static void
init_copy_prop (void)
{
basic_block bb;
n_copy_of = num_ssa_names;
copy_of = XCNEWVEC (prop_value_t, n_copy_of);
FOR_EACH_BB_FN (bb, cfun)
{
for (gimple_stmt_iterator si = gsi_start_bb (bb); !gsi_end_p (si);
gsi_next (&si))
{
gimple *stmt = gsi_stmt (si);
ssa_op_iter iter;
tree def;
/* The only statements that we care about are those that may
generate useful copies. We also need to mark conditional
jumps so that their outgoing edges are added to the work
lists of the propagator. */
if (stmt_ends_bb_p (stmt))
prop_set_simulate_again (stmt, true);
else if (stmt_may_generate_copy (stmt))
prop_set_simulate_again (stmt, true);
else
prop_set_simulate_again (stmt, false);
/* Mark all the outputs of this statement as not being
the copy of anything. */
FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS)
if (!prop_simulate_again_p (stmt))
set_copy_of_val (def, def);
}
for (gphi_iterator si = gsi_start_phis (bb); !gsi_end_p (si);
gsi_next (&si))
{
gphi *phi = si.phi ();
tree def;
def = gimple_phi_result (phi);
if (virtual_operand_p (def))
prop_set_simulate_again (phi, false);
else
prop_set_simulate_again (phi, true);
if (!prop_simulate_again_p (phi))
set_copy_of_val (def, def);
}
}
}
static void
mf_build_check_statement_for (tree base, tree limit,
gimple_stmt_iterator *instr_gsi,
location_t location, tree dirflag)
{
gimple_stmt_iterator gsi;
basic_block cond_bb, then_bb, join_bb;
edge e;
tree cond, t, u, v;
tree mf_base;
tree mf_elem;
tree mf_limit;
gimple g;
gimple_seq seq, stmts;
/* We first need to split the current basic block, and start altering
the CFG. This allows us to insert the statements we're about to
construct into the right basic blocks. */
cond_bb = gimple_bb (gsi_stmt (*instr_gsi));
gsi = *instr_gsi;
gsi_prev (&gsi);
if (! gsi_end_p (gsi))
e = split_block (cond_bb, gsi_stmt (gsi));
else
e = split_block_after_labels (cond_bb);
cond_bb = e->src;
join_bb = e->dest;
/* A recap at this point: join_bb is the basic block at whose head
is the gimple statement for which this check expression is being
built. cond_bb is the (possibly new, synthetic) basic block the
end of which will contain the cache-lookup code, and a
conditional that jumps to the cache-miss code or, much more
likely, over to join_bb. */
/* Create the bb that contains the cache-miss fallback block (mf_check). */
then_bb = create_empty_bb (cond_bb);
make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE);
make_single_succ_edge (then_bb, join_bb, EDGE_FALLTHRU);
/* Mark the pseudo-fallthrough edge from cond_bb to join_bb. */
e = find_edge (cond_bb, join_bb);
e->flags = EDGE_FALSE_VALUE;
e->count = cond_bb->count;
e->probability = REG_BR_PROB_BASE;
/* Update dominance info. Note that bb_join's data was
updated by split_block. */
if (dom_info_available_p (CDI_DOMINATORS))
{
set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb);
set_immediate_dominator (CDI_DOMINATORS, join_bb, cond_bb);
}
/* Update loop info. */
if (current_loops)
add_bb_to_loop (then_bb, cond_bb->loop_father);
/* Build our local variables. */
mf_elem = create_tmp_reg (mf_cache_structptr_type, "__mf_elem");
mf_base = create_tmp_reg (mf_uintptr_type, "__mf_base");
mf_limit = create_tmp_reg (mf_uintptr_type, "__mf_limit");
/* Build: __mf_base = (uintptr_t) <base address expression>. */
seq = NULL;
t = fold_convert_loc (location, mf_uintptr_type,
unshare_expr (base));
t = force_gimple_operand (t, &stmts, false, NULL_TREE);
gimple_seq_add_seq (&seq, stmts);
g = gimple_build_assign (mf_base, t);
gimple_set_location (g, location);
gimple_seq_add_stmt (&seq, g);
/* Build: __mf_limit = (uintptr_t) <limit address expression>. */
t = fold_convert_loc (location, mf_uintptr_type,
unshare_expr (limit));
t = force_gimple_operand (t, &stmts, false, NULL_TREE);
gimple_seq_add_seq (&seq, stmts);
g = gimple_build_assign (mf_limit, t);
gimple_set_location (g, location);
gimple_seq_add_stmt (&seq, g);
/* Build: __mf_elem = &__mf_lookup_cache [(__mf_base >> __mf_shift)
& __mf_mask]. */
t = build2 (RSHIFT_EXPR, mf_uintptr_type, mf_base,
flag_mudflap_threads ? mf_cache_shift_decl
: mf_cache_shift_decl_l);
t = build2 (BIT_AND_EXPR, mf_uintptr_type, t,
flag_mudflap_threads ? mf_cache_mask_decl
: mf_cache_mask_decl_l);
t = build4 (ARRAY_REF,
TREE_TYPE (TREE_TYPE (mf_cache_array_decl)),
mf_cache_array_decl, t, NULL_TREE, NULL_TREE);
t = build1 (ADDR_EXPR, mf_cache_structptr_type, t);
t = force_gimple_operand (t, &stmts, false, NULL_TREE);
gimple_seq_add_seq (&seq, stmts);
g = gimple_build_assign (mf_elem, t);
gimple_set_location (g, location);
gimple_seq_add_stmt (&seq, g);
/* Quick validity check.
