/* Reassociation for trees.
- Copyright (C) 2005, 2007 Free Software Foundation, Inc.
+ Copyright (C) 2005, 2007, 2008, 2009, 2010 Free Software Foundation, Inc.
Contributed by Daniel Berlin <dan@dberlin.org>
This file is part of GCC.
#include "system.h"
#include "coretypes.h"
#include "tm.h"
-#include "errors.h"
-#include "ggc.h"
#include "tree.h"
#include "basic-block.h"
-#include "diagnostic.h"
+#include "tree-pretty-print.h"
+#include "gimple-pretty-print.h"
#include "tree-inline.h"
#include "tree-flow.h"
#include "gimple.h"
mergetmp2 = d + e
and put mergetmp2 on the merge worklist.
-
+
so merge worklist = {mergetmp, c, mergetmp2}
-
+
Continue building binary ops of these operations until you have only
one operation left on the worklist.
-
+
So we have
-
+
build binary op
mergetmp3 = mergetmp + c
-
+
worklist = {mergetmp2, mergetmp3}
-
+
mergetmp4 = mergetmp2 + mergetmp3
-
+
worklist = {mergetmp4}
-
+
because we have one operation left, we can now just set the original
statement equal to the result of that operation.
-
+
This will at least expose a + b and d + e to redundancy elimination
as binary operations.
-
+
For extra points, you can reuse the old statements to build the
mergetmps, since you shouldn't run out.
So why don't we do this?
-
+
Because it's expensive, and rarely will help. Most trees we are
reassociating have 3 or less ops. If they have 2 ops, they already
will be written into a nice single binary op. If you have 3 ops, a
mergetmp = op1 + op2
newstmt = mergetmp + op3
-
+
instead of
mergetmp = op2 + op3
newstmt = mergetmp + op1
-
+
If all three are of the same rank, you can't expose them all in a
single binary operator anyway, so the above is *still* the best you
can do.
-
+
Thus, this is what we do. When we have three ops left, we check to see
what order to put them in, and call it a day. As a nod to vector sum
- reduction, we check if any of ops are a really a phi node that is a
+ reduction, we check if any of the ops are really a phi node that is a
destructive update for the associating op, and keep the destructive
update together for vector sum reduction recognition. */
typedef struct operand_entry
{
unsigned int rank;
+ int id;
tree op;
} *operand_entry_t;
static alloc_pool operand_entry_pool;
+/* This is used to assign a unique ID to each struct operand_entry
+ so that qsort results are identical on different hosts. */
+static int next_operand_entry_id;
/* Starting rank number for a given basic block, so that we can rank
operations using unmovable instructions in that BB based on the bb
find_operand_rank (tree e)
{
void **slot = pointer_map_contains (operand_rank, e);
- return slot ? (long) *slot : -1;
+ return slot ? (long) (intptr_t) *slot : -1;
}
/* Insert {E,RANK} into the operand rank hashtable. */
gcc_assert (rank > 0);
slot = pointer_map_insert (operand_rank, e);
gcc_assert (!*slot);
- *slot = (void *) rank;
+ *slot = (void *) (intptr_t) rank;
}
/* Given an expression E, return the rank of the expression. */
return 0;
if (!is_gimple_assign (stmt)
- || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VIRTUAL_DEFS))
+ || gimple_vdef (stmt))
return bb_rank[gimple_bb (stmt)->index];
/* If we already have a rank for this expression, use that. */
to fold when added/multiplied//whatever are put next to each
other. Since all constants have rank 0, order them by type. */
if (oeb->rank == 0 && oea->rank == 0)
- return constant_type (oeb->op) - constant_type (oea->op);
+ {
+ if (constant_type (oeb->op) != constant_type (oea->op))
+ return constant_type (oeb->op) - constant_type (oea->op);
+ else
+ /* To make sorting result stable, we use unique IDs to determine
+ order. */
+ return oeb->id - oea->id;
+ }
/* Lastly, make sure the versions that are the same go next to each
other. We use SSA_NAME_VERSION because it's stable. */
if ((oeb->rank - oea->rank == 0)
&& TREE_CODE (oea->op) == SSA_NAME
&& TREE_CODE (oeb->op) == SSA_NAME)
- return SSA_NAME_VERSION (oeb->op) - SSA_NAME_VERSION (oea->op);
+ {
+ if (SSA_NAME_VERSION (oeb->op) != SSA_NAME_VERSION (oea->op))
+ return SSA_NAME_VERSION (oeb->op) - SSA_NAME_VERSION (oea->op);
+ else
+ return oeb->id - oea->id;
+ }
- return oeb->rank - oea->rank;
+ if (oeb->rank != oea->rank)
+ return oeb->rank - oea->rank;
+ else
+ return oeb->id - oea->id;
}
/* Add an operand entry to *OPS for the tree operand OP. */
oe->op = op;
oe->rank = get_rank (op);
+ oe->id = next_operand_entry_id++;
VEC_safe_push (operand_entry_t, heap, *ops, oe);
}
{
VEC_free (operand_entry_t, heap, *ops);
*ops = NULL;
- add_to_ops_vec (ops, fold_convert (TREE_TYPE (last->op),
- integer_zero_node));
+ add_to_ops_vec (ops, build_zero_cst (TREE_TYPE (last->op)));
*all_done = true;
