X-Git-Url: http://git.sourceforge.jp/view?a=blobdiff_plain;f=gcc%2Fgraphite-interchange.c;h=cb4d32cc0d2f7d7734f1124bdc6c5c0fc4f95ccf;hb=16d17f31e8f374acefaac0f2b8f5b2a0c5fab8d2;hp=6ac7fca84df204543f50ee5c638d6a5605294074;hpb=87d25ca72902c6e2ddc2ffec4b985dded37f33f4;p=pf3gnuchains%2Fgcc-fork.git diff --git a/gcc/graphite-interchange.c b/gcc/graphite-interchange.c index 6ac7fca84df..cb4d32cc0d2 100644 --- a/gcc/graphite-interchange.c +++ b/gcc/graphite-interchange.c @@ -1,7 +1,7 @@ /* Interchange heuristics and transform for loop interchange on polyhedral representation. - Copyright (C) 2009 Free Software Foundation, Inc. + Copyright (C) 2009, 2010 Free Software Foundation, Inc. Contributed by Sebastian Pop and Harsha Jagasia . @@ -23,34 +23,17 @@ along with GCC; see the file COPYING3. If not see #include "config.h" #include "system.h" #include "coretypes.h" -#include "tm.h" -#include "ggc.h" -#include "tree.h" -#include "rtl.h" -#include "output.h" -#include "basic-block.h" -#include "diagnostic.h" #include "tree-flow.h" -#include "toplev.h" #include "tree-dump.h" -#include "timevar.h" #include "cfgloop.h" #include "tree-chrec.h" #include "tree-data-ref.h" #include "tree-scalar-evolution.h" -#include "tree-pass.h" -#include "domwalk.h" -#include "value-prof.h" -#include "pointer-set.h" -#include "gimple.h" -#include "params.h" +#include "sese.h" #ifdef HAVE_cloog -#include "cloog/cloog.h" #include "ppl_c.h" -#include "sese.h" #include "graphite-ppl.h" -#include "graphite.h" #include "graphite-poly.h" /* Builds a linear expression, of dimension DIM, representing PDR's @@ -78,15 +61,15 @@ build_linearized_memory_access (ppl_dimension_type offset, poly_dr_p pdr) ppl_dimension_type i; ppl_dimension_type first = pdr_subscript_dim (pdr, 0); ppl_dimension_type last = pdr_subscript_dim (pdr, PDR_NB_SUBSCRIPTS (pdr)); - Value size, sub_size; + mpz_t size, sub_size; graphite_dim_t dim = offset + pdr_dim (pdr); ppl_new_Linear_Expression_with_dimension (&res, dim); - value_init (size); - value_set_si (size, 1); - value_init (sub_size); - value_set_si (sub_size, 1); + mpz_init (size); + mpz_set_si (size, 1); + mpz_init (sub_size); + mpz_set_si (sub_size, 1); for (i = last - 1; i >= first; i--) { @@ -95,20 +78,108 @@ build_linearized_memory_access (ppl_dimension_type offset, poly_dr_p pdr) ppl_new_Linear_Expression_with_dimension (&le, dim - offset); ppl_set_coef (le, i, 1); ppl_max_for_le_pointset (PDR_ACCESSES (pdr), le, sub_size); - value_multiply (size, size, sub_size); + mpz_mul (size, size, sub_size); ppl_delete_Linear_Expression (le); } - value_clear (sub_size); - value_clear (size); + mpz_clear (sub_size); + mpz_clear (size); return res; } +/* Builds a partial difference equations and inserts them + into pointset powerset polyhedron P. Polyhedron is assumed + to have the format: T|I|T'|I'|G|S|S'|l1|l2. + + TIME_DEPTH is the time dimension w.r.t. which we are + differentiating. + OFFSET represents the number of dimensions between + columns t_{time_depth} and t'_{time_depth}. + DIM_SCTR is the number of scattering dimensions. It is + essentially the dimensionality of the T vector. + + The following equations are inserted into the polyhedron P: + | t_1 = t_1' + | ... + | t_{time_depth-1} = t'_{time_depth-1} + | t_{time_depth} = t'_{time_depth} + 1 + | t_{time_depth+1} = t'_{time_depth + 1} + | ... + | t_{dim_sctr} = t'_{dim_sctr}. */ + +static void +build_partial_difference (ppl_Pointset_Powerset_C_Polyhedron_t *p, + ppl_dimension_type time_depth, + ppl_dimension_type offset, + ppl_dimension_type dim_sctr) +{ + ppl_Constraint_t new_cstr; + ppl_Linear_Expression_t le; + ppl_dimension_type i; + ppl_dimension_type dim; + ppl_Pointset_Powerset_C_Polyhedron_t temp; + + /* Add