inner2 is -- begin -- inner(5); -- end; -- -- x : integer := 77; -- y : constant integer := 15 * 16; -- rv : integer := 10; -- -- procedure inner (bb : integer) is -- begin -- x := rv + y + bb + b; -- end; -- -- begin -- inner2; -- end case2; -- In this second example, B, X, RV are uplevel referenced. Y is not -- considered as an uplevel reference since it is a static constant -- where references are replaced by the value at compile time. -- Third, cases where there are implicit uplevel references via types -- whose bounds depend on locally declared constants or variables: -- function case3 (x, y : integer) return boolean is -- subtype dynam is integer range x .. y + 3; -- subtype static is integer range 42 .. 73; -- xx : dynam := y; -- -- type darr is array (dynam) of Integer; -- type darec is record -- A : darr; -- B : integer; -- end record; -- darecv : darec; -- -- function inner (b : integer) return boolean is -- begin -- return b in dynam and then darecv.b in static; -- end inner; -- -- begin -- return inner (42) and then inner (xx * 3 - y * 2); -- end case3; -- -- In this third example, the membership test implicitly references the -- the bounds of Dynam, which both involve uplevel references. -- ------------------ -- -- The Solution -- -- ------------------ -- Looking at the three cases above, the first case poses no problem at -- all. Indeed the subprogram could have been declared at the outer level -- (perhaps changing the name). But this style is quite common as a way -- of limiting the scope of a local procedure called only within the outer -- procedure. We could move it to the outer level (with a name change if -- needed), but we don't bother. We leave it nested, and the back end just -- translates it as though it were not nested. -- In general we leave nested procedures nested, rather than trying to move -- them to the outer level (the back end may do that, e.g. as part of the -- translation to C, but we don't do it in the tree itself). This saves a -- LOT of trouble in terms of visibility and semantics. -- But of course we have to deal with the uplevel references. The idea is -- to rewrite these nested subprograms so that they no longer have any such -- uplevel references, so by the time they reach the back end, they all are -- case 1 (no uplevel references) and thus easily handled. -- To deal with explicit uplevel references (case 2 above), we proceed with -- the following steps: -- All entities marked as being uplevel referenced are marked as aliased -- since they will be accessed indirectly via an activation record as -- described below. -- An activation record is created containing system address values -- for each uplevel referenced entity in a given scope. In the example -- given before, we would have: -- type AREC1T is record -- b : Address; -- x : Address; -- rv : Address; -- end record; -- type AREC1PT is access all AREC1T; -- AREC1 : aliased AREC1T; -- AREC1P : constant AREC1PT := AREC1'Access; -- The fields of AREC1 are set at the point the corresponding entity -- is declared (immediately for parameters). -- Note: the 1 in all these names is a unique index number. Different -- scopes requiring different ARECnT declarations will have different -- values of n to ensure uniqueness. -- Note: normally the field names in the activation record match the -- name of the entity. An exception is when the entity is declared in -- a declare block, in which case we append the entity number, to avoid -- clashes between the same name declared in different declare blocks. -- For all subprograms nested immediately within the corresponding scope, -- a parameter AREC1F is passed, and all calls to these routines have -- AREC1P added as an additional formal. -- Now within the nested procedures, any reference to an uplevel entity -- xxx is replaced by typ'Deref(AREC1.xxx) where typ is the type of the -- reference. -- Note: the reason that we use Address as the component type in the -- declaration of AREC1T is that we may create this type before we see -- the declaration of this type. -- The following shows example 2 above after this translation: -- procedure case2x (b : aliased Integer) is -- type AREC1T is record -- b : Address; -- x : Address; -- rv : Address; -- end record; -- -- type AREC1PT is access all AREC1T; -- -- AREC1 : aliased AREC1T; -- AREC1P : constant