Sven Verdoolaege [Sun, 19 May 2013 15:35:48 +0000 (17:35 +0200)]
isl_ast_expr_from_constraint: split off constant term if it avoids operation
If one hand of the comparison operation that is constructed would be
exactly zero, then we put the constant term on that side.
This avoids the operation that adds the constant term to the other side,
resulting in slightly smaller expressions that are easier to read and
that may also reduce the risk of overflows.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 22 May 2013 10:28:26 +0000 (12:28 +0200)]
Merge branch 'maint'
Sven Verdoolaege [Wed, 22 May 2013 09:52:49 +0000 (11:52 +0200)]
isl_basic_map_foreach_lexopt: properly merge partial empty solutions
b9f95db (isl_basic_map_foreach_lexopt: properly merge partial solutions,
Wed Mar 20 09:55:50 2013 +0100) fixed the merging of partial non-empty
solutions, but did not take into account that partial solutions may
be empty. If the partial solutions are empty, then M is NULL and we
cannot dereference it.
Note that the new test case only triggers on top of d7e5adb
(isl_tab_basic_map_partial_lexopt: better exploit partial solution cache,
Tue Mar 12 17:25:04 2013 +0100) in the master branch.
Reported-by: Tomofumi Yuki <tomofumi.yuki@gmail.com>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 30 Apr 2013 18:45:00 +0000 (20:45 +0200)]
merge shared isl_multi_* declarations
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 30 Apr 2013 06:21:45 +0000 (08:21 +0200)]
declare isl_multi_* together with isl_*
This will make it easier to add new isl_multi_* types.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 30 Apr 2013 19:07:28 +0000 (21:07 +0200)]
make isl_multi_* files more self-contained
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 30 Apr 2013 06:08:55 +0000 (08:08 +0200)]
isl_space_private.h: protect against multiple inclusions
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 28 Apr 2013 12:09:46 +0000 (14:09 +0200)]
detect missing gmp.h or missing/old gmp library during configure
Until now a missing gmp.h was only detected during compilation
and a missing/old library only during linking.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 11 Apr 2013 17:55:29 +0000 (19:55 +0200)]
isl_ast_codegen.c: add_nodes: use more transparent way to break up ordering SCC
The original method for breaking up an ordering SCC depended on
isl_set_make_disjoint making the basic sets rationally disjoint.
We prefer not to depend on core isl internals in the AST generator.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 28 Apr 2013 13:33:28 +0000 (15:33 +0200)]
add isl_basic_map_order_gt
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 18 Apr 2013 14:53:51 +0000 (16:53 +0200)]
bound.c: verify_point: use isl_printer
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 28 Apr 2013 11:59:08 +0000 (13:59 +0200)]
isl_input.c: read_factor: declare variables at start of code block
Reported-by: William Bader <william.bader@gmail.com>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 28 Apr 2013 07:56:35 +0000 (09:56 +0200)]
Merge branch 'maint'
Sven Verdoolaege [Tue, 29 May 2012 13:44:46 +0000 (15:44 +0200)]
isl_*alloc*: return NULL if ctx argument is NULL
Whenever an isl function needs to perform an allocation, it calls
one of the isl_*alloc* functions on an isl_ctx obtained either from
the user or from a call to one of the *_get_ctx functions.
If this isl_ctx pointer is NULL, this means that either
the original isl_ctx_alloc failed or that the pointer on which
the *_get_ctx function was called was NULL.
In both cases, we are in an error state and we should not
allocate any more memory, but instead return as soon as possible.
We therefore return NULL from the isl_*alloc* functions,
an error condition which should already be handled by the caller.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 18 Apr 2013 09:17:35 +0000 (11:17 +0200)]
isl_space.c: add missing include
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 27 Mar 2013 14:44:45 +0000 (15:44 +0100)]
declare isl_tab_get_ctx in isl_tab.h
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 27 Mar 2013 11:09:23 +0000 (12:09 +0100)]
declare isl_*_list together with isl_*
Some of the isl_*_list types were declared together in isl/list.h.
