X-Git-Url: http://git.sourceforge.jp/view?p=pf3gnuchains%2Fgcc-fork.git;a=blobdiff_plain;f=gcc%2Ftree-loop-linear.c;h=b771852c609cdd1cb8f6b49c8cab8e4fc0165367;hp=d04045ded4fe30c4d9a3e9589c22e9fe1896355b;hb=ccd582d74c95b050ef2ef0c69b45019ddd61bb18;hpb=3026d558160171c8e324325f3e31e2a4e5b1a41d diff --git a/gcc/tree-loop-linear.c b/gcc/tree-loop-linear.c index d04045ded4f..b771852c609 100644 --- a/gcc/tree-loop-linear.c +++ b/gcc/tree-loop-linear.c @@ -1,5 +1,6 @@ /* Linear Loop transforms - Copyright (C) 2003, 2004, 2005, 2007 Free Software Foundation, Inc. + Copyright (C) 2003, 2004, 2005, 2007, 2008, 2009, 2010 + Free Software Foundation, Inc. Contributed by Daniel Berlin . This file is part of GCC. @@ -23,11 +24,7 @@ along with GCC; see the file COPYING3. If not see #include "system.h" #include "coretypes.h" #include "tm.h" -#include "ggc.h" #include "tree.h" -#include "target.h" - -#include "rtl.h" #include "basic-block.h" #include "diagnostic.h" #include "obstack.h" @@ -47,7 +44,7 @@ along with GCC; see the file COPYING3. If not see scaling, skewing, and reversal. They are used to change the iteration order of loop nests in order to optimize data locality of traversals, or remove dependences that prevent - parallelization/vectorization/etc. + parallelization/vectorization/etc. TODO: Determine reuse vectors/matrix and use it to determine optimal transform matrix for locality purposes. @@ -57,7 +54,7 @@ along with GCC; see the file COPYING3. If not see considered. The first loop in the considered loop nest is FIRST_LOOP, and consequently, the index of the considered loop is obtained by LOOP->DEPTH - FIRST_LOOP->DEPTH - + Initializes: - DEPENDENCE_STEPS the sum of all the data dependence distances carried by loop LOOP, @@ -77,25 +74,25 @@ along with GCC; see the file COPYING3. If not see | A[{0, +, 1336}_1] | endloop_1 - gather_interchange_stats (in loop_1) will return + gather_interchange_stats (in loop_1) will return DEPENDENCE_STEPS = 3002 NB_DEPS_NOT_CARRIED_BY_LOOP = 5 ACCESS_STRIDES = 10694 - gather_interchange_stats (in loop_2) will return + gather_interchange_stats (in loop_2) will return DEPENDENCE_STEPS = 3000 NB_DEPS_NOT_CARRIED_BY_LOOP = 7 ACCESS_STRIDES = 8010 */ static void -gather_interchange_stats (VEC (ddr_p, heap) *dependence_relations, - VEC (data_reference_p, heap) *datarefs, - struct loop *loop, - struct loop *first_loop, - unsigned int *dependence_steps, - unsigned int *nb_deps_not_carried_by_loop, - double_int *access_strides) +gather_interchange_stats (VEC (ddr_p, heap) *dependence_relations ATTRIBUTE_UNUSED, + VEC (data_reference_p, heap) *datarefs ATTRIBUTE_UNUSED, + struct loop *loop ATTRIBUTE_UNUSED, + struct loop *first_loop ATTRIBUTE_UNUSED, + unsigned int *dependence_steps ATTRIBUTE_UNUSED, + unsigned int *nb_deps_not_carried_by_loop ATTRIBUTE_UNUSED, + double_int *access_strides ATTRIBUTE_UNUSED) { unsigned int i, j; struct data_dependence_relation *ddr; @@ -135,37 +132,35 @@ gather_interchange_stats (VEC (ddr_p, heap) *dependence_relations, { unsigned int it; tree ref = DR_REF (dr); - tree stmt = DR_STMT (dr); + gimple stmt = DR_STMT (dr); struct loop *stmt_loop = loop_containing_stmt (stmt); struct loop *inner_loop = first_loop->inner; - if (inner_loop != stmt_loop + if (inner_loop != stmt_loop && !flow_loop_nested_p (inner_loop, stmt_loop)) continue; - for (it = 0; it < DR_NUM_DIMENSIONS (dr); + for (it = 0; it < DR_NUM_DIMENSIONS (dr); it++, ref = TREE_OPERAND (ref, 0)) { - tree chrec = DR_ACCESS_FN (dr, it); - tree tstride = evolution_part_in_loop_num (chrec, loop->num); + int num = am_vector_index_for_loop (DR_ACCESS_MATRIX (dr), loop->num); + int istride = AM_GET_ACCESS_MATRIX_ELEMENT (DR_ACCESS_MATRIX (dr), it, num); tree array_size = TYPE_SIZE (TREE_TYPE (ref)); double_int dstride; - if (tstride == NULL_TREE - || array_size == NULL_TREE - || TREE_CODE (tstride) != INTEGER_CST + if (array_size == NULL_TREE || TREE_CODE (array_size) != INTEGER_CST) continue; - dstride = double_int_mul (tree_to_double_int (array_size), - tree_to_double_int (tstride)); + dstride = double_int_mul (tree_to_double_int (array_size), + shwi_to_double_int (istride)); (*access_strides) = double_int_add (*access_strides, dstride); } } } /* Attempt to apply interchange transformations to TRANS to maximize the - spatial and temporal locality of the loop. + spatial and temporal locality of the loop. Returns the new transform matrix. The smaller the reuse vector distances in the inner loops, the fewer the cache misses. FIRST_LOOP is the loop->num of the first loop in the analyzed loop @@ -173,16 +168,20 @@ gather_interchange_stats (VEC (ddr_p, heap) *dependence_relations, static lambda_trans_matrix -try_interchange_loops (lambda_trans_matrix trans, - unsigned int depth, +try_interchange_loops (lambda_trans_matrix trans, + unsigned int depth, VEC (ddr_p, heap) *dependence_relations, VEC (data_reference_p, heap) *datarefs, struct loop *first_loop) { + bool res; struct loop *loop_i; struct loop *loop_j; unsigned int dependence_steps_i, dependence_steps_j; double_int access_strides_i, access_strides_j; + double_int small, large, nb_iter; + double_int l1_cache_size, l2_cache_size; + int cmp; unsigned int nb_deps_not_carried_by_i, nb_deps_not_carried_by_j; struct data_dependence_relation *ddr; @@ -194,40 +193,64 @@ try_interchange_loops (lambda_trans_matrix trans, ddr = VEC_index (ddr_p, dependence_relations, 0); if (ddr == NULL || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) return trans; - + + l1_cache_size = uhwi_to_double_int (L1_CACHE_SIZE * 1024); + l2_cache_size = uhwi_to_double_int (L2_CACHE_SIZE * 1024); + /* LOOP_I is always the outer loop. */ - for (loop_j = first_loop->inner; - loop_j; + for (loop_j = first_loop->inner; + loop_j; loop_j = loop_j->inner) - for (loop_i = first_loop; - loop_depth (loop_i) < loop_depth (loop_j); + for (loop_i = first_loop; + loop_depth (loop_i) < loop_depth (loop_j); loop_i = loop_i->inner) { gather_interchange_stats (dependence_relations, datarefs, loop_i, first_loop, - &dependence_steps_i, + &dependence_steps_i, &nb_deps_not_carried_by_i, &access_strides_i); gather_interchange_stats (dependence_relations, datarefs, loop_j, first_loop, - &dependence_steps_j, - &nb_deps_not_carried_by_j, + &dependence_steps_j, + &nb_deps_not_carried_by_j, &access_strides_j); - + /* Heuristics for loop interchange profitability: + 0. Don't transform if the smallest stride is larger than + the L2 cache, or if the largest stride multiplied by the + number of iterations is smaller than the L1 cache. + 1. (spatial locality) Inner loops should have smallest dependence steps. 2. (spatial locality) Inner loops should contain more dependence relations not carried by the loop. - 3. (temporal locality) Inner loops should have smallest + 3. (temporal locality) Inner loops should have smallest array access strides. */ - if (dependence_steps_i < dependence_steps_j + + cmp = double_int_ucmp (access_strides_i, access_strides_j); + small = cmp < 0 ? access_strides_i : access_strides_j; + large = cmp < 0 ? access_strides_j : access_strides_i; + + if (double_int_ucmp (small, l2_cache_size) > 0) + continue; + + res = cmp < 0 ? + estimated_loop_iterations (loop_j, false, &nb_iter): + estimated_loop_iterations (loop_i, false, &nb_iter); + + if (res + && double_int_ucmp (double_int_mul (large, nb_iter), + l1_cache_size) < 0) + continue; + + if (dependence_steps_i < dependence_steps_j || nb_deps_not_carried_by_i > nb_deps_not_carried_by_j - || double_int_ucmp (access_strides_i, access_strides_j) < 0) + || cmp < 0) { lambda_matrix_row_exchange (LTM_MATRIX (trans), loop_depth (loop_i) - loop_depth (first_loop), @@ -235,8 +258,8 @@ try_interchange_loops (lambda_trans_matrix trans, /* Validate the resulting matrix. When the transformation is not valid, reverse to the previous transformation. */ if (!lambda_transform_legal_p (trans, depth, dependence_relations)) - lambda_matrix_row_exchange (LTM_MATRIX (trans), - loop_depth (loop_i) - loop_depth (first_loop), + lambda_matrix_row_exchange (LTM_MATRIX (trans), + loop_depth (loop_i) - loop_depth (first_loop), loop_depth (loop_j) - loop_depth (first_loop)); } } @@ -244,6 +267,46 @@ try_interchange_loops (lambda_trans_matrix trans, return trans; } +/* Return the number of nested loops in LOOP_NEST, or 0 if the loops + are not perfectly nested. */ + +unsigned int +perfect_loop_nest_depth (struct loop *loop_nest) +{ + struct loop *temp; + unsigned int depth = 1; + + /* If it's not a loop nest, we don't want it. We also don't handle + sibling loops properly, which are loops of the following form: + + | for (i = 0; i < 50; i++) + | { + | for (j = 0; j < 50; j++) + | { + | ... + | } + | for (j = 0; j < 50; j++) + | { + | ... + | } + | } + */ + + if (!loop_nest->inner || !single_exit (loop_nest)) + return 0; + + for (temp = loop_nest->inner; temp; temp = temp->inner) + { + /* If we have a sibling loop or multiple exit edges, jump ship. */ + if (temp->next || !single_exit (temp)) + return 0; + + depth++; + } + + return depth; +} + /* Perform a set of linear transforms on loops. */ void @@ -253,64 +316,53 @@ linear_transform_loops (void) loop_iterator li; VEC(tree,heap) *oldivs = NULL; VEC(tree,heap) *invariants = NULL; + VEC(tree,heap) *lambda_parameters = NULL; + VEC(gimple,heap) *remove_ivs = VEC_alloc (gimple, heap, 3); struct loop *loop_nest; + gimple oldiv_stmt; + unsigned i; FOR_EACH_LOOP (li, loop_nest, 0) { unsigned int depth = 0; VEC (ddr_p, heap) *dependence_relations; VEC (data_reference_p, heap) *datarefs; - struct loop *temp; + lambda_loopnest before, after; lambda_trans_matrix trans; - bool problem = false; struct obstack lambda_obstack; - gcc_obstack_init (&lambda_obstack); + struct loop *loop; + VEC(loop_p,heap) *nest; - /* If it's not a loop nest, we don't want it. - We also don't handle sibling loops properly, - which are loops of the following form: - for (i = 0; i < 50; i++) - { - for (j = 0; j < 50; j++) - { - ... - } - for (j = 0; j < 50; j++) - { - ... - } - } */ - if (!loop_nest->inner || !single_exit (loop_nest)) + depth = perfect_loop_nest_depth (loop_nest); + if (depth == 0) continue; + + nest = VEC_alloc (loop_p, heap, 3); + for (loop = loop_nest; loop; loop = loop->inner) + VEC_safe_push (loop_p, heap, nest, loop); + + gcc_obstack_init (&lambda_obstack); VEC_truncate (tree, oldivs, 0); VEC_truncate (tree, invariants, 0); - depth = 1; - for (temp = loop_nest->inner; temp; temp = temp->inner) - { - /* If we have a sibling loop or multiple exit edges, jump ship. */ - if (temp->next || !single_exit (temp)) - { - problem = true; - break; - } - depth ++; - } - if (problem) - continue; + VEC_truncate (tree, lambda_parameters, 0); - /* Analyze data references and dependence relations using scev. */ - datarefs = VEC_alloc (data_reference_p, heap, 10); dependence_relations = VEC_alloc (ddr_p, heap, 10 * 10); - compute_data_dependences_for_loop (loop_nest, true, &datarefs, - &dependence_relations); + if (!compute_data_dependences_for_loop (loop_nest, true, &datarefs, + &dependence_relations)) + goto free_and_continue; + + lambda_collect_parameters (datarefs, &lambda_parameters); + if (!lambda_compute_access_matrices (datarefs, lambda_parameters, + nest, &lambda_obstack)) + goto free_and_continue; if (dump_file && (dump_flags & TDF_DETAILS)) dump_ddrs (dump_file, dependence_relations); /* Build the transformation matrix. */ - trans = lambda_trans_matrix_new (depth, depth); + trans = lambda_trans_matrix_new (depth, depth, &lambda_obstack); lambda_matrix_id (LTM_MATRIX (trans), depth); trans = try_interchange_loops (trans, depth, dependence_relations, datarefs, loop_nest); @@ -341,7 +393,7 @@ linear_transform_loops (void) fprintf (dump_file, "Before:\n"); print_lambda_loopnest (dump_file, before, 'i'); } - + after = lambda_loopnest_transform (before, trans, &lambda_obstack); if (dump_file) @@ -351,6 +403,7 @@ linear_transform_loops (void) } lambda_loopnest_to_gcc_loopnest (loop_nest, oldivs, invariants, + &remove_ivs, after, trans, &lambda_obstack); modified = true; @@ -361,10 +414,15 @@ linear_transform_loops (void) obstack_free (&lambda_obstack, NULL); free_dependence_relations (dependence_relations); free_data_refs (datarefs); + VEC_free (loop_p, heap, nest); } + for (i = 0; VEC_iterate (gimple, remove_ivs, i, oldiv_stmt); i++) + remove_iv (oldiv_stmt); + VEC_free (tree, heap, oldivs); VEC_free (tree, heap, invariants); + VEC_free (gimple, heap, remove_ivs); scev_reset (); if (modified)