static void
tree_ssa_forward_propagate_single_use_vars (void)
{
basic_block bb;
varray_type vars_worklist, cond_worklist;
vars = BITMAP_ALLOC (NULL);
VARRAY_TREE_INIT (vars_worklist, 10, "VARS worklist");
VARRAY_TREE_INIT (cond_worklist, 10, "COND worklist");
/* Prime the COND_EXPR worklist by placing all the COND_EXPRs on the
worklist. */
FOR_EACH_BB (bb)
{
tree last = last_stmt (bb);
if (last && TREE_CODE (last) == COND_EXPR)
VARRAY_PUSH_TREE (cond_worklist, last);
}
while (VARRAY_ACTIVE_SIZE (cond_worklist) > 0)
{
/* First get a list of all the interesting COND_EXPRs and potential
single use variables which feed those COND_EXPRs. This will drain
COND_WORKLIST and initialize VARS_WORKLIST. */
record_single_argument_cond_exprs (cond_worklist, &vars_worklist, vars);
if (VARRAY_ACTIVE_SIZE (vars_worklist) > 0)
{
/* Now compute immediate uses for all the variables we care about. */
compute_immediate_uses (TDFA_USE_OPS, need_imm_uses_for);
/* We've computed immediate uses, so we can/must clear the VARS
bitmap for the next iteration. */
bitmap_clear (vars);
/* And optimize. This will drain VARS_WORKLIST and initialize
COND_WORKLIST for the next iteration. */
substitute_single_use_vars (&cond_worklist, vars_worklist);
/* We do not incrementally update the dataflow information
so we must free it here and recompute the necessary bits
on the next iteration. If this turns out to be expensive,
methods for incrementally updating the dataflow are known. */
free_df ();
}
}
/* All done. Clean up. */
BITMAP_FREE (vars);
}
/* Return source file of the statement specified by LOC. */
const char *
locator_file (int loc)
{
int max = VARRAY_ACTIVE_SIZE (file_locators_locs);
int min = 0;
if (!max || !loc)
return NULL;
while (1)
{
int pos = (min + max) / 2;
int tmp = VARRAY_INT (file_locators_locs, pos);
if (tmp <= loc && min != pos)
min = pos;
else if (tmp > loc && max != pos)
max = pos;
else
{
min = pos;
break;
}
}
return VARRAY_CHAR_PTR (file_locators_files, min);
}
/* Return line number of the statement specified by the locator. */
int
locator_line (int loc)
{
int max = VARRAY_ACTIVE_SIZE (line_locators_locs);
int min = 0;
if (!max || !loc)
return 0;
while (1)
{
int pos = (min + max) / 2;
int tmp = VARRAY_INT (line_locators_locs, pos);
if (tmp <= loc && min != pos)
min = pos;
else if (tmp > loc && max != pos)
max = pos;
else
{
min = pos;
break;
}
}
return VARRAY_INT (line_locators_lines, min);
}
/* Return lexical scope block insn belong to. */
static tree
insn_scope (rtx insn)
{
int max = VARRAY_ACTIVE_SIZE (block_locators_locs);
int min = 0;
int loc = INSN_LOCATOR (insn);
/* When block_locators_locs was initialized, the pro- and epilogue
insns didn't exist yet and can therefore not be found this way.
But we know that they belong to the outer most block of the
current function.
