/* Global, SSA-based optimizations using mathematical identities.
- Copyright (C) 2005 Free Software Foundation, Inc.
+ Copyright (C) 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
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 2, or (at your option) any
+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
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
-02110-1301, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
/* Currently, the only mini-pass in this file tries to CSE reciprocal
operations. These are common in sequences such as this one:
inserted in BB. */
tree recip_def;
- /* If non-NULL, the MODIFY_EXPR for a reciprocal computation that
+ /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that
was inserted in BB. */
- tree recip_def_stmt;
+ gimple recip_def_stmt;
/* Pointer to a list of "struct occurrence"s for blocks dominated
by BB. */
{
struct occurrence *occ;
- occ = bb->aux = pool_alloc (occ_pool);
+ bb->aux = occ = (struct occurrence *) pool_alloc (occ_pool);
memset (occ, 0, sizeof (struct occurrence));
occ->bb = bb;
/* Return whether USE_STMT is a floating-point division by DEF. */
static inline bool
-is_division_by (tree use_stmt, tree def)
+is_division_by (gimple use_stmt, tree def)
{
- return TREE_CODE (use_stmt) == MODIFY_EXPR
- && TREE_CODE (TREE_OPERAND (use_stmt, 1)) == RDIV_EXPR
- && TREE_OPERAND (TREE_OPERAND (use_stmt, 1), 1) == def;
-}
-
-/* Return the LHS of a RDIV_EXPR that computes a reciprocal in type TYPE. */
-static tree
-get_constant_one (tree type)
-{
- tree scalar, cst;
- int i;
-
- gcc_assert (FLOAT_TYPE_P (type));
- switch (TREE_CODE (type))
- {
- case REAL_TYPE:
- return build_real (type, dconst1);
-
- case VECTOR_TYPE:
- scalar = build_real (TREE_TYPE (type), dconst1);
-
- /* Create 'vect_cst_ = {cst,cst,...,cst}' */
- cst = NULL_TREE;
- for (i = TYPE_VECTOR_SUBPARTS (type); --i >= 0; )
- cst = tree_cons (NULL_TREE, scalar, cst);
-
- return build_vector (type, cst);
-
- default:
- /* Complex operations have been split already. */
- gcc_unreachable ();
- }
+ return is_gimple_assign (use_stmt)
+ && gimple_assign_rhs_code (use_stmt) == RDIV_EXPR
+ && gimple_assign_rhs2 (use_stmt) == def
+ /* Do not recognize x / x as valid division, as we are getting
+ confused later by replacing all immediate uses x in such
+ a stmt. */
+ && gimple_assign_rhs1 (use_stmt) != def;
}
/* Walk the subset of the dominator tree rooted at OCC, setting the
be used. */
static void
-insert_reciprocals (block_stmt_iterator *def_bsi, struct occurrence *occ,
+insert_reciprocals (gimple_stmt_iterator *def_gsi, struct occurrence *occ,
tree def, tree recip_def, int threshold)
{
- tree type, new_stmt;
- block_stmt_iterator bsi;
+ tree type;
+ gimple new_stmt;
+ gimple_stmt_iterator gsi;
struct occurrence *occ_child;
if (!recip_def
/* Make a variable with the replacement and substitute it. */
type = TREE_TYPE (def);
recip_def = make_rename_temp (type, "reciptmp");
- new_stmt = build2 (MODIFY_EXPR, void_type_node, recip_def,
- fold_build2 (RDIV_EXPR, type, get_constant_one (type),
- def));
-
+ new_stmt = gimple_build_assign_with_ops (RDIV_EXPR, recip_def,
+ build_one_cst (type), def);
if (occ->bb_has_division)
{
/* Case 1: insert before an existing division. */
- bsi = bsi_after_labels (occ->bb);
- while (!bsi_end_p (bsi) && !is_division_by (bsi_stmt (bsi), def))
- bsi_next (&bsi);
+ gsi = gsi_after_labels (occ->bb);
+ while (!gsi_end_p (gsi) && !is_division_by (gsi_stmt (gsi), def))
+ gsi_next (&gsi);
- bsi_insert_before (&bsi, new_stmt, BSI_SAME_STMT);
+ gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
