diff options
author | Maxim Kuvyrkov <maxim@codesourcery.com> | 2012-03-02 20:32:20 +0000 |
---|---|---|
committer | Maxim Kuvyrkov <mkuvyrkov@gcc.gnu.org> | 2012-03-02 20:32:20 +0000 |
commit | 093193bed727f13cbf10b590d13c896419f215b9 (patch) | |
tree | ad317a1431c55993729264a04f7e4d8d57e5f951 /gcc | |
parent | 501699af1603287b1b47ac450fd6eeb826aa76b1 (diff) | |
download | gcc-093193bed727f13cbf10b590d13c896419f215b9.zip gcc-093193bed727f13cbf10b590d13c896419f215b9.tar.gz gcc-093193bed727f13cbf10b590d13c896419f215b9.tar.bz2 |
invoke.texi (floop-flatten): Declare obsolete.
gcc/
* doc/invoke.texi (floop-flatten): Declare obsolete.
* toplev.c (process_options): Remove references to flag_loop_flatten.
* tree-ssa-loop.c (gate_graphite_transform): Same.
* common.opt (floop-flatten): Obsolete.
* graphite-poly.c (apply_poly_transforms): Remove reference to
flag_loop_flatten.
* Makefile.in (graphite-flattening.o): Remove.
* graphite-flattening.c: Remove.
gcc/testsuite/
* gcc.dg/graphite/pr50561.c: Update.
From-SVN: r184820
Diffstat (limited to 'gcc')
-rw-r--r-- | gcc/ChangeLog | 11 | ||||
-rw-r--r-- | gcc/Makefile.in | 4 | ||||
-rw-r--r-- | gcc/common.opt | 4 | ||||
-rw-r--r-- | gcc/doc/invoke.texi | 12 | ||||
-rw-r--r-- | gcc/graphite-flattening.c | 460 | ||||
-rw-r--r-- | gcc/graphite-poly.c | 3 | ||||
-rw-r--r-- | gcc/testsuite/ChangeLog | 4 | ||||
-rw-r--r-- | gcc/testsuite/gcc.dg/graphite/pr50561.c | 2 | ||||
-rw-r--r-- | gcc/toplev.c | 3 | ||||
-rw-r--r-- | gcc/tree-ssa-loop.c | 3 |
10 files changed, 21 insertions, 485 deletions
diff --git a/gcc/ChangeLog b/gcc/ChangeLog index 811702c..e3038fb 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,14 @@ +2012-03-02 Maxim Kuvyrkov <maxim@codesourcery.com> + + * doc/invoke.texi (floop-flatten): Remove. + * toplev.c (process_options): Remove references to flag_loop_flatten. + * tree-ssa-loop.c (gate_graphite_transform): Same. + * common.opt (floop-flatten): Obsolete. + * graphite-poly.c (apply_poly_transforms): Remove reference to + flag_loop_flatten. + * Makefile.in (graphite-flattening.o): Remove. + * graphite-flattening.c: Remove. + 2012-03-02 Uros Bizjak <ubizjak@gmail.com> * compare-elim.c (find_comparisons_in_bb): Eliminate only compares diff --git a/gcc/Makefile.in b/gcc/Makefile.in index f6cf8ff..9a35295 100644 --- a/gcc/Makefile.in +++ b/gcc/Makefile.in @@ -1241,7 +1241,6 @@ OBJS = \ graphite-clast-to-gimple.o \ graphite-cloog-util.o \ graphite-dependences.o \ - graphite-flattening.o \ graphite-interchange.o \ graphite-poly.o \ graphite-ppl.o \ @@ -2619,9 +2618,6 @@ graphite-dependences.o : graphite-dependences.c $(CONFIG_H) $(SYSTEM_H) \ coretypes.h $(TREE_FLOW_H) $(TREE_DUMP_H) $(CFGLOOP_H) $(TREE_DATA_REF_H) \ sese.h graphite-ppl.h graphite-poly.h graphite-dependences.h \ graphite-cloog-util.h -graphite-flattening.o : graphite-flattening.c $(CONFIG_H) $(SYSTEM_H) \ - coretypes.h $(TREE_FLOW_H) $(TREE_DUMP_H) $(CFGLOOP_H) $(TREE_DATA_REF_H) \ - sese.h graphite-ppl.h graphite-poly.h graphite-interchange.o : graphite-interchange.c $(CONFIG_H) $(SYSTEM_H) \ coretypes.h $(TREE_FLOW_H) $(TREE_DUMP_H) $(CFGLOOP_H) $(TREE_DATA_REF_H) \ sese.h graphite-ppl.h graphite-poly.h diff --git a/gcc/common.opt b/gcc/common.opt index 984825d..98fa2bb 100644 --- a/gcc/common.opt +++ b/gcc/common.opt @@ -1198,8 +1198,8 @@ Common Report Var(flag_tm) Enable support for GNU transactional memory floop-flatten -Common Report Var(flag_loop_flatten) Optimization -Enable Loop Flattening transformation +Common Ignore +Does nothing. Preserved for backward compatibility. fstrict-volatile-bitfields Common Report Var(flag_strict_volatile_bitfields) Init(-1) diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi index b806eb4..2963045 100644 --- a/gcc/doc/invoke.texi +++ b/gcc/doc/invoke.texi @@ -371,7 +371,7 @@ Objective-C and Objective-C++ Dialects}. -fira-loop-pressure -fno-ira-share-save-slots @gol -fno-ira-share-spill-slots -fira-verbose=@var{n} @gol -fivopts -fkeep-inline-functions -fkeep-static-consts @gol --floop-block -floop-flatten -floop-interchange -floop-strip-mine @gol +-floop-block -floop-interchange -floop-strip-mine @gol -floop-parallelize-all -flto -flto-compression-level @gol -flto-partition=@var{alg} -flto-report -fmerge-all-constants @gol -fmerge-constants -fmodulo-sched -fmodulo-sched-allow-regmoves @gol @@ -7287,16 +7287,6 @@ GIMPLE -> GRAPHITE -> GIMPLE transformation. Some minimal optimizations are also performed by the code generator CLooG, like index splitting and dead code elimination in loops. -@item -floop-flatten -@opindex floop-flatten -Removes the loop nesting structure: transforms the loop nest into a -single loop. This transformation can be useful as an enablement -transform for vectorization and parallelization. This feature -is experimental. -To use this code transformation, GCC has to be configured -with @option{--with-ppl} and @option{--with-cloog} to enable the -Graphite loop transformation infrastructure. - @item -floop-parallelize-all @opindex floop-parallelize-all Use the Graphite data dependence analysis to identify loops that can diff --git a/gcc/graphite-flattening.c b/gcc/graphite-flattening.c deleted file mode 100644 index 33f25e0..0000000 --- a/gcc/graphite-flattening.c +++ /dev/null @@ -1,460 +0,0 @@ -/* Loop flattening for Graphite. - Copyright (C) 2010 Free Software Foundation, Inc. - Contributed by Sebastian Pop <sebastian.pop@amd.com>. - -This file is part of GCC. - -GCC is free software; you can redistribute it and/or modify -it under the terms of the GNU General Public License as published by -the Free Software Foundation; either version 3, or (at your option) -any later version. - -GCC is distributed in the hope that it will be useful, -but WITHOUT ANY WARRANTY; without even the implied warranty of -MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the -GNU General Public License for more details. - -You should have received a copy of the GNU General Public License -along with GCC; see the file COPYING3. If not see -<http://www.gnu.org/licenses/>. */ - -#include "config.h" -#include "system.h" -#include "coretypes.h" -#include "tree-flow.h" -#include "tree-dump.h" -#include "cfgloop.h" -#include "tree-chrec.h" -#include "tree-data-ref.h" -#include "tree-scalar-evolution.h" -#include "sese.h" - -#ifdef HAVE_cloog -#include "ppl_c.h" -#include "graphite-ppl.h" -#include "graphite-poly.h" - -/* The loop flattening pass transforms loop nests into a single loop, - removing the loop nesting structure. The auto-vectorization can - then apply on the full loop body, without needing the outer-loop - vectorization. - - The loop flattening pass that has been described in a very Fortran - specific way in the 1992 paper by Reinhard von Hanxleden and Ken - Kennedy: "Relaxing SIMD Control Flow Constraints using Loop - Transformations" available from - http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.54.5033 - - The canonical example is as follows: suppose that we have a loop - nest with known iteration counts - - | for (i = 1; i <= 6; i++) - | for (j = 1; j <= 6; j++) - | S1(i,j); - - The loop flattening is performed by linearizing the iteration space - using the function "f (x) = 6 * i + j". In this case, CLooG would - produce this code: - - | for (c1=7;c1<=42;c1++) { - | i = floord(c1-1,6); - | S1(i,c1-6*i); - | } - - There are several limitations for loop flattening that are linked - to the expressivity of the polyhedral model. One has to take an - upper bound approximation to deal with the parametric case of loop - flattening. For example, in the loop nest: - - | for (i = 1; i <= N; i++) - | for (j = 1; j <= M; j++) - | S1(i,j); - - One would like to flatten this loop using a linearization function - like this "f (x) = M * i + j". However CLooG's schedules are not - expressive enough to deal with this case, and so the parameter M - has to be replaced by an integer upper bound approximation. If we - further know in the context of the scop that "M <= 6", then it is - possible to linearize the loop with "f (x) = 6 * i + j". In this - case, CLooG would produce this code: - - | for (c1=7;c1<=6*M+N;c1++) { - | i = ceild(c1-N,6); - | if (i <= floord(c1-1,6)) { - | S1(i,c1-6*i); - | } - | } - - For an arbitrarily complex loop nest the algorithm proceeds in two - steps. First, the LST is flattened by removing the loops structure - and by inserting the statements in the order they appear in - depth-first order. Then, the scattering of each statement is - transformed accordingly. - - Supposing that the original program is represented by the following - LST: - - | (loop_1 - | stmt_1 - | (loop_2 stmt_3 - | (loop_3 stmt_4) - | (loop_4 stmt_5 stmt_6) - | stmt_7 - | ) - | stmt_2 - | ) - - Loop flattening traverses the LST in depth-first order, and - flattens pairs of loops successively by projecting the inner loops - in the iteration domain of the outer loops: - - lst_project_loop (loop_2, loop_3, stride) - - | (loop_1 - | stmt_1 - | (loop_2 stmt_3 stmt_4 - | (loop_4 stmt_5 stmt_6) - | stmt_7 - | ) - | stmt_2 - | ) - - lst_project_loop (loop_2, loop_4, stride) - - | (loop_1 - | stmt_1 - | (loop_2 stmt_3 stmt_4 stmt_5 stmt_6 stmt_7) - | stmt_2 - | ) - - lst_project_loop (loop_1, loop_2, stride) - - | (loop_1 - | stmt_1 stmt_3 stmt_4 stmt_5 stmt_6 stmt_7 stmt_2 - | ) - - At each step, the iteration domain of the outer loop is enlarged to - contain enough points to iterate over the inner loop domain. */ - -/* Initializes RES to the number of iterations of the linearized loop - LST. RES is the cardinal of the iteration domain of LST. */ - -static void -lst_linearized_niter (lst_p lst, mpz_t res) -{ - int i; - lst_p l; - mpz_t n; - - mpz_init (n); - mpz_set_si (res, 0); - - FOR_EACH_VEC_ELT (lst_p, LST_SEQ (lst), i, l) - if (LST_LOOP_P (l)) - { - lst_linearized_niter (l, n); - mpz_add (res, res, n); - } - - if (LST_LOOP_P (lst)) - { - lst_niter_for_loop (lst, n); - - if (mpz_cmp_si (res, 0) != 0) - mpz_mul (res, res, n); - else - mpz_set (res, n); - } - - mpz_clear (n); -} - -/* Applies the translation "f (x) = x + OFFSET" to the loop containing - STMT. */ - -static void -lst_offset (lst_p stmt, mpz_t offset) -{ - lst_p inner = LST_LOOP_FATHER (stmt); - poly_bb_p pbb = LST_PBB (stmt); - ppl_Polyhedron_t poly = PBB_TRANSFORMED_SCATTERING (pbb); - int inner_depth = lst_depth (inner); - ppl_dimension_type inner_dim = psct_dynamic_dim (pbb, inner_depth); - ppl_Linear_Expression_t expr; - ppl_dimension_type dim; - ppl_Coefficient_t one; - mpz_t x; - - mpz_init (x); - mpz_set_si (x, 1); - ppl_new_Coefficient (&one); - ppl_assign_Coefficient_from_mpz_t (one, x); - - ppl_Polyhedron_space_dimension (poly, &dim); - ppl_new_Linear_Expression_with_dimension (&expr, dim); - - ppl_set_coef (expr, inner_dim, 1); - ppl_set_inhomogeneous_gmp (expr, offset); - ppl_Polyhedron_affine_image (poly, inner_dim, expr, one); - ppl_delete_Linear_Expression (expr); - ppl_delete_Coefficient (one); -} - -/* Scale by FACTOR the loop LST containing STMT. */ - -static void -lst_scale (lst_p lst, lst_p stmt, mpz_t factor) -{ - mpz_t x; - ppl_Coefficient_t one; - int outer_depth = lst_depth (lst); - poly_bb_p pbb = LST_PBB (stmt); - ppl_Polyhedron_t poly = PBB_TRANSFORMED_SCATTERING (pbb); - ppl_dimension_type outer_dim = psct_dynamic_dim (pbb, outer_depth); - ppl_Linear_Expression_t expr; - ppl_dimension_type dim; - - mpz_init (x); - mpz_set_si (x, 1); - ppl_new_Coefficient (&one); - ppl_assign_Coefficient_from_mpz_t (one, x); - - ppl_Polyhedron_space_dimension (poly, &dim); - ppl_new_Linear_Expression_with_dimension (&expr, dim); - - /* outer_dim = factor * outer_dim. */ - ppl_set_coef_gmp (expr, outer_dim, factor); - ppl_Polyhedron_affine_image (poly, outer_dim, expr, one); - ppl_delete_Linear_Expression (expr); - - mpz_clear (x); - ppl_delete_Coefficient (one); -} - -/* Project the INNER loop into the iteration domain of the OUTER loop. - STRIDE is the number of iterations between two iterations of the - outer loop. */ - -static void -lst_project_loop (lst_p outer, lst_p inner, mpz_t stride) -{ - int i; - lst_p stmt; - mpz_t x; - ppl_Coefficient_t one; - int outer_depth = lst_depth (outer); - int inner_depth = lst_depth (inner); - - mpz_init (x); - mpz_set_si (x, 1); - ppl_new_Coefficient (&one); - ppl_assign_Coefficient_from_mpz_t (one, x); - - FOR_EACH_VEC_ELT (lst_p, LST_SEQ (inner), i, stmt) - { - poly_bb_p pbb = LST_PBB (stmt); - ppl_Polyhedron_t poly = PBB_TRANSFORMED_SCATTERING (pbb); - ppl_dimension_type outer_dim = psct_dynamic_dim (pbb, outer_depth); - ppl_dimension_type inner_dim = psct_dynamic_dim (pbb, inner_depth); - ppl_Linear_Expression_t expr; - ppl_dimension_type dim; - ppl_dimension_type *ds; - - /* There should be no loops under INNER. */ - gcc_assert (!LST_LOOP_P (stmt)); - ppl_Polyhedron_space_dimension (poly, &dim); - ppl_new_Linear_Expression_with_dimension (&expr, dim); - - /* outer_dim = outer_dim * stride + inner_dim. */ - ppl_set_coef (expr, inner_dim, 1); - ppl_set_coef_gmp (expr, outer_dim, stride); - ppl_Polyhedron_affine_image (poly, outer_dim, expr, one); - ppl_delete_Linear_Expression (expr); - - /* Project on inner_dim. */ - ppl_new_Linear_Expression_with_dimension (&expr, dim - 1); - ppl_Polyhedron_affine_image (poly, inner_dim, expr, one); - ppl_delete_Linear_Expression (expr); - - /* Remove inner loop and the static schedule of its body. */ - /* FIXME: As long as we use PPL we are not able to remove the old - scattering dimensions. The reason is that these dimensions are not - entirely unused. They are not necessary as part of the scheduling - vector, as the earlier dimensions already unambiguously define the - execution time, however they may still be needed to carry modulo - constraints as introduced e.g. by strip mining. The