if (__mf_elem->low > __mf_base
|| (__mf_elem_high < __mf_limit))
{
__mf_check ();
... and only if single-threaded:
__mf_lookup_shift_1 = f...;
__mf_lookup_mask_l = ...;
}
It is expected that this body of code is rarely executed so we mark
the edge to the THEN clause of the conditional jump as unlikely. */
/* Construct t <-- '__mf_elem->low > __mf_base'. */
t = build3 (COMPONENT_REF, mf_uintptr_type,
build1 (INDIRECT_REF, mf_cache_struct_type, mf_elem),
TYPE_FIELDS (mf_cache_struct_type), NULL_TREE);
t = build2 (GT_EXPR, boolean_type_node, t, mf_base);
/* Construct '__mf_elem->high < __mf_limit'.
First build:
1) u <-- '__mf_elem->high'
2) v <-- '__mf_limit'.
Then build 'u <-- (u < v). */
u = build3 (COMPONENT_REF, mf_uintptr_type,
build1 (INDIRECT_REF, mf_cache_struct_type, mf_elem),
DECL_CHAIN (TYPE_FIELDS (mf_cache_struct_type)), NULL_TREE);
v = mf_limit;
u = build2 (LT_EXPR, boolean_type_node, u, v);
/* Build the composed conditional: t <-- 't || u'. Then store the
result of the evaluation of 't' in a temporary variable which we
can use as the condition for the conditional jump. */
t = build2 (TRUTH_OR_EXPR, boolean_type_node, t, u);
t = force_gimple_operand (t, &stmts, false, NULL_TREE);
gimple_seq_add_seq (&seq, stmts);
cond = create_tmp_reg (boolean_type_node, "__mf_unlikely_cond");
g = gimple_build_assign (cond, t);
gimple_set_location (g, location);
gimple_seq_add_stmt (&seq, g);
/* Build the conditional jump. 'cond' is just a temporary so we can
simply build a void COND_EXPR. We do need labels in both arms though. */
g = gimple_build_cond (NE_EXPR, cond, boolean_false_node, NULL_TREE,
NULL_TREE);
gimple_set_location (g, location);
gimple_seq_add_stmt (&seq, g);
/* At this point, after so much hard work, we have only constructed
the conditional jump,
if (__mf_elem->low > __mf_base
|| (__mf_elem_high < __mf_limit))
The lowered GIMPLE tree representing this code is in the statement
list starting at 'head'.
We can insert this now in the current basic block, i.e. the one that
the statement we're instrumenting was originally in. */
gsi = gsi_last_bb (cond_bb);
gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
/* Now build up the body of the cache-miss handling:
__mf_check();
refresh *_l vars.