}
- else if (def_bsi && occ->bb == def_bsi->bb)
+ else if (def_gsi && occ->bb == def_gsi->bb)
{
/* Case 2: insert right after the definition. Note that this will
never happen if the definition statement can throw, because in
that case the sole successor of the statement's basic block will
dominate all the uses as well. */
- bsi_insert_after (def_bsi, new_stmt, BSI_NEW_STMT);
+ gsi_insert_after (def_gsi, new_stmt, GSI_NEW_STMT);
}
else
{
/* Case 3: insert in a basic block not containing defs/uses. */
- bsi = bsi_after_labels (occ->bb);
- bsi_insert_before (&bsi, new_stmt, BSI_SAME_STMT);
+ gsi = gsi_after_labels (occ->bb);
+ gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
}
occ->recip_def_stmt = new_stmt;
occ->recip_def = recip_def;
for (occ_child = occ->children; occ_child; occ_child = occ_child->next)
- insert_reciprocals (def_bsi, occ_child, def, recip_def, threshold);
+ insert_reciprocals (def_gsi, occ_child, def, recip_def, threshold);
}
static inline void
replace_reciprocal (use_operand_p use_p)
{
- tree use_stmt = USE_STMT (use_p);
- basic_block bb = bb_for_stmt (use_stmt);
+ gimple use_stmt = USE_STMT (use_p);
+ basic_block bb = gimple_bb (use_stmt);
struct occurrence *occ = (struct occurrence *) bb->aux;
- if (occ->recip_def && use_stmt != occ->recip_def_stmt)
+ if (optimize_bb_for_speed_p (bb)
+ && occ->recip_def && use_stmt != occ->recip_def_stmt)
{
- TREE_SET_CODE (TREE_OPERAND (use_stmt, 1), MULT_EXPR);
+ gimple_assign_set_rhs_code (use_stmt, MULT_EXPR);
SET_USE (use_p, occ->recip_def);
fold_stmt_inplace (use_stmt);
update_stmt (use_stmt);
DEF must be a GIMPLE register of a floating-point type. */
static void
-execute_cse_reciprocals_1 (block_stmt_iterator *def_bsi, tree def)
+execute_cse_reciprocals_1 (gimple_stmt_iterator *def_gsi, tree def)
{
use_operand_p use_p;
imm_use_iterator use_iter;
FOR_EACH_IMM_USE_FAST (use_p, use_iter, def)
{
- tree use_stmt = USE_STMT (use_p);
+ gimple use_stmt = USE_STMT (use_p);
if (is_division_by (use_stmt, def))
{
- register_division_in (bb_for_stmt (use_stmt));
+ register_division_in (gimple_bb (use_stmt));
count++;
}
}
threshold = targetm.min_divisions_for_recip_mul (TYPE_MODE (TREE_TYPE (def)));
if (count >= threshold)
{
+ gimple use_stmt;
for (occ = occ_head; occ; occ = occ->next)
{
compute_merit (occ);
- insert_reciprocals (def_bsi, occ, def, NULL, threshold);
+ insert_reciprocals (def_gsi, occ, def, NULL, threshold);
}
- FOR_EACH_IMM_USE_SAFE (use_p, use_iter, def)
+ FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, def)
{
- tree use_stmt = USE_STMT (use_p);
if (is_division_by (use_stmt, def))
- replace_reciprocal (use_p);
+ {
+ FOR_EACH_IMM_USE_ON_STMT (use_p, use_iter)
+ replace_reciprocal (use_p);
+ }
}
}
occ_head = NULL;
}
-
static bool
gate_cse_reciprocals (void)
{
- return optimize && !optimize_size && flag_unsafe_math_optimizations;
+ return optimize && flag_reciprocal_math;
}
-
/* Go through all the floating-point SSA_NAMEs, and call
execute_cse_reciprocals_1 on each of them. */
static unsigned int
sizeof (struct occurrence),
n_basic_blocks / 3 + 1);
- calculate_dominance_info (CDI_DOMINATORS | CDI_POST_DOMINATORS);
+ calculate_dominance_info (CDI_DOMINATORS);
+ calculate_dominance_info (CDI_POST_DOMINATORS);
#ifdef ENABLE_CHECKING
FOR_EACH_BB (bb)
#endif
for (arg = DECL_ARGUMENTS (cfun->decl); arg; arg = TREE_CHAIN (arg))
- if (default_def (arg)
+ if (gimple_default_def (cfun, arg)
&& FLOAT_TYPE_P (TREE_TYPE (arg))
&& is_gimple_reg (arg))
- execute_cse_reciprocals_1 (NULL, default_def (arg));
