/* Induction variable optimizations.
- Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc.
+ Copyright (C) 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
This file is part of GCC.
#include "ggc.h"
#include "insn-config.h"
#include "recog.h"
+#include "pointer-set.h"
#include "hashtab.h"
#include "tree-chrec.h"
#include "tree-scalar-evolution.h"
#include "cfgloop.h"
#include "params.h"
#include "langhooks.h"
+#include "tree-affine.h"
+#include "target.h"
/* The infinite cost. */
#define INFTY 10000000
unsigned regs_used;
/* Numbers of iterations for all exits of the current loop. */
- htab_t niters;
+ struct pointer_map_t *niters;
/* The size of version_info array allocated. */
unsigned version_info_size;
edge
single_dom_exit (struct loop *loop)
{
- edge exit = loop->single_exit;
+ edge exit = single_exit (loop);
if (!exit)
return NULL;
return ver_info (data, SSA_NAME_VERSION (name));
}
-/* Checks whether there exists number X such that X * B = A, counting modulo
- 2^BITS. */
-
-static bool
-divide (unsigned bits, unsigned HOST_WIDE_INT a, unsigned HOST_WIDE_INT b,
- HOST_WIDE_INT *x)
-{
- unsigned HOST_WIDE_INT mask = ~(~(unsigned HOST_WIDE_INT) 0 << (bits - 1) << 1);
- unsigned HOST_WIDE_INT inv, ex, val;
- unsigned i;
-
- a &= mask;
- b &= mask;
-
- /* First divide the whole equation by 2 as long as possible. */
- while (!(a & 1) && !(b & 1))
- {
- a >>= 1;
- b >>= 1;
- bits--;
- mask >>= 1;
- }
-
- if (!(b & 1))
- {
- /* If b is still even, a is odd and there is no such x. */
- return false;
- }
-
- /* Find the inverse of b. We compute it as
- b^(2^(bits - 1) - 1) (mod 2^bits). */
- inv = 1;
- ex = b;
- for (i = 0; i < bits - 1; i++)
- {
- inv = (inv * ex) & mask;
- ex = (ex * ex) & mask;
- }
-
- val = (a * inv) & mask;
-
- gcc_assert (((val * b) & mask) == a);
-
- if ((val >> (bits - 1)) & 1)
- val |= ~mask;
-
- *x = val;
-
- return true;
-}
-
/* Returns true if STMT is after the place where the IP_NORMAL ivs will be
emitted in LOOP. */
contains_abnormal_ssa_name_p (tree expr)
{
enum tree_code code;
- enum tree_code_class class;
+ enum tree_code_class codeclass;
if (!expr)
return false;
code = TREE_CODE (expr);
- class = TREE_CODE_CLASS (code);
+ codeclass = TREE_CODE_CLASS (code);
if (code == SSA_NAME)
return SSA_NAME_OCCURS_IN_ABNORMAL_PHI (expr) != 0;
idx_contains_abnormal_ssa_name_p,
NULL);
- switch (class)
+ switch (codeclass)
{
case tcc_binary:
case tcc_comparison:
return false;
}
-/* Element of the table in that we cache the numbers of iterations obtained
- from exits of the loop. */
-
-struct nfe_cache_elt
-{
- /* The edge for that the number of iterations is cached. */
- edge exit;
-
- /* Number of iterations corresponding to this exit, or NULL if it cannot be
- determined. */
- tree niter;
-};
-
-/* Hash function for nfe_cache_elt E. */
-
-static hashval_t
-nfe_hash (const void *e)
-{
- const struct nfe_cache_elt *elt = e;
-
- return htab_hash_pointer (elt->exit);
-}
-
-/* Equality function for nfe_cache_elt E1 and edge E2. */
-
-static int
-nfe_eq (const void *e1, const void *e2)
-{
- const struct nfe_cache_elt *elt1 = e1;
-
- return elt1->exit == e2;
-}
-
/* Returns tree describing number of iterations determined from
EXIT of DATA->current_loop, or NULL if something goes wrong. */
static tree
niter_for_exit (struct ivopts_data *data, edge exit)
{
- struct nfe_cache_elt *nfe_desc;
struct tree_niter_desc desc;