Without this test, the prologue would be put inside the block of
the first valid instruction in the function and when that first
insn is part of an inlined function then the low_pc of that
inlined function is messed up. Likewise for the epilogue and
the last valid instruction. */
if (loc == prologue_locator || loc == epilogue_locator)
return DECL_INITIAL (cfun->decl);
if (!max || !loc)
return NULL;
while (1)
{
int pos = (min + max) / 2;
int tmp = VARRAY_INT (block_locators_locs, pos);
if (tmp <= loc && min != pos)
min = pos;
else if (tmp > loc && max != pos)
max = pos;
else
{
min = pos;
break;
}
}
return VARRAY_TREE (block_locators_blocks, min);
}
static void
gather_interchange_stats (varray_type dependence_relations,
varray_type datarefs,
struct loop *loop,
struct loop *first_loop,
unsigned int *dependence_steps,
unsigned int *nb_deps_not_carried_by_loop,
unsigned int *access_strides)
{
unsigned int i;
*dependence_steps = 0;
*nb_deps_not_carried_by_loop = 0;
*access_strides = 0;
for (i = 0; i < VARRAY_ACTIVE_SIZE (dependence_relations); i++)
{
int dist;
struct data_dependence_relation *ddr =
(struct data_dependence_relation *)
VARRAY_GENERIC_PTR (dependence_relations, i);
/* If we don't know anything about this dependence, or the distance
vector is NULL, or there is no dependence, then there is no reuse of
data. */
if (DDR_DIST_VECT (ddr) == NULL
|| DDR_ARE_DEPENDENT (ddr) == chrec_dont_know
|| DDR_ARE_DEPENDENT (ddr) == chrec_known)
continue;
dist = DDR_DIST_VECT (ddr)[loop->depth - first_loop->depth];
if (dist == 0)
(*nb_deps_not_carried_by_loop) += 1;
else if (dist < 0)
(*dependence_steps) += -dist;
else
(*dependence_steps) += dist;
}
/* Compute the access strides. */
for (i = 0; i < VARRAY_ACTIVE_SIZE (datarefs); i++)
{
unsigned int it;
struct data_reference *dr = VARRAY_GENERIC_PTR (datarefs, i);
tree stmt = DR_STMT (dr);
struct loop *stmt_loop = loop_containing_stmt (stmt);
struct loop *inner_loop = first_loop->inner;
if (inner_loop != stmt_loop
&& !flow_loop_nested_p (inner_loop, stmt_loop))
continue;
for (it = 0; it < DR_NUM_DIMENSIONS (dr); it++)
{
tree chrec = DR_ACCESS_FN (dr, it);
tree tstride = evolution_part_in_loop_num
(chrec, loop->num);
if (tstride == NULL_TREE
|| TREE_CODE (tstride) != INTEGER_CST)
continue;
(*access_strides) += int_cst_value (tstride);
}
}
}
void
linear_transform_loops (struct loops *loops)
{
unsigned int i;
compute_immediate_uses (TDFA_USE_OPS | TDFA_USE_VOPS, NULL);
for (i = 1; i < loops->num; i++)
{
unsigned int depth = 0;
varray_type datarefs;
varray_type dependence_relations;
struct loop *loop_nest = loops->parray[i];
struct loop *temp;
VEC (tree) *oldivs = NULL;
VEC (tree) *invariants = NULL;
lambda_loopnest before, after;
lambda_trans_matrix trans;
bool problem = false;
bool need_perfect_nest = false;
/* If it's not a loop nest, we don't want it.
We also don't handle sibling loops properly,
which are loops of the following form:
for (i = 0; i < 50; i++)
{
for (j = 0; j < 50; j++)
{
...
}
for (j = 0; j < 50; j++)
{
...
}
} */
if (!loop_nest->inner)
continue;
depth = 1;
for (temp = loop_nest->inner; temp; temp = temp->inner)
{
flow_loop_scan (temp, LOOP_ALL);
/* If we have a sibling loop or multiple exit edges, jump ship. */
if (temp->next || temp->num_exits != 1)
{
problem = true;
break;
}
depth ++;
}
if (problem)
continue;
/* Analyze data references and dependence relations using scev. */
VARRAY_GENERIC_PTR_INIT (datarefs, 10, "datarefs");
VARRAY_GENERIC_PTR_INIT (dependence_relations, 10,
"dependence_relations");
compute_data_dependences_for_loop (depth, loop_nest,
&datarefs, &dependence_relations);