the equality: t_{time_depth} = t'_{time_depth} + 1. + This is the core part of this alogrithm, since this + constraint asks for the memory access stride (difference) + between two consecutive points in time dimensions. */ + + ppl_Pointset_Powerset_C_Polyhedron_space_dimension (*p, &dim); + ppl_new_Linear_Expression_with_dimension (&le, dim); + ppl_set_coef (le, time_depth, 1); + ppl_set_coef (le, time_depth + offset, -1); + ppl_set_inhomogeneous (le, 1); + ppl_new_Constraint (&new_cstr, le, PPL_CONSTRAINT_TYPE_EQUAL); + ppl_Pointset_Powerset_C_Polyhedron_add_constraint (*p, new_cstr); + ppl_delete_Linear_Expression (le); + ppl_delete_Constraint (new_cstr); + + /* Add equalities: + | t1 = t1' + | ... + | t_{time_depth-1} = t'_{time_depth-1} + | t_{time_depth+1} = t'_{time_depth+1} + | ... + | t_{dim_sctr} = t'_{dim_sctr} + + This means that all the time dimensions are equal except for + time_depth, where the constraint is t_{depth} = t'_{depth} + 1 + step. More to this: we should be carefull not to add equalities + to the 'coupled' dimensions, which happens when the one dimension + is stripmined dimension, and the other dimension corresponds + to the point loop inside stripmined dimension. */ + + ppl_new_Pointset_Powerset_C_Polyhedron_from_Pointset_Powerset_C_Polyhedron (&temp, *p); + + for (i = 0; i < dim_sctr; i++) + if (i != time_depth) + { + ppl_new_Linear_Expression_with_dimension (&le, dim); + ppl_set_coef (le, i, 1); + ppl_set_coef (le, i + offset, -1); + ppl_new_Constraint (&new_cstr, le, PPL_CONSTRAINT_TYPE_EQUAL); + ppl_Pointset_Powerset_C_Polyhedron_add_constraint (temp, new_cstr); + + if (ppl_Pointset_Powerset_C_Polyhedron_is_empty (temp)) + { + ppl_delete_Pointset_Powerset_C_Polyhedron (temp); + ppl_new_Pointset_Powerset_C_Polyhedron_from_Pointset_Powerset_C_Polyhedron (&temp, *p); + } + else + ppl_Pointset_Powerset_C_Polyhedron_add_constraint (*p, new_cstr); + ppl_delete_Linear_Expression (le); + ppl_delete_Constraint (new_cstr); + } + + ppl_delete_Pointset_Powerset_C_Polyhedron (temp); +} + + /* Set STRIDE to the stride of PDR in memory by advancing by one in - time dimension DEPTH. */ + the loop at DEPTH. */ static void -memory_stride_in_loop (Value stride, graphite_dim_t depth, poly_dr_p pdr) +pdr_stride_in_loop (mpz_t stride, graphite_dim_t depth, poly_dr_p pdr) { ppl_dimension_type time_depth; ppl_Linear_Expression_t le, lma; @@ -215,48 +286,11 @@ memory_stride_in_loop (Value stride, graphite_dim_t depth, poly_dr_p pdr) free (map); } - /* Add equalities: - | t1 = t1' - | ... - | t_{depth-1} = t'_{depth-1} - | t_{depth+1} = t'_{depth+1} - | ... - | t_{dim_sctr} = t'_{dim_sctr} - - This means that all the time dimensions are equal except for - depth, where we will add t_{depth} = t'_{depth} + 1 in the next - step. */ - time_depth = psct_dynamic_dim (pbb, depth); - for (i = 0; i < dim_sctr; i++) - if (i != time_depth) - { - ppl_new_Linear_Expression_with_dimension (&le, new_dim); - ppl_set_coef (le, i, 1); - ppl_set_coef (le, i + offset, -1); - ppl_new_Constraint (&new_cstr, le, PPL_CONSTRAINT_TYPE_EQUAL); - ppl_Pointset_Powerset_C_Polyhedron_add_constraint (p2, new_cstr); - ppl_delete_Linear_Expression (le); - ppl_delete_Constraint (new_cstr); - } - - /* Add equality : t_{depth} = t'_{depth} + 1. - This is the core part of this alogrithm, since this - constraint asks for the memory access stride (difference) - between two consecutive points in time dimensions. */ - { - ppl_new_Linear_Expression_with_dimension (&le, new_dim); - ppl_set_coef (le, time_depth, 1); - ppl_set_coef (le, time_depth + offset, -1); - ppl_set_inhomogeneous (le, 1); - ppl_new_Constraint (&new_cstr, le, PPL_CONSTRAINT_TYPE_EQUAL); - ppl_Pointset_Powerset_C_Polyhedron_add_constraint (p2, new_cstr); - ppl_delete_Linear_Expression (le); - ppl_delete_Constraint (new_cstr); - } /* P1 = P1 inter P2. */ ppl_Pointset_Powerset_C_Polyhedron_intersection_assign (p1, p2); + build_partial_difference (&p1, time_depth, offset, dim_sctr); /* Maximise the expression L2 - L1. */ { @@ -266,13 +300,72 @@ memory_stride_in_loop (Value stride, graphite_dim_t depth, poly_dr_p pdr) ppl_max_for_le_pointset (p1, le, stride); } + if (dump_file && (dump_flags & TDF_DETAILS)) + { + char *str; + void (*gmp_free) (void *, size_t); + + fprintf (dump_file, "\nStride in BB_%d, DR_%d, depth %d:", + pbb_index (pbb), PDR_ID (pdr), (int) depth); + str = mpz_get_str (0, 10, stride); + fprintf (dump_file, " %s ", str); + mp_get_memory_functions (NULL, NULL, &gmp_free); + (*gmp_free) (str, strlen (str) + 1); + } + ppl_delete_Pointset_Powerset_C_Polyhedron (p1); ppl_delete_Pointset_Powerset_C_Polyhedron (p2); ppl_delete_Linear_Expression (le); } -/* Returns true when it is profitable to interchange time dimensions DEPTH1 - and DEPTH2 with DEPTH1 < DEPTH2 for PBB. + +/* Sets STRIDES to the sum of all the strides of the data references + accessed in LOOP at DEPTH. */ + +static void +memory_strides_in_loop_1 (lst_p loop, graphite_dim_t depth, mpz_t strides) +{ + int i, j; + lst_p l; + poly_dr_p pdr; + mpz_t s, n; + + mpz_init (s); + mpz_init (n); + + FOR_EACH_VEC_ELT (lst_p, LST_SEQ (loop), j, l) + if (LST_LOOP_P (l)) + memory_strides_in_loop_1 (l, depth, strides); + else + FOR_EACH_VEC_ELT (poly_dr_p, PBB_DRS (LST_PBB (l)), i, pdr) + { + pdr_stride_in_loop (s, depth, pdr); + mpz_set_si (n, PDR_NB_REFS (pdr)); + mpz_mul (s, s, n); + mpz_add (strides, strides, s); + } + + mpz_clear (s); + mpz_clear (n); +} + +/* Sets STRIDES to the sum of all the strides of the data references + accessed in LOOP at DEPTH. */ + +static void +memory_strides_in_loop (lst_p loop, graphite_dim_t depth, mpz_t strides) +{ + if (mpz_cmp_si (loop->memory_strides, -1) == 0) + { + mpz_set_si (strides, 0); + memory_strides_in_loop_1 (loop, depth, strides); + } + else + mpz_set (strides, loop->memory_strides); +} + +/* Return true when the interchange of loops LOOP1 and LOOP2 is + profitable. Example: @@ -353,42 +446,23 @@ memory_stride_in_loop (Value stride, graphite_dim_t depth, poly_dr_p pdr) profitable to interchange the loops at DEPTH1 and DEPTH2. */ static bool -pbb_interchange_profitable_p (graphite_dim_t depth1, graphite_dim_t depth2, - poly_bb_p pbb) +lst_interchange_profitable_p (lst_p nest, int depth1, int depth2) { - int i; - poly_dr_p pdr; - Value d1, d2, s, n; + mpz_t d1, d2; bool res; gcc_assert (depth1 < depth2); - value_init (d1); - value_set_si (d1, 0); - value_init (d2); - value_set_si (d2, 0); - value_init (s); - value_init (n); - - for (i = 0; VEC_iterate (poly_dr_p, PBB_DRS (pbb), i, pdr); i++) - { - value_set_si (n, PDR_NB_REFS (pdr)); - - memory_stride_in_loop (s, depth1, pdr); - value_multiply (s, s, n); - value_addto (d1, d1, s); + mpz_init (d1); + mpz_init (d2); - memory_stride_in_loop (s, depth2, pdr); - value_multiply (s, s, n); - value_addto (d2, d2, s); - } + memory_strides_in_loop (nest, depth1, d1); + memory_strides_in_loop (nest, depth2, d2); - res = value_lt (d1, d2); + res = mpz_cmp (d1, d2) < 0; - value_clear (d1); - value_clear (d2); - value_clear (s); - value_clear (n); + mpz_clear (d1); + mpz_clear (d2); return res; } @@ -434,47 +508,13 @@ lst_apply_interchange (lst_p lst, int depth1, int depth2) int i; lst_p l; - for (i = 0; VEC_iterate (lst_p, LST_SEQ (lst), i, l); i++) + FOR_EACH_VEC_ELT (lst_p, LST_SEQ (lst), i, l) lst_apply_interchange (l, depth1, depth2); } else pbb_interchange_loop_depths (depth1, depth2, LST_PBB (lst)); } -/* Return true when the interchange