It seems cleaner to declare them together with the corresponding isl_* instead.
Similarly, these types were defined together in isl_list_private.h and
are now defined together with the corresponding isl_*.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 21 Apr 2013 15:56:13 +0000 (17:56 +0200)]
isl_ast_graft.c: remove spurious include
The include has been redundant ever since it was introduced in
5888ac1 (add support for generating ASTs from schedule relations,
Mon Sep 17 22:11:06 2012 +0200).
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 2 Apr 2013 09:00:25 +0000 (11:00 +0200)]
isl_ast_codegen.c: remove dependence on internals of isl_constraint_list
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 1 Apr 2013 09:47:03 +0000 (11:47 +0200)]
extract generic isl_*_list_foreach_scc from isl_ast_codegen.c
isl_ast_codegen.c contained two implementations of this function
that have now been replaced by two calls to the extracted functions.
This makes isl_ast_codegen.c less dependent on the internals of
isl_basic_set_list. We should also be able to later reuse the
extracted function in other contexts. Once all users of isl_tarjan
have been converted to this interface, we probably want to remove
the old interface and perform the calls to callback as soon as
each SCC is found rather than waiting until all of them have been found.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 19 Apr 2013 14:57:00 +0000 (16:57 +0200)]
isl_map_affine_hull: make stride information explicit before dropping divs
In a2cbe44 (isl_map_affine_hull: avoid computing explicit representations
for divs, Sun Jan 15 16:12:59 2012 +0100), isl_map_affine_hull was changed
to drop unknown divs rather than computing an explicit representation
for them. This caused some stride information that involves multiple
unknown divs to get dropped.
Try and recover this lost information by making stride information
explicitly available before dropping the unknown divs.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 19 Apr 2013 14:54:18 +0000 (16:54 +0200)]
isl_morph.c: extract out isl_mat_parameter_compression_ext
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 19 Apr 2013 09:42:09 +0000 (11:42 +0200)]
isl_ast_build_get_stride_constraint: combine multiple strides
If we find multiple constraints from which can derive a stride,
then we would only keep the stride information from the last
of these constraints. We now combine multiple strides into
a single stride. Note that this may not have any effect in practice
because multiple strides are usually already combined into one
at the level of the constraints of a basic set.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 17 Apr 2013 15:51:55 +0000 (17:51 +0200)]
isl_ast_build.c: use isl_int_gcdext instead of open-coded version
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 17 Apr 2013 15:51:28 +0000 (17:51 +0200)]
add isl_int_gcdext
We will use it in the next commit.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 19 Apr 2013 08:08:11 +0000 (10:08 +0200)]
isl_ast_build_ast_from_schedule: also add stride guard in generic case
Any guard implied by the stride constraint in the AST build domain
needs to be enforced by the generated AST and may not necessarily
be implied by the strided loop itself.
We were already adding this guard to the graft in the degenerate case,
but not in the generic case.
The implied guard is usually already available in loop bounds constraints,
so it is not clear if this fixes any bug that can appear in practice,
However, this may just be a consequence of the internal representation of sets
and we should not depend too much on this internal representation.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 18 Apr 2013 14:12:00 +0000 (16:12 +0200)]
isl_ast_build_set_loop_bounds: compute divs after eliminating strided dimension
Although it is unlikely that the elimination would result in any unknown divs,
we had better make sure to preserve the invariant that build->domain does not
have any unknown divs.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 18 Apr 2013 13:02:47 +0000 (15:02 +0200)]
isl_map_simplify.c: normalize_divs_in_context: improve error handling
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 17 Apr 2013 14:20:14 +0000 (16:20 +0200)]
isl_ast_build_expr.c: remove left-over from refactoring
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 11 Apr 2013 15:36:38 +0000 (17:36 +0200)]
isl_ast_codegen.c: create_node_scaled: reflect current domain slice in schedule
If we detect that a given schedule dimension has a fixed (possibly
parametric) value, then this value is recorded in the build and
the schedule dimension is eliminated. This elimination process may
drop constraints on the part of the domain that we are currently considering,
which may cause all sorts of problems. In particular, we may fail
to discover at the end of the AST generation that the original schedule
was single valued and insert spurious loops.
