>> However, this form doesn't have loop guards which are generated by >> graphite_create_new_loop_guard in gcc/graphite-isl-ast-to-gimple.c and >> by graphite_create_new_loop_guard in graphite-clast-to-gimple.c. > > > Maybe the guards are directly constant folded? Can you try with:
I've tried this. It seems that the result is the same. If we consider the following code, we'll see that the guards are generated: int main (int n, int *a) { int i; for (i = n; i < 100; i++) a[i] = i; return 0; } gcc/graphite-isl-ast-to-gimple.c loop_0 (header = 0, latch = 1, niter = ) { bb_2 (preds = {bb_0 }, succs = {bb_4 bb_3 }) { <bb 2>: if (n_3(D) <= 99) goto <bb 4>; else goto <bb 3>; } bb_3 (preds = {bb_2 bb_8 }, succs = {bb_1 }) { <bb 3>: # .MEM_12 = PHI <.MEM_4(D)(2), .MEM_11(8)> # VUSE <.MEM_12> return 0; } bb_4 (preds = {bb_2 }, succs = {bb_5 bb_8 }) { <bb 4>: _2 = n_3(D) <= 99; if (_2 != 0) goto <bb 5>; else goto <bb 8>; } bb_5 (preds = {bb_4 }, succs = {bb_6 }) { <bb 5>: _1 = (__int128) n_3(D); _16 = 99 - _1; } bb_8 (preds = {bb_4 bb_6 }, succs = {bb_3 }) { <bb 8>: # .MEM_11 = PHI <.MEM_4(D)(4), .MEM_4(D)(6)> goto <bb 3>; } loop_2 (header = 6, latch = 7, niter = (uint128_t) MAX_EXPR <_16, 0>, upper_bound = 0x7ffffffffffffffffffffffffffffffe) { bb_6 (preds = {bb_5 bb_7 }, succs = {bb_7 bb_8 }) { <bb 6>: # graphite_IV.3_17 = PHI <0(5), graphite_IV.3_18(7)> graphite_IV.3_18 = graphite_IV.3_17 + 1; if (graphite_IV.3_17 < _16) goto <bb 7>; else goto <bb 8>; } bb_7 (preds = {bb_6 }, succs = {bb_6 }) { <bb 7>: goto <bb 6>; } } } graphite-clast-to-gimple.c loop_0 (header = 0, latch = 1, niter = ) { bb_2 (preds = {bb_0 }, succs = {bb_4 bb_3 }) { <bb 2>: if (n_3(D) <= 99) goto <bb 4>; else goto <bb 3>; } bb_3 (preds = {bb_2 bb_8 }, succs = {bb_1 }) { <bb 3>: # .MEM_12 = PHI <.MEM_4(D)(2), .MEM_25(8)> # VUSE <.MEM_12> return 0; } bb_4 (preds = {bb_2 }, succs = {bb_5 bb_8 }) { <bb 4>: _2 = n_3(D) <= 99; if (_2 != 0) goto <bb 5>; else goto <bb 8>; } bb_5 (preds = {bb_4 }, succs = {bb_6 }) { <bb 5>: _1 = 99 - n_3(D); } bb_8 (preds = {bb_6 bb_4 }, succs = {bb_3 }) { <bb 8>: # .MEM_25 = PHI <.MEM_18(6), .MEM_4(D)(4)> goto <bb 3>; } loop_2 (header = 6, latch = 7, niter = (unsigned int) MAX_EXPR <_1, 0>, upper_bound = 2147483646) { bb_6 (preds = {bb_5 bb_7 }, succs = {bb_7 bb_8 }) { <bb 6>: # graphite_IV.3_16 = PHI <0(5), graphite_IV.3_17(7)> # .MEM_26 = PHI <.MEM_4(D)(5), .MEM_18(7)> _19 = (sizetype) n_3(D); _20 = (sizetype) graphite_IV.3_16; _21 = _19 + _20; _22 = _21 * 4; _23 = a_7(D) + _22; _24 = n_3(D) + graphite_IV.3_16; # .MEM_18 = VDEF <.MEM_26> *_23 = _24; graphite_IV.3_17 = graphite_IV.3_16 + 1; if (graphite_IV.3_16 < _1) goto <bb 7>; else goto <bb 8>; } bb_7 (preds = {bb_6 }, succs = {bb_6 }) { <bb 7>: goto <bb 6>; } } } >> Below is the code of this generation (It still uses isl_int for >> generation of isl_expr_int, because the error related to isl/val_gmp.h >> still arises. I've tried to use isl 0.12.2 and 0.13, but gotten the >> same error). > > > Did using 'extern "C"' around the include statement not help? I can't build gcc without using 'extern "C"'. After the successful building the error is arising when I am trying to compile, for example, the following code: int main (int n, int *a) { int i; for (i = n; i < 100; i++) a[i] = i; return 0; } >> +/* Stores the INDEX in a vector and the loop nesting LEVEL for a given >> + isl_id NAME. BOUND_ONE and BOUND_TWO represent the exact lower and >> + upper bounds that can be inferred from the polyhedral representation. >> */ > > > Why do you mention BOUND_ONE & BOUND_TWO? I do not see any use of them? This is a