This is the body of the conditional. */
seq = NULL;
/* u is a string, so it is already a gimple value. */
u = mf_file_function_line_tree (location);
/* NB: we pass the overall [base..limit] range to mf_check. */
v = fold_build2_loc (location, PLUS_EXPR, mf_uintptr_type,
fold_build2_loc (location,
MINUS_EXPR, mf_uintptr_type, mf_limit, mf_base),
build_int_cst (mf_uintptr_type, 1));
v = force_gimple_operand (v, &stmts, true, NULL_TREE);
gimple_seq_add_seq (&seq, stmts);
g = gimple_build_call (mf_check_fndecl, 4, mf_base, v, dirflag, u);
gimple_seq_add_stmt (&seq, g);
if (! flag_mudflap_threads)
{
if (stmt_ends_bb_p (g))
{
gsi = gsi_start_bb (then_bb);
gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
e = split_block (then_bb, g);
then_bb = e->dest;
seq = NULL;
}
g = gimple_build_assign (mf_cache_shift_decl_l, mf_cache_shift_decl);
gimple_seq_add_stmt (&seq, g);
g = gimple_build_assign (mf_cache_mask_decl_l, mf_cache_mask_decl);
gimple_seq_add_stmt (&seq, g);
}
/* Insert the check code in the THEN block. */
gsi = gsi_start_bb (then_bb);
gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
*instr_gsi = gsi_start_bb (join_bb);
}
static bool
init_dont_simulate_again (void)
{
basic_block bb;
block_stmt_iterator bsi;
tree phi;
bool saw_a_complex_op = false;
FOR_EACH_BB (bb)
{
for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
DONT_SIMULATE_AGAIN (phi) = !is_complex_reg (PHI_RESULT (phi));
for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
{
tree orig_stmt, stmt, rhs = NULL;
bool dsa;
orig_stmt = stmt = bsi_stmt (bsi);
/* Most control-altering statements must be initially
simulated, else we won't cover the entire cfg. */
dsa = !stmt_ends_bb_p (stmt);
switch (TREE_CODE (stmt))
{
case RETURN_EXPR:
/* We don't care what the lattice value of <retval> is,
since it's never used as an input to another computation. */
dsa = true;
stmt = TREE_OPERAND (stmt, 0);
if (!stmt || TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
break;
/* FALLTHRU */
case GIMPLE_MODIFY_STMT:
dsa = !is_complex_reg (GIMPLE_STMT_OPERAND (stmt, 0));
rhs = GIMPLE_STMT_OPERAND (stmt, 1);
break;
case COND_EXPR:
rhs = TREE_OPERAND (stmt, 0);
break;
default:
break;
}
if (rhs)
switch (TREE_CODE (rhs))
{
case EQ_EXPR:
case NE_EXPR:
rhs = TREE_OPERAND (rhs, 0);
/* FALLTHRU */
case PLUS_EXPR:
case MINUS_EXPR:
case MULT_EXPR:
case TRUNC_DIV_EXPR:
case CEIL_DIV_EXPR:
case FLOOR_DIV_EXPR:
case ROUND_DIV_EXPR:
case RDIV_EXPR:
case NEGATE_EXPR:
case CONJ_EXPR:
if (TREE_CODE (TREE_TYPE (rhs)) == COMPLEX_TYPE)
saw_a_complex_op = true;
break;
case REALPART_EXPR:
case IMAGPART_EXPR:
/* The total store transformation performed during
gimplification creates such uninitialized loads
and we need to lower the statement to be able
to fix things up. */
if (TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
&& ssa_undefined_value_p (TREE_OPERAND (rhs, 0)))
saw_a_complex_op = true;
break;
default:
break;
}
DONT_SIMULATE_AGAIN (orig_stmt) = dsa;
}
}
return saw_a_complex_op;
}
static void
init_copy_prop (void)
{
basic_block bb;
copy_of = XCNEWVEC (prop_value_t, num_ssa_names);
FOR_EACH_BB (bb)
{
gimple_stmt_iterator si;
int depth = bb_loop_depth (bb);
for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
{
gimple stmt = gsi_stmt (si);
ssa_op_iter iter;
tree def;
/* The only statements that we care about are those that may
generate useful copies. We also need to mark conditional
jumps so that their outgoing edges are added to the work
lists of the propagator.
Avoid copy propagation from an inner into an outer loop.
Otherwise, this may move loop variant variables outside of
their loops and prevent coalescing opportunities. If the
value was loop invariant, it will be hoisted by LICM and
exposed for copy propagation.
??? This doesn't make sense. */
if (stmt_ends_bb_p (stmt))
prop_set_simulate_again (stmt, true);
else if (stmt_may_generate_copy (stmt)
/* Since we are iterating over the statements in
BB, not the phi nodes, STMT will always be an
assignment. */
&& loop_depth_of_name (gimple_assign_rhs1 (stmt)) <= depth)
prop_set_simulate_again (stmt, true);
else
prop_set_simulate_again (stmt, false);
/* Mark all the outputs of this statement as not being
the copy of anything. */
FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS)
if (!prop_simulate_again_p (stmt))
set_copy_of_val (def, def);
}
for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
{
gimple phi = gsi_stmt (si);
tree def;
def = gimple_phi_result (phi);
if (virtual_operand_p (def))
prop_set_simulate_again (phi, false);
else
prop_set_simulate_again (phi, true);
if (!prop_simulate_again_p (phi))
set_copy_of_val (def, def);
}
}
}
static void
init_copy_prop (void)
{
basic_block bb;
copy_of = XCNEWVEC (prop_value_t, num_ssa_names);
cached_last_copy_of = XCNEWVEC (tree, num_ssa_names);
FOR_EACH_BB (bb)
{
gimple_stmt_iterator si;
int depth = bb->loop_depth;
for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
{
gimple stmt = gsi_stmt (si);
ssa_op_iter iter;
tree def;
/* The only statements that we care about are those that may
generate useful copies. We also need to mark conditional
jumps so that their outgoing edges are added to the work
lists of the propagator.
Avoid copy propagation from an inner into an outer loop.