Plug in the value we found instead of simply eliminating the schedule
dimension.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 12 Apr 2013 16:18:31 +0000 (18:18 +0200)]
add isl_union_map_preimage_domain_multi_aff
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 16 Apr 2013 13:01:00 +0000 (15:01 +0200)]
add isl_map_preimage_domain_multi_aff
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 16 Apr 2013 12:15:41 +0000 (14:15 +0200)]
isl_basic_map_preimage_multi_aff: micro-refactoring
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 16 Apr 2013 09:49:05 +0000 (11:49 +0200)]
isl_map_simplify.c: div_is_redundant: ignore coefficients of unknown divs
If a div is unknown, then its coefficients are undefined so we should
not take them into account while checking if some other div is redundant.
In rare cases, looking at those undefined coefficients could result
in system dependent behavior.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 12 Apr 2013 15:26:10 +0000 (17:26 +0200)]
generalize isl_basic_set_preimage_multi_aff to isl_basic_map_preimage_multi_aff
This will allow us to implement isl_map_preimage_domain_multi_aff.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 19 Apr 2013 08:08:11 +0000 (10:08 +0200)]
isl_ast_build_ast_from_schedule: also add stride guard in generic case
Any guard implied by the stride constraint in the AST build domain
needs to be enforced by the generated AST and may not necessarily
be implied by the strided loop itself.
We were already adding this guard to the graft in the degenerate case,
but not in the generic case.
The implied guard is usually already available in loop bounds constraints,
so it is not clear if this fixes any bug that can appear in practice,
However, this may just be a consequence of the internal representation of sets
and we should not depend too much on this internal representation.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 18 Apr 2013 14:12:00 +0000 (16:12 +0200)]
isl_ast_build_set_loop_bounds: compute divs after eliminating strided dimension
Although it is unlikely that the elimination would result in any unknown divs,
we had better make sure to preserve the invariant that build->domain does not
have any unknown divs.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 18 Apr 2013 13:02:47 +0000 (15:02 +0200)]
isl_map_simplify.c: normalize_divs_in_context: improve error handling
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 16 Apr 2013 11:42:09 +0000 (13:42 +0200)]
isl_map.c: fix typo in comment
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 12 Apr 2013 14:57:31 +0000 (16:57 +0200)]
extend isl_seq_preimage to allow taking preimage of a subset of dimensions
This will allow us to reuse isl_seq_preimage for computing the preimage
of the domain or range of a basic map.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 16 Apr 2013 09:49:05 +0000 (11:49 +0200)]
isl_map_simplify.c: div_is_redundant: ignore coefficients of unknown divs
If a div is unknown, then its coefficients are undefined so we should
not take them into account while checking if some other div is redundant.
In rare cases, looking at those undefined coefficients could result
in system dependent behavior.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 15 Apr 2013 20:36:51 +0000 (22:36 +0200)]
Merge branch 'maint'
Sven Verdoolaege [Mon, 15 Apr 2013 20:08:53 +0000 (22:08 +0200)]
isl_ast_codegen.c: fix typo in comment
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 15 Apr 2013 12:02:05 +0000 (14:02 +0200)]
isl_ast_build_ast_from_schedule: properly handle union atomic options
In particular, if the domain for which a given dimension should
be atomic is a union of basic sets, then the actual atomic domain
may be larger than this union and we need to make sure that this
larger set is removed from the remainder domain to preserve
the invariant that the elements of the list computed by compute_domains
are disjoint.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 15 Apr 2013 10:55:59 +0000 (12:55 +0200)]
isl_ast_codegen.c: update_unrolling_lower_bound: check for overflow
If the number of slices along a given lower bound exceeds the maximal
value of an integer, then it would mistakenly be replaced by its
least signficant bits, resulting in slices being dropped if the lower
bound ends up being selected as the best (based on the wrong number).