AREC1PT := AREC1'Access; -- -- AREC1.b := b'Address; -- -- procedure inner (bb : integer; AREC1F : AREC1PT); -- -- procedure inner2 (AREC1F : AREC1PT) is -- begin -- inner(5, AREC1F); -- end; -- -- x : aliased integer := 77; -- AREC1.x := X'Address; -- -- y : constant Integer := 15 * 16; -- -- rv : aliased Integer; -- AREC1.rv := rv'Address; -- -- procedure inner (bb : integer; AREC1F : AREC1PT) is -- begin -- Integer'Deref(AREC1F.x) := -- Integer'Deref(AREC1F.rv) + y + b + Integer_Deref(AREC1F.b); -- end; -- -- begin -- inner2 (AREC1P); -- end case2x; -- And now the inner procedures INNER2 and INNER have no uplevel references -- so they have been reduced to case 1, which is the case easily handled by -- the back end. Note that the generated code is not strictly legal Ada -- because of the assignments to AREC1 in the declarative sequence, but the -- GNAT tree always allows such mixing of declarations and statements, so -- the back end must be prepared to handle this in any case. -- Case 3 where we have uplevel references to types is a bit more complex. -- That would especially be the case if we did a full transformation that -- completely eliminated such uplevel references as we did for case 2. But -- instead of trying to do that, we rewrite the subprogram so that the code -- generator can easily detect and deal with these uplevel type references. -- First we distinguish two cases -- Static types are one of the two following cases: -- Discrete types whose bounds are known at compile time. This is not -- quite the same as what is tested by Is_OK_Static_Subtype, in that -- it allows compile time known values that are not static expressions. -- Composite types, whose components are (recursively) static types. -- Dynamic types are one of the two following cases: -- Discrete types with at least one bound not known at compile time. -- Composite types with at least one component that is (recursively) -- a dynamic type. -- Uplevel references to static types are not a problem, the front end -- or the code generator fetches the bounds as required, and since they -- are compile time known values, this value can just be extracted and -- no actual uplevel reference is required. -- Uplevel references to dynamic types are a potential problem, since -- such references may involve an implicit access to a dynamic bound, -- and this reference is an implicit uplevel access. -- To fully unnest such references would be messy, since we would have -- to create local copies of the dynamic types involved, so that the -- front end or code generator could generate an explicit uplevel -- reference to the bound involved. Rather than do that, we set things -- up so that this situation can be easily detected and dealt with when -- there is an implicit reference to the bounds. -- What we do is to always generate a local constant for any dynamic -- bound in a dynamic subtype xx with name xx_FIRST or xx_LAST. The one -- case where we can skip this is where the bound is already a constant. -- E.g. in the third example above, subtype dynam is expanded as -- dynam_LAST : constant Integer := y + 3; -- subtype dynam is integer range x .. dynam_LAST; -- Now if type dynam is uplevel referenced (as it is in this case), then -- the bounds x and dynam_LAST are marked as uplevel references -- so that appropriate entries are made in the activation record. Any -- explicit reference to such a bound in the front end generated code -- will be handled by the normal uplevel reference mechanism which we -- described above for case 2. For implicit references by a back end -- that needs to unnest things, any such implicit reference to one of -- these bounds can be replaced by an appropriate reference to the entry -- in the activation record for xx_FIRST or xx_LAST. Thus the back end -- can eliminate the problematical uplevel reference without the need to -- do the heavy tree modification to do that at the code expansion level. -- Looking at case 3 again, here is the normal -gnatG expanded code -- function case3 (x : integer; y : integer) return boolean is -- dynam_LAST : constant integer := y {+} 3; -- subtype dynam is integer range x .. dynam_LAST; -- subtype static is integer range 42 .. 