Otherwise, this may move loop variant variables outside of
their loops and prevent coalescing opportunities. If the
value was loop invariant, it will be hoisted by LICM and
exposed for copy propagation. */
if (stmt_ends_bb_p (stmt))
prop_set_simulate_again (stmt, true);
else if (stmt_may_generate_copy (stmt)
/* Since we are iterating over the statements in
BB, not the phi nodes, STMT will always be an
assignment. */
&& loop_depth_of_name (gimple_assign_rhs1 (stmt)) <= depth)
prop_set_simulate_again (stmt, true);
else
prop_set_simulate_again (stmt, false);
/* Mark all the outputs of this statement as not being
the copy of anything. */
FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS)
if (!prop_simulate_again_p (stmt))
set_copy_of_val (def, def);
else
cached_last_copy_of[SSA_NAME_VERSION (def)] = def;
}
for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
{
gimple phi = gsi_stmt (si);
tree def;
def = gimple_phi_result (phi);
if (!is_gimple_reg (def)
/* In loop-closed SSA form do not copy-propagate through
PHI nodes. Technically this is only needed for loop
exit PHIs, but this is difficult to query. */
|| (current_loops
&& gimple_phi_num_args (phi) == 1
&& loops_state_satisfies_p (LOOP_CLOSED_SSA)))
prop_set_simulate_again (phi, false);
else
prop_set_simulate_again (phi, true);
if (!prop_simulate_again_p (phi))
set_copy_of_val (def, def);
else
cached_last_copy_of[SSA_NAME_VERSION (def)] = def;
}
}
}
static bool
gimple_find_edge_insert_loc (edge e, gimple_stmt_iterator *gsi,
basic_block *new_bb)
{
basic_block dest, src;
gimple *tmp;
dest = e->dest;
/* If the destination has one predecessor which has no PHI nodes,
insert there. Except for the exit block.
The requirement for no PHI nodes could be relaxed. Basically we
would have to examine the PHIs to prove that none of them used
the value set by the statement we want to insert on E. That
hardly seems worth the effort. */
restart:
if (single_pred_p (dest)
&& gimple_seq_empty_p (phi_nodes (dest))
&& dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
{
*gsi = gsi_start_bb (dest);
if (gsi_end_p (*gsi))
return true;
/* Make sure we insert after any leading labels. */
tmp = gsi_stmt (*gsi);
while (gimple_code (tmp) == GIMPLE_LABEL)
{
gsi_next (gsi);
if (gsi_end_p (*gsi))
break;
tmp = gsi_stmt (*gsi);
}
if (gsi_end_p (*gsi))
{
*gsi = gsi_last_bb (dest);
return true;
}
else
return false;
}
/* If the source has one successor, the edge is not abnormal and
the last statement does not end a basic block, insert there.
Except for the entry block. */
src = e->src;
if ((e->flags & EDGE_ABNORMAL) == 0
&& single_succ_p (src)
&& src != ENTRY_BLOCK_PTR_FOR_FN (cfun))
{
*gsi = gsi_last_bb (src);
if (gsi_end_p (*gsi))
return true;
tmp = gsi_stmt (*gsi);
if (!stmt_ends_bb_p (tmp))
return true;
switch (gimple_code (tmp))
{
case GIMPLE_RETURN:
case GIMPLE_RESX:
return false;
default:
break;
}
}
/* Otherwise, create a new basic block, and split this edge. */
dest = split_edge (e);
if (new_bb)
*new_bb = dest;
e = single_pred_edge (dest);
goto restart;
}
static bool
init_dont_simulate_again (void)
{
basic_block bb;
gimple_stmt_iterator gsi;
gimple phi;
bool saw_a_complex_op = false;
FOR_EACH_BB (bb)
{
for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
phi = gsi_stmt (gsi);
prop_set_simulate_again (phi,
is_complex_reg (gimple_phi_result (phi)));
}
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
gimple stmt;
tree op0, op1;
bool sim_again_p;
stmt = gsi_stmt (gsi);
op0 = op1 = NULL_TREE;
/* Most control-altering statements must be initially
simulated, else we won't cover the entire cfg. */
sim_again_p = stmt_ends_bb_p (stmt);
switch (gimple_code (stmt))
{
case GIMPLE_CALL:
if (gimple_call_lhs (stmt))
sim_again_p = is_complex_reg (gimple_call_lhs (stmt));
break;
case GIMPLE_ASSIGN:
sim_again_p = is_complex_reg (gimple_assign_lhs (stmt));
if (gimple_assign_rhs_code (stmt) == REALPART_EXPR
|| gimple_assign_rhs_code (stmt) == IMAGPART_EXPR)
op0 = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
else