- PTR *slot;
-
- slot = htab_find_slot_with_hash (data->niters, exit,
- htab_hash_pointer (exit),
- INSERT);
+ tree niter;
+ void **slot;
- if (!*slot)
+ if (!data->niters)
{
- nfe_desc = xmalloc (sizeof (struct nfe_cache_elt));
- nfe_desc->exit = exit;
+ data->niters = pointer_map_create ();
+ slot = NULL;
+ }
+ else
+ slot = pointer_map_contains (data->niters, exit);
+ if (!slot)
+ {
/* Try to determine number of iterations. We must know it
unconditionally (i.e., without possibility of # of iterations
being zero). Also, we cannot safely work with ssa names that
overlapping life ranges for them (PR 27283). */
if (number_of_iterations_exit (data->current_loop,
exit, &desc, true)
- && zero_p (desc.may_be_zero)
+ && integer_zerop (desc.may_be_zero)
&& !contains_abnormal_ssa_name_p (desc.niter))
- nfe_desc->niter = desc.niter;
+ niter = desc.niter;
else
- nfe_desc->niter = NULL_TREE;
+ niter = NULL_TREE;
+
+ *pointer_map_insert (data->niters, exit) = niter;
}
else
- nfe_desc = *slot;
+ niter = (tree) *slot;
- return nfe_desc->niter;
+ return niter;
}
/* Returns tree describing number of iterations determined from
data->relevant = BITMAP_ALLOC (NULL);
data->important_candidates = BITMAP_ALLOC (NULL);
data->max_inv_id = 0;
- data->niters = htab_create (10, nfe_hash, nfe_eq, free);
+ data->niters = NULL;
data->iv_uses = VEC_alloc (iv_use_p, heap, 20);
data->iv_candidates = VEC_alloc (iv_cand_p, heap, 20);
decl_rtl_to_reset = VEC_alloc (tree, heap, 20);
determine_base_object (tree expr)
{
enum tree_code code = TREE_CODE (expr);
- tree base, obj, op0, op1;
+ tree base, obj;
/* If this is a pointer casted to any type, we need to determine
the base object for the pointer; so handle conversions before
return fold_convert (ptr_type_node,
build_fold_addr_expr (base));
+ case POINTER_PLUS_EXPR:
+ return determine_base_object (TREE_OPERAND (expr, 0));
+
case PLUS_EXPR:
case MINUS_EXPR:
- op0 = determine_base_object (TREE_OPERAND (expr, 0));
- op1 = determine_base_object (TREE_OPERAND (expr, 1));
-
- if (!op1)
- return op0;
-
- if (!op0)
- return (code == PLUS_EXPR
- ? op1
- : fold_build1 (NEGATE_EXPR, ptr_type_node, op1));
-
- return fold_build2 (code, ptr_type_node, op0, op1);
+ /* Pointer addition is done solely using POINTER_PLUS_EXPR. */
+ gcc_unreachable ();
default:
return fold_convert (ptr_type_node, expr);
alloc_iv (tree base, tree step)
{
struct iv *iv = XCNEW (struct iv);
-
- if (step && integer_zerop (step))
- step = NULL_TREE;
+ gcc_assert (step != NULL_TREE);
iv->base = base;
iv->base_object = determine_base_object (base);
get_iv (struct ivopts_data *data, tree var)
{
basic_block bb;
-
+ tree type = TREE_TYPE (var);
+
+ if (!POINTER_TYPE_P (type)
+ && !INTEGRAL_TYPE_P (type))
+ return NULL;
+
if (!name_info (data, var)->iv)
{
bb = bb_for_stmt (SSA_NAME_DEF_STMT (var));
if (!bb
|| !flow_bb_inside_loop_p (data->current_loop, bb))
- set_iv (data, var, var, NULL_TREE);
+ set_iv (data, var, var, build_int_cst (type, 0));
}
return name_info (data, var)->iv;
if (!simple_iv (loop, phi, name, &iv, true))
return NULL_TREE;
- return (zero_p (iv.step) ? NULL_TREE : iv.step);
+ return integer_zerop (iv.step) ? NULL_TREE : iv.step;
}
/* Finds basic ivs. */
iv->base = NULL_TREE;
iv->step = NULL_TREE;
- if (TREE_CODE (stmt) != MODIFY_EXPR)
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- lhs = TREE_OPERAND (stmt, 0);
