Files
mercury/compiler/mark_static_terms.m
Peter Wang b86f973fa9 Allow the use of Mercury abstract machine float registers for passing
Branches: main

Allow the use of Mercury abstract machine float registers for passing
double-precision float arguments in higher order calls.

In of itself this is not so useful for typical Mercury code.  However, as
all non-local procedures are potentially the targets of higher order calls,
without this change first order calls to non-local procedures could not use
float registers either.  That is the actual motivation for this change.

The basic mechanism is straightforward.  As before, do_call_closure_* is
invoked to place the closure's hidden arguments into r1, ..., rN, and extra
input arguments shifted into rN+1, etc.  With float registers, extra input
arguments may also be in f1, f2, etc. and the closure may also have hidden
float arguments.  Optimising for calls, we order the closure's hidden
arguments so that all float register arguments come after all regular
register arguments in the vector.  Having the arguments out of order does
complicate code which needs to deconstruct closures, but that is not so
important.

Polymorphism complicates things.  A closure with type pred(float) may be
passed to a procedure expecting pred(T).  Due to the `float' argument type,
the closure expects its argument in a float register.  But when passed to the
procedure, the polymorphic argument type means it would be called with the
argument in a regular register.

Higher-order insts already contain information about the calling convention,
without which a higher-order term cannot be called.  We extend higher-order
insts to include information about the register class required for each
argument.  For example, we can distinguish between:

	pred(in) is semidet /* arg regs: [reg_f] */
and
	pred(in) is semidet /* arg regs: [reg_r] */

Using this information, we can create a wrapper around a higher-order
variable if it appears in a context requiring a different calling convention.
We do this in a new HLDS pass, called float_regs.m.

Note: Mercury code has a tendency to lose insts for higher-order terms, then
"recover" them by hacky means.  The float_regs pass depends on higher-order
insts; it is impossible to create a wrapper for a procedure without knowing
how to call it.  The float_regs pass will report errors which we otherwise
accepted, due to higher-order insts being unavailable.  It should be possible
for the user to adjust the code to satisfy the pass, though the user may not
understand why it should be necessary.  In most cases, it probably really
*is* unnecessary.  We may be able to make the float_regs pass more tolerant
of missing higher-order insts in the future.

Class method calls do not use float registers because I didn't want to deal
with them yet.


compiler/options.m:
compiler/handle_options.m:
	Always enable float registers in low-level C grades when floats are
	wider than a word.

compiler/make_hlds_passes.m:
	Always allow double word floats to be stored unboxed in cells on C
	grades.

compiler/hlds_goal.m:
	Add an extra field to `generic_call' which gives the register class
	to use for each argument.  This is set by the float_regs pass.

compiler/prog_data.m:
	Add an extra field to `pred_inst_info' which records the register class
	to use for each argument.  This is set by the float_regs pass.

compiler/hlds_pred.m:
	Add a field to `proc_sub_info' which lists the headvars which must be
	passed via regular registers despite their types.

	Add a field to `pred_sub_info' to record the original unsubstituted
	argument types for instance method predicates.

compiler/check_typeclass.m:
	In the pred_info of an instance method predicate, record the original
	argument types before substituting the type variables for the instance.

compiler/float_regs.m:
compiler/transform_hlds.m:
	Add the new HLDS pass.

compiler/mercury_compile_middle_passes.m:
	Run the new pass if float registers are enabled.

compiler/lambda.m:
	Export the predicate to produce a predicate from a lambda.
	This is reused by float_regs.m to create wrapper closures.

	Add an argument to `expand_lambda' to set the reg_r_headvars field on
	the newly created procedure.

	Delete some unused fields from `lambda_info'.

compiler/arg_info.m:
	Make `generate_proc_arg_info' no longer always use regular registers
	for calls to exported procedures.  Do always use regular registers for
	class methods calls.

	Add a version of `make_arg_infos' which takes an explicit list of
	argument registers.  Rename the previous version.