}
else
return false;
}
-/* If OPCODE is PLUS_EXPR, CURR->OP is really a negate expression,
- look in OPS for a corresponding positive operation to cancel it
- out. If we find one, remove the other from OPS, replace
- OPS[CURRINDEX] with 0, and return true. Otherwise, return
- false. */
+static VEC(tree, heap) *plus_negates;
+
+/* If OPCODE is PLUS_EXPR, CURR->OP is a negate expression or a bitwise not
+ expression, look in OPS for a corresponding positive operation to cancel
+ it out. If we find one, remove the other from OPS, replace
+ OPS[CURRINDEX] with 0 or -1, respectively, and return true. Otherwise,
+ return false. */
static bool
eliminate_plus_minus_pair (enum tree_code opcode,
operand_entry_t curr)
{
tree negateop;
+ tree notop;
unsigned int i;
operand_entry_t oe;
return false;
negateop = get_unary_op (curr->op, NEGATE_EXPR);
- if (negateop == NULL_TREE)
+ notop = get_unary_op (curr->op, BIT_NOT_EXPR);
+ if (negateop == NULL_TREE && notop == NULL_TREE)
return false;
/* Any non-negated version will have a rank that is one less than
}
VEC_ordered_remove (operand_entry_t, *ops, i);
- add_to_ops_vec (ops, fold_convert(TREE_TYPE (oe->op),
- integer_zero_node));
+ add_to_ops_vec (ops, build_zero_cst (TREE_TYPE (oe->op)));
+ VEC_ordered_remove (operand_entry_t, *ops, currindex);
+ reassociate_stats.ops_eliminated ++;
+
+ return true;
+ }
+ else if (oe->op == notop)
+ {
+ tree op_type = TREE_TYPE (oe->op);
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Equivalence: ");
+ print_generic_expr (dump_file, notop, 0);
+ fprintf (dump_file, " + ~");
+ print_generic_expr (dump_file, oe->op, 0);
+ fprintf (dump_file, " -> -1\n");
+ }
+
+ VEC_ordered_remove (operand_entry_t, *ops, i);
+ add_to_ops_vec (ops, build_int_cst_type (op_type, -1));
VEC_ordered_remove (operand_entry_t, *ops, currindex);
reassociate_stats.ops_eliminated ++;
}
}
+ /* CURR->OP is a negate expr in a plus expr: save it for later
+ inspection in repropagate_negates(). */
+ if (negateop != NULL_TREE)
+ VEC_safe_push (tree, heap, plus_negates, curr->op);
+
return false;
}
}
if (opcode == BIT_AND_EXPR)
- oe->op = fold_convert (TREE_TYPE (oe->op), integer_zero_node);
+ oe->op = build_zero_cst (TREE_TYPE (oe->op));
else if (opcode == BIT_IOR_EXPR)
oe->op = build_low_bits_mask (TREE_TYPE (oe->op),
TYPE_PRECISION (TREE_TYPE (oe->op)));
- reassociate_stats.ops_eliminated
+ reassociate_stats.ops_eliminated
+= VEC_length (operand_entry_t, *ops) - 1;
VEC_free (operand_entry_t, heap, *ops);
*ops = NULL;
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found & 0, removing all other ops\n");
- reassociate_stats.ops_eliminated
+ reassociate_stats.ops_eliminated
+= VEC_length (operand_entry_t, *ops) - 1;
-
+
VEC_free (operand_entry_t, heap, *ops);
*ops = NULL;
VEC_safe_push (operand_entry_t, heap, *ops, oelast);
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found | -1, removing all other ops\n");
- reassociate_stats.ops_eliminated
+ reassociate_stats.ops_eliminated
+= VEC_length (operand_entry_t, *ops) - 1;
-
+
VEC_free (operand_entry_t, heap, *ops);
*ops = NULL;
VEC_safe_push (operand_entry_t, heap, *ops, oelast);
return;
}
- }
+ }
else if (integer_zerop (oelast->op))
{
if (VEC_length (operand_entry_t, *ops) != 1)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found * 0, removing all other ops\n");
-
- reassociate_stats.ops_eliminated
+
+ reassociate_stats.ops_eliminated
+= VEC_length (operand_entry_t, *ops) - 1;
VEC_free (operand_entry_t, heap, *ops);
*ops = NULL;
}
}
+
+static void linearize_expr_tree (VEC(operand_entry_t, heap) **, gimple,
+ bool, bool);
+
+/* Structure for tracking and counting operands. */
+typedef struct oecount_s {
+ int cnt;
+ int id;
+ enum tree_code oecode;
+ tree op;
+} oecount;
+
+DEF_VEC_O(oecount);
+DEF_VEC_ALLOC_O(oecount,heap);
+
+/* The heap for the oecount hashtable and the sorted list of operands. */
+static VEC (oecount, heap) *cvec;
+
+/* Hash function for oecount. */
+
+static hashval_t
+oecount_hash (const void *p)
+{
+ const oecount *c = VEC_index (oecount, cvec, (size_t)p - 42);
+ return htab_hash_pointer (c->op) ^ (hashval_t)c->oecode;
+}
+
+/* Comparison function for oecount. */
+
+static int
+oecount_eq (const void *p1, const void *p2)
+{
+ const oecount *c1 = VEC_index (oecount, cvec, (size_t)p1 - 42);
+ const oecount *c2 = VEC_index (oecount, cvec, (size_t)p2 - 42);
+ return (c1->oecode == c2->oecode
+ && c1->op == c2->op);
+}
+
+/* Comparison function for qsort sorting oecount elements by count. */
+
+static int
+oecount_cmp (const void *p1, const void *p2)
+{
+ const oecount *c1 = (const oecount *)p1;
+ const oecount *c2 = (const oecount *)p2;
+ if (c1->cnt != c2->cnt)
+ return c1->cnt - c2->cnt;