misprint. I've fixed it. > Any reason this is not a simpel std::map<isl_id *, tree>, but you instead > have this manually implemented hash table? I think that using of std::map may create installation dependency, which requires libstdc++. I've asked the community about this. >Does gcc have a policy that forbids std::map? I think, that it is normal to use std::map from the GNU C++ Library. > Is there a reason we have separat newivs_index and params_index maps? > This was necessary for CLooG as far as I remember, but for isl a simple > isl_id -> tree map should be sufficient, no? > Is there aneed for the level at all? I think in the current > implementation and for isl in general it may not be needed as we have > isl_ids to find the corresponding loop ivs. Yes, this is redundant and removed in the new version. I wanted to write the code, which reuses the code from graphite-clast-to-gimple.c as much as possible, to reduce a number of errors. After this, I planned to eliminate redundant parts. > Please explain why we need a special function to get the upper bound. > Specifically, explain that isl in some configurations can generate loop > exit conditions such as: > > for (i = ?; i + 2 >= 3 && 3 <= i + m; i++) > ; > > This get upper bound assumes a certain shape of loop exit condition( > > for (i = ?; i < expr; i++) > > Also, you need to set the option --no-isl-ast-build-atomic-upper-bound > in isl_ast_build to be able to rely on this behavior. > (You did this below) > > Do you have a specific motivation, why you don't want to support > arbitrary expressions? I assume this is necessary for the if - do-while > optimization. If this is the case please state this. I haven't found another function for generation loops in Gimple form. The create_empty_loop_on_edge needs, which is being used now, requires upper bound. /* create_empty_loop_on_edge | | - pred_bb - -------- pred_bb ------ | | | | iv0 = initial_value | | -----|----- --------------|----------- | | ___ | entry_edge | | entry_edge / | | | | ====> / -----V---V- loop_header ------------- | V | | iv_before = phi (iv0, iv_after) | | - succ_bb - | ----|----------------------------------------- | | | | | | ----------- | ---V--- loop_body -------------------- | | | iv_after = iv_before + stride | | | | if (iv_before < upper_bound) | | | ---|--------------------------\-------------- | | | \ exit_e | | V \ | | - loop_latch - ---V- succ_bb - | | | | | | | | /----------------- ------------------------- | \ _ / Furthermore, at the moment of loop generation we don't have the induction variable, which is need for generation of a loop condition in case of the option –no-isl-ast-build-atomic-upper-bound is unset. The induction variable is returned by create_empty_loop_on_edge. Could you please advise me another function to generate them? > Please explain why we do not just generate a loop that has the loop > bound at the top, but instead create a structure of the form > > if (lb > ub) > do { > > } while (lb ..) > > (Such a figure, but completed might help). > > (I think the original motivation was that later we may be able to prove > that a loop is executed at least once which allows us to remove the if > which again enables better loop invariant code motion) I didn't have special intentions for this. As was mentioned above, I haven't found another way for generation of loops. > Why are the previous two functions necessary? Yes, they are unimportant. I've replaced them with add_parameters_to_ivs_params, which add tree representations and names of parameters to ivs_param > Why did you switch back to an isl_map? > This seems incorrect for scops with more than two statements. Yes, this is a mistake. I've fixed it. > P.S.: I just wanted to let you know that your work is amazing. Almost > fully unsupervised you are always providing high-quality patches! I am > very impressed. Thank you! -- Cheers, Roman Gareev
