Tobias Grosser [Sun, 19 Feb 2012 13:05:50 +0000 (14:05 +0100)]
scheduler: allow to bound the coefficients in the calculated schedule
isl already supports the bounding of the constant term of the calculated
schedule, but within polybench there are several cases where bounding
the coefficients of parameter and variable dimensions is also needed.
With a bound of 20 Polly can run the isl scheduler on all polybench 2.0 kernels
and takes for each less than 5 seconds, whereas without the bound five test
cases had scheduling times larger than 40 seconds. This is still slower, than
scheduling with gist simplified dependences but it yields better results.
The bounded, non-gist-simplified schedule yields for dynprog, lu, reg_detect
and trmm to faster code compared to scheduling with gist simplified
dependences. For gramschmidt it yields slower code.
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Sat, 18 Feb 2012 15:13:23 +0000 (16:13 +0100)]
isl_tab_basic_set_non_trivial_lexmin: do not add cuts for all variables
We found a test case where adding cuts for all variables can increase
the time of the scheduling algorithm significantly (from ms to never
seen terminating). Even though we did not understand the problem
entirely, the intuition we have is the following:
When using isl_tab_basic_set_non_trivial_lexmin, the ILP problem is
(usually) not empty, but infinite. Adding all cuts, seems to increase
the objective function only in very small steps. Adding only one cut at
a time allows us to increase it with a step as big as possible.
This patch does not change the behaviour when calling
check_integer_feasible. In case we're calling cut_to_integer_lexmin
from check_integer_feasible, if it takes a long time to compute, then
it's "usually" because the context is empty and it's just taking a long
time for us to discover that it's empty. We believe in that case it
could still make sense to add all cuts.
This patch was tested on polybench 2.0 with Polly and minimal fusion.
On the larger kernels, it could not solve any infinite scheduling time
problems, but it was able to speed up one test case by 2x and did not
yield increased scheduling time for the other test cases.
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 14 Feb 2012 23:18:59 +0000 (00:18 +0100)]
isl_map_fix_si: drop basic maps that have become empty
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 14 Feb 2012 23:18:21 +0000 (00:18 +0100)]
isl_map.c: extract out common remove_if_empty
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 14 Feb 2012 23:15:55 +0000 (00:15 +0100)]
isl_map_coalesce: drop empty parts before checking the number of parts
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 14 Feb 2012 23:22:47 +0000 (00:22 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Tue, 14 Feb 2012 22:55:22 +0000 (23:55 +0100)]
isl_convex_hull.c: uset_convex_hull_wrap_bounded: remove redundant constraints
The caller compute_facet expects the result to not contain any redundant
constraints. The input to uset_convex_hull_wrap_bounded, however, might
contain redundant constraints so if this input consists of only one
basic set, we need to make sure redundant constraints are removed.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 14 Feb 2012 19:02:43 +0000 (20:02 +0100)]
isl_basic_map_realign: preserve (most) flags
In particular, preserve the ISL_BASIC_MAP_EMPTY flag.
Many parts of the code check for this flags so they do not have
to deal with the "1 = 0" constraint.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:26:43 +0000 (12:26 +0100)]
add isl_printer_print_pw_aff test
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:26:20 +0000 (12:26 +0100)]
isl_basic_map_foreach_lexopt: normalize isl_affs
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:19:20 +0000 (12:19 +0100)]
isl_pw_*_union_add: don't subtract non-overlapping cells
There is no point in subtracting a cell from another if their intersection
is empty. It can only lead to a further subdivision of the other cell.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:17:03 +0000 (12:17 +0100)]
isl_printer_print_pw_aff: skip constraints implied by div expression
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:11:26 +0000 (12:11 +0100)]
add isl_basic_set_is_div_constraint
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:58:43 +0000 (12:58 +0100)]
Merge branch 'maint'
Conflicts:
isl_output.c
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:13:23 +0000 (12:13 +0100)]
isl_printer_print_pw_aff: fix printing in C format
In 3280c05 (make isl_pw_* object live in a map space,
Tue Aug 30 16:47:59 2011 +0200), the space in which objects live was
changed from that of its domain to that of a map, but
isl_printer_print_pw_aff was not changed accordingly.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:09:03 +0000 (12:09 +0100)]
isl_stream_read_pw_aff: call isl_pw_aff_union_add instead of isl_pw_aff_add
We missed this in 7a86dd8 (rename isl_pw_aff_add to isl_pw_aff_union_add,
Thu Oct 13 12:16:45 2011 +0200).