+ lhs = GIMPLE_STMT_OPERAND (stmt, 0);
if (TREE_CODE (lhs) != SSA_NAME)
return false;
- if (!simple_iv (loop, stmt, TREE_OPERAND (stmt, 1), iv, true))
+ if (!simple_iv (loop, stmt, GIMPLE_STMT_OPERAND (stmt, 1), iv, true))
return false;
iv->base = expand_simple_operations (iv->base);
if (!find_givs_in_stmt_scev (data, stmt, &iv))
return;
- set_iv (data, TREE_OPERAND (stmt, 0), iv.base, iv.step);
+ set_iv (data, GIMPLE_STMT_OPERAND (stmt, 0), iv.base, iv.step);
}
/* Finds general ivs in basic block BB. */
return use;
}
- if (zero_p (iv->step))
+ if (integer_zerop (iv->step))
{
record_invariant (data, op, true);
return NULL;
stmt = SSA_NAME_DEF_STMT (op);
gcc_assert (TREE_CODE (stmt) == PHI_NODE
- || TREE_CODE (stmt) == MODIFY_EXPR);
+ || TREE_CODE (stmt) == GIMPLE_MODIFY_STMT);
use = record_use (data, NULL, civ, stmt, USE_NONLINEAR_EXPR);
iv->use_id = use->id;
return use;
}
-/* Checks whether the condition *COND_P in STMT is interesting
- and if so, records it. */
+/* Given a condition *COND_P, checks whether it is a compare of an induction
+ variable and an invariant. If this is the case, CONTROL_VAR is set
+ to location of the iv, BOUND to the location of the invariant,
+ IV_VAR and IV_BOUND are set to the corresponding induction variable
+ descriptions, and true is returned. If this is not the case,
+ CONTROL_VAR and BOUND are set to the arguments of the condition and
+ false is returned. */
-static void
-find_interesting_uses_cond (struct ivopts_data *data, tree stmt, tree *cond_p)
+static bool
+extract_cond_operands (struct ivopts_data *data, tree *cond_p,
+ tree **control_var, tree **bound,
+ struct iv **iv_var, struct iv **iv_bound)
{
- tree *op0_p;
- tree *op1_p;
- struct iv *iv0 = NULL, *iv1 = NULL, *civ;
- struct iv const_iv;
- tree zero = integer_zero_node;
-
- const_iv.step = NULL_TREE;
+ /* The nodes returned when COND has just one operand. Note that you should
+ not modify anything in BOUND or IV_BOUND because of this. */
+ static struct iv const_iv;
+ static tree zero;
+ tree cond = *cond_p;
+ tree *op0 = &zero, *op1 = &zero, *tmp_op;
+ struct iv *iv0 = &const_iv, *iv1 = &const_iv, *tmp_iv;
+ bool ret = false;
- if (TREE_CODE (*cond_p) != SSA_NAME
- && !COMPARISON_CLASS_P (*cond_p))
- return;
+ zero = integer_zero_node;
+ const_iv.step = integer_zero_node;
- if (TREE_CODE (*cond_p) == SSA_NAME)
+ if (TREE_CODE (cond) == SSA_NAME)
{
- op0_p = cond_p;
- op1_p = &zero;
+ op0 = cond_p;
+ iv0 = get_iv (data, cond);
+ ret = (iv0 && !integer_zerop (iv0->step));
+ goto end;
}
- else
+
+ if (!COMPARISON_CLASS_P (cond))
{
- op0_p = &TREE_OPERAND (*cond_p, 0);
- op1_p = &TREE_OPERAND (*cond_p, 1);
+ op0 = cond_p;
+ goto end;
}
- if (TREE_CODE (*op0_p) == SSA_NAME)
- iv0 = get_iv (data, *op0_p);
- else
- iv0 = &const_iv;
+ op0 = &TREE_OPERAND (cond, 0);
+ op1 = &TREE_OPERAND (cond, 1);
+ if (TREE_CODE (*op0) == SSA_NAME)
+ iv0 = get_iv (data, *op0);
+ if (TREE_CODE (*op1) == SSA_NAME)
+ iv1 = get_iv (data, *op1);
- if (TREE_CODE (*op1_p) == SSA_NAME)
- iv1 = get_iv (data, *op1_p);
- else
- iv1 = &const_iv;
+ /* Exactly one of the compared values must be an iv, and the other one must
+ be an invariant. */
+ if (!iv0 || !iv1)
+ goto end;
- if (/* When comparing with non-invariant value, we may not do any senseful
- induction variable elimination. */
- (!iv0 || !iv1)
- /* Eliminating condition based on two ivs would be nontrivial.