Index: gcc/graphite-isl-ast-to-gimple.c =================================================================== --- gcc/graphite-isl-ast-to-gimple.c (revision 212194) +++ gcc/graphite-isl-ast-to-gimple.c (working copy) @@ -42,16 +42,610 @@ #include "cfgloop.h" #include "tree-data-ref.h" #include "sese.h" +#include "tree-ssa-loop-manip.h" +#include "tree-scalar-evolution.h" #ifdef HAVE_cloog #include "graphite-poly.h" #include "graphite-isl-ast-to-gimple.h" +#include "graphite-htab.h" /* This flag is set when an error occurred during the translation of ISL AST to Gimple. */ static bool graphite_regenerate_error; +/* We always use signed 128, until the isl is able to give information about +types */ + +static tree *graphite_temporary_tree_type = &int128_integer_type_node; + +/* Converts a GMP constant VAL to a tree and returns it. */ + +static tree +gmp_cst_to_tree (tree type, mpz_t val) +{ + tree t = type ? type : integer_type_node; + mpz_t tmp; + + mpz_init (tmp); + mpz_set (tmp, val); + wide_int wi = wi::from_mpz (t, tmp, true); + mpz_clear (tmp); + + return wide_int_to_tree (t, wi); +} + +/* Verifies properties that GRAPHITE should maintain during translation. */ + +static inline void +graphite_verify (void) +{ +#ifdef ENABLE_CHECKING + verify_loop_structure (); + verify_loop_closed_ssa (true); +#endif +} + +/* Stores the INDEX in a vector for a given isl_id NAME. */ + +typedef struct ast_isl_name_index { + int index; + const char *name; + /* If free_name is set, the content of name was allocated by us and needs + to be freed. */ + char *free_name; +} *ast_isl_name_index_p; + +/* Helper for hashing ast_isl_name_index. */ + +struct ast_isl_index_hasher +{ + typedef ast_isl_name_index value_type; + typedef ast_isl_name_index compare_type; + static inline hashval_t hash (const value_type *); + static inline bool equal (const value_type *, const compare_type *); + static inline void remove (value_type *); +}; + +/* Computes a hash function for database element E. */ + +inline hashval_t +ast_isl_index_hasher::hash (const value_type *e) +{ + hashval_t hash = 0; + + int length = strlen (e->name); + int i; + + for (i = 0; i < length; ++i) + hash = hash | (e->name[i] << (i % 4)); + + return hash; +} + +/* Compares database elements ELT1 and ELT2. */ + +inline bool +ast_isl_index_hasher::equal (const value_type *elt1, const compare_type *elt2) +{ + return strcmp (elt1->name, elt2->name) == 0; +} + +/* Free the memory taken by a ast_isl_name_index struct. */ + +inline void +ast_isl_index_hasher::remove (value_type *c) +{ + if (c->free_name) + free (c->free_name); + free (c); +} + +typedef hash_table<ast_isl_index_hasher> ast_isl_index_htab_type; + +/* Returns a pointer to a new element of type ast_isl_name_index_p built + from NAME, INDEX. */ + +static inline ast_isl_name_index_p +new_ast_isl_name_index (const char *name, int index) +{ + ast_isl_name_index_p res = XNEW (struct ast_isl_name_index); + char *new_name = XNEWVEC (char, strlen (name) + 1); + strcpy (new_name, name); + + res->name = new_name; + res->free_name = new_name; + res->index = index; + return res; +} + +/* For a given name of the isl_id, which is stored in the isl_ast_expr EXPR_ID, + returns -1 if it does not correspond to any parameter, or otherwise, returns + the index in the PARAMS or SCATTERING_DIMENSIONS