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 7 Feb 2012 11:55:43 +0000 (12:55 +0100)]
isl_vertices.c: compute_chambers: avoid access to freed data structure
The problem was introduced in 328c78f (isl_tab_from_basic_map: preserve all
constraints in input when tracking, Mon Jan 16 16:55:44 2012 +0100).
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 29 Jan 2012 15:52:15 +0000 (16:52 +0100)]
export isl_pw_multi_aff_coalesce
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 29 Jan 2012 17:15:51 +0000 (18:15 +0100)]
export isl_pw_multi_aff_domain
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 30 Jan 2012 12:18:09 +0000 (13:18 +0100)]
isl_aff_normalize: remove unused divs
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 2 Feb 2012 11:01:53 +0000 (12:01 +0100)]
isl_pw_multi_aff_from_map: add a special case where result can be read off
If all output dimensions are uniquely defined in terms of the parameters
and input dimensions, we can directly read off the desired isl_pw_multi_aff.
This usually results in much simpler expressions.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 2 Feb 2012 09:29:43 +0000 (10:29 +0100)]
privately export isl_basic_map_plain_is_single_valued
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 2 Feb 2012 09:45:48 +0000 (10:45 +0100)]
isl_printer_print_pw_aff: simplify domain expression in C format
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 1 Feb 2012 16:11:38 +0000 (17:11 +0100)]
isl_morph.c: fix typo in comment
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 1 Feb 2012 21:50:00 +0000 (22:50 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Wed, 1 Feb 2012 17:01:31 +0000 (18:01 +0100)]
isl_basic_set_full_compression: detect equalities in input first
Without this explicit detection of equalities, some equalities
may be discovered in the middle of the computation, possibly
leading to the introduction of existentially quantified variables,
while callers of isl_basic_set_full_compression typically do not
want such variables.
Since the detection of equalities can in some cases be fairly expensive,
we may have to find a better way of dealing with this problem at some point.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 1 Feb 2012 11:34:30 +0000 (12:34 +0100)]
isl_stream_read_pw_qpolynomial: accept products of integer divisions
Before, we called accept_affine to parse an integer division that appears
inside a function, but accept_affine also parses a subsequent '*' and
then expects this '*' to be followed by a constant.
Call accept_div directly intead.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 31 Jan 2012 18:18:23 +0000 (19:18 +0100)]
isl_union_set_product: call isl_set_product on entries
Before, we would simply delegate control to isl_union_map_product,
which calls isl_map_product on the entries. This doesn't work
on isl_sets because they only have a "range" and no "domain".
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 31 Jan 2012 14:09:12 +0000 (15:09 +0100)]
isl_tab_pip.c: sol_add: skip empty contexts
If the context becomes non-empty inside find_solutions, we stop
looking for solutions since a89841a (isl_tab_pip.c: find_solutions: break
when context becomes empty, Sun Jan 31 18:29:12 2010 +0100), but
we still call sol_add. We need to return from sol_add early in
such cases because there is nothing to do and, more importantly,
when a basic set is marked empty, its existentials are removed,
resulting in a mismatch in the number of dimensions. In particular,
the basic set returned by sol_domain may not satisfy all assumptions
made by the remainder of sol_add.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 31 Jan 2012 11:27:09 +0000 (12:27 +0100)]
isl_space_set_dim_id: also set id of parameter in nested spaces
A nested space is supposed to have the same parameters as
the nesting space, so if the id of a parameter in the nesting
space space changes, it has to be changed in the nested spaces too.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 31 Jan 2012 11:24:38 +0000 (12:24 +0100)]
isl_space_set_dim_id: rename "dim" variable to "space"
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 29 Jan 2012 13:42:16 +0000 (14:42 +0100)]
scheduler: replace split_parallel by split_scaled option
The split_parallel option was ill-conceived.
The condition split_parallel looked for was too specific
and the implementation was wrong. Part of this has been fixed
in 1b2fdb6 (schedule.c: split_parallel: avoid invalid memory accesses,
Mon Jan 30 15:23:18 2012 +0100), but the code could in principle
still produce an incorrect schedule because it didn't check the sizes
of the constant terms.
The new option essentially applies strip-mining and should therefore
be safe. In particular, it is applied when the linear parts of the
schedules have a non-trivial common divisor. The strip-mining is
then performed with respect to this common divisor.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 29 Jan 2012 13:42:01 +0000 (14:42 +0100)]
add isl_mat_scale_down_row
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Mon, 30 Jan 2012 12:19:38 +0000 (13:19 +0100)]
Add isl_union_map_gist_range
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 26 Nov 2011 14:52:07 +0000 (15:52 +0100)]
scheduler: add an option to configure the level of fusing
We currently only support maximal (the old behavior) and minimal fusion.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 30 Jan 2012 14:28:04 +0000 (15:28 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Mon, 30 Jan 2012 14:23:18 +0000 (15:23 +0100)]
schedule.c: split_parallel: avoid invalid memory accesses
The original code incorrectly assumed that the schedules for
all statements have the same number of columns, i.e., that the
dimensions of the iteration domains of all statements are the same.