73; -- -- [constraint_error when -- not (y in x .. dynam_LAST) -- "range check failed"] -- -- xx : dynam := y; -- -- type darr is array (x .. dynam_LAST) of integer; -- type darec is record -- a : darr; -- b : integer; -- end record; -- [type TdarrB is array (x .. dynam_LAST range <>) of integer] -- freeze TdarrB [] -- darecv : darec; -- -- function inner (b : integer) return boolean is -- begin -- return b in x .. dynam_LAST and then darecv.b in 42 .. 73; -- end inner; -- begin -- return inner (42) and then inner (xx {*} 3 {-} y {*} 2); -- end case3; -- Note: the actual expanded code has fully qualified names so for -- example function inner is actually function case3__inner. For now -- we ignore that detail to clarify the examples. -- Here we see that some of the bounds references are expanded by the -- front end, so that we get explicit references to y or dynam_Last. These -- cases are handled by the normal uplevel reference mechanism described -- above for case 2. This is the case for the constraint check for the -- initialization of xx, and the range check in function inner. -- But the reference darecv.b in the return statement of function -- inner has an implicit reference to the bounds of dynam, since to -- compute the location of b in the record, we need the length of a. -- Here is the full translation of the third example: -- function case3x (x, y : integer) return boolean is -- type AREC1T is record -- x : Address; -- dynam_LAST : Address; -- end record; -- -- type AREC1PT is access all AREC1T; -- -- AREC1 : aliased AREC1T; -- AREC1P : constant AREC1PT := AREC1'Access; -- -- AREC1.x := x'Address; -- -- dynam_LAST : constant integer := y {+} 3; -- AREC1.dynam_LAST := dynam_LAST'Address; -- subtype dynam is integer range x .. dynam_LAST; -- xx : dynam := y; -- -- [constraint_error when -- not (y in x .. dynam_LAST) -- "range check failed"] -- -- subtype static is integer range 42 .. 73; -- -- type darr is array (x .. dynam_LAST) of Integer; -- type darec is record -- A : darr; -- B : integer; -- end record; -- darecv : darec; -- -- function inner (b : integer; AREC1F : AREC1PT) return boolean is -- begin -- return b in x .. Integer'Deref(AREC1F.dynam_LAST) -- and then darecv.b in 42 .. 73; -- end inner; -- -- begin -- return inner (42, AREC1P) and then inner (xx * 3, AREC1P); -- end case3x; -- And now the back end when it processes darecv.b will access the bounds -- of darecv.a by referencing the d and dynam_LAST fields of AREC1P. ----------------------------- -- Multiple Nesting Levels -- ----------------------------- -- In our examples so far, we have only nested to a single level, but the -- scheme generalizes to multiple levels of nesting and in this section we -- discuss how this generalization works. -- Consider this example with two nesting levels -- To deal with elimination of uplevel references, we follow the same basic -- approach described above for case 2, except that we need an activation -- record at each nested level. Basically the rule is that any procedure -- that has nested procedures needs an activation record. When we do this, -- the inner activation records have a pointer (uplink) to the immediately -- enclosing activation record, the normal arrangement of static links. The -- following shows the full translation of this fourth case. -- function case4x (x : integer) return integer is -- type AREC1T is record -- v1 : Address; -- end record; -- -- type AREC1PT is access all AREC1T; -- -- AREC1 : aliased AREC1T; -- AREC1P : constant AREC1PT := AREC1'Access; -- -- v1 : integer := x; -- AREC1.v1 := v1'Address; -- -- function inner1 (y : integer; AREC1F : AREC1PT) return integer is -- type AREC2T is record -- AREC1U : AREC1PT; -- v2 : Address; -- end record; -- -- type AREC2PT is access all AREC2T; -- -- AREC2 : aliased AREC2T; -- AREC2P : constant AREC2PT := AREC2'Access; -- -- AREC2.AREC1U := AREC1F; -- -- v2 : integer := Integer'Deref (AREC1F.v1) {+} 1; -- AREC2.v2 := v2'Address; -- -- function inner2 -- (z : integer; AREC2F : AREC2PT) return integer -- is -- begin -- return integer(z {+} -- Integer'Deref (AREC2F.AREC1U.v1) {+} -- Integer'Deref (AREC2F.v2).all); -- end inner2; -- begin -- return integer(y {+} -- inner2 (Integer'Deref (AREC1F.v1), AREC2P)); -- end inner1; -- begin -- return inner1 (x, AREC1P); -- end case4x; -- As can be seen in this example, the index numbers following AREC in the -- generated names avoid confusion between AREC names at different