mirror of
https://github.com/Mercury-Language/mercury.git
synced 2025-12-15 22:03:26 +00:00
712027f307
Estimated hours taken: 100 Branches: main This patch changes the parallel execution mechanism in the low level backend. The main idea is that, even in programs with only moderate parallelism, we won't have enough processors to exploit it all. We should try to reduce the cost in the common case, i.e. when a parallel conjunction gets executed sequentially. This patch does two things along those lines: (1) Instead of unconditionally executing all parallel conjuncts (but the last) in separate Mercury contexts, we allow a context to continue execution of the next conjunct of a parallel conjunction if it has just finished executing the previous conjunct. This saves on allocating unnecessary contexts, which can be a big reduction in memory usage. We also try to execute conjuncts left-to-right so as to minimise the need to suspend contexts when there are dependencies between conjuncts. (2) Conjuncts that *are* executed in parallel still need separate contexts. We used to pass variable bindings to those conjuncts by flushing input variable values to stack slots and copying the procedure's stack frame to the new context. When the conjunct finished, we would copy new variable bindings back to stack slots in the original context. What happens now is that we don't do any copying back and forth. We introduce a new abstract machine register `parent_sp' which points to the location of the stack pointer at the time that a parallel conjunction began. In parallel conjuncts we refer to all stack slots via the `parent_sp' pointer, since we could be running on a different context altogether and `sp' would be pointing into a new detstack. Since parallel conjuncts now share the procedure's stack frame, we have to allocate stack slots such that all parallel conjuncts in a procedure that could be executing simultaneously have distinct sets of stack slots. We currently use the simplest possible strategy, i.e. don't allow variables in parallel conjuncts to reuse stack slots. Note: in effect parent_sp is a frame pointer which is only set for and used by the code of parallel conjuncts. We don't call it a frame pointer as it can be confused with "frame variables" which have to do with the nondet stack. compiler/code_info.m: Add functionality to keep track of how deep inside of nested parallel conjunctions the code generator is. Add functionality to acquire and release "persistent" temporary stack slots. Unlike normal temporary stack slots, these don't get implicitly released when the code generator's location-dependent state is reset. Conform to additions of `parent_sp' and parent stack variables. compiler/exprn_aux.m: Generalise the `substitute_lval_in_*' predicates by `transform_lval_in_*' predicates. Instead of performing a fixed substitution, these take a higher order predicate which performs some operation on each lval. Redefine the substitution predicates in terms of the transformation predicates. Conform to changes in `fork', `join_and_terminate' and `join_and_continue' instructions. Conform to additions of `parent_sp' and parent stack variables. Remove `substitute_rval_in_args' and `substitute_rval_in_arg' which were unused. compiler/live_vars.m: Introduce a new type `parallel_stackvars' which is threaded through `build_live_sets_in_goal'. We accumulate the sets of variables which are assigned stack slots in each parallel conjunct. At the end of processing a parallel conjunction, use this information to force variables which are assigned stack slots to have distinct slots. compiler/llds.m: Change the semantics of the `fork' instruction. It now takes a single argument: the label of the next conjunct after the current one. The instruction now "sparks" the next conjunct to be run, either in a different context (possibly in parallel, on another Mercury engine) or is queued to be executed in the current context after the current conjunct is finished. Change the semantics of the `join_and_continue' instruction. This instruction now serves to end all parallel conjuncts, not just the last one in a parallel conjunction. Remove the `join_and_terminate' instruction (no longer used). Add the new abstract machine register `parent_sp'. Introduce "parent stack slots", which are similar to normal stack slots but relative to the `parent_sp' register. compiler/par_conj_gen.m: Change the code generated for parallel conjunctions. That is: - use the new `fork' instruction at the beginning of a parallel conjunct; - use the `join_and_continue' instruction at the end of all parallel conjuncts; - keep track of how deep the code generator currently is in parallel conjunctions; - set and restore