+ else
+ /* If counts are identical, use unique IDs to stabilize qsort. */
+ return c1->id - c2->id;
+}
+
+/* Walks the linear chain with result *DEF searching for an operation
+ with operand OP and code OPCODE removing that from the chain. *DEF
+ is updated if there is only one operand but no operation left. */
+
+static void
+zero_one_operation (tree *def, enum tree_code opcode, tree op)
+{
+ gimple stmt = SSA_NAME_DEF_STMT (*def);
+
+ do
+ {
+ tree name = gimple_assign_rhs1 (stmt);
+
+ /* If this is the operation we look for and one of the operands
+ is ours simply propagate the other operand into the stmts
+ single use. */
+ if (gimple_assign_rhs_code (stmt) == opcode
+ && (name == op
+ || gimple_assign_rhs2 (stmt) == op))
+ {
+ gimple use_stmt;
+ use_operand_p use;
+ gimple_stmt_iterator gsi;
+ if (name == op)
+ name = gimple_assign_rhs2 (stmt);
+ gcc_assert (has_single_use (gimple_assign_lhs (stmt)));
+ single_imm_use (gimple_assign_lhs (stmt), &use, &use_stmt);
+ if (gimple_assign_lhs (stmt) == *def)
+ *def = name;
+ SET_USE (use, name);
+ if (TREE_CODE (name) != SSA_NAME)
+ update_stmt (use_stmt);
+ gsi = gsi_for_stmt (stmt);
+ gsi_remove (&gsi, true);
+ release_defs (stmt);
+ return;
+ }
+
+ /* Continue walking the chain. */
+ gcc_assert (name != op
+ && TREE_CODE (name) == SSA_NAME);
+ stmt = SSA_NAME_DEF_STMT (name);
+ }
+ while (1);
+}
+
+/* Builds one statement performing OP1 OPCODE OP2 using TMPVAR for
+ the result. Places the statement after the definition of either
+ OP1 or OP2. Returns the new statement. */
+
+static gimple
+build_and_add_sum (tree tmpvar, tree op1, tree op2, enum tree_code opcode)
+{
+ gimple op1def = NULL, op2def = NULL;
+ gimple_stmt_iterator gsi;
+ tree op;
+ gimple sum;
+
+ /* Create the addition statement. */
+ sum = gimple_build_assign_with_ops (opcode, tmpvar, op1, op2);
+ op = make_ssa_name (tmpvar, sum);
+ gimple_assign_set_lhs (sum, op);
+
+ /* Find an insertion place and insert. */
+ if (TREE_CODE (op1) == SSA_NAME)
+ op1def = SSA_NAME_DEF_STMT (op1);
+ if (TREE_CODE (op2) == SSA_NAME)
+ op2def = SSA_NAME_DEF_STMT (op2);
+ if ((!op1def || gimple_nop_p (op1def))
+ && (!op2def || gimple_nop_p (op2def)))
+ {
+ gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR));
+ gsi_insert_before (&gsi, sum, GSI_NEW_STMT);
+ }
+ else if ((!op1def || gimple_nop_p (op1def))
+ || (op2def && !gimple_nop_p (op2def)
+ && stmt_dominates_stmt_p (op1def, op2def)))
+ {
+ if (gimple_code (op2def) == GIMPLE_PHI)
+ {
+ gsi = gsi_after_labels (gimple_bb (op2def));
+ gsi_insert_before (&gsi, sum, GSI_NEW_STMT);
+ }
+ else
+ {
+ if (!stmt_ends_bb_p (op2def))
+ {
+ gsi = gsi_for_stmt (op2def);
+ gsi_insert_after (&gsi, sum, GSI_NEW_STMT);
+ }
+ else
+ {
+ edge e;
+ edge_iterator ei;
+
+ FOR_EACH_EDGE (e, ei, gimple_bb (op2def)->succs)
+ if (e->flags & EDGE_FALLTHRU)
+ gsi_insert_on_edge_immediate (e, sum);
+ }
+ }
+ }
+ else
+ {
+ if (gimple_code (op1def) == GIMPLE_PHI)
+ {
+ gsi = gsi_after_labels (gimple_bb (op1def));
+ gsi_insert_before (&gsi, sum, GSI_NEW_STMT);
+ }
+ else
+ {
+ if (!stmt_ends_bb_p (op1def))
+ {
+ gsi = gsi_for_stmt (op1def);
+ gsi_insert_after (&gsi, sum, GSI_NEW_STMT);
+ }
+ else
+ {
+ edge e;
+ edge_iterator ei;
+
+ FOR_EACH_EDGE (e, ei, gimple_bb (op1def)->succs)
+ if (e->flags & EDGE_FALLTHRU)
+ gsi_insert_on_edge_immediate (e, sum);
+ }
+ }
+ }
+ update_stmt (sum);
+
+ return sum;
+}
+
+/* Perform un-distribution of divisions and multiplications.
+ A * X + B * X is transformed into (A + B) * X and A / X + B / X
+ to (A + B) / X for real X.
+
+ The algorithm is organized as follows.
+
+ - First we walk the addition chain *OPS looking for summands that
+ are defined by a multiplication or a real division. This results
+ in the candidates bitmap with relevant indices into *OPS.
+
+ - Second we build the chains of multiplications or divisions for
+ these candidates, counting the number of occurences of (operand, code)
+ pairs in all of the candidates chains.
+
+ - Third we sort the (operand, code) pairs by number of occurence and
+ process them starting with the pair with the most uses.
+
+ * For each such pair we walk the candidates again to build a
+ second candidate bitmap noting all multiplication/division chains
+ that have at least one occurence of (operand, code).
+
+ * We build an alternate addition chain only covering these
+ candidates with one (operand, code) operation removed from their
+ multiplication/division chain.
+
+ * The first candidate gets replaced by the alternate addition chain
+ multiplied/divided by the operand.
+
+ * All candidate chains get disabled for further processing and
+ processing of (operand, code) pairs continues.