If the number of slices really exceeds the maximal value of an integer,
then we do not want to honor the request of the user anyway.
We should probably set an even smaller limit.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 9 Apr 2013 14:22:53 +0000 (16:22 +0200)]
isl_mat_free: return NULL
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 11 Apr 2013 15:47:21 +0000 (17:47 +0200)]
isl_basic_map_insert_dims: plug memory leak when called on empty input
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 11 Apr 2013 10:32:23 +0000 (12:32 +0200)]
ChangeLog: drop some text that appears to have been copied by mistake
In particular, the text was copied by 9fe3e28 (isl 0.10,
Sun Jun 3 15:47:09 2012 +0200).
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 9 Apr 2013 09:51:29 +0000 (11:51 +0200)]
isl 0.11.2
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 9 Apr 2013 16:36:35 +0000 (18:36 +0200)]
add missing interface/all.h and interface/isl.py.top to distribution
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 9 Apr 2013 15:37:46 +0000 (17:37 +0200)]
isl_schedule.c: sort_band_list: take into account that parent may be NULL
We missed this in 81bf4c2 (isl_schedule_get_band_forest: sort bands
in band list, Tue Feb 26 12:35:16 2013 +0100).
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 9 Apr 2013 15:37:35 +0000 (17:37 +0200)]
isl_schedule_get_map: pad schedule map with zeros if reconstructed from bands
The schedule map extracted from the isl_schedule_nodes is padded with zeros
by construction. However, if the schedule map is reconstructed from the
band forest, then the schedule map was not be padded with zeros.
Pad this schedule as well.
We may want to change isl_band_list_get_suffix_schedule to perform this
padded instead.
Reported-by: Ali El Moussawi <elmoussawi.ali@gmail.com>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 4 Apr 2013 13:22:10 +0000 (15:22 +0200)]
isl_ast_graft.c: store_guard: coalesce guard after taking gist
On rare occasions, coalescing may simplify the gisted guard.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 7 Apr 2013 15:38:40 +0000 (17:38 +0200)]
Merge branch 'maint'
Sven Verdoolaege [Fri, 29 Mar 2013 13:53:13 +0000 (14:53 +0100)]
isl_ast_node_print: avoid duplicate declarations in printed code
If a loop has only one iteration but if the expression
for this single iteration is too complicated to be substituted into the AST,
then the expression is assigned to a variable corresponding to the loop
iterator.
If two such loops occur in the same block, then we would end up with
two declarations of the same variable. Avoid this problem by introducing
a block around the declaration if it appears in a sequence of statements.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 7 Apr 2013 15:13:12 +0000 (17:13 +0200)]
isl_ast.c: extract out start_block and end_block from print_ast_node_c
We will be able to reuse them in the next commit.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 7 Apr 2013 12:07:27 +0000 (14:07 +0200)]
configure.ac: explicitly link in llvm bitreader component in extract_interface
Apparently, the bitreader component is required for ASTReader and
is no longer included implicitly.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 7 Apr 2013 11:20:41 +0000 (13:20 +0200)]
update for change in arguments of clang's HeaderSearchOptions::AddPath
In particular, the IsUserSpecified argument has been removed.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 7 Apr 2013 11:13:14 +0000 (13:13 +0200)]
update for change in arguments of clang's CompilerInstance::createDiagnostics
In particular, CompilerInstance::createDiagnostics no longer takes
argc and argv as arugments.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 1 Apr 2013 10:13:31 +0000 (12:13 +0200)]
isl_ast_codegen.c: fix typo in comment
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 27 Mar 2013 11:31:32 +0000 (12:31 +0100)]
isl_ast_graft_list_sort_guard: use isl_ast_graft_list_sort
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 27 Mar 2013 11:27:40 +0000 (12:27 +0100)]
add isl_*_list_sort
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 27 Mar 2013 11:27:23 +0000 (12:27 +0100)]
rename isl_ast_graft_list_sort to isl_ast_graft_list_sort_guard
We will introduce a generic isl_*_list_sort in the next commit.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 15 Jan 2012 15:12:59 +0000 (16:12 +0100)]
isl_map_affine_hull: avoid computing explicit representations for divs
If an existentially quantified variable is not determined by an equality
then it should have little effect on the overall affine hull.
