5 Commits

Author	SHA1	Message	Date
Zoltan Somogyi	72c3829fb9	Fix a subtle bug in the new way of handling failure. ... Estimated hours taken: 8 Fix a subtle bug in the new way of handling failure. When we enter the last disjunct of a model_non disjunction inside the model_non condition of an if-then-else, set resume_point_known to unknown. This forces code in the last disjunct and after the last disjunction to fail by invoking redo(), not by a direct branch. The reason this is necessary that a direct branch to the start of the else case is incorrect if the condition has already returned some answers. The bug was causing tests/general/complex_failure to fail at optimization levels that caused the inlining of a predicate, which resulted in a disjunction inside the last disjunct of a disjunction inside a model_non condition. compiler/notes/failure.html: Modify the description of failure handling to encompass the solution of the problem. compiler/code_info.m: Implement the solution. tests/general/semi_fail_in_non_ite.{m,exp}: Add a new test case, a variant of complex_failure, that tests for the bug at all optimization levels. tests/general/Mmakefile: Enable the new test case. tests/general/complex_failure.m: Remove an obsolete comment.	1998-08-07 05:13:17 +00:00
Zoltan Somogyi	ed063bcc00	Extend the new failure handling method to optionally preserve an invariant ... Estimated hours taken: 30 Extend the new failure handling method to optionally preserve an invariant required by accurate gc: always being able to tell with respect to what MR_sp or MR_curfr to interpret the stackvars and framevars referred to by the label whose address is the redoip slot of a nondet stack frame. This basically requires limitations on the hijacking of redoip/redofr slot pairs. compiler/notes/failure.html: Describe the new extension. Change the terminology to conform to what is used in the code. compiler/llds.m: For each temporary frame on the nondet stack, specify the native stack of the procedure that created it. This is so that we know whether the temporary frame ought to have the fourth slot that specifies the right value of MR_sp. (The fourth slot is included only in temporary nondet stack frames created procedures that live on the det stack; procedures that live on the nondet stack never have any of their variables on the det stack.) Remove the modframe llds instruction, since it does not specify what frame's redoip slot it is assigning to. This is error-prone (see peephole.m below). We were not using modframe much anyway. compiler/llds_out.m: compiler/opt_debug.m: Emit either mktempframe or mkdettempframe depending on the new field in temp_frame. compiler/code_info.m: Add a fourth item to the failure state, which states whether the top frame of the nondet stack may be hijacked. Initialize it from the option --allow-hijacks. If about to do a hijack but the failure state says no, create a temporary frame instead. Separate out the code for creating temporary frames, since it is now called from more than one place. Generalize the code to handle the new type of temp frame. compiler/code_info.m: compiler/ite_gen.m: Simplify the way we transmit information about the location of the address of the nondet stack frame in which the soft cut is performed from the start of the condition to its end. Remove the predicate code_info__maybe_push_temp_frame; its functionality is now achieved in code_info.m by disabling and restoring --allow-hijacks around the generation of code for the condition. This also allows us to get rid of the code that finds out whether the condition can do any hijacking. compiler/opt_util.m: Rename next_modframe as next_assign_to_redoip, and add an additional argument that says which frame's redoip we are after. compiler/peephole.m: Use the new argument of opt_util__next_assign_to_redoip to fix a bug where a mkframe of a temporary frame (which leaves curfr unchanged) that was followed a modframe (which assigns to the redoip of curfr) was incorrectly optimized (the assignment to the redoip slot was applied to the temporary frame, not the ordinary one). compiler/*.m: Minor changes to accommodate the disappearance of modframe. runtime/mercury_stacks.h: Add macros to support four-word temp frames on the nondet stack. Make the macros that access nondet stack frame slots start with MR_, while keeping (redefined) macros needed for backward compatibility. Remove the modframe macro. Fix a dangling reference to PREDNM instead of MR_PREDNM. runtime/mercury_stack_trace.c: Modify the stack tracing code to allow for the new four-word temp frames. Use the new MR_ prefixed variants of the macros. runtime/mercury_debug.h: runtime/mercury_misc.