+
+ The alternate addition chains built are re-processed by the main
+ reassociation algorithm which allows optimizing a * x * y + b * y * x
+ to (a + b ) * x * y in one invocation of the reassociation pass. */
+
+static bool
+undistribute_ops_list (enum tree_code opcode,
+ VEC (operand_entry_t, heap) **ops, struct loop *loop)
+{
+ unsigned int length = VEC_length (operand_entry_t, *ops);
+ operand_entry_t oe1;
+ unsigned i, j;
+ sbitmap candidates, candidates2;
+ unsigned nr_candidates, nr_candidates2;
+ sbitmap_iterator sbi0;
+ VEC (operand_entry_t, heap) **subops;
+ htab_t ctable;
+ bool changed = false;
+ int next_oecount_id = 0;
+
+ if (length <= 1
+ || opcode != PLUS_EXPR)
+ return false;
+
+ /* Build a list of candidates to process. */
+ candidates = sbitmap_alloc (length);
+ sbitmap_zero (candidates);
+ nr_candidates = 0;
+ FOR_EACH_VEC_ELT (operand_entry_t, *ops, i, oe1)
+ {
+ enum tree_code dcode;
+ gimple oe1def;
+
+ if (TREE_CODE (oe1->op) != SSA_NAME)
+ continue;
+ oe1def = SSA_NAME_DEF_STMT (oe1->op);
+ if (!is_gimple_assign (oe1def))
+ continue;
+ dcode = gimple_assign_rhs_code (oe1def);
+ if ((dcode != MULT_EXPR
+ && dcode != RDIV_EXPR)
+ || !is_reassociable_op (oe1def, dcode, loop))
+ continue;
+
+ SET_BIT (candidates, i);
+ nr_candidates++;
+ }
+
+ if (nr_candidates < 2)
+ {
+ sbitmap_free (candidates);
+ return false;
+ }
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "searching for un-distribute opportunities ");
+ print_generic_expr (dump_file,
+ VEC_index (operand_entry_t, *ops,
+ sbitmap_first_set_bit (candidates))->op, 0);
+ fprintf (dump_file, " %d\n", nr_candidates);
+ }
+
+ /* Build linearized sub-operand lists and the counting table. */
+ cvec = NULL;
+ ctable = htab_create (15, oecount_hash, oecount_eq, NULL);
+ subops = XCNEWVEC (VEC (operand_entry_t, heap) *,
+ VEC_length (operand_entry_t, *ops));
+ EXECUTE_IF_SET_IN_SBITMAP (candidates, 0, i, sbi0)
+ {
+ gimple oedef;
+ enum tree_code oecode;
+ unsigned j;
+
+ oedef = SSA_NAME_DEF_STMT (VEC_index (operand_entry_t, *ops, i)->op);
+ oecode = gimple_assign_rhs_code (oedef);
+ linearize_expr_tree (&subops[i], oedef,
+ associative_tree_code (oecode), false);
+
+ FOR_EACH_VEC_ELT (operand_entry_t, subops[i], j, oe1)
+ {
+ oecount c;
+ void **slot;
+ size_t idx;
+ c.oecode = oecode;
+ c.cnt = 1;
+ c.id = next_oecount_id++;
+ c.op = oe1->op;
+ VEC_safe_push (oecount, heap, cvec, &c);
+ idx = VEC_length (oecount, cvec) + 41;
+ slot = htab_find_slot (ctable, (void *)idx, INSERT);
+ if (!*slot)
+ {
+ *slot = (void *)idx;
+ }
+ else
+ {
+ VEC_pop (oecount, cvec);
+ VEC_index (oecount, cvec, (size_t)*slot - 42)->cnt++;
+ }
+ }
+ }
+ htab_delete (ctable);
+
+ /* Sort the counting table. */
+ VEC_qsort (oecount, cvec, oecount_cmp);
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ oecount *c;
+ fprintf (dump_file, "Candidates:\n");
+ FOR_EACH_VEC_ELT (oecount, cvec, j, c)
+ {
+ fprintf (dump_file, " %u %s: ", c->cnt,
+ c->oecode == MULT_EXPR
+ ? "*" : c->oecode == RDIV_EXPR ? "/" : "?");
+ print_generic_expr (dump_file, c->op, 0);
+ fprintf (dump_file, "\n");
+ }
+ }
+
+ /* Process the (operand, code) pairs in order of most occurence. */
+ candidates2 = sbitmap_alloc (length);
+ while (!VEC_empty (oecount, cvec))
+ {
+ oecount *c = VEC_last (oecount, cvec);
+ if (c->cnt < 2)
+ break;
+
+ /* Now collect the operands in the outer chain that contain
+ the common operand in their inner chain. */
+ sbitmap_zero (candidates2);
+ nr_candidates2 = 0;
+ EXECUTE_IF_SET_IN_SBITMAP (candidates, 0, i, sbi0)
+ {
+ gimple oedef;
+ enum tree_code oecode;
+ unsigned j;
+ tree op = VEC_index (operand_entry_t, *ops, i)->op;
+
+ /* If we undistributed in this chain already this may be
+ a constant. */
+ if (TREE_CODE (op) != SSA_NAME)
+ continue;
+
+ oedef = SSA_NAME_DEF_STMT (op);
+ oecode = gimple_assign_rhs_code (oedef);
+ if (oecode != c->oecode)
+ continue;
+
+ FOR_EACH_VEC_ELT (operand_entry_t, subops[i], j, oe1)
+ {
+ if (oe1->op == c->op)
+ {
+ SET_BIT (candidates2, i);
+ ++nr_candidates2;
+ break;
+ }
+ }
+ }
+
+ if (nr_candidates2 >= 2)
+ {
+ operand_entry_t oe1, oe2;
+ tree tmpvar;
+ gimple prod;
+ int first = sbitmap_first_set_bit (candidates2);
+
+ /* Build the new addition chain. */
+ oe1 = VEC_index (operand_entry_t, *ops, first);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Building (");
+ print_generic_expr (dump_file, oe1->op, 0);
+ }
+ tmpvar = create_tmp_reg (TREE_TYPE (oe1->op), NULL);
+ add_referenced_var (tmpvar);
+ zero_one_operation (&oe1->op, c->oecode, c->op);
+ EXECUTE_IF_SET_IN_SBITMAP (candidates2, first+1, i, sbi0)
+ {
+ gimple sum;
+ oe2 = VEC_index (operand_entry_t, *ops, i);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, " + ");
+ print_generic_expr (dump_file, oe2->op, 0);
+ }
+ zero_one_operation (&oe2->op, c->oecode, c->op);
+ sum = build_and_add_sum (tmpvar, oe1->op, oe2->op, opcode);
+ oe2->op = build_zero_cst (TREE_TYPE (oe2->op));
+ oe2->rank = 0;
+ oe1->op = gimple_get_lhs (sum);
+ }
+
+ /* Apply the multiplication/division. */
+ prod = build_and_add_sum (tmpvar, oe1->op, c->op, c->oecode);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, ") %s ", c->oecode == MULT_EXPR ? "*" : "/");
+ print_generic_expr (dump_file, c->op, 0);
+ fprintf (dump_file, "\n");
+ }
+
+ /* Record it in the addition chain and disable further
+ undistribution with this op. */
+ oe1->op = gimple_assign_lhs (prod);
+ oe1->rank = get_rank (oe1->op);
+ VEC_free (operand_entry_t, heap, subops[first]);
+
+ changed = true;
+ }
+
+ VEC_pop (oecount, cvec);
+ }
+
+ for (i = 0; i < VEC_length (operand_entry_t, *ops); ++i)
+ VEC_free (operand_entry_t, heap, subops[i]);
+ free (subops);
+ VEC_free (oecount, heap, cvec);
+ sbitmap_free (candidates);
+ sbitmap_free (candidates2);
+
+ return changed;
+}
+
+/* If OPCODE is BIT_IOR_EXPR or BIT_AND_EXPR and CURR is a comparison
+ expression, examine the other OPS to see if any of them are comparisons
+ of the same values, which we may be able to combine or eliminate.
