static void
execute_cse_reciprocals_1 (gimple_stmt_iterator *def_gsi, tree def)
{
use_operand_p use_p;
imm_use_iterator use_iter;
struct occurrence *occ;
int count = 0, threshold;
gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def)) && is_gimple_reg (def));
FOR_EACH_IMM_USE_FAST (use_p, use_iter, def)
{
gimple use_stmt = USE_STMT (use_p);
if (is_division_by (use_stmt, def))
{
register_division_in (gimple_bb (use_stmt));
count++;
}
}
static bool
process_assignment (gimple stmt, gimple_stmt_iterator call, tree *m,
tree *a, tree *ass_var)
{
tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE;
tree dest = gimple_assign_lhs (stmt);
enum tree_code code = gimple_assign_rhs_code (stmt);
enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
tree src_var = gimple_assign_rhs1 (stmt);
/* See if this is a simple copy operation of an SSA name to the function
result. In that case we may have a simple tail call. Ignore type
conversions that can never produce extra code between the function
call and the function return. */
if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
&& (TREE_CODE (src_var) == SSA_NAME))
{
/* Reject a tailcall if the type conversion might need
additional code. */
if (gimple_assign_cast_p (stmt))
{
if (TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
return false;
/* Even if the type modes are the same, if the precision of the
type is smaller than mode's precision,
reduce_to_bit_field_precision would generate additional code. */
if (INTEGRAL_TYPE_P (TREE_TYPE (dest))
&& (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (dest)))
> TYPE_PRECISION (TREE_TYPE (dest))))
return false;
}
if (src_var != *ass_var)
return false;
*ass_var = dest;
return true;
}
switch (rhs_class)
{
case GIMPLE_BINARY_RHS:
op1 = gimple_assign_rhs2 (stmt);
/* Fall through. */
case GIMPLE_UNARY_RHS:
op0 = gimple_assign_rhs1 (stmt);
break;
default:
return false;
}
/* Accumulator optimizations will reverse the order of operations.
We can only do that for floating-point types if we're assuming
that addition and multiplication are associative. */
if (!flag_associative_math)
if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
return false;
if (rhs_class == GIMPLE_UNARY_RHS)
;
else if (op0 == *ass_var
&& (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
;
else if (op1 == *ass_var
&& (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
;
else
return false;
switch (code)
{
case PLUS_EXPR:
*a = non_ass_var;
*ass_var = dest;
return true;
case POINTER_PLUS_EXPR:
if (op0 != *ass_var)
return false;
*a = non_ass_var;
*ass_var = dest;
return true;
case MULT_EXPR:
*m = non_ass_var;
*ass_var = dest;
return true;
case NEGATE_EXPR:
*m = build_minus_one_cst (TREE_TYPE (op0));
*ass_var = dest;
return true;
case MINUS_EXPR:
if (*ass_var == op0)
*a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
else
{
*m = build_minus_one_cst (TREE_TYPE (non_ass_var));
*a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
}
*ass_var = dest;
return true;
/* TODO -- Handle POINTER_PLUS_EXPR. */
default:
return false;
}
}
static bool
abs_replacement (basic_block cond_bb, basic_block middle_bb,
edge e0 ATTRIBUTE_UNUSED, edge e1,
gimple phi, tree arg0, tree arg1)
{
tree result;
gimple new_stmt, cond;
gimple_stmt_iterator gsi;
edge true_edge, false_edge;
gimple assign;
edge e;
tree rhs, lhs;
bool negate;
enum tree_code cond_code;
/* If the type says honor signed zeros we cannot do this
optimization. */
if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
return false;
/* OTHER_BLOCK must have only one executable statement which must have the
form arg0 = -arg1 or arg1 = -arg0. */
assign = last_and_only_stmt (middle_bb);
/* If we did not find the proper negation assignment, then we can not