We fix this up by simply bailing out if we find this assumption not to hold.
We will improve the applicability of the transformation performed by
split_parallel in a subsequent patch.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 18 Jan 2012 13:18:26 +0000 (14:18 +0100)]
update for change in clang's HandleTopLevelDecl return type
In recent versions of clang, HandleTopLevelDecl should return a boolean
indicating whether parsing should continue.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 18 Jan 2012 15:47:38 +0000 (16:47 +0100)]
configure.ac: only link against libraries from LLVM analysis component
Before, we would link against all LLVM components, which would make
us depend on google-test.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 18 Jan 2012 15:49:19 +0000 (16:49 +0100)]
put annotations after "struct"
Recent versions of clang require the annotations to be placed after "struct".
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 18 Jan 2012 17:41:37 +0000 (18:41 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Tue, 17 Jan 2012 12:52:09 +0000 (13:52 +0100)]
isl_pw_aff_non_zero_set: don't assume isl_pw_aff is total
Before, we would return the domain where the function is not zero
or undefined, whereas we should only return the domain where the function
is not zero.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 18 Jan 2012 17:36:53 +0000 (18:36 +0100)]
isl_aff_nonneg_basic_set and isl_aff_zero_basic_set: simplify result
The constructed constraint may be a tautology in which case it needs
to be simplified away, especially if it is an equality.
We probably want to perform this simplification in
isl_basic_set_from_constraint instead, but CLooG currently depends
on the resulting basic set not being simplified.
We would have to fix CLooG first before we can change this behavior.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 16 Jan 2012 15:55:44 +0000 (16:55 +0100)]
isl_tab_from_basic_map: preserve all constraints in input when tracking
If the user wants to keep track of a basic map representation of the tableau,
then we need to make sure that the initial tableau corresponds exactly to
the input. In particular, we need to make sure that no (redundant) constraints
are removed.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 16 Jan 2012 15:46:55 +0000 (16:46 +0100)]
isl_tab_track_bmap: drop sanity checks in case of empty tab
If tab is empty, then we may not have added all constraints of "bmap"
and so the number of constraints in tab and bmap may be different.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 16 Jan 2012 15:21:49 +0000 (16:21 +0100)]
isl_tab_track_bmap: make sure we have a unique copy of the basic map
The basic map will be modified in place by isl_tab_add_ineq.
Some callers were passing a unique copy to isl_tab_track_bmap
already, but some others apparently were not.
It's safer to just let isl_tab_track_bmap take care of this.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 16 Jan 2012 11:59:33 +0000 (12:59 +0100)]
isl_tab_track_bmap: fix memory leak on error path
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 15 Jan 2012 15:12:49 +0000 (16:12 +0100)]
isl_basic_map_remove_unknown_divs: restart whenever div has been removed
The call to isl_basic_map_remove_dims may end up calling
isl_basic_map_simplify, which may in turn remove additional divs
if they no longer occur in any constraint.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 15 Jan 2012 14:21:17 +0000 (15:21 +0100)]
isl_map_simple_hull: improve error handling
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 12 Jan 2012 14:49:38 +0000 (15:49 +0100)]
export isl_multi_aff_drop_dims
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 12 Jan 2012 14:49:28 +0000 (15:49 +0100)]
isl_multi_aff_drop_dims: handle removal of output dimensions
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 12 Jan 2012 14:48:27 +0000 (15:48 +0100)]
isl_multi_aff_set_aff: improve error handling
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 12 Jan 2012 14:24:01 +0000 (15:24 +0100)]
add isl_multi_aff_set_tuple_id
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 12 Jan 2012 10:10:04 +0000 (11:10 +0100)]
isl_pw_aff_on_shared_domain: perform gist on results wrt shared domains
The shared domain may have extra constraints compared to the original
domains, which may help to simplify some of the integer divisions involved.
As a side effect, the gist operation may also simplify integer divisions
that could have been simplified even on a universe domain.