There should therefor be no need to compute explicit representations
for these existentially quantified variables, which is a possibly costly
operation and which can result in large number of basic maps.
Instead, we first compute the affine hull of each basic map separately,
extract an explicit representation for existentially quantified variables
determined by an equality and then remove all unknown divs before aligning
the divs over the basic maps.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 23 Mar 2013 14:55:02 +0000 (15:55 +0100)]
isl_map_coalesce: extend the handling of a pair adjacent inequalities
In particular, one of the adjacent inequalities may imply an equality
on the integer points in the corresponding set and this equality will
not be valid for the other set.
For example, the set
{ [x,y] : 0 <= x <= 2 and y >= 0 and x + 2y <= 4;
[x,y] : 3 <= x <= 4 and y >= 0 and x + 2y <= 4 }
may be simplified to
{ [x,y] : 0 <= x <= 2 and y >= 0 and x + 2y <= 4;
[x,0] : 3 <= x <= 4 }
and the equality y = 0 is not valid for the other component.
However, we may still be able to fuse in such cases if all the constraints
of the other component are valid for the component with the equality and if
we can show that replacing the adjacent inequality x <= 2 by its opposite
(x >= 3) and adding the valid constraints (x <= 4) of the component with
the equality results in a subset of this component. Since it is also
a superset of this component (all its constraints are valid for the component),
we can fuse the two components by combining their valid constraints.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 23 Mar 2013 14:53:50 +0000 (15:53 +0100)]
add isl_tab_unrestrict
This will be useful for another special case of coalescing.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 27 Mar 2013 08:03:24 +0000 (09:03 +0100)]
isl_coalesce.c: fix typo in comment
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sat, 23 Mar 2013 11:20:52 +0000 (12:20 +0100)]
isl_test: generalize coalesce tests
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 28 Mar 2013 13:06:25 +0000 (14:06 +0100)]
add isl_set_{lexmin,lexmax}_pw_multi_aff
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 22 Mar 2013 12:00:34 +0000 (13:00 +0100)]
add isl_tab_get_ctx
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 22 Mar 2013 11:58:46 +0000 (12:58 +0100)]
isl_mat_variable_compression: add memory management annotations
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Fri, 22 Mar 2013 11:58:05 +0000 (12:58 +0100)]
isl_mat_product: add memory management annotations
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 24 Mar 2013 13:00:21 +0000 (14:00 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Sun, 24 Mar 2013 12:46:44 +0000 (13:46 +0100)]
isl_tab_basic_map_partial_lexopt: disable use of shifted tableau on equalities
If the input context already involves equalities then the shfited tableau
is not used as it would be empty. However, if equalities are added at
a later stage (e.g., transferred for the input map), then we would fail
to disable the use of the shifted tableau, possibly resulting in invalid
sample values being computing from this shifted tableau.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 24 Mar 2013 09:20:24 +0000 (10:20 +0100)]
isl_tab_pip.c: context_gbr_detect_equalities: improve error handling
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 20 Mar 2013 09:27:28 +0000 (10:27 +0100)]
update ax_create_pkgconfig_info.m4 to latest upstream version
The main change is that user CPPFLAGS and LDFLAGS are no longer added
to the generated .pc file. These flags are specified by the user for
compiling isl and may not make sense for compiling libraries or applications
that depend on isl.