op0 = gimple_assign_rhs1 (stmt);
if (gimple_num_ops (stmt) > 2)
op1 = gimple_assign_rhs2 (stmt);
break;
case GIMPLE_COND:
op0 = gimple_cond_lhs (stmt);
op1 = gimple_cond_rhs (stmt);
break;
default:
break;
}
if (op0 || op1)
switch (gimple_expr_code (stmt))
{
case EQ_EXPR:
case NE_EXPR:
case PLUS_EXPR:
case MINUS_EXPR:
case MULT_EXPR:
case TRUNC_DIV_EXPR:
case CEIL_DIV_EXPR:
case FLOOR_DIV_EXPR:
case ROUND_DIV_EXPR:
case RDIV_EXPR:
if (TREE_CODE (TREE_TYPE (op0)) == COMPLEX_TYPE
|| TREE_CODE (TREE_TYPE (op1)) == COMPLEX_TYPE)
saw_a_complex_op = true;
break;
case NEGATE_EXPR:
case CONJ_EXPR:
if (TREE_CODE (TREE_TYPE (op0)) == COMPLEX_TYPE)
saw_a_complex_op = true;
break;
case REALPART_EXPR:
case IMAGPART_EXPR:
/* The total store transformation performed during
gimplification creates such uninitialized loads
and we need to lower the statement to be able
to fix things up. */
if (TREE_CODE (op0) == SSA_NAME
&& ssa_undefined_value_p (op0))
saw_a_complex_op = true;
break;
default:
break;
}
prop_set_simulate_again (stmt, sim_again_p);
}
}
return saw_a_complex_op;
}
enum move_pos
movement_possibility (tree stmt)
{
tree lhs, rhs;
if (flag_unswitch_loops
&& TREE_CODE (stmt) == COND_EXPR)
{
/* If we perform unswitching, force the operands of the invariant
condition to be moved out of the loop. */
return MOVE_POSSIBLE;
}
if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return MOVE_IMPOSSIBLE;
if (stmt_ends_bb_p (stmt))
return MOVE_IMPOSSIBLE;
if (stmt_ann (stmt)->has_volatile_ops)
return MOVE_IMPOSSIBLE;
lhs = GIMPLE_STMT_OPERAND (stmt, 0);
if (TREE_CODE (lhs) == SSA_NAME
&& SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
return MOVE_IMPOSSIBLE;
rhs = GIMPLE_STMT_OPERAND (stmt, 1);
if (TREE_SIDE_EFFECTS (rhs)
|| tree_could_throw_p (rhs))
return MOVE_IMPOSSIBLE;
if (TREE_CODE (lhs) != SSA_NAME
|| tree_could_trap_p (rhs))
return MOVE_PRESERVE_EXECUTION;
if (get_call_expr_in (stmt))
{
/* While pure or const call is guaranteed to have no side effects, we
cannot move it arbitrarily. Consider code like
char *s = something ();
while (1)
{
if (s)
t = strlen (s);
else
t = 0;
}
Here the strlen call cannot be moved out of the loop, even though
s is invariant. In addition to possibly creating a call with
invalid arguments, moving out a function call that is not executed
may cause performance regressions in case the call is costly and
not executed at all. */
return MOVE_PRESERVE_EXECUTION;
}
return MOVE_POSSIBLE;
}
static bool
shrink_wrap_one_built_in_call (gimple bi_call)
{
gimple_stmt_iterator bi_call_bsi;
basic_block bi_call_bb, join_tgt_bb, guard_bb, guard_bb0;
edge join_tgt_in_edge_from_call, join_tgt_in_edge_fall_thru;
edge bi_call_in_edge0, guard_bb_in_edge;
unsigned tn_cond_stmts, nconds;
unsigned ci;
gimple cond_expr = NULL;
gimple cond_expr_start;
tree bi_call_label_decl;
gimple bi_call_label;
auto_vec<gimple, 12> conds;
gen_shrink_wrap_conditions (bi_call, conds, &nconds);
/* This can happen if the condition generator decides
it is not beneficial to do the transformation. Just
return false and do not do any transformation for
the call. */
if (nconds == 0)
return false;
bi_call_bb = gimple_bb (bi_call);
/* Now find the join target bb -- split bi_call_bb if needed. */
if (stmt_ends_bb_p (bi_call))
{
/* If the call must be the last in the bb, don't split the block,
it could e.g. have EH edges. */
join_tgt_in_edge_from_call = find_fallthru_edge (bi_call_bb->succs);
if (join_tgt_in_edge_from_call == NULL)
return false;
}
else
join_tgt_in_edge_from_call = split_block (bi_call_bb, bi_call);
bi_call_bsi = gsi_for_stmt (bi_call);
join_tgt_bb = join_tgt_in_edge_from_call->dest;
/* Now it is time to insert the first conditional expression
into bi_call_bb and split this bb so that bi_call is
shrink-wrapped. */
tn_cond_stmts = conds.length ();
cond_expr = NULL;
cond_expr_start = conds[0];
for (ci = 0; ci < tn_cond_stmts; ci++)
{
gimple c = conds[ci];
gcc_assert (c || ci != 0);
if (!c)
break;
gsi_insert_before (&bi_call_bsi, c, GSI_SAME_STMT);
cond_expr = c;
}
nconds--;