	Add `generic_call_arg_reg_types' to return the argument registers
	for a generic call.

	Add a version of `compute_in_and_out_vars' which additionally separates
	arguments for float and regular registers.

compiler/call_gen.m:
	Use float registers for argument passing in higher-order calls, as
	directed by the new field in `generic_call'.

compiler/code_util.m:
	Add a function to encode the number of regular and float register
	arguments when making a higher-order call.

compiler/llds.m:
	Say that the `do_call_closure_N' functions only work for zero float
	register arguments.

compiler/follow_vars.m:
compiler/interval.m:
	Account for the use of float registers by generic call goals in these
	passes.

compiler/unify_gen.m:
	Move float register arguments to the end of a closure's hidden
	arguments vector, after regular register arguments.

	Count hidden regular and float register arguments separately, but
	encode them in the same word in the closure.  This is preferable to
	using two words because it reduces the differences between grades
	with and without float registers present.

	Disable generating code which creates a closure from an existing
	closure, if float registers exist.  That code does not understand the
	reordered hidden arguments vector yet.

compiler/continuation_info.m:
	Replace an argument's type_info in the closure layout if the argument
	is a float *and* is passed via a regular register, when floats are
	normally passed via float registers.  Instead, give it the type_info
	for `private_builtin.float_box'.

compiler/builtin_lib_types.m:
	Add function to return the type of `private_builtin.float_box/0'.

compiler/hlds_out_goal.m:
compiler/hlds_out_pred.m:
compiler/mercury_to_mercury.m:
	Dump the new fields added to `generic_call', `pred_inst_info' and
	`proc_sub_info'.

compiler/prog_type.m:
	Add helper predicate.

compiler/*.m:
	Conform to changes.

library/private_builtin.m:
	Add a type `float_box'.

runtime/mercury_ho_call.h:
	Describe the modified closure representation.

	Rename the field which counts the number of hidden arguments to prevent
	it being used incorrectly, as it now encodes two numbers (potentially).

	Add macros to unpack the encoded field.

runtime/mercury_ho_call.c:
	Update the description of how higher-order calls work.

	Update code which extracts closure arguments to take account the
	arguments being reordered in the hidden arguments vector.

runtime/mercury_deep_copy.c:
runtime/mercury_deep_copy_body.h:
runtime/mercury_layout_util.c:
runtime/mercury_ml_expand_body.h:
	Update code which extracts closure arguments to take account the
	arguments being reordered in the hidden arguments vector.

runtime/mercury_type_info.c:
runtime/mercury_type_info.h:
	Add helper function.

tools/make_spec_ho_call:
	Update the generated do_call_closure_* functions to place float
	register arguments.

tests/hard_coded/Mercury.options:
tests/hard_coded/Mmakefile:
tests/hard_coded/ho_float_reg.exp:
tests/hard_coded/ho_float_reg.m:
	Add new test case.

tests/hard_coded/copy_pred.exp:
tests/hard_coded/copy_pred.m:
tests/hard_coded/deconstruct_arg.exp:
tests/hard_coded/deconstruct_arg.exp2:
tests/hard_coded/deconstruct_arg.m:
	Extend test cases with float arguments in closures.

tests/debugger/higher_order.exp2:
	Add alternative output, changed due to closure wrapping.

tests/hard_coded/ho_univ_to_type.m:
	Adjust test case so that the float_regs pass does not report errors
	about missing higher-order insts.

compiler/notes/compiler_design.html:
	Describe the new module.

