mirror of
https://github.com/Mercury-Language/mercury.git
synced 2025-12-16 14:25:56 +00:00
25b8b1abc3
Estimated hours taken: 20 Branches: main Fix several performance bugs that showed up when the compiler was invoked on Douglas Auclair's training_cars example. Also fix some minor problems that made it harder to find the information needed to localize those problems. training_cars.m is hard to compile quickly because it is big in two dimensions: it has lots of clauses, and each clause has big terms. My laptop still tries to swap itself to death on the full version of training_cars.m (it has only 512 Mb), but the compiler now works fine on a version containing about 20% of its clauses, whereas previously it couldn't compile it at all. In most cases, the changes convert N^2 algorithms to NlogN algorithms. They probably have higher constant factors and may yield small slowdowns for small N, but this is probably not noticeable. Avoiding bad worst case behavior is more important. compiler/superhomogeneous.m: Record the number of goals inserted in each goal being converted to superhomogeneous form. If this exceeds a threshold, wrap a from_ground_term scope around it. Put the predicates into a more cohesive sequence. compiler/field_access.m: Work with the code in superhomogeneous to record the number of inserted goals. Reorder the arguments of some performances to be consistent with the predicates in superhomogeneous.m. compiler/modes.m: Use the from_ground_term scope to reverse the list of inserted unifications if necessary. It is much more efficient to do this here than to let it happen by sequences of delays and wakeups. That would have quadratic complexity; this is linear. This is what I originally introduced from_ground_term scopes for. Then, the overhead was too high, because I added one scope per function symbol. This version should be fine, since there is at most one scope added per argument of an atom (clause head or call). compiler/modes.m: compiler/unique_modes.m: When we are processing goals inside a from_ground_term scope, record this fact. compiler/mode_info.m: Make it possible to record this fact. compiler/modecheck_unify.m: When we are inside a from_ground_term scope, don't try to update the insts of vars on the right hand sides of construction unifications. Since these variables came from expansion to superhomogeneous form, those variables won't occur in any following code, so updating their state is useless, and the algorithm we used to do so is linear in the size of the inst. Since the size of an inst of a variable that results from superhomogeneous expansion is itself on average proportional to the size of the original term, this change turns a quadratic algorithm into a linear one. compiler/inst_match.m: Use balanced trees instead of ordered lists to represents sets of expansions, since these sets can be large. Note an opportunity for further improvement. compiler/inst_util.m: Note another opportunity for further improvement. compiler/instmap.m: Rename several predicates to avoid ambiguities. compiler/cse_detection.m: We used to print statistics for the processing of each procedure without saying which procedure it is for; fix this. compiler/switch_detection.m: Don't print progress messages for predicates with no procedures, since they would be misleading. compiler/higher_order.m: Change an algorithm that was quadratic in the number of arms for merging the information from the different arms of disjunctions and switches to an NlogN algorithm. Change the algorithm for merging the info from two branches that quadratic in the number of variables per arm to an NlogN algorithm. Changed some type equivalences to notag types to aid robustness. compiler/quantification.m: Rename several predicates to avoid ambiguities. The sets of variables in different arms of disjunctions and switches tend to have relatively small intersections. Yet the algorithms we used to compute the set of variables free in the disjunction or switch included the variables from the already processed arms in the sets being accumulated when processing later arms, leading to the quadratic behavior. This diff changes the algorithm to process each arm independently, and then use a more balanced algorithm to summarize the result. Specialize the predicates that compute sets of free vars in various HLDS fragments to work either with ordinary_nonlocals or code_gen_nonlocals without making the same decision repeatedly. Move some code out of large predicates into predicates of their own. compiler/Mercury.options: Specify the compiler option that can exploit this specialization to make the code run faster. compiler/simplify.m: Use a more efficient data structure for recording the parameters of an invocation of simplification. Change some predicate names and function symbol names to avoid ambiguity. compiler/common.m: compiler/deforest.m: compiler/deforest.m: compiler/make_hlds_warn.m: compiler/mercury_compile.m: compiler/pd_util.m: compiler/stack_opt.m: compiler/term_constr_build.m: Conform to the changes in simplify.m and/or instmap.m. compiler/mercury_compile.m: Fix a bug in progress messages for polymorphism.m. compiler/equiv_type_hlds.m: Most of the time, substitutions inside insts have no effect, because very few insts include any reference to a types. Instead of the old approach of building new insts and then throwing them away if they are the same as the old ones, don't build new insts at all if the old inst contains no types. compiler/common.m: Change some predicate names to make them clearer. compiler/hlds_clauses.m: Record the number of clauses so far, to allow a more informative progress message to be printed. compiler/add_clause.m: Print this more informative progress message. Conform to the changes in superhomogeneous.m. compiler/code_gen.m: Use the context of the predicate's first clause (which will be the context of the first clause head) as the context of the predicate's interface events. Unlike the context of the body goal, this won't be affected by program transformations such as wrapping a from_ground_term scope around some goals. It is better for users anyway, since the old policy lead to contexts in the middle of procedure bodies if the top level goal was a disjunction, switch or if-then-else. tests/debugger/*.exp: Update the expected outputs to conform to the change to code_gen.m.