ci++;
gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND);
/* Now the label. */
bi_call_label_decl = create_artificial_label (gimple_location (bi_call));
bi_call_label = gimple_build_label (bi_call_label_decl);
gsi_insert_before (&bi_call_bsi, bi_call_label, GSI_SAME_STMT);
bi_call_in_edge0 = split_block (bi_call_bb, cond_expr);
bi_call_in_edge0->flags &= ~EDGE_FALLTHRU;
bi_call_in_edge0->flags |= EDGE_TRUE_VALUE;
guard_bb0 = bi_call_bb;
bi_call_bb = bi_call_in_edge0->dest;
join_tgt_in_edge_fall_thru = make_edge (guard_bb0, join_tgt_bb,
EDGE_FALSE_VALUE);
bi_call_in_edge0->probability = REG_BR_PROB_BASE * ERR_PROB;
bi_call_in_edge0->count =
apply_probability (guard_bb0->count,
bi_call_in_edge0->probability);
join_tgt_in_edge_fall_thru->probability =
inverse_probability (bi_call_in_edge0->probability);
join_tgt_in_edge_fall_thru->count =
guard_bb0->count - bi_call_in_edge0->count;
/* Code generation for the rest of the conditions */
guard_bb = guard_bb0;
while (nconds > 0)
{
unsigned ci0;
edge bi_call_in_edge;
gimple_stmt_iterator guard_bsi = gsi_for_stmt (cond_expr_start);
ci0 = ci;
cond_expr_start = conds[ci0];
for (; ci < tn_cond_stmts; ci++)
{
gimple c = conds[ci];
gcc_assert (c || ci != ci0);
if (!c)
break;
gsi_insert_before (&guard_bsi, c, GSI_SAME_STMT);
cond_expr = c;
}
nconds--;
ci++;
gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND);
guard_bb_in_edge = split_block (guard_bb, cond_expr);
guard_bb_in_edge->flags &= ~EDGE_FALLTHRU;
guard_bb_in_edge->flags |= EDGE_FALSE_VALUE;
bi_call_in_edge = make_edge (guard_bb, bi_call_bb, EDGE_TRUE_VALUE);
bi_call_in_edge->probability = REG_BR_PROB_BASE * ERR_PROB;
bi_call_in_edge->count =
apply_probability (guard_bb->count,
bi_call_in_edge->probability);
guard_bb_in_edge->probability =
inverse_probability (bi_call_in_edge->probability);
guard_bb_in_edge->count = guard_bb->count - bi_call_in_edge->count;
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
location_t loc;
loc = gimple_location (bi_call);
fprintf (dump_file,
"%s:%d: note: function call is shrink-wrapped"
" into error conditions.\n",
LOCATION_FILE (loc), LOCATION_LINE (loc));
}
return true;
}
static void
use_internal_fn (gcall *call)
{
/* We'll be inserting another call with the same arguments after the
lhs has been set, so prevent any possible coalescing failure from
having both values live at once. See PR 71020. */
replace_abnormal_ssa_names (call);
unsigned nconds = 0;
auto_vec<gimple *, 12> conds;
if (can_test_argument_range (call))
{
gen_shrink_wrap_conditions (call, conds, &nconds);
gcc_assert (nconds != 0);
}
else
gcc_assert (edom_only_function (call));
internal_fn ifn = replacement_internal_fn (call);
gcc_assert (ifn != IFN_LAST);
/* Construct the new call, with the same arguments as the original one. */
auto_vec <tree, 16> args;
unsigned int nargs = gimple_call_num_args (call);
for (unsigned int i = 0; i < nargs; ++i)
args.safe_push (gimple_call_arg (call, i));
gcall *new_call = gimple_build_call_internal_vec (ifn, args);
gimple_set_location (new_call, gimple_location (call));
gimple_call_set_nothrow (new_call, gimple_call_nothrow_p (call));
/* Transfer the LHS to the new call. */
tree lhs = gimple_call_lhs (call);
gimple_call_set_lhs (new_call, lhs);
gimple_call_set_lhs (call, NULL_TREE);
SSA_NAME_DEF_STMT (lhs) = new_call;
/* Insert the new call. */
gimple_stmt_iterator gsi = gsi_for_stmt (call);
gsi_insert_before (&gsi, new_call, GSI_SAME_STMT);
if (nconds == 0)
{
/* Skip the call if LHS == LHS. If we reach here, EDOM is the only
valid errno value and it is used iff the result is NaN. */
conds.quick_push (gimple_build_cond (EQ_EXPR, lhs, lhs,
NULL_TREE, NULL_TREE));
nconds++;
/* Try replacing the original call with a direct assignment to
errno, via an internal function. */
if (set_edom_supported_p () && !stmt_ends_bb_p (call))
{
gimple_stmt_iterator gsi = gsi_for_stmt (call);
gcall *new_call = gimple_build_call_internal (IFN_SET_EDOM, 0);
gimple_set_vuse (new_call, gimple_vuse (call));
gimple_set_vdef (new_call, gimple_vdef (call));