+ For example, we can rewrite (a < b) | (a == b) as (a <= b). */
+
+static bool
+eliminate_redundant_comparison (enum tree_code opcode,
+ VEC (operand_entry_t, heap) **ops,
+ unsigned int currindex,
+ operand_entry_t curr)
+{
+ tree op1, op2;
+ enum tree_code lcode, rcode;
+ gimple def1, def2;
+ int i;
+ operand_entry_t oe;
+
+ if (opcode != BIT_IOR_EXPR && opcode != BIT_AND_EXPR)
+ return false;
+
+ /* Check that CURR is a comparison. */
+ if (TREE_CODE (curr->op) != SSA_NAME)
+ return false;
+ def1 = SSA_NAME_DEF_STMT (curr->op);
+ if (!is_gimple_assign (def1))
+ return false;
+ lcode = gimple_assign_rhs_code (def1);
+ if (TREE_CODE_CLASS (lcode) != tcc_comparison)
+ return false;
+ op1 = gimple_assign_rhs1 (def1);
+ op2 = gimple_assign_rhs2 (def1);
+
+ /* Now look for a similar comparison in the remaining OPS. */
+ for (i = currindex + 1;
+ VEC_iterate (operand_entry_t, *ops, i, oe);
+ i++)
+ {
+ tree t;
+
+ if (TREE_CODE (oe->op) != SSA_NAME)
+ continue;
+ def2 = SSA_NAME_DEF_STMT (oe->op);
+ if (!is_gimple_assign (def2))
+ continue;
+ rcode = gimple_assign_rhs_code (def2);
+ if (TREE_CODE_CLASS (rcode) != tcc_comparison)
+ continue;
+
+ /* If we got here, we have a match. See if we can combine the
+ two comparisons. */
+ if (opcode == BIT_IOR_EXPR)
+ t = maybe_fold_or_comparisons (lcode, op1, op2,
+ rcode, gimple_assign_rhs1 (def2),
+ gimple_assign_rhs2 (def2));
+ else
+ t = maybe_fold_and_comparisons (lcode, op1, op2,
+ rcode, gimple_assign_rhs1 (def2),
+ gimple_assign_rhs2 (def2));
+ if (!t)
+ continue;
+
+ /* maybe_fold_and_comparisons and maybe_fold_or_comparisons
+ always give us a boolean_type_node value back. If the original
+ BIT_AND_EXPR or BIT_IOR_EXPR was of a wider integer type,
+ we need to convert. */
+ if (!useless_type_conversion_p (TREE_TYPE (curr->op), TREE_TYPE (t)))
+ t = fold_convert (TREE_TYPE (curr->op), t);
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Equivalence: ");
+ print_generic_expr (dump_file, curr->op, 0);
+ fprintf (dump_file, " %s ", op_symbol_code (opcode));
+ print_generic_expr (dump_file, oe->op, 0);
+ fprintf (dump_file, " -> ");
+ print_generic_expr (dump_file, t, 0);
+ fprintf (dump_file, "\n");
+ }
+
+ /* Now we can delete oe, as it has been subsumed by the new combined
+ expression t. */
+ VEC_ordered_remove (operand_entry_t, *ops, i);
+ reassociate_stats.ops_eliminated ++;
+
+ /* If t is the same as curr->op, we're done. Otherwise we must
+ replace curr->op with t. Special case is if we got a constant
+ back, in which case we add it to the end instead of in place of
+ the current entry. */
+ if (TREE_CODE (t) == INTEGER_CST)
+ {
+ VEC_ordered_remove (operand_entry_t, *ops, currindex);
+ add_to_ops_vec (ops, t);
+ }
+ else if (!operand_equal_p (t, curr->op, 0))
+ {
+ tree tmpvar;
+ gimple sum;
+ enum tree_code subcode;
+ tree newop1;
+ tree newop2;
+ gcc_assert (COMPARISON_CLASS_P (t));
+ tmpvar = create_tmp_var (TREE_TYPE (t), NULL);
+ add_referenced_var (tmpvar);
+ extract_ops_from_tree (t, &subcode, &newop1, &newop2);
+ STRIP_USELESS_TYPE_CONVERSION (newop1);
+ STRIP_USELESS_TYPE_CONVERSION (newop2);
+ gcc_checking_assert (is_gimple_val (newop1)
+ && is_gimple_val (newop2));
+ sum = build_and_add_sum (tmpvar, newop1, newop2, subcode);
+ curr->op = gimple_get_lhs (sum);
+ }
+ return true;
+ }
+
+ return false;
+}
+
/* Perform various identities and other optimizations on the list of
operand entries, stored in OPS. The tree code for the binary
operation between all the operands is OPCODE. */
if (eliminate_not_pairs (opcode, ops, i, oe))
return;
if (eliminate_duplicate_pair (opcode, ops, &done, i, oe, oelast)
- || (!done && eliminate_plus_minus_pair (opcode, ops, i, oe)))
+ || (!done && eliminate_plus_minus_pair (opcode, ops, i, oe))
+ || (!done && eliminate_redundant_comparison (opcode, ops, i, oe)))
{
if (done)
return;
return false;
}
+/* Remove def stmt of VAR if VAR has zero uses and recurse
+ on rhs1 operand if so. */
+
+static void
+remove_visited_stmt_chain (tree var)
+{
+ gimple stmt;
+ gimple_stmt_iterator gsi;
+
+ while (1)
+ {
+ if (TREE_CODE (var) != SSA_NAME || !has_zero_uses (var))
+ return;
+ stmt = SSA_NAME_DEF_STMT (var);
+ if (!is_gimple_assign (stmt)
+ || !gimple_visited_p (stmt))
+ return;
+ var = gimple_assign_rhs1 (stmt);
+ gsi = gsi_for_stmt (stmt);
+ gsi_remove (&gsi, true);
+ release_defs (stmt);
+ }
+}
+