of loops at depths DEPTH1 and - DEPTH2 to all the statements below LST is profitable. */ - -static bool -lst_interchange_profitable_p (lst_p lst, int depth1, int depth2) -{ - if (!lst) - return false; - - if (LST_LOOP_P (lst)) - { - int i; - lst_p l; - bool res = false; - - for (i = 0; VEC_iterate (lst_p, LST_SEQ (lst), i, l); i++) - { - bool profitable = lst_interchange_profitable_p (l, depth1, depth2); - - if (profitable && !LST_LOOP_P (lst) - && dump_file && (dump_flags & TDF_DETAILS)) - fprintf (dump_file, - "Interchanging loops at depths %d and %d is profitable for stmt_%d.\n", - depth1, depth2, pbb_index (LST_PBB (lst))); - - res |= profitable; - } - - return res; - } - else - return pbb_interchange_profitable_p (depth1, depth2, LST_PBB (lst)); -} - /* Return true when the nest starting at LOOP1 and ending on LOOP2 is perfect: i.e. there are no sequence of statements. */ @@ -519,39 +559,51 @@ lst_perfect_nestify (lst_p loop1, lst_p loop2, lst_p *before, lst_remove_all_before_excluding_pbb (*nest, first, true); lst_remove_all_before_excluding_pbb (*nest, last, false); + + if (lst_empty_p (*before)) + { + free_lst (*before); + *before = NULL; + } + if (lst_empty_p (*after)) + { + free_lst (*after); + *after = NULL; + } + if (lst_empty_p (*nest)) + { + free_lst (*nest); + *nest = NULL; + } } /* Try to interchange LOOP1 with LOOP2 for all the statements of the body of LOOP2. LOOP1 contains LOOP2. Return true if it did the - interchange. CREATED_LOOP_BEFORE/CREATED_LOOP_AFTER are set to - true if the loop distribution created a loop before/after LOOP1. */ + interchange. */ static bool -lst_try_interchange_loops (scop_p scop, lst_p loop1, lst_p loop2, - lst_p *before, lst_p *nest, lst_p *after) +lst_try_interchange_loops (scop_p scop, lst_p loop1, lst_p loop2) { int depth1 = lst_depth (loop1); int depth2 = lst_depth (loop2); lst_p transformed; - *before = NULL; - *after = NULL; - *nest = NULL; + lst_p before = NULL, nest = NULL, after = NULL; + + if (!lst_perfectly_nested_p (loop1, loop2)) + lst_perfect_nestify (loop1, loop2, &before, &nest, &after); if (!lst_interchange_profitable_p (loop2, depth1, depth2)) return false; - if (!lst_perfectly_nested_p (loop1, loop2)) - lst_perfect_nestify (loop1, loop2, before, nest, after); - lst_apply_interchange (loop2, depth1, depth2); /* Sync the transformed LST information and the PBB scatterings before using the scatterings in the data dependence analysis. */ - if (*before || *nest || *after) + if (before || nest || after) { transformed = lst_substitute_3 (SCOP_TRANSFORMED_SCHEDULE (scop), loop1, - *before, *nest, *after); + before, nest, after); lst_update_scattering (transformed); free_lst (transformed); } @@ -564,12 +616,12 @@ lst_try_interchange_loops (scop_p scop, lst_p loop1, lst_p loop2, depth1, depth2); /* Transform the SCOP_TRANSFORMED_SCHEDULE of the SCOP. */ - lst_insert_in_sequence (*before, loop1, true); - lst_insert_in_sequence (*after, loop1, false); + lst_insert_in_sequence (before, loop1, true); + lst_insert_in_sequence (after, loop1, false); - if (*nest) + if (nest) { - lst_replace (loop1, *nest); + lst_replace (loop1, nest); free_lst (loop1); } @@ -577,109 +629,82 @@ lst_try_interchange_loops (scop_p scop, lst_p loop1, lst_p loop2, } /* Undo the transform. */ + free_lst (before); + free_lst (nest); + free_lst (after); lst_apply_interchange (loop2, depth2, depth1); - *before = NULL; - *after = NULL; - *nest = NULL; return false; } -static bool lst_do_interchange_1 (scop_p, lst_p, int *); - -/* Try to interchange LOOP with all the loops contained in the body of - LST. Return true if it did interchanged some loops. INDEX points - to the next element to be processed by lst_do_interchange. */ +/* Selects