if (dump_file && (dump_flags & TDF_DETAILS))
{
unsigned int j;
for (j = 0; j < VARRAY_ACTIVE_SIZE (dependence_relations); j++)
{
struct data_dependence_relation *ddr =
(struct data_dependence_relation *)
VARRAY_GENERIC_PTR (dependence_relations, j);
if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
{
fprintf (dump_file, "DISTANCE_V (");
print_lambda_vector (dump_file, DDR_DIST_VECT (ddr),
DDR_SIZE_VECT (ddr));
fprintf (dump_file, ")\n");
fprintf (dump_file, "DIRECTION_V (");
print_lambda_vector (dump_file, DDR_DIR_VECT (ddr),
DDR_SIZE_VECT (ddr));
fprintf (dump_file, ")\n");
}
}
fprintf (dump_file, "\n\n");
}
/* Build the transformation matrix. */
trans = lambda_trans_matrix_new (depth, depth);
lambda_matrix_id (LTM_MATRIX (trans), depth);
trans = try_interchange_loops (trans, depth, dependence_relations,
datarefs, loop_nest);
if (lambda_trans_matrix_id_p (trans))
{
if (dump_file)
fprintf (dump_file, "Won't transform loop. Optimal transform is the identity transform\n");
continue;
}
/* Check whether the transformation is legal. */
if (!lambda_transform_legal_p (trans, depth, dependence_relations))
{
if (dump_file)
fprintf (dump_file, "Can't transform loop, transform is illegal:\n");
continue;
}
if (!perfect_nest_p (loop_nest))
need_perfect_nest = true;
before = gcc_loopnest_to_lambda_loopnest (loops,
loop_nest, &oldivs,
&invariants,
need_perfect_nest);
if (!before)
continue;
if (dump_file)
{
fprintf (dump_file, "Before:\n");
print_lambda_loopnest (dump_file, before, 'i');
}
after = lambda_loopnest_transform (before, trans);
if (dump_file)
{
fprintf (dump_file, "After:\n");
print_lambda_loopnest (dump_file, after, 'u');
}
lambda_loopnest_to_gcc_loopnest (loop_nest, oldivs, invariants,
after, trans);
if (dump_file)
fprintf (dump_file, "Successfully transformed loop.\n");
oldivs = NULL;
invariants = NULL;
free_dependence_relations (dependence_relations);
free_data_refs (datarefs);
}
free_df ();
scev_reset ();
rewrite_into_loop_closed_ssa ();
#ifdef ENABLE_CHECKING
verify_loop_closed_ssa ();
#endif
}
static void
substitute_single_use_vars (varray_type *cond_worklist,
varray_type vars_worklist)
{
while (VARRAY_ACTIVE_SIZE (vars_worklist) > 0)
{
tree test_var = VARRAY_TOP_TREE (vars_worklist);
tree def = SSA_NAME_DEF_STMT (test_var);
dataflow_t df;
int j, num_uses, propagated_uses;
VARRAY_POP (vars_worklist);
/* Now compute the immediate uses of TEST_VAR. */
df = get_immediate_uses (def);
num_uses = num_immediate_uses (df);
propagated_uses = 0;
/* If TEST_VAR is used more than once and is not a boolean set
via TRUTH_NOT_EXPR with another SSA_NAME as its argument, then
we can not optimize. */
if (num_uses == 1
|| (TREE_CODE (TREE_TYPE (test_var)) == BOOLEAN_TYPE
&& TREE_CODE (TREE_OPERAND (def, 1)) == TRUTH_NOT_EXPR
&& (TREE_CODE (TREE_OPERAND (TREE_OPERAND (def, 1), 0))
== SSA_NAME)))
;
else
continue;
/* Walk over each use and try to forward propagate the RHS of
DEF into the use. */
for (j = 0; j < num_uses; j++)
{
tree cond_stmt;
tree cond;
enum tree_code cond_code;
tree def_rhs;
enum tree_code def_rhs_code;
tree new_cond;
cond_stmt = immediate_use (df, j);
/* For now we can only propagate into COND_EXPRs. */
if (TREE_CODE (cond_stmt) != COND_EXPR)
continue;
cond = COND_EXPR_COND (cond_stmt);
cond_code = TREE_CODE (cond);
def_rhs = TREE_OPERAND (def, 1);
def_rhs_code = TREE_CODE (def_rhs);
/* If the definition of the single use variable was from an
arithmetic operation, then we just need to adjust the
constant in the COND_EXPR_COND and update the variable tested. */
if (def_rhs_code == PLUS_EXPR || def_rhs_code == MINUS_EXPR)
{