/* Recursively rewrite our linearized statements so that the operators
match those in OPS[OPINDEX], putting the computation in rank
order. */
static void
rewrite_expr_tree (gimple stmt, unsigned int opindex,
- VEC(operand_entry_t, heap) * ops)
+ VEC(operand_entry_t, heap) * ops, bool moved)
{
tree rhs1 = gimple_assign_rhs1 (stmt);
tree rhs2 = gimple_assign_rhs2 (stmt);
gimple_assign_set_rhs1 (stmt, oe1->op);
gimple_assign_set_rhs2 (stmt, oe2->op);
update_stmt (stmt);
+ if (rhs1 != oe1->op && rhs1 != oe2->op)
+ remove_visited_stmt_chain (rhs1);
if (dump_file && (dump_flags & TDF_DETAILS))
{
if (oe->op != rhs2)
{
+ if (!moved)
+ {
+ gimple_stmt_iterator gsinow, gsirhs1;
+ gimple stmt1 = stmt, stmt2;
+ unsigned int count;
+
+ gsinow = gsi_for_stmt (stmt);
+ count = VEC_length (operand_entry_t, ops) - opindex - 2;
+ while (count-- != 0)
+ {
+ stmt2 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt1));
+ gsirhs1 = gsi_for_stmt (stmt2);
+ gsi_move_before (&gsirhs1, &gsinow);
+ gsi_prev (&gsinow);
+ stmt1 = stmt2;
+ }
+ moved = true;
+ }
if (dump_file && (dump_flags & TDF_DETAILS))
{
}
/* Recurse on the LHS of the binary operator, which is guaranteed to
be the non-leaf side. */
- rewrite_expr_tree (SSA_NAME_DEF_STMT (rhs1), opindex + 1, ops);
+ rewrite_expr_tree (SSA_NAME_DEF_STMT (rhs1), opindex + 1, ops, moved);
}
/* Transform STMT, which is really (A +B) + (C + D) into the left
return NULL;
}
-static VEC(tree, heap) *broken_up_subtracts;
-
/* Recursively negate the value of TONEGATE, and return the SSA_NAME
representing the negated value. Insertions of any necessary
instructions go before GSI.
tonegate = fold_build1 (NEGATE_EXPR, TREE_TYPE (tonegate), tonegate);
resultofnegate = force_gimple_operand_gsi (gsi, tonegate, true,
NULL_TREE, true, GSI_SAME_STMT);
- VEC_safe_push (tree, heap, broken_up_subtracts, resultofnegate);
return resultofnegate;
}
if (TREE_CODE (lhs) == SSA_NAME
&& (immusestmt = get_single_immediate_use (lhs))
&& is_gimple_assign (immusestmt)
- && gimple_assign_rhs_code (immusestmt) == PLUS_EXPR)
+ && (gimple_assign_rhs_code (immusestmt) == PLUS_EXPR
+ || gimple_assign_rhs_code (immusestmt) == MULT_EXPR))
return true;
return false;
}
Place the operands of the expression tree in the vector named OPS. */
static void
-linearize_expr_tree (VEC(operand_entry_t, heap) **ops, gimple stmt)
+linearize_expr_tree (VEC(operand_entry_t, heap) **ops, gimple stmt,
+ bool is_associative, bool set_visited)
{
- gimple_stmt_iterator gsinow, gsilhs;
tree binlhs = gimple_assign_rhs1 (stmt);
tree binrhs = gimple_assign_rhs2 (stmt);
gimple binlhsdef, binrhsdef;
enum tree_code rhscode = gimple_assign_rhs_code (stmt);
struct loop *loop = loop_containing_stmt (stmt);
- gimple_set_visited (stmt, true);
+ if (set_visited)
+ gimple_set_visited (stmt, true);
if (TREE_CODE (binlhs) == SSA_NAME)
{
binlhsdef = SSA_NAME_DEF_STMT (binlhs);
- binlhsisreassoc = is_reassociable_op (binlhsdef, rhscode, loop);
+ binlhsisreassoc = (is_reassociable_op (binlhsdef, rhscode, loop)
+ && !stmt_could_throw_p (binlhsdef));
}
if (TREE_CODE (binrhs) == SSA_NAME)
{
binrhsdef = SSA_NAME_DEF_STMT (binrhs);
- binrhsisreassoc = is_reassociable_op (binrhsdef, rhscode, loop);
+ binrhsisreassoc = (is_reassociable_op (binrhsdef, rhscode, loop)
+ && !stmt_could_throw_p (binrhsdef));
}
/* If the LHS is not reassociable, but the RHS is, we need to swap
{
tree temp;
+ /* If this is not a associative operation like division, give up. */
+ if (!is_associative)
+ {
+ add_to_ops_vec (ops, binrhs);
+ return;
+ }
+
if (!binrhsisreassoc)
{
add_to_ops_vec (ops, binrhs);
gcc_assert (TREE_CODE (binrhs) != SSA_NAME
|| !is_reassociable_op (SSA_NAME_DEF_STMT (binrhs),
rhscode, loop));
- gsinow = gsi_for_stmt (stmt);
- gsilhs = gsi_for_stmt (SSA_NAME_DEF_STMT (binlhs));
- gsi_move_before (&gsilhs, &gsinow);
- linearize_expr_tree (ops, SSA_NAME_DEF_STMT (binlhs));
+ linearize_expr_tree (ops, SSA_NAME_DEF_STMT (binlhs),
+ is_associative, set_visited);
add_to_ops_vec (ops, binrhs);
}
unsigned int i = 0;
tree negate;
- for (i = 0; VEC_iterate (tree, broken_up_subtracts, i, negate); i++)
+ FOR_EACH_VEC_ELT (tree, plus_negates, i, negate)
{
gimple user = get_single_immediate_use (negate);
+ if (!user || !is_gimple_assign (user))
+ continue;
+
/* The negate operand can be either operand of a PLUS_EXPR
(it can be the LHS if the RHS is a constant for example).