vector. */ + +static inline int +ast_expr_id_to_index (__isl_keep isl_ast_expr *expr_id, + ast_isl_index_htab_type *index_table) +{ + struct ast_isl_name_index tmp; + ast_isl_name_index **slot; + + isl_id *tmp_isl_id = isl_ast_expr_get_id (expr_id); + tmp.name = isl_id_get_name (tmp_isl_id); + tmp.free_name = NULL; + + slot = index_table->find_slot (&tmp, NO_INSERT); + + isl_id_free (tmp_isl_id); + if (slot && *slot) + return (*slot)->index; + + return -1; +} + +/* Records in INDEX_TABLE the INDEX and for NAME. */ + +static inline void +save_isl_id_name_index (ast_isl_index_htab_type *index_table, + const char *name, int index) +{ + struct ast_isl_name_index tmp; + ast_isl_name_index **slot; + + tmp.name = name; + tmp.free_name = NULL; + slot = index_table->find_slot (&tmp, INSERT); + + if (slot) + { + free (*slot); + *slot = new_ast_isl_name_index (name, index); + } +} + +/* INDEX binds ISL's scattering and parameter name to the index of the tree + induction variable and paramatere in IVS_PARAMS_VEC. + + PARAMS_INDEX binds ISL's parameter name to the index of the tree + parameter in PARAMS. */ + +typedef struct ivs_params { + vec<tree> ivs_params_vec; + ast_isl_index_htab_type *ivs_params_index; + sese region; +} *ivs_params_p; + +static tree +gcc_expression_from_isl_expression (tree type, __isl_keep isl_ast_expr *, + ivs_params_p ip); + +/* Returns the tree variable from the name of isl_id, which is stored + in the isl_ast_expr EXPR_ID that was given in ISL representation. */ + +static tree +gcc_expression_from_isl_ast_expr_id (__isl_keep isl_ast_expr *expr_id, + ivs_params_p ip) +{ + gcc_assert (isl_ast_expr_get_type (expr_id) == isl_ast_expr_id); + int index; + gcc_assert (ip->ivs_params_vec.exists () && ip->ivs_params_index); + index = ast_expr_id_to_index (expr_id, ip->ivs_params_index); + gcc_assert (index >= 0); + return (ip->ivs_params_vec)[index]; +} + +/* Converts a isl_ast_expr_int expression E to a GCC expression tree of + type TYPE. */ + +static tree +gcc_expression_from_isl_expr_int (tree type, __isl_keep isl_ast_expr *expr) +{ + gcc_assert (isl_ast_expr_get_type (expr) == isl_ast_expr_int); + isl_int val; + isl_int_init (val); + if (isl_ast_expr_get_int (expr, &val) == -1) + { + isl_int_clear (val); + return NULL_TREE; + } + else + return gmp_cst_to_tree (type, val); +} + +/* Converts a binary isl_ast_expr_op expression E to a GCC expression tree of + type TYPE. */ + +static tree +binary_op_to_tree (tree type, __isl_keep isl_ast_expr *expr, ivs_params_p ip) +{ + isl_ast_expr *arg_expr = isl_ast_expr_get_op_arg (expr, 0); + tree tree_lhs_expr = gcc_expression_from_isl_expression (type, arg_expr, ip); + isl_ast_expr_free (arg_expr); + arg_expr = isl_ast_expr_get_op_arg (expr, 1); + tree tree_rhs_expr = gcc_expression_from_isl_expression (type, arg_expr, ip); + isl_ast_expr_free (arg_expr); + switch (isl_ast_expr_get_op_type (expr)) + { + case isl_ast_op_add: + return fold_build2 (PLUS_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_sub: + return fold_build2 (MINUS_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_mul: + return fold_build2 (MULT_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_div: + return fold_build2 (EXACT_DIV_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_fdiv_q: + return fold_build2 (FLOOR_DIV_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_and: + return fold_build2 (TRUTH_ANDIF_EXPR, type, + tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_or: + return fold_build2 (TRUTH_ORIF_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_eq: + return fold_build2 (EQ_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_le: + return fold_build2 (LE_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_lt: + return fold_build2 (LT_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_ge: + return fold_build2 (GE_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + case isl_ast_op_gt: + return fold_build2 (GT_EXPR, type, tree_lhs_expr, tree_rhs_expr); + + default: + gcc_unreachable (); + } +} + +/* Converts a ternary isl_ast_expr_op expression E to a GCC expression tree of + type TYPE. */ + +static tree +ternary_op_to_tree (tree type, __isl_keep isl_ast_expr *expr, ivs_params_p ip) +{ + gcc_assert (isl_ast_expr_get_op_type (expr) == isl_ast_op_minus); + isl_ast_expr *arg_expr = isl_ast_expr_get_op_arg (expr, 0); + tree tree_first_expr + = gcc_expression_from_isl_expression (type, arg_expr, ip); + isl_ast_expr_free (arg_expr); + arg_expr = isl_ast_expr_get_op_arg (expr, 1); + tree tree_second_expr + = gcc_expression_from_isl_expression (type, arg_expr, ip); + isl_ast_expr_free (arg_expr); + arg_expr = isl_ast_expr_get_op_arg (expr, 2); + tree tree_third_expr + = gcc_expression_from_isl_expression (type, arg_expr, ip); + isl_ast_expr_free (arg_expr); + return fold_build3 (COND_EXPR, type, tree_first_expr, + tree_second_expr, tree_third_expr); +} + +/* Converts a unary isl_ast_expr_op expression E to a GCC expression tree of + type TYPE. */ + +static tree +unary_op_to_tree (tree type, __isl_keep isl_ast_expr *expr, ivs_params_p ip) +{ + gcc_assert (isl_ast_expr_get_op_type (expr) == isl_ast_op_minus); + isl_ast_expr *arg_expr = isl_ast_expr_get_op_arg (expr, 0); + tree tree_expr = gcc_expression_from_isl_expression (type, arg_expr, ip); + isl_ast_expr_free (arg_expr); + return fold_build1 (NEGATE_EXPR, type, tree_expr); +} + +/* Converts a isl_ast_expr_op expression E with unknown number of arguments + to a GCC expression tree of type TYPE. */ + +static tree +nary_op_to_tree (tree type, __isl_keep isl_ast_expr *expr, ivs_params_p ip) +{ + enum tree_code op_code; + switch (isl_ast_expr_get_op_type (expr)) + { + case isl_ast_op_max: + op_code = MAX_EXPR; + break; + + case isl_ast_op_min: + op_code = MIN_EXPR; + break; + + default: + gcc_unreachable (); + } + isl_ast_expr *arg_expr = isl_ast_expr_get_op_arg (expr, 0); + tree res = gcc_expression_from_isl_expression (type, arg_expr, ip); + isl_ast_expr_free (arg_expr); + int i; + for (i = 1; i < isl_ast_expr_get_op_n_arg (expr); i++) + { + arg_expr = isl_ast_expr_get_op_arg (expr, i); + tree t = gcc_expression_from_isl_expression (type, arg_expr, ip); + res = fold_build2 (op_code, type, res, t); + isl_ast_expr_free (arg_expr); + } + return res; +} + + +/* Converts an isl_ast_expr_op expression E to a GCC expression tree of + type TYPE. */ + +static tree +gcc_expression_from_isl_expr_op (tree type, __isl_keep isl_ast_expr *expr, + ivs_params_p ip) +{ + gcc_assert (isl_ast_expr_get_type (expr) == isl_ast_expr_op); + switch (isl_ast_expr_get_op_type (expr)) + { + /* These isl ast expressions are not supported yet. */ + case isl_ast_op_error: + case isl_ast_op_call: + case isl_ast_op_and_then: + case isl_ast_op_or_else: + case isl_ast_op_pdiv_q: + case isl_ast_op_pdiv_r: + case isl_ast_op_select: + gcc_unreachable (); + + case isl_ast_op_max: + case isl_ast_op_min: + return nary_op_to_tree (type, expr, ip); + + case isl_ast_op_add: + case isl_ast_op_sub: + case isl_ast_op_mul: + case isl_ast_op_div: + case isl_ast_op_fdiv_q: + case isl_ast_op_and: + case isl_ast_op_or: + case isl_ast_op_eq: + case isl_ast_op_le: + case isl_ast_op_lt: + case isl_ast_op_ge: + case isl_ast_op_gt: + return binary_op_to_tree (type, expr, ip); + + case isl_ast_op_minus: + return unary_op_to_tree (type, expr, ip); + + case isl_ast_op_cond: + return ternary_op_to_tree (type, expr, ip); + + default: + gcc_unreachable (); + } + + return NULL_TREE; +} + +/* Converts a ISL AST expression E back to a GCC expression tree of + type TYPE. */ + +static tree +gcc_expression_from_isl_expression (tree type, __isl_keep isl_ast_expr *expr, + ivs_params_p ip) +{ + switch (isl_ast_expr_get_type (expr)) + { + case isl_ast_expr_id: + return gcc_expression_from_isl_ast_expr_id (expr, ip); + + case isl_ast_expr_int: + return gcc_expression_from_isl_expr_int (type, expr); + + case isl_ast_expr_op: + return gcc_expression_from_isl_expr_op (type, expr, ip); + + default: + gcc_unreachable (); + } + + return NULL_TREE; +} + +/* Creates a new LOOP corresponding to isl_ast_node_for. Inserts an + induction variable for the new LOOP. New LOOP is attached to CFG + starting at ENTRY_EDGE. LOOP is inserted into the loop tree and + becomes the child loop of the OUTER_LOOP. NEWIVS_INDEX binds + ISL's scattering name to the induction variable created for the + loop of STMT. The new induction variable is inserted in the NEWIVS + vector and is of type TYPE. */ + +static struct loop * +graphite_create_new_loop (edge entry_edge, __isl_keep isl_ast_node *node_for, + loop_p outer, tree type, tree lb, tree ub, + ivs_params_p ip) +{ + isl_ast_expr *for_inc = isl_ast_node_for_get_inc (node_for); + tree stride = gcc_expression_from_isl_expression (type, for_inc, ip); + isl_ast_expr_free (for_inc); + tree ivvar = create_tmp_var (type, "graphite_IV"); + tree iv, iv_after_increment; + loop_p loop = create_empty_loop_on_edge + (entry_edge, lb, stride, ub, ivvar, &iv, &iv_after_increment, + outer ? outer : entry_edge->src->loop_father); + + isl_ast_expr *for_iterator = isl_ast_node_for_get_iterator (node_for); + isl_id *id = isl_ast_expr_get_id (for_iterator); + save_isl_id_name_index (ip->ivs_params_index, isl_id_get_name (id), + (ip->ivs_params_vec).length ()); + isl_id_free (id); + isl_ast_expr_free (for_iterator); + (ip->ivs_params_vec).safe_push (iv); + return loop; +} + +static edge +translate_isl_ast (loop_p context_loop, __isl_keep isl_ast_node *node, + edge next_e, ivs_params_p ip); + +/* Create the loop for a isl_ast_node_for. + + - NEXT_E is the edge where new generated code should be attached. */ + +static edge +translate_isl_ast_for_loop (loop_p context_loop, + __isl_keep isl_ast_node *node_for, edge next_e, + tree type, tree lb, tree ub, + ivs_params_p ip) +{ + gcc_assert (isl_ast_node_get_type (node_for) == isl_ast_node_for); + struct loop *loop = graphite_create_new_loop (next_e, node_for, context_loop, + type, lb, ub, ip); + edge last_e = single_exit (loop); + edge to_body = single_succ_edge (loop->header); + basic_block after = to_body->dest; + + /* Create a basic block for loop close phi nodes. */ + last_e = single_succ_edge (split_edge (last_e)); + + /* Translate the body of the loop. */ + isl_ast_node *for_body = isl_ast_node_for_get_body (node_for); + next_e = translate_isl_ast (loop, for_body, to_body, ip); + isl_ast_node_free (for_body); + redirect_edge_succ_nodup (next_e, after); + set_immediate_dominator (CDI_DOMINATORS, next_e->dest, next_e->src); + + /* TODO: Add checking for the loop parallelism. */ + + return last_e; +} + +/* Get upper bound from the cond of the for */ + +static __isl_give isl_ast_expr * +get_upper_bound (__isl_keep isl_ast_node *node_for) +{ + gcc_assert (isl_ast_node_get_type (node_for) == isl_ast_node_for); + isl_ast_expr *for_cond = isl_ast_node_for_get_cond (node_for); + gcc_assert (isl_ast_expr_get_type (for_cond) == isl_ast_expr_op); + isl_ast_expr *res; + switch (isl_ast_expr_get_op_type (for_cond)) + { + case isl_ast_op_le: + res = isl_ast_expr_get_op_arg (for_cond, 1); + break; + + case isl_ast_op_lt: + { + // (iteraotr < ub) => (iterator <= ub - 1) + isl_val *one = isl_val_int_from_si (isl_ast_expr_get_ctx (for_cond), 1); + isl_ast_expr *ub = isl_ast_expr_get_op_arg (for_cond, 1); + res = isl_ast_expr_sub (ub, isl_ast_expr_from_val (one)); + break; + } + + default: + gcc_unreachable (); + } + isl_ast_expr_free (for_cond); + return res; +} + +/* Creates a new if region protecting the loop to be executed, if the execution + count is zero (lb > ub). */ + +static edge +graphite_create_new_loop_guard (edge entry_edge, + __isl_keep isl_ast_node *node_for, tree *type, + tree *lb, tree *ub, ivs_params_p ip) +{ + gcc_assert (isl_ast_node_get_type (node_for) == isl_ast_node_for); + tree cond_expr; + edge exit_edge; + + *type = *graphite_temporary_tree_type; + isl_ast_expr *for_init = isl_ast_node_for_get_init (node_for); + *lb = gcc_expression_from_isl_expression (*type, for_init, ip); + isl_ast_expr_free (for_init); + isl_ast_expr *upper_bound = get_upper_bound (node_for); + *ub = gcc_expression_from_isl_expression (*type, upper_bound, ip); + isl_ast_expr_free (upper_bound); + + /* When ub is simply a constant or a parameter, use lb <= ub. */ + if (TREE_CODE (*ub) == INTEGER_CST || TREE_CODE (*ub) == SSA_NAME) + cond_expr = fold_build2 (LE_EXPR, boolean_type_node, *lb, *ub); + else + { + tree one = (POINTER_TYPE_P (*type) + ? convert_to_ptrofftype (integer_one_node) + : fold_convert (*type, integer_one_node)); + /* Adding +1 and using LT_EXPR helps with loop latches that have a + loop iteration count of "PARAMETER - 1". For PARAMETER == 0 this + becomes 2^k-1 due to integer overflow, and the condition lb <= ub + is true, even if we do not want this. However lb < ub + 1 is false, + as expected. */ + tree ub_one = fold_build2 (POINTER_TYPE_P (*type) ? POINTER_PLUS_EXPR + : PLUS_EXPR, *type, *ub, one); + + cond_expr = fold_build2 (LT_EXPR, boolean_type_node, *lb, ub_one); + } + + exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr); + + return exit_edge; +} + +/* Translates an isl_ast_node_for to Gimple. */ + +static edge +translate_isl_ast_node_for (loop_p context_loop, __isl_keep isl_ast_node *node, + edge next_e, ivs_params_p ip) +{ + gcc_assert (isl_ast_node_get_type (node) == isl_ast_node_for); + tree type, lb, ub; + edge last_e = graphite_create_new_loop_guard (next_e, node, &type, + &lb, &ub, ip); + edge true_e = get_true_edge_from_guard_bb (next_e->dest); + + translate_isl_ast_for_loop (context_loop, node, true_e, + type, lb, ub, ip); + return last_e; +} + +/* Translates an ISL AST node NODE to GCC representation in the + context of a SESE. */ + +static edge +translate_isl_ast (loop_p context_loop, __isl_keep isl_ast_node *node, + edge next_e, ivs_params_p ip) +{ + switch (isl_ast_node_get_type (node)) + { + case isl_ast_node_error: + gcc_unreachable (); + + case isl_ast_node_for: + return translate_isl_ast_node_for (context_loop, node, + next_e, ip); + + case isl_ast_node_if: + return next_e; + + case isl_ast_node_user: + return next_e; + + case isl_ast_node_block: + return next_e; + + default: + gcc_unreachable (); + } +} + /* Prints NODE to FILE. */ void @@ -65,9 +659,33 @@ isl_printer_free (prn); } +/* Add tree representations and names of parameters to ivs_params */ + +static void +add_parameters_to_ivs_params (scop_p scop, ivs_params_p ip) +{ + sese region = SCOP_REGION (scop); + int i; + int nb_parameters = SESE_PARAMS (region).length (); + gcc_assert (nb_parameters == (short) isl_set_dim (scop->context, + isl_dim_param)); + + for (i = 0; i < nb_parameters; i++) + { + tree param = SESE_PARAMS (region)[i]; + const char *name = get_name (param); + + if (!name) + name = "T"; + save_isl_id_name_index (ip->ivs_params_index, name, i); + (ip->ivs_params_vec).safe_push (param); + } +} + + /* Generates a build, which specifies the constraints on the parameters. */ -static isl_ast_build * +static __isl_give isl_ast_build * generate_isl_context (scop_p scop) { isl_set *context_isl = isl_set_params (isl_set_copy (scop->context)); @@ -77,7 +695,7 @@ /* Generates a schedule, which specifies an order used to visit elements in a domain. */ -static isl_union_map * +static __isl_give isl_union_map * generate_isl_schedule (scop_p scop) { int i; @@ -102,9 +720,16 @@ return schedule_isl; } -static isl_ast_node * -scop_to_isl_ast (scop_p scop) +static __isl_give isl_ast_node * +scop_to_isl_ast (scop_p scop, ivs_params_p ip) { + /* Generate loop upper bounds that consist of the current loop iterator, + an operator (< or <=) and an expression not involving the iterator. + If this option is not set, then the current loop iterator may appear several + times in the upper bound. See the isl manual for more details. */ + isl_options_set_ast_build_atomic_upper_bound (scop->ctx, true); + + add_parameters_to_ivs_params (scop, ip); isl_union_map *schedule_isl = generate_isl_schedule (scop); isl_ast_build *context_isl = generate_isl_context (scop); isl_ast_node *ast_isl = isl_ast_build_ast_from_schedule (context_isl, @@ -117,21 +742,68 @@ the given SCOP. Return true if code generation succeeded. FIXME: This is not yet a full implementation of the code generator - with ISL ASTs. Generation of GIMPLE code is have to be added. */ + with ISL ASTs. Generation of GIMPLE code has to be completed. */ bool graphite_regenerate_ast_isl (scop_p scop) { + loop_p context_loop; + sese region = SCOP_REGION (scop); + ifsese if_region = NULL; + ast_isl_index_htab_type *ivs_params_index; + isl_ast_node *root_node; + struct ivs_params ip; + timevar_push (TV_GRAPHITE_CODE_GEN); graphite_regenerate_error = false; - isl_ast_node *root_node = scop_to_isl_ast (scop); + + auto_vec<tree, 10> ivs_params_vec; + ivs_params_index = new ast_isl_index_htab_type (10); + ip.ivs_params_vec = ivs_params_vec; + ip.ivs_params_index = ivs_params_index; + ip.region = region; + + root_node = scop_to_isl_ast (scop, &ip); + if (dump_file && (dump_flags & TDF_DETAILS)) { fprintf (dump_file, "\nISL AST generated by ISL: \n"); print_isl_ast_node (dump_file, root_node, scop->ctx); + fprintf (dump_file, "\n"); } + + recompute_all_dominators (); + graphite_verify (); + + if_region = move_sese_in_condition (region); + sese_insert_phis_for_liveouts (region, + if_region->region->exit->src, + if_region->false_region->exit, + if_region->true_region->exit); + recompute_all_dominators (); + graphite_verify (); + + context_loop = SESE_ENTRY (region)->src->loop_father; + + translate_isl_ast (context_loop, root_node, if_region->true_region->entry, + &ip); + graphite_verify (); + scev_reset (); + recompute_all_dominators (); + graphite_verify (); + + if (graphite_regenerate_error) + set_ifsese_condition (if_region, integer_zero_node); + + free (if_region->true_region); + free (if_region->region); + free (if_region); + + delete ivs_params_index; + ivs_params_index = NULL; isl_ast_node_free (root_node); timevar_pop (TV_GRAPHITE_CODE_GEN); + /* TODO: Add dump */ return !graphite_regenerate_error; } #endif