[ch]: Remove references to modframe. tests/general/complex_failure.{m,exp}: A new test case to tickle the various ways of handling nested disjunctions and if-then-elses in the new code generator. tests/general/Mmakefile: Enable the new test case. tests/{general,hard_coded}/space.{m,exp}: Move this test case from general to hard_coded. Although NU-Prolog can execute it, it does not give the same answers as Mercury due to a different default ordering and a difference in integer size (26-bit integers in NU-Prolog) that changes the behavior of the pseudo random-number generator. tests/hard_coded/cycles2.exp: Add the missing .exp file for this existing test case. tests/hard_coded/Mmakefile: Enable the old test cases cycles, cycles2 and space, since we now pass them.	1998-07-29 08:57:09 +00:00
Zoltan Somogyi	d1855187e5	Implement new methods of handling failures and the end points of branched ... Estimated hours taken: 260 Implement new methods of handling failures and the end points of branched control structures. compiler/notes/failure.html: Fix an omission about the handling of resume_is_known in if-then-elses. (This omission lead to a bug in the implementation.) Optimize cuts across multi goals when curfr is known to be equal to maxfr. Clarify the wording in several places. compiler/code_info.m: Completely rewrite the methods for handling failure. Separate the fields of code_info into three classes: those which do not change after initialization, those which record state that depends on where in the HLDS goal we are, and those which contain persistent data such as label and cell counters. Rename grab_code_info and slap_code_info as remember_position and reset_to_position, and add a wrapper around the remembered code_info to make it harder to make mistakes in its use. (Only the location-dependent fields of the remembered code_info are used, but putting only them into a separate data structure would result in more, not less, memory being allocated.) Gather the predicates that deal with handling branched control structures into a submodule. Reorder the declarations and definitions of access predicates to conform to the new order of fields. Reorder the declarations and definitions of the failure handling submodule to better reflect the separation of higher-level and lower-level predicates. compiler/code_gen.m: Replace code_gen__generate_{det,semi,non}_goal_2 with a single predicate, since for most HLDS constructs the code here is the same anyway (the called preds check the code model when needed). Move classification of the various kinds of unifications to unify_gen, since that is where it belongs. Move responsibility for initializing the code generator's trace info to code_info. Move the generation of code for negations to ite_gen, since the handling of negations is a cut-down version of the handling of negations. This should make the required double maintenance easier, and more likely to happen. compiler/disj_gen.m: compiler/ite_gen.m: These are the two modules that handle most failures; they have undergone a significant rewrite. As part of this rewrite, factor out the remaining common code between model_non and model_{det,semi} goals. compiler/unify_gen.m: Move classification of the various kinds of unifications here from code_gen. This allows us to keep several previously exported predicates private. compiler/call_gen.m: Factor out some code that was common to ordinary calls, higher order calls and method calls. Move the common code that checks whether we are doing tracing to trace.m. Replace call_gen__generate_{det,semi,nondet}_builtin with a single predicate. Delete the commented out call_gen__generate_complicated_unify, since it will never be needed and in any case suffered from significant code rot. compiler/llds.m: Change the mkframe instruction so that depending on one of its arguments, it can create either ordinary frames, or the cut-down frames used by the new failure handling algorithm (they have only three fixed fields: prevfr, redoip and redofr). compiler/llds_out.m: Emit a #define MR_USE_REDOFR before including mercury_imp.h, to tell the runtime we are using the new failure handling scheme. This effectively changes the grade of the compiled module. Emit MR_stackvar and MR_framevar instead of detstackvar and framevar. This is a step towards cleaning up the name-space, and a step towards making both start numbering at 0. For the time being, the compiler internally still starts counting framevars at 0; the code in llds_out.m adds a +1 offset. compiler/trace.m: Change the way trace info is initialized to fit in with the new requirements of code_info.m. Move the "are we tracing" check from the callers to the implementation of