tree op0 = TREE_OPERAND (def_rhs, 0);
tree op1 = TREE_OPERAND (def_rhs, 1);
enum tree_code new_code;
tree t;
/* If the variable was defined via X + C, then we must subtract
C from the constant in the conditional. Otherwise we add
C to the constant in the conditional. The result must fold
into a valid gimple operand to be optimizable. */
new_code = def_rhs_code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR;
t = int_const_binop (new_code, TREE_OPERAND (cond, 1), op1, 0);
if (!is_gimple_val (t))
continue;
new_cond = build (cond_code, boolean_type_node, op0, t);
}
/* If the variable is defined by a conditional expression... */
else if (TREE_CODE_CLASS (def_rhs_code) == tcc_comparison)
{
/* TEST_VAR was set from a relational operator. */
tree op0 = TREE_OPERAND (def_rhs, 0);
tree op1 = TREE_OPERAND (def_rhs, 1);
new_cond = build (def_rhs_code, boolean_type_node, op0, op1);
/* Invert the conditional if necessary. */
if ((cond_code == EQ_EXPR
&& integer_zerop (TREE_OPERAND (cond, 1)))
|| (cond_code == NE_EXPR
&& integer_onep (TREE_OPERAND (cond, 1))))
{
new_cond = invert_truthvalue (new_cond);
/* If we did not get a simple relational expression or
bare SSA_NAME, then we can not optimize this case. */
if (!COMPARISON_CLASS_P (new_cond)
&& TREE_CODE (new_cond) != SSA_NAME)
continue;
}
}
else
{
bool invert = false;
enum tree_code new_code;
tree new_arg;
/* TEST_VAR was set from a TRUTH_NOT_EXPR or a NOP_EXPR. */
if (def_rhs_code == TRUTH_NOT_EXPR)
invert = true;
/* If we don't have <NE_EXPR/EQ_EXPR x INT_CST>, then we cannot
optimize this case. */
if ((cond_code == NE_EXPR || cond_code == EQ_EXPR)
&& TREE_CODE (TREE_OPERAND (cond, 1)) != INTEGER_CST)
continue;
if (cond_code == SSA_NAME
|| (cond_code == NE_EXPR
&& integer_zerop (TREE_OPERAND (cond, 1)))
|| (cond_code == EQ_EXPR
&& integer_onep (TREE_OPERAND (cond, 1))))
new_code = NE_EXPR;
else
new_code = EQ_EXPR;
if (invert)
new_code = (new_code == EQ_EXPR ? NE_EXPR : EQ_EXPR);
new_arg = TREE_OPERAND (def_rhs, 0);
new_cond = build2 (new_code, boolean_type_node, new_arg,
fold_convert (TREE_TYPE (new_arg),
integer_zero_node));
}
/* Dump details. */
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, " Replaced '");
print_generic_expr (dump_file, cond, dump_flags);
fprintf (dump_file, "' with '");
print_generic_expr (dump_file, new_cond, dump_flags);
fprintf (dump_file, "'\n");
}
/* Replace the condition. */
COND_EXPR_COND (cond_stmt) = new_cond;
modify_stmt (cond_stmt);
propagated_uses++;
VARRAY_PUSH_TREE (*cond_worklist, cond_stmt);
}
/* If we propagated into all the uses, then we can delete DEF.
Unfortunately, we have to find the defining statement in
whatever block it might be in. */
if (num_uses && num_uses == propagated_uses)
{
block_stmt_iterator bsi = bsi_for_stmt (def);
bsi_remove (&bsi);
}
}
}
static void
record_single_argument_cond_exprs (varray_type cond_worklist,
varray_type *vars_worklist,
bitmap vars)
{
/* The first pass over the blocks gathers the set of variables we need
immediate uses for as well as the set of interesting COND_EXPRs.
A simpler implementation may be appropriate if/when we have a lower
overhead means of getting immediate use information. */
while (VARRAY_ACTIVE_SIZE (cond_worklist) > 0)
{
tree last = VARRAY_TOP_TREE (cond_worklist);
VARRAY_POP (cond_worklist);
/* See if this block ends in a COND_EXPR. */
if (last && TREE_CODE (last) == COND_EXPR)
{
tree cond = COND_EXPR_COND (last);
enum tree_code cond_code = TREE_CODE (cond);
/* If the condition is a lone variable or an equality test of
an SSA_NAME against an integral constant, then we may have an
optimizable case.