Reported-by: Samuli Suominen <ssuominen.private@gmail.com>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 20 Mar 2013 09:17:34 +0000 (10:17 +0100)]
isl_mat_drop_cols: special case dropping of zero columns
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 20 Mar 2013 08:55:50 +0000 (09:55 +0100)]
isl_basic_map_foreach_lexopt: properly merge partial solutions
If we detect two identical solutions in adjacent parts of the search space,
then we merge them into a single solution. The combined domain may however
have fewer existentially quantified variables and so we also need to adjust
the matrix representing the solution on that domain to no longer refer
to the extra existentially quantified variables. Otherwise, the code
in sol_for_add could end up accessing elements of aff->v outside its bounds.
Note that it is safe to remove the columns that refer to the extra
existentially quantified variables since their coefficients need to be zero
for the two partial solutions to be considered identical.
Reported-by: Andreas Kloeckner <lists@informa.tiker.net>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 12 Mar 2013 16:25:04 +0000 (17:25 +0100)]
isl_tab_basic_map_partial_lexopt: better exploit partial solution cache
In 4fff507 (isl_tab_pip: keep cache of partial solutions,
Fri Oct 16 14:46:28 2009 +0200) we introduced a cache of partial solutions
so that we could attempt to merge them before adding them to the result.
The merging happens during backtracking, however, and once we reached
the end of the search tree, we would not backtrack back to the root anymore.
We now do track back to the root of the search tree such that partial
solution in the final part of the search tree could also potentially
be merged. We only do this if there is some chance that we may find
identical partial solutions since the rollback of the context tableau
does not come for free.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 12 Mar 2013 16:23:49 +0000 (17:23 +0100)]
isl_basic_map_lexopt: preinitialize domain
The original code would pass a universe domain to isl_basic_map_partial_lexopt,
but this means that the parametric integer programming engine will consider
all inequalities as splitting (with respect to the universe context),
resulting in a fair amount of wasted effort.
More importantly, the border constraints may be transferred to the
context in the middle of constraints that actually split the context
into parts with possibly different solutions. The current heuristic
for merging solutions that turn out to be identical will not be able
to transfer solution across the border constraints since they only
allow for a solution on one side of the constraint.
By prepopulating the domain with some of the constraints from the input map,
we can avoid some redundant computations and we improve the opportunities
for merging identical partial solutions.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 12 Mar 2013 16:27:44 +0000 (17:27 +0100)]
export isl_basic_map_drop_constraints_involving_dims
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 12 Mar 2013 13:56:36 +0000 (14:56 +0100)]
isl_basic_map_compute_divs: compress existentially quantified variables
This compression should simplify the computations during parametric
integer programming.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 19 Mar 2013 10:23:20 +0000 (11:23 +0100)]
isl_basic_map_compute_divs: recursively remove parameter equalities
The removal of explicit equalities among the parameters may uncover
additional equalities. If so, we now recurse in parameter_compute_divs
so that also these extra equalities are removed.
This will be important for the next commit where we will try to compress
the existentially quantified variables and this requires the absence
of equalities among the parameters.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Wed, 13 Mar 2013 07:21:36 +0000 (08:21 +0100)]
isl_map.c: parameter_compute_divs: extract out first_parameter_equality
We will be able to reuse this function in the next commit.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 12 Mar 2013 13:55:51 +0000 (14:55 +0100)]
isl_basic_set_parameter_compression: allow divs in input
For the purpose of compressing parameters, there should be no need
to make a distinction between set variables and existentially quantified
variables.