Force the negate operand to the RHS of the PLUS_EXPR, then
transform the PLUS_EXPR into a MINUS_EXPR. */
- if (user
- && is_gimple_assign (user)
- && gimple_assign_rhs_code (user) == PLUS_EXPR)
+ if (gimple_assign_rhs_code (user) == PLUS_EXPR)
{
/* If the negated operand appears on the LHS of the
PLUS_EXPR, exchange the operands of the PLUS_EXPR
update_stmt (user);
}
}
+ else if (gimple_assign_rhs_code (user) == MINUS_EXPR)
+ {
+ if (gimple_assign_rhs1 (user) == negate)
+ {
+ /* We have
+ x = -a
+ y = x - b
+ which we transform into
+ x = a + b
+ y = -x .
+ This pushes down the negate which we possibly can merge
+ into some other operation, hence insert it into the
+ plus_negates vector. */
+ gimple feed = SSA_NAME_DEF_STMT (negate);
+ tree a = gimple_assign_rhs1 (feed);
+ tree rhs2 = gimple_assign_rhs2 (user);
+ gimple_stmt_iterator gsi = gsi_for_stmt (feed), gsi2;
+ gimple_replace_lhs (feed, negate);
+ gimple_assign_set_rhs_with_ops (&gsi, PLUS_EXPR, a, rhs2);
+ update_stmt (gsi_stmt (gsi));
+ gsi2 = gsi_for_stmt (user);
+ gimple_assign_set_rhs_with_ops (&gsi2, NEGATE_EXPR, negate, NULL);
+ update_stmt (gsi_stmt (gsi2));
+ gsi_move_before (&gsi, &gsi2);
+ VEC_safe_push (tree, heap, plus_negates,
+ gimple_assign_lhs (gsi_stmt (gsi2)));
+ }
+ else
+ {
+ /* Transform "x = -a; y = b - x" into "y = b + a", getting
+ rid of one operation. */
+ gimple feed = SSA_NAME_DEF_STMT (negate);
+ tree a = gimple_assign_rhs1 (feed);
+ tree rhs1 = gimple_assign_rhs1 (user);
+ gimple_stmt_iterator gsi = gsi_for_stmt (user);
+ gimple_assign_set_rhs_with_ops (&gsi, PLUS_EXPR, rhs1, a);
+ update_stmt (gsi_stmt (gsi));
+ }
+ }
}
}
+/* Returns true if OP is of a type for which we can do reassociation.
+ That is for integral or non-saturating fixed-point types, and for
+ floating point type when associative-math is enabled. */
+
+static bool
+can_reassociate_p (tree op)
+{
+ tree type = TREE_TYPE (op);
+ if ((INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_WRAPS (type))
+ || NON_SAT_FIXED_POINT_TYPE_P (type)
+ || (flag_associative_math && FLOAT_TYPE_P (type)))
+ return true;
+ return false;
+}
+
/* Break up subtract operations in block BB.
We do this top down because we don't know whether the subtract is
part of a possible chain of reassociation except at the top.
-
+
IE given
d = f + g
c = a + e
b = c - d
q = b - r
k = t - q
-
+
we want to break up k = t - q, but we won't until we've transformed q
= b - r, which won't be broken up until we transform b = c - d.
gimple stmt = gsi_stmt (gsi);
gimple_set_visited (stmt, false);
+ if (!is_gimple_assign (stmt)
+ || !can_reassociate_p (gimple_assign_lhs (stmt)))
+ continue;
+
/* Look for simple gimple subtract operations. */
- if (is_gimple_assign (stmt)
- && gimple_assign_rhs_code (stmt) == MINUS_EXPR)
+ if (gimple_assign_rhs_code (stmt) == MINUS_EXPR)
{
- tree lhs = gimple_assign_lhs (stmt);
- tree rhs1 = gimple_assign_rhs1 (stmt);
- tree rhs2 = gimple_assign_rhs2 (stmt);
-
- /* If associative-math we can do reassociation for
- non-integral types. Or, we can do reassociation for
- non-saturating fixed-point types. */
- if ((!INTEGRAL_TYPE_P (TREE_TYPE (lhs))
- || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
- || !INTEGRAL_TYPE_P (TREE_TYPE (rhs2)))
- && (!SCALAR_FLOAT_TYPE_P (TREE_TYPE (lhs))
- || !SCALAR_FLOAT_TYPE_P (TREE_TYPE(rhs1))
- || !SCALAR_FLOAT_TYPE_P (TREE_TYPE(rhs2))
- || !flag_associative_math)
- && (!NON_SAT_FIXED_POINT_TYPE_P (TREE_TYPE (lhs))
- || !NON_SAT_FIXED_POINT_TYPE_P (TREE_TYPE(rhs1))
- || !NON_SAT_FIXED_POINT_TYPE_P (TREE_TYPE(rhs2))))
+ if (!can_reassociate_p (gimple_assign_rhs1 (stmt))
+ || !can_reassociate_p (gimple_assign_rhs2 (stmt)))
continue;
/* Check for a subtract used only in an addition. If this
if (should_break_up_subtract (stmt))
break_up_subtract (stmt, &gsi);
}
+ else if (gimple_assign_rhs_code (stmt) == NEGATE_EXPR
+ && can_reassociate_p (gimple_assign_rhs1 (stmt)))
+ VEC_safe_push (tree, heap, plus_negates, gimple_assign_lhs (stmt));
}
for (son = first_dom_son (CDI_DOMINATORS, bb);
son;
{
gimple stmt = gsi_stmt (gsi);
- if (is_gimple_assign (stmt))
+ if (is_gimple_assign (stmt)
+ && !stmt_could_throw_p (stmt))
{
tree lhs, rhs1, rhs2;
enum tree_code rhs_code = gimple_assign_rhs_code (stmt);
/* If this was part of an already processed statement,
we don't need to touch it again. */
if (gimple_visited_p (stmt))
- continue;
+ {
+ /* This statement might have become dead because of previous
+ reassociations. */
+ if (has_zero_uses (gimple_get_lhs (stmt)))
+ {
+ gsi_remove (&gsi, true);
+ release_defs (stmt);
+ /* We might end up removing the last stmt above which
+ places the iterator to the end of the sequence.