We should probably handle such cases much earlier during the construction
of isl_affs.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 3 Jan 2012 09:35:43 +0000 (10:35 +0100)]
add isl_aff_project_domain_on_params
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 3 Jan 2012 09:42:33 +0000 (10:42 +0100)]
add isl_aff_align_params
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 3 Jan 2012 10:30:04 +0000 (11:30 +0100)]
isl_printer_print_aff: only print () when needed in C format
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 3 Jan 2012 10:25:47 +0000 (11:25 +0100)]
isl_aff_gist: normalize result
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 3 Jan 2012 10:43:51 +0000 (11:43 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Tue, 3 Jan 2012 09:05:50 +0000 (10:05 +0100)]
isl_schedule.c: mark row constructed by compute_split_schedule as non-zero
compute_split_schedule is only called when no zero-distance schedule row
could be found on the entire graph.
We missed this part in ee1ca56 (isl_schedule.c: keep track of parallel loops,
Sun May 29 19:41:55 2011 +0200).
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 2 Jan 2012 13:10:50 +0000 (14:10 +0100)]
isl_map_project_out: properly handle named or nested spaces
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 1 Jan 2012 17:24:58 +0000 (18:24 +0100)]
isl_arg_parse: reset all flags with same option name whenever a flag is set
Otherwise, there is no way to turn off a flag that is set by default.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 1 Jan 2012 17:00:49 +0000 (18:00 +0100)]
isl_ctx_peek_options: special case isl_options_args
The isl_options_args options are always available, even
when they do not appear in ctx->user_opt.
Just return them instead of looking through ctx->user_opt.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 30 Dec 2011 23:19:47 +0000 (00:19 +0100)]
isl_basic_set_compute_vertices: set activity domain on single vertex
isl_basic_set_compute_vertices always sets the activity domains of
the vertices, except when there is only a single vertex.
Chnage vertices_0D to also set the activity domain in this case.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 30 Dec 2011 23:28:50 +0000 (00:28 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Fri, 30 Dec 2011 23:18:11 +0000 (00:18 +0100)]
isl_morph_{dom,ran}_params: add missing isl_morph_cow
Usually, isl_morph_remove_{dom,ran}_dims will call isl_morph_cow for us,
but not when n == 0.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 28 Nov 2011 14:28:37 +0000 (15:28 +0100)]
isl_access_info: allow user interaction to restrict potential sources
In particular, the user can register a callback that will be called
any time right before the last potential source is computed.
This should be useful for fuzzy array dataflow analysis.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 26 Dec 2011 15:07:48 +0000 (16:07 +0100)]
isl_access_info_compute_flow: store sink domain map in isl_access_info
We will be able to reuse this map in the next commit.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 26 Dec 2011 14:59:57 +0000 (15:59 +0100)]
isl_access_info_alloc: use isl_calloc to allocate structure
This makes it easier to add new (zero-by-default) fields.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 26 Dec 2011 14:28:46 +0000 (15:28 +0100)]
isl_access_info_compute_flow: don't pass ownership of "acc" down
This slightly simplifies the error handling.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 17 Dec 2011 17:19:52 +0000 (18:19 +0100)]
isl 0.09
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 18 Dec 2011 08:41:15 +0000 (09:41 +0100)]
avoid "errno" argument name
On mingw, errno is defined as (*_errno()) in stdlib.h
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 17 Dec 2011 09:14:59 +0000 (10:14 +0100)]
add isl_morph_dump
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 17 Dec 2011 09:13:43 +0000 (10:13 +0100)]
rename isl_morph_dump to isl_morph_print_internal
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 16 Dec 2011 08:20:34 +0000 (09:20 +0100)]
isl_basic_set_compute_vertices: fix handling of lower-dimensional input
Support was broken by 9043b59 (explicitly differentiate between spaces
of maps, sets and parameter sets, Thu Aug 25 19:21:05 2011 +0200).
This fix should have gone in before that commit.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 14 Dec 2011 10:45:51 +0000 (11:45 +0100)]
doc: improve consistency in list of schedule options
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven van Haastregt [Wed, 14 Dec 2011 09:55:18 +0000 (10:55 +0100)]
add Feautrier's scheduling algorithm
Offer Feautrier's algorithm as an alternative to isl's default scheduling
algorithm. The schedule is extended to a full-dimensional schedule.
To select the scheduling algorithm to use, we introduce isl option
schedule-algorithm.
Signed-off-by: Sven van Haastregt <svhaastr@liacs.nl>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Mon, 12 Dec 2011 16:07:18 +0000 (17:07 +0100)]
add option to bound the constant scheduling coefficients
If larger coefficients appear as part of the input dependences, the
schedule calculation can take a very long time. We observed that the
main overhead in this calculation is due to optimizing the constant
coefficients. They are misused to increase locality by merging several
unrelated dimensions into a single dimension. This unwanted optimization
increases the complexity of the code and slows down the generated code.