- ??? TODO -- it is not really important to handle this case. */
- || (!zero_p (iv0->step) && !zero_p (iv1->step)))
+ if (integer_zerop (iv0->step))
{
- find_interesting_uses_op (data, *op0_p);
- find_interesting_uses_op (data, *op1_p);
- return;
+ /* Control variable may be on the other side. */
+ tmp_op = op0; op0 = op1; op1 = tmp_op;
+ tmp_iv = iv0; iv0 = iv1; iv1 = tmp_iv;
}
+ ret = !integer_zerop (iv0->step) && integer_zerop (iv1->step);
+
+end:
+ if (control_var)
+ *control_var = op0;;
+ if (iv_var)
+ *iv_var = iv0;;
+ if (bound)
+ *bound = op1;
+ if (iv_bound)
+ *iv_bound = iv1;
+
+ return ret;
+}
- if (zero_p (iv0->step) && zero_p (iv1->step))
+/* Checks whether the condition *COND_P in STMT is interesting
+ and if so, records it. */
+
+static void
+find_interesting_uses_cond (struct ivopts_data *data, tree stmt, tree *cond_p)
+{
+ tree *var_p, *bound_p;
+ struct iv *var_iv, *civ;
+
+ if (!extract_cond_operands (data, cond_p, &var_p, &bound_p, &var_iv, NULL))
{
- /* If both are invariants, this is a work for unswitching. */
+ find_interesting_uses_op (data, *var_p);
+ find_interesting_uses_op (data, *bound_p);
return;
}
civ = XNEW (struct iv);
- *civ = zero_p (iv0->step) ? *iv1: *iv0;
+ *civ = *var_iv;
record_use (data, cond_p, civ, stmt, USE_COMPARE);
}
return true;
}
- if (!EXPR_P (expr))
+ if (!EXPR_P (expr) && !GIMPLE_STMT_P (expr))
return false;
- len = TREE_CODE_LENGTH (TREE_CODE (expr));
+ len = TREE_OPERAND_LENGTH (expr);
for (i = 0; i < len; i++)
if (!expr_invariant_in_loop_p (loop, TREE_OPERAND (expr, i)))
return false;
{
struct ivopts_data *ivopts_data;
tree stmt;
- tree *step_p;
+ tree step;
};
static bool
idx_find_step (tree base, tree *idx, void *data)
{
- struct ifs_ivopts_data *dta = data;
+ struct ifs_ivopts_data *dta = (struct ifs_ivopts_data *) data;
struct iv *iv;
tree step, iv_base, iv_step, lbound, off;
struct loop *loop = dta->ivopts_data->current_loop;
ARRAY_REF path below. */
*idx = iv->base;
- if (!iv->step)
+ if (integer_zerop (iv->step))
return true;
if (TREE_CODE (base) == ARRAY_REF)
}
step = fold_build2 (MULT_EXPR, sizetype, step, iv_step);
-
- if (!*dta->step_p)
- *dta->step_p = step;
- else
- *dta->step_p = fold_build2 (PLUS_EXPR, sizetype, *dta->step_p, step);
+ dta->step = fold_build2 (PLUS_EXPR, sizetype, dta->step, step);
return true;
}
static bool
idx_record_use (tree base, tree *idx,
- void *data)
+ void *vdata)
{
+ struct ivopts_data *data = (struct ivopts_data *) vdata;
find_interesting_uses_op (data, *idx);
if (TREE_CODE (base) == ARRAY_REF)
{
static void
find_interesting_uses_address (struct ivopts_data *data, tree stmt, tree *op_p)
{
- tree base = *op_p, step = NULL;
+ tree base = *op_p, step = build_int_cst (sizetype, 0);
struct iv *civ;
struct ifs_ivopts_data ifs_ivopts_data;
if (TMR_STEP (base))
astep = fold_build2 (MULT_EXPR, type, TMR_STEP (base), astep);
- if (step)
- step = fold_build2 (PLUS_EXPR, type, step, astep);
- else
- step = astep;
+ step = fold_build2 (PLUS_EXPR, type, step, astep);