+ Reset it to the last stmt in this case which might
+ be the end of the sequence as well if we removed
+ the last statement of the sequence. In which case
+ we need to bail out. */
+ if (gsi_end_p (gsi))
+ {
+ gsi = gsi_last_bb (bb);
+ if (gsi_end_p (gsi))
+ break;
+ }
+ }
+ continue;
+ }
lhs = gimple_assign_lhs (stmt);
rhs1 = gimple_assign_rhs1 (stmt);
rhs2 = gimple_assign_rhs2 (stmt);
- /* If associative-math we can do reassociation for
- non-integral types. Or, we can do reassociation for
- non-saturating fixed-point types. */
- if ((!INTEGRAL_TYPE_P (TREE_TYPE (lhs))
- || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
- || !INTEGRAL_TYPE_P (TREE_TYPE (rhs2)))
- && (!SCALAR_FLOAT_TYPE_P (TREE_TYPE (lhs))
- || !SCALAR_FLOAT_TYPE_P (TREE_TYPE(rhs1))
- || !SCALAR_FLOAT_TYPE_P (TREE_TYPE(rhs2))
- || !flag_associative_math)
- && (!NON_SAT_FIXED_POINT_TYPE_P (TREE_TYPE (lhs))
- || !NON_SAT_FIXED_POINT_TYPE_P (TREE_TYPE(rhs1))
- || !NON_SAT_FIXED_POINT_TYPE_P (TREE_TYPE(rhs2))))
+ /* For non-bit or min/max operations we can't associate
+ all types. Verify that here. */
+ if (rhs_code != BIT_IOR_EXPR
+ && rhs_code != BIT_AND_EXPR
+ && rhs_code != BIT_XOR_EXPR
+ && rhs_code != MIN_EXPR
+ && rhs_code != MAX_EXPR
+ && (!can_reassociate_p (lhs)
+ || !can_reassociate_p (rhs1)
+ || !can_reassociate_p (rhs2)))
continue;
if (associative_tree_code (rhs_code))
continue;
gimple_set_visited (stmt, true);
- linearize_expr_tree (&ops, stmt);
- qsort (VEC_address (operand_entry_t, ops),
- VEC_length (operand_entry_t, ops),
- sizeof (operand_entry_t),
- sort_by_operand_rank);
+ linearize_expr_tree (&ops, stmt, true, true);
+ VEC_qsort (operand_entry_t, ops, sort_by_operand_rank);
optimize_ops_list (rhs_code, &ops);
+ if (undistribute_ops_list (rhs_code, &ops,
+ loop_containing_stmt (stmt)))
+ {
+ VEC_qsort (operand_entry_t, ops, sort_by_operand_rank);
+ optimize_ops_list (rhs_code, &ops);
+ }
if (VEC_length (operand_entry_t, ops) == 1)
{
fprintf (dump_file, "Transforming ");
print_gimple_stmt (dump_file, stmt, 0, 0);
}
-
+
+ rhs1 = gimple_assign_rhs1 (stmt);
gimple_assign_set_rhs_from_tree (&gsi,
VEC_last (operand_entry_t,
ops)->op);
update_stmt (stmt);
+ remove_visited_stmt_chain (rhs1);
if (dump_file && (dump_flags & TDF_DETAILS))
{
}
}
else
- {
- rewrite_expr_tree (stmt, 0, ops);
- }
+ rewrite_expr_tree (stmt, 0, ops, false);
VEC_free (operand_entry_t, heap, ops);
}
operand_entry_t oe;
unsigned int i;
- for (i = 0; VEC_iterate (operand_entry_t, ops, i, oe); i++)
+ FOR_EACH_VEC_ELT (operand_entry_t, ops, i, oe)
{
fprintf (file, "Op %d -> rank: %d, tree: ", i, oe->rank);
print_generic_expr (file, oe->op, 0);
/* Dump the operand entry vector OPS to STDERR. */
-void
+DEBUG_FUNCTION void
debug_ops_vector (VEC (operand_entry_t, heap) *ops)
{
dump_ops_vector (stderr, ops);
operand_entry_pool = create_alloc_pool ("operand entry pool",
sizeof (struct operand_entry), 30);
+ next_operand_entry_id = 0;
/* Reverse RPO (Reverse Post Order) will give us something where
deeper loops come later. */
/* Give each argument a distinct rank. */
for (param = DECL_ARGUMENTS (current_function_decl);
param;
- param = TREE_CHAIN (param))
+ param = DECL_CHAIN (param))
{
if (gimple_default_def (cfun, param) != NULL)
{
free (bbs);
calculate_dominance_info (CDI_POST_DOMINATORS);
- broken_up_subtracts = NULL;
+ plus_negates = NULL;
}
/* Cleanup after the reassociation pass, and print stats if
pointer_map_destroy (operand_rank);
free_alloc_pool (operand_entry_pool);
free (bb_rank);
- VEC_free (tree, heap, broken_up_subtracts);
+ VEC_free (tree, heap, plus_negates);
free_dominance_info (CDI_POST_DOMINATORS);
loop_optimizer_finalize ();
}
NULL, /* next */
0, /* static_pass_number */
TV_TREE_REASSOC, /* tv_id */
- PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
+ PROP_cfg | PROP_ssa, /* properties_required */
0, /* properties_provided */
0, /* properties_destroyed */
0, /* todo_flags_start */