levels. ------------------------- -- Name Disambiguation -- ------------------------- -- As described above, the translation scheme would raise issues when the -- code generator did the actual unnesting if identically named nested -- subprograms exist. Similarly overloading would cause a naming issue. -- In fact, the expanded code includes qualified names which eliminate this -- problem. We omitted the qualification from the exapnded examples above -- for simplicity. But to see this in action, consider this example: -- function Mnames return Boolean is -- procedure Inner is -- procedure Inner is -- begin -- null; -- end; -- begin -- Inner; -- end; -- function F (A : Boolean) return Boolean is -- begin -- return not A; -- end; -- function F (A : Integer) return Boolean is -- begin -- return A > 42; -- end; -- begin -- Inner; -- return F (42) or F (True); -- end; -- The expanded code actually looks like: -- function mnames return boolean is -- procedure mnames__inner is -- procedure mnames__inner__inner is -- begin -- null; -- return; -- end mnames__inner__inner; -- begin -- mnames__inner__inner; -- return; -- end mnames__inner; -- function mnames__f (a : boolean) return boolean is -- begin -- return not a; -- end mnames__f; -- function mnames__f__2 (a : integer) return boolean is -- begin -- return a > 42; -- end mnames__f__2; -- begin -- mnames__inner; -- return mnames__f__2 (42) or mnames__f (true); -- end mnames; -- As can be seen from studying this example, the qualification deals both -- with the issue of clashing names (mnames__inner, mnames__inner__inner), -- and with overloading (mnames__f, mnames__f__2). -- In addition, the declarations of ARECnT and ARECnPT get moved to the -- outer level when we actually generate C code, so we suffix these names -- with the corresponding entity name to make sure they are unique. --------------------------- -- Terminology for Calls -- --------------------------- -- The level of a subprogram in the nest being analyzed is defined to be -- the level of nesting, so the outer level subprogram (the one passed to -- Unnest_Subprogram) is 1, subprograms immediately nested within this -- outer level subprogram have a level of 2, etc. -- Calls within the nest being analyzed are of three types: -- Downward call: this is a call from a subprogram to a subprogram that -- is immediately nested with in the caller, and thus has a level that -- is one greater than the caller. It is a fundamental property of the -- nesting structure and visibility that it is not possible to make a -- call from level N to level M, where M is greater than N + 1. -- Parallel call: this is a call from a nested subprogram to another -- nested subprogram that is at the same level. -- Upward call: this is a call from a subprogram to a subprogram that -- encloses the caller. The level of the callee is less than the level -- of the caller, and there is no limit on the difference, e.g. for an -- uplevel call, a subprogram at level 5 can call one at level 2 or even -- the outer level subprogram at level 1. ----------- -- Subps -- ----------- -- Table to record subprograms within the nest being currently analyzed. -- Entries in this table are made for each subprogram expanded, and do not -- get cleared as we complete the expansion, since we want the table info -- around in Cprint for the actual unnesting operation. Subps_First in this -- unit records the starting entry in the table for the entries for Subp -- and this is also recorded in the Subps_Index field of the outer level -- subprogram in the nest. The last subps index for the nest can be found -- in the Subp_Entry Last field of this first entry. subtype SI_Type is Nat; -- Index type for the table Subps_First : SI_Type; -- Record starting index for entries in the current nest (this is the table -- index of the entry for Subp itself, and is recorded in the Subps_Index -- field of the entity for this subprogram). type Subp_Entry is record Ent : Entity_Id; -- Entity of the subprogram Bod : Node_Id; -- Subprogram_Body node for this subprogram Lev : Nat; -- Subprogram level (1 = outer subprogram (Subp argument), 2 = nested -- immediately within this outer subprogram etc.) Reachable : Boolean; -- This flag is set True if there is a call path from the outer level -- subprogram to this subprogram. If Reachable is False, it means that -- the subprogram is declared but not actually referenced. We remove -- such subprograms from the tree, which simplifies our task, because -- we don't have to worry about e.g. uplevel references from such an -- unreferenced subpogram, which might require (useless) activation -- records to be created. This is computed by setting the outer level -- subprogram (Subp itself) as reachable, and then doing a transitive -- closure following all calls. Uplevel_Ref : Nat; -- The outermost level which defines entities which this subprogram -- references either directly or indirectly via a call. This cannot -- be greater than Lev. If it is equal to Lev, then it means that the -- subprogram does not make any uplevel references and that thus it -- does not need an activation record pointer passed. If it is less than -- Lev, then an activation record pointer is needed, since there is at -- least one uplevel reference. This is computed by initially setting -- Uplevel_Ref to Lev for all subprograms. Then on the initial tree -- traversal, decreasing Uplevel_Ref for an explicit uplevel reference, -- and finally by doing a transitive closure that follows calls (if A -- calls B and B has an uplevel reference to level X, then A references -- level X indirectly). Declares_AREC : Boolean; -- This is set True for a subprogram which include the declarations -- for a local activation record to be passed on downward calls. It -- is set True for the target level of an uplevel reference, and for -- all intervening nested subprograms. For example, if a subprogram X -- at level 5 makes an uplevel reference to an entity declared in a -- level 2 subprogram, then the subprograms at levels 4,3,2 enclosing -- the level 5 subprogram will have this flag set True. Uents : Elist_Id; -- This is a list of entities declared in this subprogram which are -- uplevel referenced. It contains both objects (which will be put in -- the corresponding AREC activation record), and types. The types are -- not put in the AREC activation record, but referenced bounds (i.e. -- generated _FIRST and _LAST entites, and formal parameters) will be -- in the list in their own right. Last : SI_Type; -- This field is set only in the entry for the outer level subprogram -- in a nest, and records the last index in the Subp table for all the -- entries for subprograms in this nest. ARECnF : Entity_Id; -- This entity is defined for all subprograms which need an extra formal -- that contains a pointer to the activation record needed for uplevel -- references. ARECnF must be defined for any subprogram which has a -- direct or indirect uplevel reference (i.e. Reference_Level < Lev). ARECn : Entity_Id; ARECnT : Entity_Id; ARECnPT : Entity_Id; ARECnP : Entity_Id; -- These AREC entities are defined only for subprograms for which we -- generate an activation record declaration, i.e. for subprograms for -- which the Declares_AREC flag is set True. ARECnU : Entity_Id; -- This AREC entity is the uplink component. It is other than Empty only -- for nested subprograms that declare an activation record as indicated -- by Declares_AREC being Ture, and which have uplevel references (Lev -- greater than Uplevel_Ref). It is the additional component in the -- activation record that references the ARECnF pointer (which points -- the activation record one level higher, thus forming the chain). end record; package Subps is new Table.Table ( Table_Component_Type => Subp_Entry, Table_Index_Type => SI_Type, Table_Low_Bound => 1, Table_Initial => 1000, Table_Increment => 200, Table_Name => "Unnest_Subps"); -- Records the subprograms in the nest whose outer subprogram is Subp ----------------- -- Subprograms -- ----------------- function Get_Level (Subp : Entity_Id; Sub : Entity_Id) return Nat; -- Sub is either Subp itself, or a subprogram nested within Subp. This -- function returns the level of nesting (Subp = 1, subprograms that -- are immediately nested within Subp = 2, etc.). function Subp_Index (Sub : Entity_Id) return SI_Type; -- Given the entity for a subprogram, return corresponding Subp's index procedure Unnest_Subprograms (N : Node_Id); -- Called to unnest subprograms. If we are in unnest subprogram mode, this -- is the call that traverses the tree N and locates all the library level -- subprograms with nested subprograms to process them. end Exp_Unst;