+ execute_cse_reciprocals_1 (NULL, gimple_default_def (cfun, arg));
FOR_EACH_BB (bb)
{
- block_stmt_iterator bsi;
- tree phi, def;
+ gimple_stmt_iterator gsi;
+ gimple phi;
+ tree def;
- for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
+ for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
+ phi = gsi_stmt (gsi);
def = PHI_RESULT (phi);
if (FLOAT_TYPE_P (TREE_TYPE (def))
&& is_gimple_reg (def))
execute_cse_reciprocals_1 (NULL, def);
}
- for (bsi = bsi_after_labels (bb); !bsi_end_p (bsi); bsi_next (&bsi))
+ for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
- tree stmt = bsi_stmt (bsi);
- if (TREE_CODE (stmt) == MODIFY_EXPR
+ gimple stmt = gsi_stmt (gsi);
+
+ if (gimple_has_lhs (stmt)
&& (def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF)) != NULL
&& FLOAT_TYPE_P (TREE_TYPE (def))
&& TREE_CODE (def) == SSA_NAME)
- execute_cse_reciprocals_1 (&bsi, def);
+ execute_cse_reciprocals_1 (&gsi, def);
+ }
+
+ if (optimize_bb_for_size_p (bb))
+ continue;
+
+ /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */
+ for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
+ {
+ gimple stmt = gsi_stmt (gsi);
+ tree fndecl;
+
+ if (is_gimple_assign (stmt)
+ && gimple_assign_rhs_code (stmt) == RDIV_EXPR)
+ {
+ tree arg1 = gimple_assign_rhs2 (stmt);
+ gimple stmt1;
+
+ if (TREE_CODE (arg1) != SSA_NAME)
+ continue;
+
+ stmt1 = SSA_NAME_DEF_STMT (arg1);
+
+ if (is_gimple_call (stmt1)
+ && gimple_call_lhs (stmt1)
+ && (fndecl = gimple_call_fndecl (stmt1))
+ && (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
+ || DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD))
+ {
+ enum built_in_function code;
+ bool md_code;
+
+ code = DECL_FUNCTION_CODE (fndecl);
+ md_code = DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD;
+
+ fndecl = targetm.builtin_reciprocal (code, md_code, false);
+ if (!fndecl)
+ continue;
+
+ gimple_call_set_fndecl (stmt1, fndecl);
+ update_stmt (stmt1);
+
+ gimple_assign_set_rhs_code (stmt, MULT_EXPR);
+ fold_stmt_inplace (stmt);
+ update_stmt (stmt);
+ }
+ }
}
}
- free_dominance_info (CDI_DOMINATORS | CDI_POST_DOMINATORS);
+ free_dominance_info (CDI_DOMINATORS);
+ free_dominance_info (CDI_POST_DOMINATORS);
free_alloc_pool (occ_pool);
return 0;
}
-struct tree_opt_pass pass_cse_reciprocals =
+struct gimple_opt_pass pass_cse_reciprocals =
{
+ {
+ GIMPLE_PASS,
"recip", /* name */
gate_cse_reciprocals, /* gate */
execute_cse_reciprocals, /* execute */
0, /* properties_destroyed */
0, /* todo_flags_start */
TODO_dump_func | TODO_update_ssa | TODO_verify_ssa
- | TODO_verify_stmts, /* todo_flags_finish */
- 0 /* letter */
+ | TODO_verify_stmts /* todo_flags_finish */
+ }
+};
+
+/* Records an occurrence at statement USE_STMT in the vector of trees
+ STMTS if it is dominated by *TOP_BB or dominates it or this basic block
+ is not yet initialized. Returns true if the occurrence was pushed on
+ the vector. Adjusts *TOP_BB to be the basic block dominating all
+ statements in the vector. */
+
+static bool
+maybe_record_sincos (VEC(gimple, heap) **stmts,
+ basic_block *top_bb, gimple use_stmt)
+{
+ basic_block use_bb = gimple_bb (use_stmt);
+ if (*top_bb
+ && (*top_bb == use_bb
+ || dominated_by_p (CDI_DOMINATORS, use_bb, *top_bb)))
+ VEC_safe_push (gimple, heap, *stmts, use_stmt);
+ else if (!*top_bb
+ || dominated_by_p (CDI_DOMINATORS, *top_bb, use_bb))
+ {
+ VEC_safe_push (gimple, heap, *stmts, use_stmt);
+ *top_bb = use_bb;
+ }
+ else
+ return false;
+
+ return true;
+}
+
+/* Look for sin, cos and cexpi calls with the same argument NAME and
+ create a single call to cexpi CSEing the result in this case.