the inner loop in LST_SEQ (INNER_FATHER) to be interchanged + with the loop OUTER in LST_SEQ (OUTER_FATHER). */ static bool -lst_try_interchange (scop_p scop, lst_p loop, lst_p lst, int *index) +lst_interchange_select_inner (scop_p scop, lst_p outer_father, int outer, + lst_p inner_father) { - int i; - lst_p l; - lst_p before, nest, after; - bool res; - - if (!lst || !LST_LOOP_P (lst)) - return false; - - res = lst_try_interchange_loops (scop, loop, lst, &before, &nest, &after); - - if (before) - { - res |= lst_do_interchange_1 (scop, before, index); - (*index)++; - } - - if (nest) - res |= lst_do_interchange_1 (scop, nest, index); - else - for (i = 0; VEC_iterate (lst_p, LST_SEQ (lst), i, l); i++) - res |= lst_try_interchange (scop, loop, l, index); - - if (after) - { - res |= lst_do_interchange_1 (scop, after, index); - (*index)++; - } - - (*index)++; - return res; -} - -/* Interchanges all the loops of LOOP that are considered profitable - to interchange. Return true if it did interchanged some loops. - INDEX points to the next element to be processed by - lst_do_interchange. */ + int inner; + lst_p loop1, loop2; -static bool -lst_do_interchange_1 (scop_p scop, lst_p loop, int *index) -{ - int i; - lst_p l; - bool res = false; + gcc_assert (outer_father + && LST_LOOP_P (outer_father) + && LST_LOOP_P (VEC_index (lst_p, LST_SEQ (outer_father), outer)) + && inner_father + && LST_LOOP_P (inner_father)); - if (!loop || !LST_LOOP_P (loop)) - return false; + loop1 = VEC_index (lst_p, LST_SEQ (outer_father), outer); - for (i = 0; VEC_iterate (lst_p, LST_SEQ (loop), i, l); i++) - res |= lst_try_interchange (scop, loop, l, index); + FOR_EACH_VEC_ELT (lst_p, LST_SEQ (inner_father), inner, loop2) + if (LST_LOOP_P (loop2) + && (lst_try_interchange_loops (scop, loop1, loop2) + || lst_interchange_select_inner (scop, outer_father, outer, loop2))) + return true; - return res; + return false; } /* Interchanges all the loops of LOOP and the loops of its body that - are considered profitable to interchange. Return true if it did - interchanged some loops. INDEX points to the next element to be - processed in the LST_SEQ (LOOP) vector. */ + are considered profitable to interchange. Return the number of + interchanged loops. OUTER is the index in LST_SEQ (LOOP) that + points to the next outer loop to be considered for interchange. */ -static bool -lst_do_interchange (scop_p scop, lst_p loop, int *index) +static int +lst_interchange_select_outer (scop_p scop, lst_p loop, int outer) { lst_p l; - bool res = false; + int res = 0; + int i = 0; + lst_p father; if (!loop || !LST_LOOP_P (loop)) - return false; + return 0; - if (lst_depth (loop) >= 0) - res = lst_do_interchange_1 (scop, loop, index); + father = LST_LOOP_FATHER (loop); + if (father) + { + while (lst_interchange_select_inner (scop, father, outer, loop)) + { + res++; + loop = VEC_index (lst_p, LST_SEQ (father), outer); + } + } - while (VEC_iterate (lst_p, LST_SEQ (loop), *index, l)) - if (LST_LOOP_P (l)) - res |= lst_do_interchange (scop, l, index); - else - (*index)++; + if (LST_LOOP_P (loop)) + FOR_EACH_VEC_ELT (lst_p, LST_SEQ (loop), i, l) + if (LST_LOOP_P (l)) + res += lst_interchange_select_outer (scop, l, i); - (*index)++; return res; } -/* Interchanges all the loop depths that are considered profitable for SCOP. */ +/* Interchanges all the loop depths that are considered profitable for + SCOP. Return the number of interchanged loops. */ -bool +int scop_do_interchange (scop_p scop) { - int i = 0; - bool res = lst_do_interchange (scop, SCOP_TRANSFORMED_SCHEDULE (scop), &i); + int res = lst_interchange_select_outer + (scop, SCOP_TRANSFORMED_SCHEDULE (scop), 0); lst_update_scattering (SCOP_TRANSFORMED_SCHEDULE (scop));