correct solution - would be to project these dimensions out of the scattering polyhedra. - In case they are still required to carry modulo constraints they should be kept - internally as existentially quantified dimensions. PPL does only support - projection of rational polyhedra, however in this case we need an integer - projection. With isl this will be trivial to implement. For now we just - leave the dimensions. This is a little ugly, but should be correct. */ - if (0) { - ds = XNEWVEC (ppl_dimension_type, 2); - ds[0] = inner_dim; - ds[1] = inner_dim + 1; - ppl_Polyhedron_remove_space_dimensions (poly, ds, 2); - PBB_NB_SCATTERING_TRANSFORM (pbb) -= 2; - free (ds); - } - } - - mpz_clear (x); - ppl_delete_Coefficient (one); -} - -/* Flattens the loop nest LST. Return true when something changed. - OFFSET is used to compute the number of iterations of the outermost - loop before the current LST is executed. */ - -static bool -lst_flatten_loop (lst_p lst, mpz_t init_offset) -{ - int i; - lst_p l; - bool res = false; - mpz_t n, one, offset, stride; - - mpz_init (n); - mpz_init (one); - mpz_init (offset); - mpz_init (stride); - mpz_set (offset, init_offset); - mpz_set_si (one, 1); - - lst_linearized_niter (lst, stride); - lst_niter_for_loop (lst, n); - mpz_tdiv_q (stride, stride, n); - - FOR_EACH_VEC_ELT (lst_p, LST_SEQ (lst), i, l) - if (LST_LOOP_P (l)) - { - res = true; - - lst_flatten_loop (l, offset); - lst_niter_for_loop (l, n); - - lst_project_loop (lst, l, stride); - - /* The offset is the number of iterations minus 1, as we want - to execute the next statements at the same iteration as the - last iteration of the loop. */ - mpz_sub (n, n, one); - mpz_add (offset, offset, n); - } - else - { - lst_scale (lst, l, stride); - if (mpz_cmp_si (offset, 0) != 0) - lst_offset (l, offset); - } - - FOR_EACH_VEC_ELT (lst_p, LST_SEQ (lst), i, l) - if (LST_LOOP_P (l)) - lst_remove_loop_and_inline_stmts_in_loop_father (l); - - mpz_clear (n); - mpz_clear (one); - mpz_clear (offset); - mpz_clear (stride); - return res; -} - -/* Remove all but the first 3 dimensions of the scattering: - - dim0: the static schedule for the loop - - dim1: the dynamic schedule of the loop - - dim2: the static schedule for the loop body. */ - -static void -remove_unused_scattering_dimensions (lst_p lst) -{ - int i; - lst_p stmt; - mpz_t x; - ppl_Coefficient_t one; - - mpz_init (x); - mpz_set_si (x, 1); - ppl_new_Coefficient (&one); - ppl_assign_Coefficient_from_mpz_t (one, x); - - FOR_EACH_VEC_ELT (lst_p, LST_SEQ (lst), i, stmt) - { - poly_bb_p pbb = LST_PBB (stmt); - ppl_Polyhedron_t poly = PBB_TRANSFORMED_SCATTERING (pbb); - int j, nb_dims_to_remove = PBB_NB_SCATTERING_TRANSFORM (pbb) - 3; - ppl_dimension_type *ds; - - /* There should be no loops inside LST after flattening. */ - gcc_assert (!LST_LOOP_P (stmt)); - - if (!nb_dims_to_remove) - continue; - - ds = XNEWVEC (ppl_dimension_type, nb_dims_to_remove); - for (j = 0; j < nb_dims_to_remove; j++) - ds[j] = j + 3; - - ppl_Polyhedron_remove_space_dimensions (poly, ds, nb_dims_to_remove); - PBB_NB_SCATTERING_TRANSFORM (pbb) -= nb_dims_to_remove; - free (ds); - } - - mpz_clear (x); - ppl_delete_Coefficient (one); -} - -/* Flattens all the loop nests of LST. Return true when something - changed. */ - -static bool -lst_do_flatten (lst_p lst) -{ - int i; - lst_p l; - bool res = false; - mpz_t zero; - - if (!lst - || !LST_LOOP_P (lst)) - return false; - - mpz_init (zero); - mpz_set_si (zero, 0); - - FOR_EACH_VEC_ELT (lst_p, LST_SEQ (lst), i, l) - if (LST_LOOP_P (l)) - { - res |= lst_flatten_loop (l, zero); - - /* FIXME: As long as we use PPL we are not able to remove the old - scattering dimensions. The reason is that these dimensions are not - entirely unused. They are not necessary as part of the scheduling - vector, as the earlier dimensions already unambiguously define the - execution time, however they may still be needed to carry modulo - constraints as introduced e.g. by strip mining. The correct solution - would be to project these dimensions out of the scattering polyhedra. - In case they are still required to carry modulo constraints they should be kept - internally as existentially quantified dimensions. PPL does only support - projection of rational polyhedra, however in this case we need an integer - projection. With isl this will be trivial to implement. For now we just - leave the dimensions. This is a little ugly, but should be correct. */ - if (0) - remove_unused_scattering_dimensions (l); - } - - lst_update_scattering (lst); - mpz_clear (zero); - return res; -} - -/* Flatten all the loop nests in SCOP. Returns true when something - changed. */ - -bool -flatten_all_loops (scop_p scop) -{ - return lst_do_flatten (SCOP_TRANSFORMED_SCHEDULE (scop)); -} - -#endif diff --git a/gcc/graphite-poly.c b/gcc/graphite-poly.c index db5b0cb..9f3ba1d 100644 --- a/gcc/graphite-poly.c +++ b/gcc/graphite-poly.c @@ -771,9 +771,6 @@ apply_poly_transforms (scop_p scop) transform_done |= scop_do_interchange (scop); } - if (flag_loop_flatten) - transform_done |= flatten_all_loops (scop); - /* This feature is only enabled in the Graphite branch. */ if (0) { diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog index 4e11259..4868f33 100644 --- a/gcc/testsuite/ChangeLog +++ b/gcc/testsuite/ChangeLog @@ -1,3 +1,7 @@ +2012-03-02 Maxim Kuvyrkov <maxim@codesourcery.com> + + * gcc.dg/graphite/pr50561.c: Update. + 2012-03-02 Peter Bergner <bergner@vnet.ibm.com> * gcc.target/powerpc/pr52457.c: New test. diff --git a/gcc/testsuite/gcc.dg/graphite/pr50561.c b/gcc/testsuite/gcc.dg/graphite/pr50561.c index 70c6bbc..5fbad4a 100644 --- a/gcc/testsuite/gcc.dg/graphite/pr50561.c +++ b/gcc/testsuite/gcc.dg/graphite/pr50561.c @@ -1,5 +1,5 @@ /* { dg-do compile } */ -/* { dg-options "-O2 -floop-flatten -floop-strip-mine" } */ +/* { dg-options "-O2 -floop-strip-mine" } */ void f (unsigned *s) { diff --git a/gcc/toplev.c b/gcc/toplev.c index c145af4..6656640 100644 --- a/gcc/toplev.c +++ b/gcc/toplev.c @@ -1315,12 +1315,11 @@ process_options (void) if (flag_graphite || flag_graphite_identity || flag_loop_block - || flag_loop_flatten || flag_loop_interchange || flag_loop_strip_mine || flag_loop_parallelize_all) sorry ("Graphite loop optimizations cannot be used (-fgraphite, " - "-fgraphite-identity, -floop-block, -floop-flatten, " + "-fgraphite-identity, -floop-block, " "-floop-interchange, -floop-strip-mine, -floop-parallelize-all, " "and -ftree-loop-linear)"); #endif diff --git a/gcc/tree-ssa-loop.c b/gcc/tree-ssa-loop.c index 1a9c2be..91eeb16 100644 --- a/gcc/tree-ssa-loop.c +++ b/gcc/tree-ssa-loop.c @@ -266,8 +266,7 @@ gate_graphite_transforms (void) || flag_loop_interchange || flag_loop_strip_mine || flag_graphite_identity - || flag_loop_parallelize_all - || flag_loop_flatten) + || flag_loop_parallelize_all) flag_graphite = 1; return flag_graphite != 0; |