	Delete a duplicated paragraph.

compiler/notes/todo.html:
TODO:
	Delete one hundred billion year old todos.
2012-02-13 00:11:57 +00:00

212 lines
8.3 KiB
Mathematica

%-----------------------------------------------------------------------------%
% vim: ft=mercury ts=4 sw=4 et
%-----------------------------------------------------------------------------%
% Copyright (C) 2000-2012 The University of Melbourne.
% This file may only be copied under the terms of the GNU General
% Public License - see the file COPYING in the Mercury distribution.
%-----------------------------------------------------------------------------%
%
% File: mark_static_terms.m.
% Main author: fjh.
%
% This module traverses the HLDS, updating the `how_to_construct' field of
% construction unifications. For each construction which can be done
% statically, i.e. whose arguments are all static, it replaces this field with
% `construct_statically'. The main use of information is in the MLDS back-end,
% to determine when we can generate static initialized constants instead of
% calling new_object(). However, other parts of the compiler also use this
% information.
%
%-----------------------------------------------------------------------------%
:- module hlds.mark_static_terms.
:- interface.
:- import_module hlds.hlds_module.
:- import_module hlds.hlds_pred.
%-----------------------------------------------------------------------------%
:- pred mark_static_terms(module_info::in, proc_info::in, proc_info::out)
is det.
%-----------------------------------------------------------------------------%
%-----------------------------------------------------------------------------%
:- implementation.
:- import_module hlds.hlds_goal.
:- import_module parse_tree.prog_data.
:- import_module list.
:- import_module require.
:- import_module set_tree234.
%-----------------------------------------------------------------------------%
% As we traverse the goal, we keep track of which variables are static at
% the current program point.
%
:- type static_info == set_tree234(prog_var).
mark_static_terms(_ModuleInfo, !Proc) :-
% The ModuleInfo argument is there just for passes_aux.
proc_info_get_goal(!.Proc, Goal0),
StaticInfo0 = set_tree234.init,
goal_mark_static_terms(Goal0, Goal, StaticInfo0, _StaticInfo),
proc_info_set_goal(Goal, !Proc).
:- pred goal_mark_static_terms(hlds_goal::in, hlds_goal::out,
static_info::in, static_info::out) is det.
goal_mark_static_terms(Goal0, Goal, !SI) :-
Goal0 = hlds_goal(GoalExpr0, GoalInfo),
Goal = hlds_goal(GoalExpr, GoalInfo),
(
GoalExpr0 = conj(ConjType, Goals0),
% It's OK to treat parallel conjunctions as if they were sequential
% here, since if we mark any variables as static, the computation
% of those variables will be done at compile time.
conj_mark_static_terms(Goals0, Goals, !SI),
GoalExpr = conj(ConjType, Goals)
;
GoalExpr0 = disj(Goals0),
% We revert to the original static_info at the end of branched goals.
disj_mark_static_terms(Goals0, Goals, !.SI),
GoalExpr = disj(Goals)
;
GoalExpr0 = switch(Var, CanFail, Cases0),
% We revert to the original static_info at the end of branched goals.
cases_mark_static_terms(Cases0, Cases, !.SI),
GoalExpr = switch(Var, CanFail, Cases)
;
GoalExpr0 = negation(SubGoal0),
% We revert to the original static_info at the end of the negation.
goal_mark_static_terms(SubGoal0, SubGoal, !.SI, _SI),
GoalExpr = negation(SubGoal)
;
GoalExpr0 = scope(Reason, SubGoal0),
( Reason = from_ground_term(TermVar, from_ground_term_construct) ->
% These scopes already have all their unifications marked
% as construct_statically.
set_tree234.insert(TermVar, !SI),
GoalExpr = GoalExpr0
;
goal_mark_static_terms(SubGoal0, SubGoal, !SI),
GoalExpr = scope(Reason, SubGoal)
)
;
GoalExpr0 = if_then_else(Vars, Cond0, Then0, Else0),
