[m-rev.] for review: fix another dead proc elim problem

[Julien Fischer]

On Fri, 10 Aug 2007, Zoltan Somogyi wrote:

> My previous fix to dead proc elimination helped fixed some compiler aborts,
> but a related problem remained.
>
> The problem involved an unused procedure that was kept around so that the
> code generator would create the table associated with it. Since the
> procedure was unused, its body was thought to be unused too. If it
> contained a reference to a procedure that wasn't referred to from anywhere
> else, that procedure would be removed, leaving a dangling reference. This
> would cause a code generator abort.
>
> We can't fix the abort by replacing the kept-around procedure's body with
> "true", since that would cause a different code generator abort when
> moving the (now unbound) output variables to their argument registers. We
> could avoid generating any code for the procedure at all by e.g. marking
> it as opt_imported, but this would (a) be inconsistent and (b) require
> special case coding to still generate the table structure.
>
> The fix is to generate the global variable used for tabling
> *independently* of the procedure that enters things in the table.
>
> compiler/hlds_module.m:
> 	Add a field to the module_info (actually module_sub_info) that records
> 	the information the backends need to create the global variables
> 	representing call tables.
>
> 	Name all the fields of the module_info and module_sub_info during
> 	initialization, to make it easier to know where to add a new field.
> 	Put the initializations of the fields in the same order as the fields
> 	themselves.
>
> compiler/hlds_pred.m:
> 	Keep only the info for I/O tabling in procedures, since such tabling
> 	does not require defining a per-procedure global variable.
>
> 	Since the info for the forms of tabling that *do* require a
> 	per-procedure global variable are now divorced from the procedure,
> 	change their definition to avoid storing prog_vars in them, since
> 	those prog_vars would be separated from their varset. Instead, we
> 	record their numbers and their names (both are used only for debug
> 	support).
>
> 	On the other hand, some info from the pred_info and proc_info are
> 	to create the global variable; copy them into the data structure stored
> 	in hlds_module.
>
> 	Rename some fields to avoid ambiguities.
>
> compiler/table_gen.m:
> 	Continue to record information about I/O tabling in the proc_info,
> 	but record information about other forms of tabling in the new field
> 	in the module_info.
>
> compiler/rtti.m:
> compiler/hlds_rtti.m:
> 	Move the functions for constructing and deconstructing rtti_proc_labels
> 	from rtti.m (which is in backend_libs) to hlds_rtti.m (which is in
> 	hlds); the definition of rtti_proc_label was already in hlds_rtti.m.
> 	The move is needed to allow table_gen to put an rtti_proc_label
> 	in the data structures it puts in the module_info.
>
> compiler/hlds_out.m:
> 	Print out the new module_info field, and conform to the change to
> 	hlds_pred and table_arg_info.
>
> 	Always print variable numbers for type variables in table_arg_infos.
>
> compiler/continuation_info.m:
> 	Make room for either kind of tabling info for a procedure.
> 	(While the LLDS code generator doesn't need to know about the global
> 	variable representing the call table in order to create it, it does
> 	need to know about it in order to describe it to the debugger.)
>
> 	Conform to the change in table_arg_info.
>
> 	Rename some fields to avoid ambiguities.
>
> compiler/proc_gen.m:
> 	When generating code for procedures, do not try to create a
> 	per-procedure tabling struct, but do fill in the slot describing it
> 	in the continuation_info.
>
> 	Add a predicate to define all the tabling structs in a module.
>
> compiler/mercury_compile.m:
> 	Call proc_gen separately to define all the tabling structs.
>
> compiler/ml_code_gen.m:
> 	As with proc_gen, define tabling structs directly from the module_info
> 	and not when generating code from each proc_info.
>
> 	(The code for handling each proc is now logically not contiguous;
> 	I will address that in a separate change, to make the diff for this one
> 	easier to read.)
>
> compiler/dead_proc_elim.m:
> 	Don't keep unused tabled procedures alive, since that leads to the
> 	problem described up top.
>
> 	Keep track of which tabling structs are live, but don't yet act on that
> 	information, since some uses are hidden (for now).
>
> 	Add conditionally compiled tracing code that helped me trace down the
> 	problem.
>
> 	Fix an oversight in the severity level of an error spec.
>
> compiler/base_typeclass_info.m:
> compiler/code_util.m:
> compiler/deep_profiling.m:
> compiler/ml_code_util.m:
> compiler/proc_label.m:
> compiler/type_ctor_info.m:
> 	Conform to the move of make_rtti_proc_label.
>
> compiler/optimize.m:
> 	Conform to the change to continuation_info.
>
> compiler/stack_layout.m:
> 	Conform to the data structure changes above.
>
> doc/user_guide.texi:
> 	Document 'Z' as the character in -D arguments that tells hlds_out
> 	to dump the global structures needed for tabling.
>
> 	Fix an old oversight: document 'S' as the character in -D arguments
> 	that tells hlds_out to dump info about structure sharing.
>
> compiler/handle_options.m:
> 	Include 'Z' in -DALL -and -Dall.
>
> tests/tabling/mercury_java_parser_dead_proc_elim_bug.{m,exp}:
> 	Move this test case here from valid, since compiling all the way to
> 	executable doesn't work in valid (in yields link errors unrelated to
> 	the bug we are testing for).
>
> tests/tabling/mercury_java_parser_dead_proc_elim_bug2.{m,exp}:
> 	Add this new test case that in unfixed compilers gives the problem
> 	described up top.
>
> tests/tabling/Mmakefile:
> 	Enable the new tests.
>
> tests/valid/Mmakefile:
> tests/valid/Mercury.options:
> 	Remove references to the moved tests.

That looks ok.

Julien.
--------------------------------------------------------------------------
mercury-reviews mailing list
Post messages to:       mercury-reviews at csse.unimelb.edu.au
Administrative Queries: owner-mercury-reviews at csse.unimelb.edu.au
Subscriptions:          mercury-reviews-request at csse.unimelb.edu.au
--------------------------------------------------------------------------