Note these conditions also ensure the COND_EXPR has no
virtual operands or other side effects. */
if (cond_code == SSA_NAME
|| ((cond_code == EQ_EXPR || cond_code == NE_EXPR)
&& TREE_CODE (TREE_OPERAND (cond, 0)) == SSA_NAME
&& CONSTANT_CLASS_P (TREE_OPERAND (cond, 1))
&& INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (cond, 1)))))
{
tree def;
tree test_var;
/* Extract the single variable used in the test into TEST_VAR. */
if (cond_code == SSA_NAME)
test_var = cond;
else
test_var = TREE_OPERAND (cond, 0);
/* If we have already recorded this SSA_NAME as interesting,
do not do so again. */
if (bitmap_bit_p (vars, SSA_NAME_VERSION (test_var)))
continue;
/* Now get the defining statement for TEST_VAR and see if it
something we are interested in. */
def = SSA_NAME_DEF_STMT (test_var);
if (TREE_CODE (def) == MODIFY_EXPR)
{
tree def_rhs = TREE_OPERAND (def, 1);
/* If TEST_VAR is set by adding or subtracting a constant
from an SSA_NAME, then it is interesting to us as we
can adjust the constant in the conditional and thus
eliminate the arithmetic operation. */
if (TREE_CODE (def_rhs) == PLUS_EXPR
|| TREE_CODE (def_rhs) == MINUS_EXPR)
{
tree op0 = TREE_OPERAND (def_rhs, 0);
tree op1 = TREE_OPERAND (def_rhs, 1);
/* The first operand must be an SSA_NAME and the second
operand must be a constant. */
if (TREE_CODE (op0) != SSA_NAME
|| !CONSTANT_CLASS_P (op1)
|| !INTEGRAL_TYPE_P (TREE_TYPE (op1)))
continue;
/* Don't propagate if the first operand occurs in
an abnormal PHI. */
if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0))
continue;
}
/* These cases require comparisons of a naked SSA_NAME or
comparison of an SSA_NAME against zero or one. */
else if (TREE_CODE (cond) == SSA_NAME
|| integer_zerop (TREE_OPERAND (cond, 1))
|| integer_onep (TREE_OPERAND (cond, 1)))
{
/* If TEST_VAR is set from a relational operation
between two SSA_NAMEs or a combination of an SSA_NAME
and a constant, then it is interesting. */
if (COMPARISON_CLASS_P (def_rhs))
{
tree op0 = TREE_OPERAND (def_rhs, 0);
tree op1 = TREE_OPERAND (def_rhs, 1);
/* Both operands of DEF_RHS must be SSA_NAMEs or
constants. */
if ((TREE_CODE (op0) != SSA_NAME
&& !is_gimple_min_invariant (op0))
|| (TREE_CODE (op1) != SSA_NAME
&& !is_gimple_min_invariant (op1)))
continue;
/* Don't propagate if the first operand occurs in
an abnormal PHI. */
if (TREE_CODE (op0) == SSA_NAME
&& SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0))
continue;
/* Don't propagate if the second operand occurs in
an abnormal PHI. */
if (TREE_CODE (op1) == SSA_NAME
&& SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op1))
continue;
}
/* If TEST_VAR is set from a TRUTH_NOT_EXPR, then it
is interesting. */
else if (TREE_CODE (def_rhs) == TRUTH_NOT_EXPR)
{
def_rhs = TREE_OPERAND (def_rhs, 0);
/* DEF_RHS must be an SSA_NAME or constant. */
if (TREE_CODE (def_rhs) != SSA_NAME
&& !is_gimple_min_invariant (def_rhs))
continue;
/* Don't propagate if the operand occurs in
an abnormal PHI. */
if (TREE_CODE (def_rhs) == SSA_NAME
&& SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_rhs))
continue;
}
/* If TEST_VAR was set from a cast of an integer type
to a boolean type or a cast of a boolean to an
integral, then it is interesting. */
else if (TREE_CODE (def_rhs) == NOP_EXPR
|| TREE_CODE (def_rhs) == CONVERT_EXPR)
{
tree outer_type;
tree inner_type;
outer_type = TREE_TYPE (def_rhs);
inner_type = TREE_TYPE (TREE_OPERAND (def_rhs, 0));
if ((TREE_CODE (outer_type) == BOOLEAN_TYPE
&& INTEGRAL_TYPE_P (inner_type))
|| (TREE_CODE (inner_type) == BOOLEAN_TYPE
&& INTEGRAL_TYPE_P (outer_type)))
;
else
continue;
/* Don't propagate if the operand occurs in
an abnormal PHI. */
if (TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME
&& SSA_NAME_OCCURS_IN_ABNORMAL_PHI (TREE_OPERAND
(def_rhs, 0)))
continue;
}
else
continue;
}
else
continue;
/* All the tests passed, record TEST_VAR as interesting. */
VARRAY_PUSH_TREE (*vars_worklist, test_var);
bitmap_set_bit (vars, SSA_NAME_VERSION (test_var));
}
}
}
}
}