optimize. */
if (assign == NULL)
return false;
/* If we got here, then we have found the only executable statement
in OTHER_BLOCK. If it is anything other than arg = -arg1 or
arg1 = -arg0, then we can not optimize. */
if (gimple_code (assign) != GIMPLE_ASSIGN)
return false;
lhs = gimple_assign_lhs (assign);
if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
return false;
rhs = gimple_assign_rhs1 (assign);
/* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
if (!(lhs == arg0 && rhs == arg1)
&& !(lhs == arg1 && rhs == arg0))
return false;
cond = last_stmt (cond_bb);
result = PHI_RESULT (phi);
/* Only relationals comparing arg[01] against zero are interesting. */
cond_code = gimple_cond_code (cond);
if (cond_code != GT_EXPR && cond_code != GE_EXPR
&& cond_code != LT_EXPR && cond_code != LE_EXPR)
return false;
/* Make sure the conditional is arg[01] OP y. */
if (gimple_cond_lhs (cond) != rhs)
return false;
if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond)))
? real_zerop (gimple_cond_rhs (cond))
: integer_zerop (gimple_cond_rhs (cond)))
;
else
return false;
/* We need to know which is the true edge and which is the false
edge so that we know if have abs or negative abs. */
extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
/* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
the false edge goes to OTHER_BLOCK. */
if (cond_code == GT_EXPR || cond_code == GE_EXPR)
e = true_edge;
else
e = false_edge;
if (e->dest == middle_bb)
negate = true;
else
negate = false;
result = duplicate_ssa_name (result, NULL);
if (negate)
{
tree tmp = create_tmp_var (TREE_TYPE (result), NULL);
add_referenced_var (tmp);
lhs = make_ssa_name (tmp, NULL);
}
else
lhs = result;
/* Build the modify expression with abs expression. */
new_stmt = gimple_build_assign_with_ops (ABS_EXPR, lhs, rhs, NULL);
gsi = gsi_last_bb (cond_bb);
gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
if (negate)
{
/* Get the right GSI. We want to insert after the recently
added ABS_EXPR statement (which we know is the first statement
in the block. */
new_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, result, lhs, NULL);
gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
}
replace_phi_edge_with_variable (cond_bb, e1, phi, result);
/* Note that we optimized this PHI. */
return true;
}
static inline bool
is_replaceable_p (gimple stmt)
{
use_operand_p use_p;
tree def;
gimple use_stmt;
location_t locus1, locus2;
tree block1, block2;
/* Only consider modify stmts. */
if (!is_gimple_assign (stmt))
return false;
/* If the statement may throw an exception, it cannot be replaced. */
if (stmt_could_throw_p (stmt))
return false;
/* Punt if there is more than 1 def. */
def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF);
if (!def)
return false;
/* Only consider definitions which have a single use. */
if (!single_imm_use (def, &use_p, &use_stmt))
return false;
/* If the use isn't in this block, it wont be replaced either. */
if (gimple_bb (use_stmt) != gimple_bb (stmt))
return false;
locus1 = gimple_location (stmt);
block1 = gimple_block (stmt);
if (gimple_code (use_stmt) == GIMPLE_PHI)
{
locus2 = 0;
block2 = NULL_TREE;
}
else
{
locus2 = gimple_location (use_stmt);
block2 = gimple_block (use_stmt);
}
if (!optimize
&& ((locus1 && locus1 != locus2) || (block1 && block1 != block2)))
return false;
/* Used in this block, but at the TOP of the block, not the end. */
if (gimple_code (use_stmt) == GIMPLE_PHI)
return false;
/* There must be no VDEFs. */
if (!(ZERO_SSA_OPERANDS (stmt, SSA_OP_VDEF)))
return false;
/* Without alias info we can't move around loads. */
if (gimple_references_memory_p (stmt) && !optimize)
return false;