55 lines
2.2 KiB
Plaintext
55 lines
2.2 KiB
Plaintext
#-----------------------------------------------------------------------------#
|
|
# Copyright (C) 2002-2006 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.
|
|
#-----------------------------------------------------------------------------#
|
|
# Mercury.options - module-specific flags for Mmake and `mmc --make'.
|
|
#-----------------------------------------------------------------------------#
|
|
|
|
# llds_out.m contains some sanity checking code that is usually turned off.
|
|
# However, the cost of checking whether they turned off exceeds the cost of
|
|
# performing them unless inlining is enabled.
|
|
MCFLAGS-ll_backend.llds_out = --inlining
|
|
|
|
# Several predicates in quantification.m contain an argument which is needed
|
|
# only to select the right mode. We want to make sure that we don't take up
|
|
# code space or execution time to pass these arguments.
|
|
MCFLAGS-hlds.quantification = --optimize-unused-args
|
|
|
|
# process_util.m uses `kill' and `struct sigaction' from <signal.h>,
|
|
# which are not available with `--ansi'.
|
|
MGNUCFLAGS-libs.process_util = --no-ansi
|
|
|
|
# If intermodule optimization is enabled, we also need to use `--no-ansi'
|
|
# for the following modules, because they import process_util.
|
|
MGNUCFLAGS-make.module_dep_file = --no-ansi
|
|
MGNUCFLAGS-make.module_target = --no-ansi
|
|
MGNUCFLAGS-make.program_target = --no-ansi
|
|
|
|
MCFLAGS-libs.process_util = --no-ansi-c
|
|
MCFLAGS-make.module_dep_file = --no-ansi-c
|
|
MCFLAGS-make.module_target = --no-ansi-c
|
|
MCFLAGS-make.program_target = --no-ansi-c
|
|
|
|
# Bug workarounds
|
|
|
|
# This is the same bug as tests/valid/ho_and_type_spec_bug
|
|
MCFLAGS-mode_robdd.tfeirn = -O3
|
|
MCFLAGS-mode_robdd.implications = -O0
|
|
|
|
# The c_code in the module gcc.m needs the header files from the GNU C
|
|
# distribution. Note that we need to compile these with
|
|
# -DMR_NO_BACKWARDS_COMPAT, because otherwise there are name
|
|
# conflicts with symbols defined in the GCC headers.
|
|
CFLAGS-gcc = -DMR_NO_BACKWARDS_COMPAT \
|
|
-DIN_GCC -DHAVE_CONFIG_H \
|
|
-I. \
|
|
-I$(GCC_SRC_DIR)/gcc \
|
|
-I$(GCC_SRC_DIR)/gcc/mercury \
|
|
-I$(GCC_SRC_DIR)/gcc/config \
|
|
-I$(GCC_SRC_DIR)/include \
|
|
-I$(GCC_SRC_DIR)
|
|
# Likewise for mlds_to_gcc.m
|
|
CFLAGS-mlds_to_gcc = $(CFLAGS-gcc)
|
|
|