the `parent_sp' register when entering a non-nested parallel conjunction; - after generating the code of a parallel conjunct, replace all references to stack slots by parent stack slots; - remove code to copy back output variables when a parallel conjunct finishes. Update some comments. runtime/mercury_context.c: runtime/mercury_context.h: Add the type `MR_Spark'. Sparks are allocated on the heap and contain enough information to begin execution of a single parallel conjunct. Add globals `MR_spark_queue_head' and `MR_spark_queue_tail'. These are pointers to the start and end of a global queue of sparks. Idle engines can pick up work from this queue in the same way that they can pick up work from the global context queue (the "run queue"). Add new fields to the MR_Context structure. `MR_ctxt_parent_sp' is a saved copy of the `parent_sp' register for when the context is suspended. `MR_ctxt_spark_stack' is a stack of sparks that we decided not to put on the global spark queue. Update `MR_load_context' and `MR_save_context' to save and restore `MR_ctxt_parent_sp'. Add the counters `MR_num_idle_engines' and `MR_num_outstanding_contexts_and_sparks'. These are used to decide, when a `fork' instruction is reached, whether a spark should be put on the global spark queue (with potential for parallelism but also more overhead) or on the calling context's spark stack (no parallelism and less overhead). Rename `MR_init_context' to `MR_init_context_maybe_generator'. When initialising contexts, don't reset redzones of already allocated stacks. It seems to be unnecessary (and the reset implementation is buggy anyway, though it's fine on Linux). Rename `MR_schedule' to `MR_schedule_context'. Add new functions `MR_schedule_spark_globally' and `MR_schedule_spark_locally'. In `MR_do_runnext', add code for idle engines to get work from the global spark queue. Resuming contexts are prioritised over sparks. Rename `MR_fork_new_context' to `MR_fork_new_child'. Change the definitions of `MR_fork_new_child' and `MR_join_and_continue' as per the new behaviour of the `fork' and `join_and_continue' instructions. Delete `MR_join_and_terminate'. Add a new field `MR_st_orig_context' to the MR_SyncTerm structure to record which context originated the parallel conjunction instance represented by a MR_SyncTerm instance, and update `MR_init_sync_term'. This is needed by the new behaviour of `MR_join_and_continue'. Update some comments. runtime/mercury_engine.h: runtime/mercury_regs.c: runtime/mercury_regs.h: runtime/mercury_stacks.h: Add the abstract machine register `parent_sp' and code to copy it to and from the fake_reg array. Add a macro `MR_parent_sv' to access stack slots via `parent_sp'. Add `MR_eng_parent_sp' to the MercuryEngine structure. runtime/mercury_wrapper.c: runtime/mercury_wrapper.h: Add Mercury runtime option `--max-contexts-per-thread' which is saved in the global variable `MR_max_contexts_per_thread'. The number `MR_max_outstanding_contexts' is derived from this. It sets a soft limit on the number of sparks we put in the global spark queue, relative to the number of threads we are running. We don't want to put too many sparks on the global queue if there are plenty of ready contexts or sparks already on the global queues, as they are likely to result in new contexts being allocated. When initially creating worker engines, wait until all the worker engines have acknowledged that they are idle before continuing. This is mainly so programs (especially benchmarks and test cases) with only a few fork instructions near the beginning of the program don't execute the forks before any worker engines are ready, resulting in no parallelism. runtime/mercury_engine.c: runtime/mercury_thread.c: Don't allocate a context at the time a Mercury engine is created. An engine only needs a new context when it is about to pick up a spark. configure.in: compiler/options.m: scripts/Mercury.config.in: Update to reflect the extra field in MR_SyncTerm. Add the option `--sync-term-size' and actually make use the result of the sync term size calculated during configuration. compiler/code_util.m: compiler/continuation_info.m: compiler/dupelim.m: compiler/dupproc.m: compiler/global_data.m: compiler/hlds_llds.m: compiler/jumpopt.m: compiler/livemap.m: compiler/llds_out.m: compiler/middle_rec.m: compiler/opt_debug.m: compiler/opt_util.m: compiler/reassign.m: compiler/stack_layout.m: compiler/use_local_vars.m: compiler/var_locn.m: Conform to changes in `fork', `join_and_terminate' and `join_and_continue' instructions. Conform to additions of `parent_sp' and parent stack variables. XXX not sure about the changes in stack_layout.m library/par_builtin.m: Conform to changes in the runtime system.