+ We first walk over all immediate uses of the argument collecting
+ statements that we can CSE in a vector and in a second pass replace
+ the statement rhs with a REALPART or IMAGPART expression on the
+ result of the cexpi call we insert before the use statement that
+ dominates all other candidates. */
+
+static void
+execute_cse_sincos_1 (tree name)
+{
+ gimple_stmt_iterator gsi;
+ imm_use_iterator use_iter;
+ tree fndecl, res, type;
+ gimple def_stmt, use_stmt, stmt;
+ int seen_cos = 0, seen_sin = 0, seen_cexpi = 0;
+ VEC(gimple, heap) *stmts = NULL;
+ basic_block top_bb = NULL;
+ int i;
+
+ type = TREE_TYPE (name);
+ FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, name)
+ {
+ if (gimple_code (use_stmt) != GIMPLE_CALL
+ || !gimple_call_lhs (use_stmt)
+ || !(fndecl = gimple_call_fndecl (use_stmt))
+ || DECL_BUILT_IN_CLASS (fndecl) != BUILT_IN_NORMAL)
+ continue;
+
+ switch (DECL_FUNCTION_CODE (fndecl))
+ {
+ CASE_FLT_FN (BUILT_IN_COS):
+ seen_cos |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0;
+ break;
+
+ CASE_FLT_FN (BUILT_IN_SIN):
+ seen_sin |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0;
+ break;
+
+ CASE_FLT_FN (BUILT_IN_CEXPI):
+ seen_cexpi |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0;
+ break;
+
+ default:;
+ }
+ }
+
+ if (seen_cos + seen_sin + seen_cexpi <= 1)
+ {
+ VEC_free(gimple, heap, stmts);
+ return;
+ }
+
+ /* Simply insert cexpi at the beginning of top_bb but not earlier than
+ the name def statement. */
+ fndecl = mathfn_built_in (type, BUILT_IN_CEXPI);
+ if (!fndecl)
+ return;
+ res = make_rename_temp (TREE_TYPE (TREE_TYPE (fndecl)), "sincostmp");
+ stmt = gimple_build_call (fndecl, 1, name);
+ gimple_call_set_lhs (stmt, res);
+
+ def_stmt = SSA_NAME_DEF_STMT (name);
+ if (!SSA_NAME_IS_DEFAULT_DEF (name)
+ && gimple_code (def_stmt) != GIMPLE_PHI
+ && gimple_bb (def_stmt) == top_bb)
+ {
+ gsi = gsi_for_stmt (def_stmt);
+ gsi_insert_after (&gsi, stmt, GSI_SAME_STMT);
+ }
+ else
+ {
+ gsi = gsi_after_labels (top_bb);
+ gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
+ }
+ update_stmt (stmt);
+
+ /* And adjust the recorded old call sites. */
+ for (i = 0; VEC_iterate(gimple, stmts, i, use_stmt); ++i)
+ {
+ tree rhs = NULL;
+ fndecl = gimple_call_fndecl (use_stmt);
+
+ switch (DECL_FUNCTION_CODE (fndecl))
+ {
+ CASE_FLT_FN (BUILT_IN_COS):
+ rhs = fold_build1 (REALPART_EXPR, type, res);
+ break;
+
+ CASE_FLT_FN (BUILT_IN_SIN):
+ rhs = fold_build1 (IMAGPART_EXPR, type, res);
+ break;
+
+ CASE_FLT_FN (BUILT_IN_CEXPI):
+ rhs = res;
+ break;
+
+ default:;
+ gcc_unreachable ();
+ }
+
+ /* Replace call with a copy. */
+ stmt = gimple_build_assign (gimple_call_lhs (use_stmt), rhs);
+
+ gsi = gsi_for_stmt (use_stmt);
+ gsi_insert_after (&gsi, stmt, GSI_SAME_STMT);
+ gsi_remove (&gsi, true);
+ }
+
+ VEC_free(gimple, heap, stmts);
+}
+
+/* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1
+ on the SSA_NAME argument of each of them. */
+
+static unsigned int
+execute_cse_sincos (void)
+{
+ basic_block bb;
+
+ calculate_dominance_info (CDI_DOMINATORS);
+
+ FOR_EACH_BB (bb)
+ {