SSA_NAME_DEF_STMT (gimple_vdef (new_call)) = new_call;
gimple_set_location (new_call, gimple_location (call));
gsi_replace (&gsi, new_call, false);
call = new_call;
}
}
shrink_wrap_one_built_in_call_with_conds (call, conds, nconds);
}
static void
shrink_wrap_one_built_in_call_with_conds (gcall *bi_call, vec <gimple *> conds,
unsigned int nconds)
{
gimple_stmt_iterator bi_call_bsi;
basic_block bi_call_bb, join_tgt_bb, guard_bb;
edge join_tgt_in_edge_from_call, join_tgt_in_edge_fall_thru;
edge bi_call_in_edge0, guard_bb_in_edge;
unsigned tn_cond_stmts;
unsigned ci;
gimple *cond_expr = NULL;
gimple *cond_expr_start;
/* The cfg we want to create looks like this:
[guard n-1] <- guard_bb (old block)
| \
| [guard n-2] }
| / \ }
| / ... } new blocks
| / [guard 0] }
| / / | }
[ call ] | <- bi_call_bb }
| \ |
| \ |
| [ join ] <- join_tgt_bb (old iff call must end bb)
|
possible EH edges (only if [join] is old)
When [join] is new, the immediate dominators for these blocks are:
1. [guard n-1]: unchanged
2. [call]: [guard n-1]
3. [guard m]: [guard m+1] for 0 <= m <= n-2
4. [join]: [guard n-1]
We punt for the more complex case case of [join] being old and
simply free the dominance info. We also punt on postdominators,
which aren't expected to be available at this point anyway. */
bi_call_bb = gimple_bb (bi_call);
/* Now find the join target bb -- split bi_call_bb if needed. */
if (stmt_ends_bb_p (bi_call))
{
/* We checked that there was a fallthrough edge in
can_guard_call_p. */
join_tgt_in_edge_from_call = find_fallthru_edge (bi_call_bb->succs);
gcc_assert (join_tgt_in_edge_from_call);
/* We don't want to handle PHIs. */
if (EDGE_COUNT (join_tgt_in_edge_from_call->dest->preds) > 1)
join_tgt_bb = split_edge (join_tgt_in_edge_from_call);
else
{
join_tgt_bb = join_tgt_in_edge_from_call->dest;
/* We may have degenerate PHIs in the destination. Propagate
those out. */
for (gphi_iterator i = gsi_start_phis (join_tgt_bb); !gsi_end_p (i);)
{
gphi *phi = i.phi ();
replace_uses_by (gimple_phi_result (phi),
gimple_phi_arg_def (phi, 0));
remove_phi_node (&i, true);
}
}
}
else
{
join_tgt_in_edge_from_call = split_block (bi_call_bb, bi_call);
join_tgt_bb = join_tgt_in_edge_from_call->dest;
}
bi_call_bsi = gsi_for_stmt (bi_call);
/* Now it is time to insert the first conditional expression
into bi_call_bb and split this bb so that bi_call is
shrink-wrapped. */
tn_cond_stmts = conds.length ();
cond_expr = NULL;
cond_expr_start = conds[0];
for (ci = 0; ci < tn_cond_stmts; ci++)
{
gimple *c = conds[ci];
gcc_assert (c || ci != 0);
if (!c)
break;
gsi_insert_before (&bi_call_bsi, c, GSI_SAME_STMT);
cond_expr = c;
}
ci++;
gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND);
typedef std::pair<edge, edge> edge_pair;
auto_vec<edge_pair, 8> edges;
bi_call_in_edge0 = split_block (bi_call_bb, cond_expr);
bi_call_in_edge0->flags &= ~EDGE_FALLTHRU;
bi_call_in_edge0->flags |= EDGE_FALSE_VALUE;
guard_bb = bi_call_bb;
bi_call_bb = bi_call_in_edge0->dest;
join_tgt_in_edge_fall_thru = make_edge (guard_bb, join_tgt_bb,
EDGE_TRUE_VALUE);
edges.reserve (nconds);
edges.quick_push (edge_pair (bi_call_in_edge0, join_tgt_in_edge_fall_thru));
/* Code generation for the rest of the conditions */
for (unsigned int i = 1; i < nconds; ++i)
{
unsigned ci0;
edge bi_call_in_edge;
gimple_stmt_iterator guard_bsi = gsi_for_stmt (cond_expr_start);
ci0 = ci;
cond_expr_start = conds[ci0];
for (; ci < tn_cond_stmts; ci++)
{
gimple *c = conds[ci];
gcc_assert (c || ci != ci0);
if (!c)
break;
gsi_insert_before (&guard_bsi, c, GSI_SAME_STMT);
cond_expr = c;
}
ci++;
gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND);
guard_bb_in_edge = split_block (guard_bb, cond_expr);
guard_bb_in_edge->flags &= ~EDGE_FALLTHRU;
guard_bb_in_edge->flags |= EDGE_TRUE_VALUE;
bi_call_in_edge = make_edge (guard_bb, bi_call_bb, EDGE_FALSE_VALUE);
edges.quick_push (edge_pair (bi_call_in_edge, guard_bb_in_edge));
}
/* Now update the probability and profile information, processing the
guards in order of execution.