/* Float expressions must go through memory if float-store is on. */
if (flag_float_store
&& FLOAT_TYPE_P (gimple_expr_type (stmt)))
return false;
/* An assignment with a register variable on the RHS is not
replaceable. */
if (gimple_assign_rhs_code (stmt) == VAR_DECL
&& DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt)))
return false;
/* No function calls can be replaced. */
if (is_gimple_call (stmt))
return false;
/* Leave any stmt with volatile operands alone as well. */
if (gimple_has_volatile_ops (stmt))
return false;
return true;
}
void
record_conditions (vec<cond_equivalence> *p, tree cond, tree inverted)
{
tree op0, op1;
cond_equivalence c;
if (!COMPARISON_CLASS_P (cond))
return;
op0 = TREE_OPERAND (cond, 0);
op1 = TREE_OPERAND (cond, 1);
switch (TREE_CODE (cond))
{
case LT_EXPR:
case GT_EXPR:
if (FLOAT_TYPE_P (TREE_TYPE (op0)))
{
build_and_record_new_cond (ORDERED_EXPR, op0, op1, p);
build_and_record_new_cond (LTGT_EXPR, op0, op1, p);
}
build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
? LE_EXPR : GE_EXPR),
op0, op1, p);
build_and_record_new_cond (NE_EXPR, op0, op1, p);
build_and_record_new_cond (EQ_EXPR, op0, op1, p, false);
break;
case GE_EXPR:
case LE_EXPR:
if (FLOAT_TYPE_P (TREE_TYPE (op0)))
{
build_and_record_new_cond (ORDERED_EXPR, op0, op1, p);
}
break;
case EQ_EXPR:
if (FLOAT_TYPE_P (TREE_TYPE (op0)))
{
build_and_record_new_cond (ORDERED_EXPR, op0, op1, p);
}
build_and_record_new_cond (LE_EXPR, op0, op1, p);
build_and_record_new_cond (GE_EXPR, op0, op1, p);
break;
case UNORDERED_EXPR:
build_and_record_new_cond (NE_EXPR, op0, op1, p);
build_and_record_new_cond (UNLE_EXPR, op0, op1, p);
build_and_record_new_cond (UNGE_EXPR, op0, op1, p);
build_and_record_new_cond (UNEQ_EXPR, op0, op1, p);
build_and_record_new_cond (UNLT_EXPR, op0, op1, p);
build_and_record_new_cond (UNGT_EXPR, op0, op1, p);
break;
case UNLT_EXPR:
case UNGT_EXPR:
build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
? UNLE_EXPR : UNGE_EXPR),
op0, op1, p);
build_and_record_new_cond (NE_EXPR, op0, op1, p);
break;
case UNEQ_EXPR:
build_and_record_new_cond (UNLE_EXPR, op0, op1, p);
build_and_record_new_cond (UNGE_EXPR, op0, op1, p);
break;
case LTGT_EXPR:
build_and_record_new_cond (NE_EXPR, op0, op1, p);
build_and_record_new_cond (ORDERED_EXPR, op0, op1, p);
break;
default:
break;
}
/* Now store the original true and false conditions into the first
two slots. */
initialize_expr_from_cond (cond, &c.cond);
c.value = boolean_true_node;
p->safe_push (c);
/* It is possible for INVERTED to be the negation of a comparison,
and not a valid RHS or GIMPLE_COND condition. This happens because
invert_truthvalue may return such an expression when asked to invert
a floating-point comparison. These comparisons are not assumed to
obey the trichotomy law. */
initialize_expr_from_cond (inverted, &c.cond);
c.value = boolean_false_node;
p->safe_push (c);
}
/* The core routine of conditional store replacement and normal
phi optimizations. Both share much of the infrastructure in how
to match applicable basic block patterns. DO_STORE_ELIM is true
when we want to do conditional store replacement, false otherwise.
DO_HOIST_LOADS is true when we want to hoist adjacent loads out
of diamond control flow patterns, false otherwise. */
static unsigned int
tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads)
{
basic_block bb;
basic_block *bb_order;
unsigned n, i;
bool cfgchanged = false;
hash_set<tree> *nontrap = 0;
if (do_store_elim)
/* Calculate the set of non-trapping memory accesses. */
nontrap = get_non_trapping ();
/* Search every basic block for COND_EXPR we may be able to optimize.
We walk the blocks in order that guarantees that a block with
a single predecessor is processed before the predecessor.