We introduce a new option that bounds the values in the constant
dimension by a user defined value. If the right value is choosen,
costly overoptimization can be prevented.
This solution works, but requires the user to specify the value by
which the constants are bound. For the moment, this is our best
solution, but we hope to to find a more generic one later on.
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Sun, 11 Dec 2011 23:21:55 +0000 (00:21 +0100)]
Introduce new ISL_CTX_(GET|SET)_INT_DEF macro
This macro uses the same implementation as the corresponding
BOOL and CHOICE macros. Hence, we simplify BOOL and CHOICE by
forwarding to the new INT macro.
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Sun, 11 Dec 2011 23:21:53 +0000 (00:21 +0100)]
Fix typo in documentation
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Sun, 11 Dec 2011 23:21:52 +0000 (00:21 +0100)]
Add interface to set schedule_maximize_band_depth
Previously, schedule_maximize_band_depth could only be set with the
corresponding command line option. This patch adds two functions
to change this option when using isl as a library.
int isl_options_set_schedule_maximize_band_depth(
isl_ctx *ctx, int val);
int isl_options_get_schedule_maximize_band_depth(
isl_ctx *ctx);
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven van Haastregt [Mon, 12 Dec 2011 15:28:32 +0000 (16:28 +0100)]
doc/SubmittingPatches: fix typo in email address
Signed-off-by: Sven van Haastregt <svhaastr@liacs.nl>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven van Haastregt [Mon, 12 Dec 2011 10:45:17 +0000 (11:45 +0100)]
export schedule_split_parallel option
Signed-off-by: Sven van Haastregt <svhaastr@liacs.nl>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 11 Dec 2011 12:12:29 +0000 (13:12 +0100)]
add isl_pw_multi_aff_get_pw_aff
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 10 Dec 2011 18:43:38 +0000 (19:43 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Fri, 9 Dec 2011 18:51:06 +0000 (19:51 +0100)]
isl_map_from_pw_aff: handle NULL input
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 6 Dec 2011 08:17:03 +0000 (09:17 +0100)]
add isl_space_map_from_domain_and_range
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Sat, 3 Dec 2011 15:52:35 +0000 (16:52 +0100)]
Introduce option 'on_error'
The option on_error defines the behaviour of isl in case of an error.
By default, the option is set to 'warn'. This is the current behaviour
and means isl prints a warning and continues the execution of the
program. We also introduce two new modes: 'continue' and 'abort'.
'continue' does not print any warning, but continues execution and 'abort'
abort the program from within isl. 'abort' is useful for debuggers as
they automatically break at the place of the error condition.
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 1 Dec 2011 15:02:22 +0000 (16:02 +0100)]
export isl_map_eliminate
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Thu, 1 Dec 2011 14:15:45 +0000 (15:15 +0100)]
Improve message when ctx is freed, but still referenced
Instead of a cryptic 'Assertion "ctx->ref == 0" failed' we report
'isl_ctx freed, but some objects still reference it'.
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Tobias Grosser [Thu, 1 Dec 2011 14:15:44 +0000 (15:15 +0100)]
add isl_local_space_range
Signed-off-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 2 Dec 2011 15:39:58 +0000 (16:39 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Thu, 1 Dec 2011 10:34:31 +0000 (11:34 +0100)]
isl_pw_aff_scale: don't special case multiplication with zero
isl_pw_*_scale returns an empty result if a multiplication with zero
is being performed. We do this because zero is the default value
for polynomials. Affine expressions, by contrast, have no default value
and so we should not return an empty result in this case.
Reported-by: Tomofumi Yuki <tomofumi.yuki@gmail.com>
Tested-by: Tomofumi Yuki <tomofumi.yuki@gmail.com>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 28 Nov 2011 15:45:16 +0000 (16:45 +0100)]
add isl_set_find_dim_by_name
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 29 Nov 2011 09:17:46 +0000 (10:17 +0100)]
isl_basic_map_involves_dims: double-check divs
In principle, the extra test is not needed at this moment, but
this may change in future.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 26 Nov 2011 09:40:31 +0000 (10:40 +0100)]
isl_schedule.c: fix typo in comment
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 25 Nov 2011 10:00:17 +0000 (11:00 +0100)]
configure.ac: only create interface/Makefile if user has clang
This allows users without clang to perform a "make dist".
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 21 Nov 2011 21:27:28 +0000 (22:27 +0100)]
add isl_pw_aff_has_dim_id
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 22 Nov 2011 22:28:39 +0000 (23:28 +0100)]
add missing namespace qualification on llvm::ArrayRef
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 23 Nov 2011 09:33:26 +0000 (10:33 +0100)]
Merge branch 'maint'