There are two approaches we could take here. On the one hand we
could assign a probability of X to the call block and distribute
that probability among its incoming edges. On the other hand we
could assign a probability of X to each individual call edge.
The choice only affects calls that have more than one condition.
In those cases, the second approach would give the call block
a greater probability than the first. However, the difference
is only small, and our chosen X is a pure guess anyway.
Here we take the second approach because it's slightly simpler
and because it's easy to see that it doesn't lose profile counts. */
bi_call_bb->count = profile_count::zero ();
while (!edges.is_empty ())
{
edge_pair e = edges.pop ();
edge call_edge = e.first;
edge nocall_edge = e.second;
basic_block src_bb = call_edge->src;
gcc_assert (src_bb == nocall_edge->src);
call_edge->probability = profile_probability::very_unlikely ();
nocall_edge->probability = profile_probability::always ()
- call_edge->probability;
bi_call_bb->count += call_edge->count ();
if (nocall_edge->dest != join_tgt_bb)
nocall_edge->dest->count = src_bb->count - bi_call_bb->count;
}
if (dom_info_available_p (CDI_DOMINATORS))
{
/* The split_blocks leave [guard 0] as the immediate dominator
of [call] and [call] as the immediate dominator of [join].
Fix them up. */
set_immediate_dominator (CDI_DOMINATORS, bi_call_bb, guard_bb);
set_immediate_dominator (CDI_DOMINATORS, join_tgt_bb, guard_bb);
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
location_t loc;
loc = gimple_location (bi_call);
fprintf (dump_file,
"%s:%d: note: function call is shrink-wrapped"
" into error conditions.\n",
LOCATION_FILE (loc), LOCATION_LINE (loc));
}
}
static bool
can_guard_call_p (gimple *call)
{
return (!stmt_ends_bb_p (call)
|| find_fallthru_edge (gimple_bb (call)->succs));
}
static bool
init_dont_simulate_again (void)
{
basic_block bb;
block_stmt_iterator bsi;
tree phi;
bool saw_a_complex_op = false;
FOR_EACH_BB (bb)
{
for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
DONT_SIMULATE_AGAIN (phi) = !is_complex_reg (PHI_RESULT (phi));
for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
{
tree orig_stmt, stmt, rhs = NULL;
bool dsa;
orig_stmt = stmt = bsi_stmt (bsi);
/* Most control-altering statements must be initially
simulated, else we won't cover the entire cfg. */
dsa = !stmt_ends_bb_p (stmt);
switch (TREE_CODE (stmt))
{
case RETURN_EXPR:
/* We don't care what the lattice value of <retval> is,
since it's never used as an input to another computation. */
dsa = true;
stmt = TREE_OPERAND (stmt, 0);
if (!stmt || TREE_CODE (stmt) != MODIFY_EXPR)
break;
/* FALLTHRU */
case MODIFY_EXPR:
dsa = !is_complex_reg (TREE_OPERAND (stmt, 0));
rhs = TREE_OPERAND (stmt, 1);
break;
case COND_EXPR:
rhs = TREE_OPERAND (stmt, 0);
break;
default:
break;
}
if (rhs)
switch (TREE_CODE (rhs))
{
case EQ_EXPR:
case NE_EXPR:
rhs = TREE_OPERAND (rhs, 0);
/* FALLTHRU */
case PLUS_EXPR:
case MINUS_EXPR:
case MULT_EXPR:
case TRUNC_DIV_EXPR:
case CEIL_DIV_EXPR:
case FLOOR_DIV_EXPR:
case ROUND_DIV_EXPR:
case RDIV_EXPR:
case NEGATE_EXPR:
case CONJ_EXPR:
if (TREE_CODE (TREE_TYPE (rhs)) == COMPLEX_TYPE)
saw_a_complex_op = true;
break;
default:
break;
}
DONT_SIMULATE_AGAIN (orig_stmt) = dsa;
}
}
return saw_a_complex_op;
}