This ensures that we collapse inner ifs before visiting the
outer ones, and also that we do not try to visit a removed
block. */
bb_order = single_pred_before_succ_order ();
n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS;
for (i = 0; i < n; i++)
{
gimple cond_stmt;
gphi *phi;
basic_block bb1, bb2;
edge e1, e2;
tree arg0, arg1;
bb = bb_order[i];
cond_stmt = last_stmt (bb);
/* Check to see if the last statement is a GIMPLE_COND. */
if (!cond_stmt
|| gimple_code (cond_stmt) != GIMPLE_COND)
continue;
e1 = EDGE_SUCC (bb, 0);
bb1 = e1->dest;
e2 = EDGE_SUCC (bb, 1);
bb2 = e2->dest;
/* We cannot do the optimization on abnormal edges. */
if ((e1->flags & EDGE_ABNORMAL) != 0
|| (e2->flags & EDGE_ABNORMAL) != 0)
continue;
/* If either bb1's succ or bb2 or bb2's succ is non NULL. */
if (EDGE_COUNT (bb1->succs) == 0
|| bb2 == NULL
|| EDGE_COUNT (bb2->succs) == 0)
continue;
/* Find the bb which is the fall through to the other. */
if (EDGE_SUCC (bb1, 0)->dest == bb2)
;
else if (EDGE_SUCC (bb2, 0)->dest == bb1)
{
std::swap (bb1, bb2);
std::swap (e1, e2);
}
else if (do_store_elim
&& EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
{
basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
if (!single_succ_p (bb1)
|| (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0
|| !single_succ_p (bb2)
|| (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0
|| EDGE_COUNT (bb3->preds) != 2)
continue;
if (cond_if_else_store_replacement (bb1, bb2, bb3))
cfgchanged = true;
continue;
}
else if (do_hoist_loads
&& EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
{
basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt)))
&& single_succ_p (bb1)
&& single_succ_p (bb2)
&& single_pred_p (bb1)
&& single_pred_p (bb2)
&& EDGE_COUNT (bb->succs) == 2
&& EDGE_COUNT (bb3->preds) == 2
/* If one edge or the other is dominant, a conditional move
is likely to perform worse than the well-predicted branch. */
&& !predictable_edge_p (EDGE_SUCC (bb, 0))
&& !predictable_edge_p (EDGE_SUCC (bb, 1)))
hoist_adjacent_loads (bb, bb1, bb2, bb3);
continue;
}
else
continue;
e1 = EDGE_SUCC (bb1, 0);
/* Make sure that bb1 is just a fall through. */
if (!single_succ_p (bb1)
|| (e1->flags & EDGE_FALLTHRU) == 0)
continue;
/* Also make sure that bb1 only have one predecessor and that it
is bb. */
if (!single_pred_p (bb1)
|| single_pred (bb1) != bb)
continue;
if (do_store_elim)
{
/* bb1 is the middle block, bb2 the join block, bb the split block,
e1 the fallthrough edge from bb1 to bb2. We can't do the
optimization if the join block has more than two predecessors. */
if (EDGE_COUNT (bb2->preds) > 2)
continue;
if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
cfgchanged = true;
}
else
{
gimple_seq phis = phi_nodes (bb2);
gimple_stmt_iterator gsi;
bool candorest = true;
/* Value replacement can work with more than one PHI
so try that first. */
for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
{
phi = as_a <gphi *> (gsi_stmt (gsi));
arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2)
{
candorest = false;
cfgchanged = true;
break;
}
}
if (!candorest)
continue;
phi = single_non_singleton_phi_for_edges (phis, e1, e2);
if (!phi)
continue;
arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
/* Something is wrong if we cannot find the arguments in the PHI
node. */
gcc_assert (arg0 != NULL && arg1 != NULL);
if (factor_out_conditional_conversion (e1, e2, phi, arg0, arg1))
{
/* factor_out_conditional_conversion may create a new PHI in
BB2 and eliminate an existing PHI in BB2. Recompute values
that may be affected by that change. */
phis = phi_nodes (bb2);
phi = single_non_singleton_phi_for_edges (phis, e1, e2);
gcc_assert (phi);
arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
gcc_assert (arg0 != NULL && arg1 != NULL);
}
/* Do the replacement of conditional if it can be done. */
if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
cfgchanged = true;
else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
cfgchanged = true;
else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
cfgchanged = true;
}
}
free (bb_order);
if (do_store_elim)
delete nontrap;
/* If the CFG has changed, we should cleanup the CFG. */
if (cfgchanged && do_store_elim)
{
/* In cond-store replacement we have added some loads on edges
and new VOPS (as we moved the store, and created a load). */
gsi_commit_edge_inserts ();
return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
}
else if (cfgchanged)
return TODO_cleanup_cfg;
return 0;
}