There does not appear to be any reason for not allowing existentially
quantified variables before other than the fact that we didn't need
to apply this function of basic sets with existentially quantified
variables before. However, this will turn out to be useful when
we want to compress the unknown divs within isl_basic_map_compute_divs.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 12 Mar 2013 10:09:21 +0000 (11:09 +0100)]
isl_basic_map_compute_divs: preserve already known divs
The purpose of isl_basic_map_compute_divs is to ensure that all divs
have an explicit representation. If we already have an explicit
representation for some of the divs, then there is no need to
recompute an explicit representation for them. Instead, we preserve
the explicit representations of the known divs and only compute
explicit representations for the currently unknown divs.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 17 Mar 2013 19:47:42 +0000 (20:47 +0100)]
isl_basic_map_project_out: extract out insert_div_rows
The new insert_div_rows function will be reused in the next commit.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 17 Mar 2013 19:49:31 +0000 (20:49 +0100)]
isl_basic_map_swap_div: extract out swap_div
The new swap_div will be reused in a subsequent commit.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Sun, 17 Mar 2013 18:49:06 +0000 (19:49 +0100)]
Merge branch 'maint'
Sven Verdoolaege [Sat, 16 Mar 2013 13:40:56 +0000 (14:40 +0100)]
isl_map_coalesce: avoid dropping constraints redundant wrt implicit equalities
isl_map_coalesce first detects implicit equalities in the basic maps.
If two basic maps can indeed be coalesced, then fuse() would
only add this constraint to the fused basic map as an inequality.
Other constraints that imply that this constraint is actually an equality
are redundant with respect to the equality and would not get copied
inside fuse() either. The end result is that we would lose information.
In principle, we could just modify fuse() to add the implicit equalities
to the fused basic map as equalities, but it seems more prudent to
turn them into explicit equalities from the start.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 14 Feb 2013 20:25:58 +0000 (21:25 +0100)]
add isl_band_split
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 26 Feb 2013 11:35:16 +0000 (12:35 +0100)]
isl_schedule_get_band_forest: sort bands in band list
In particular if the partial schedule of the parent of a band list
assigns different constant values to the domains of the bands in the list,
then sort the bands according to those constant values.
This should provide for a more "natural" order to the user.
The need for sorting could be avoided entirely if we were to
construct a forest of bands directly from isl_union_set_compute_schedule.
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Tue, 26 Feb 2013 11:04:18 +0000 (12:04 +0100)]
declare isl_union_pw_multi_aff_intersect_domain
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Mon, 25 Feb 2013 12:56:24 +0000 (13:56 +0100)]
add tile_shift_point_loops option
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 14 Mar 2013 17:21:47 +0000 (18:21 +0100)]
add isl_union_pw_multi_aff_scale_vec
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 14 Mar 2013 17:12:24 +0000 (18:12 +0100)]
add isl_pw_multi_aff_scale_vec
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 14 Mar 2013 16:07:52 +0000 (17:07 +0100)]
add isl_multi_aff_scale_vec
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 14 Mar 2013 16:47:14 +0000 (17:47 +0100)]
isl_basic_map_simplify: avoid infinite loop on empty input
eliminate_divs_ineq skips the step of actually removing the div
if the basic map is marked empty. This is needed because in the
process of marking the basic map empty, the constraints may have
been replaced by the canonical representation of an empty basic map
which does not have any divs.
However, the basic map may have already been marked empty before,
possibly without having the constraints replaced by the canonical
representation. In such cases, we would also skip the step that
remove the div, but it would actually still be present and we would
keep making "progress" in every iteration of isl_basic_map_simplify,
resulting in an infinite loop.
Reported-by: Tobias Grosser <tobias@grosser.es>
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 14 Mar 2013 08:31:01 +0000 (09:31 +0100)]
generalize isl_union_pw_qpolynomial_sub to include isl_union_pw_multi_aff_sub
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
Sven Verdoolaege [Thu, 14 Mar 2013 08:20:01 +0000 (09:20 +0100)]
add isl_pw_multi_aff_sub
Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>