trace__prepare_for_call. compiler/*.m: Minor changes in accordance with the major ones above. compiler/options.m: Introduce a new option, allow_hijacks, which is set to "yes" by default. It is not used yet, but the idea is that when it is set to no, the code generator will not generate code that hijacks the nondet stack frame of another procedure invocation; instead, it will create a new temporary nondet stack frame. If the current procedure is model_non, it will have three fields: prevfr, redoip and redofr. If the current procedure is model_det or model_semi, it will have a fourth field that is set to the value of MR_sp. The idea is that the runtime system, which will be able to distinguish between ordinary frames (whose size is at least 5 words), 3-word and 4-word temporary frames, will now be able to use the redofr slots of all three kinds of frames and the fourth slot values of 4-word temporary frames as the addresses relative to which framevars and detstackvars respectively ought to be offset in stack layouts. compiler/handle_options.m: Turn off allow_hijacks if the gc method is accurate. runtime/mercury_stacks.h: Change the definitions for the nondet stack handling macros to accommodate the new nondet stack handling discipline. Define a new macro for creating temp nondet frames. Define MR_based_stackvar and MR_based_framevar (both of which start numbering slots at 1), and express other references, including MR_stackvar and MR_framevar and backward compatible definitions of detstackvar and framevar for hand-written C code, in terms of those two. runtime/mercury_stack_trace.[ch]: Add a new function to print a dump of the fixed elements nondet stack, for debugging my changes. (The dump does not include variable values.) runtime/mercury_trace_internal.c: Add a new undocumented command "D" for dumping the nondet stack (users should not know about this command, since the output is intelligible only to implementors). Add a new command "toggle_echo" that can cause the debugger to echo all commands. When the input to the debugger is redirected, this echo causes the output of the session to be much more readable. runtime/mercury_wrapper.c: Save the address of the artificial bottom nondet stack frame, so that the new function in mercury_stack_trace.c can find out where to stop. runtime/mercury_engine.c: runtime/mercury_wrapper.c: Put MR_STACK_TRACE_THIS_MODULE at the tops of these modules, so that the labels they define (e.g. do_fail and global_success) are registered in the label table when their module initialization functions are called. This is necessary for a meaningful nondet stack dump. runtime/mercury_grade.h: Add a new component to the grade string that specifies whether the code was compiled with the old or the new method of handling the nondet stack. This is important, because modules compiled with different nondet stack handling disciplines are not compatible. This component depends on whether MR_USE_REDOFR is defined or not. runtime/mercury_imp.h: If MR_DISABLE_REDOFR is defined, undefine off MR_USE_REDOFR before including mercury_grade.h. This is to allow people to continue working on un-updated workspaces after this change is installed; they should put "EXTRA_CFLAGS = -DMR_DISABLE_REDOFR" into Mmake.stage.params. (This way their stage1 will use the new method of handling failure, while their stage2 2&3 will use the old one.) This change should be undone once all our workspaces have switched over to the new failure handling method. tests/hard_coded/cut_test.{m,exp}: A new test case to tickle the various ways of handling cuts in the new code generator. tests/hard_coded/Mmakefile: Enable the new test case.	1998-07-20 10:04:02 +00:00
Zoltan Somogyi	10aa4af064	Fix the grammar in a couple of places. ... Estimated hours taken: 0.05 failure.html: Fix the grammar in a couple of places.	1997-11-05 08:22:26 +00:00
Zoltan Somogyi	c564706e32	A new file that describes a new way of handling failure. ... Estimated hours taken: 12 failure.html: A new file that describes a new way of handling failure. The advantages of the proposed new approach are: - it is significantly simpler than the existing approach, which makes it easier to get it correct - it will create temporary nondet frames significantly less frequently - commits inside nondet code do not need to use the det stack, which makes gc easier, and results in faster commits - it should make it easier to eventually delay the setting of redoip slots in disjuncts, instead of setting them on entry to each nonlast disjunct the way we do now The disadvantage is that nondet stack frames have an extra fixed slot, which has to be filled in and referenced.	1997-10-31 04:34:24 +00:00