+ gimple_stmt_iterator gsi;
+
+ for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
+ {
+ gimple stmt = gsi_stmt (gsi);
+ tree fndecl;
+
+ if (is_gimple_call (stmt)
+ && gimple_call_lhs (stmt)
+ && (fndecl = gimple_call_fndecl (stmt))
+ && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
+ {
+ tree arg;
+
+ switch (DECL_FUNCTION_CODE (fndecl))
+ {
+ CASE_FLT_FN (BUILT_IN_COS):
+ CASE_FLT_FN (BUILT_IN_SIN):
+ CASE_FLT_FN (BUILT_IN_CEXPI):
+ arg = gimple_call_arg (stmt, 0);
+ if (TREE_CODE (arg) == SSA_NAME)
+ execute_cse_sincos_1 (arg);
+ break;
+
+ default:;
+ }
+ }
+ }
+ }
+
+ free_dominance_info (CDI_DOMINATORS);
+ return 0;
+}
+
+static bool
+gate_cse_sincos (void)
+{
+ /* Make sure we have either sincos or cexp. */
+ return (TARGET_HAS_SINCOS
+ || TARGET_C99_FUNCTIONS)
+ && optimize;
+}
+
+struct gimple_opt_pass pass_cse_sincos =
+{
+ {
+ GIMPLE_PASS,
+ "sincos", /* name */
+ gate_cse_sincos, /* gate */
+ execute_cse_sincos, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ 0, /* tv_id */
+ PROP_ssa, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_dump_func | TODO_update_ssa | TODO_verify_ssa
+ | TODO_verify_stmts /* todo_flags_finish */
+ }
+};
+
+/* Find all expressions in the form of sqrt(a/b) and
+ convert them to rsqrt(b/a). */
+
+static unsigned int
+execute_convert_to_rsqrt (void)
+{
+ basic_block bb;
+
+ FOR_EACH_BB (bb)
+ {
+ gimple_stmt_iterator gsi;
+
+ for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
+ {
+ gimple stmt = gsi_stmt (gsi);
+ tree fndecl;
+
+ if (is_gimple_call (stmt)
+ && gimple_call_lhs (stmt)
+ && (fndecl = gimple_call_fndecl (stmt))
+ && (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
+ || DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD))
+ {
+ enum built_in_function code;
+ bool md_code;
+ tree arg1;
+ gimple stmt1;
+
+ code = DECL_FUNCTION_CODE (fndecl);
+ md_code = DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD;
+
+ fndecl = targetm.builtin_reciprocal (code, md_code, true);
+ if (!fndecl)
+ continue;
+
+ arg1 = gimple_call_arg (stmt, 0);
+
+ if (TREE_CODE (arg1) != SSA_NAME)
+ continue;
+
+ stmt1 = SSA_NAME_DEF_STMT (arg1);
+
+ if (is_gimple_assign (stmt1)
+ && gimple_assign_rhs_code (stmt1) == RDIV_EXPR)
+ {
+ tree arg10, arg11;
+
+ arg10 = gimple_assign_rhs1 (stmt1);
+ arg11 = gimple_assign_rhs2 (stmt1);
+
+ /* Swap operands of RDIV_EXPR. */
+ gimple_assign_set_rhs1 (stmt1, arg11);
+ gimple_assign_set_rhs2 (stmt1, arg10);
+ fold_stmt_inplace (stmt1);
+ update_stmt (stmt1);
+
+ gimple_call_set_fndecl (stmt, fndecl);
+ update_stmt (stmt);
+ }
+ }
+ }
+ }
+
+ return 0;
+}
+
+static bool
+gate_convert_to_rsqrt (void)
+{
+ return flag_unsafe_math_optimizations && optimize;
+}
+
+struct gimple_opt_pass pass_convert_to_rsqrt =
+{
+ {
+ GIMPLE_PASS,
+ "rsqrt", /* name */
+ gate_convert_to_rsqrt, /* gate */
+ execute_convert_to_rsqrt, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ 0, /* tv_id */
+ PROP_ssa, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_dump_func | TODO_update_ssa | TODO_verify_ssa
+ | TODO_verify_stmts /* todo_flags_finish */
+ }
};
OSDN Git Service