SI0 = !.SI,
% We run the Cond and the Then in sequence, and we run the Else
% in parallel with that, and then we throw away the static_infos
% we computed and revert to the original static_info at the end,
% since this was a branched goal.
goal_mark_static_terms(Cond0, Cond, SI0, SI_Cond),
goal_mark_static_terms(Then0, Then, SI_Cond, _SI_Then),
goal_mark_static_terms(Else0, Else, SI0, _SI_Else),
GoalExpr = if_then_else(Vars, Cond, Then, Else)
;
( GoalExpr0 = plain_call(_, _, _, _, _, _)
; GoalExpr = generic_call(_, _, _, _, _)
; GoalExpr = call_foreign_proc(_, _, _, _, _, _, _)
),
GoalExpr = GoalExpr0
;
GoalExpr0 = unify(LHS, RHS, Mode, Unification0, Context),
unification_mark_static_terms(Unification0, Unification, !SI),
GoalExpr = unify(LHS, RHS, Mode, Unification, Context)
;
GoalExpr0 = shorthand(_),
% These should have been expanded out by now.
unexpected($module, $pred, "shorthand")
).
:- pred conj_mark_static_terms(hlds_goals::in, hlds_goals::out,
static_info::in, static_info::out) is det.
conj_mark_static_terms(Goals0, Goals, !SI) :-
list.map_foldl(goal_mark_static_terms, Goals0, Goals, !SI).
:- pred disj_mark_static_terms(hlds_goals::in, hlds_goals::out,
static_info::in) is det.
disj_mark_static_terms([], [], _).
disj_mark_static_terms([Goal0 | Goals0], [Goal | Goals], SI0) :-
% We throw away the static_info obtained after each branch.
goal_mark_static_terms(Goal0, Goal, SI0, _SI),
disj_mark_static_terms(Goals0, Goals, SI0).
:- pred cases_mark_static_terms(list(case)::in, list(case)::out,
static_info::in) is det.
cases_mark_static_terms([], [], _SI0).
cases_mark_static_terms([Case0 | Cases0], [Case | Cases], SI0) :-
Case0 = case(MainConsId, OtherConsIds, Goal0),
% We throw away the static_info obtained after each branch.
goal_mark_static_terms(Goal0, Goal, SI0, _SI),
Case = case(MainConsId, OtherConsIds, Goal),
cases_mark_static_terms(Cases0, Cases, SI0).
:- pred unification_mark_static_terms(unification::in, unification::out,
static_info::in, static_info::out) is det.
unification_mark_static_terms(Unification0, Unification, !StaticVars) :-
(
Unification0 = construct(Var, ConsId, ArgVars, ArgModes,
HowToConstruct0, Unique, SubInfo),
% If all the arguments are static, then the newly constructed variable
% is static too.
( list.all_true(set_tree234.contains(!.StaticVars), ArgVars) ->
HowToConstruct = construct_statically,
set_tree234.insert(Var, !StaticVars),
% This is a minor optimization to improve the efficiency of the
% compiler: don't bother allocating memory if we don't need to.
( HowToConstruct = HowToConstruct0 ->
Unification = Unification0
;
Unification = construct(Var, ConsId, ArgVars, ArgModes,
HowToConstruct, Unique, SubInfo)
)
;
Unification = Unification0
)
;
Unification0 = deconstruct(_Var, _ConsId, _ArgVars, _UniModes,
_CanFail, _CanCGC),
Unification = Unification0
% (
% % if the variable being deconstructed is static,
% % and the deconstruction cannot fail,
% % then the newly extracted argument variables
% % are static too
% % (XXX is the "cannot fail" bit really necessary?)
% map.search(StaticVars0, Var, Data),
% CanFail = cannot_fail
% ->
% XXX insert ArgVars into StaticVars0
% ;
% true
% )
;
Unification0 = assign(TargetVar, SourceVar),
Unification = Unification0,
% If the variable being assigned from is static, then the variable
% being assigned to is static too.
( set_tree234.contains(!.StaticVars, SourceVar) ->
set_tree234.insert(TargetVar, !StaticVars)
;
true
)
;
( Unification0 = simple_test(_, _)
; Unification0 = complicated_unify(_, _, _)
),
Unification = Unification0
).
%-----------------------------------------------------------------------------%
:- end_module hlds.mark_static_terms.
%-----------------------------------------------------------------------------%