}
}
- if (zero_p (step))
+ if (integer_zerop (step))
goto fail;
base = tree_mem_ref_addr (type, base);
}
{
ifs_ivopts_data.ivopts_data = data;
ifs_ivopts_data.stmt = stmt;
- ifs_ivopts_data.step_p = &step;
+ ifs_ivopts_data.step = build_int_cst (sizetype, 0);
if (!for_each_index (&base, idx_find_step, &ifs_ivopts_data)
- || zero_p (step))
+ || integer_zerop (ifs_ivopts_data.step))
goto fail;
+ step = ifs_ivopts_data.step;
gcc_assert (TREE_CODE (base) != ALIGN_INDIRECT_REF);
gcc_assert (TREE_CODE (base) != MISALIGNED_INDIRECT_REF);
return;
}
- if (TREE_CODE (stmt) == MODIFY_EXPR)
+ if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
{
- lhs = TREE_OPERAND (stmt, 0);
- rhs = TREE_OPERAND (stmt, 1);
+ lhs = GIMPLE_STMT_OPERAND (stmt, 0);
+ rhs = GIMPLE_STMT_OPERAND (stmt, 1);
if (TREE_CODE (lhs) == SSA_NAME)
{
iv = get_iv (data, lhs);
- if (iv && !zero_p (iv->step))
+ if (iv && !integer_zerop (iv->step))
return;
}
switch (TREE_CODE_CLASS (TREE_CODE (rhs)))
{
case tcc_comparison:
- find_interesting_uses_cond (data, stmt, &TREE_OPERAND (stmt, 1));
+ find_interesting_uses_cond (data, stmt,
+ &GIMPLE_STMT_OPERAND (stmt, 1));
return;
case tcc_reference:
- find_interesting_uses_address (data, stmt, &TREE_OPERAND (stmt, 1));
+ find_interesting_uses_address (data, stmt,
+ &GIMPLE_STMT_OPERAND (stmt, 1));
if (REFERENCE_CLASS_P (lhs))
- find_interesting_uses_address (data, stmt, &TREE_OPERAND (stmt, 0));
+ find_interesting_uses_address (data, stmt,
+ &GIMPLE_STMT_OPERAND (stmt, 0));
return;
default: ;
if (REFERENCE_CLASS_P (lhs)
&& is_gimple_val (rhs))
{
- find_interesting_uses_address (data, stmt, &TREE_OPERAND (stmt, 0));
+ find_interesting_uses_address (data, stmt,
+ &GIMPLE_STMT_OPERAND (stmt, 0));
find_interesting_uses_op (data, rhs);
return;
}
lhs = PHI_RESULT (stmt);
iv = get_iv (data, lhs);
- if (iv && !zero_p (iv->step))
+ if (iv && !integer_zerop (iv->step))
return;
}
for (phi = phi_nodes (exit->dest); phi; phi = PHI_CHAIN (phi))
{
def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
- find_interesting_uses_op (data, def);
+ if (is_gimple_reg (def))
+ find_interesting_uses_op (data, def);
}
}
{
case INTEGER_CST:
if (!cst_and_fits_in_hwi (expr)
- || zero_p (expr))
+ || integer_zerop (expr))
return orig_expr;
*offset = int_cst_value (expr);
&& op1 == TREE_OPERAND (expr, 1))
return orig_expr;
- if (zero_p (op1))
+ if (integer_zerop (op1))
expr = op0;
- else if (zero_p (op0))
+ else if (integer_zerop (op0))
{
if (code == PLUS_EXPR)
expr = op1;
*offset = off1 * st;
if (top_compref
- && zero_p (op1))
+ && integer_zerop (op1))
{
/* Strip the component reference completely. */
op0 = TREE_OPERAND (expr, 0);
static tree
find_depends (tree *expr_p, int *ws ATTRIBUTE_UNUSED, void *data)
{
- bitmap *depends_on = data;
+ bitmap *depends_on = (bitmap *) data;
struct version_info *info;
if (TREE_CODE (*expr_p) != SSA_NAME)
if (type != orig_type)
{
base = fold_convert (type, base);
- if (step)
- step = fold_convert (type, step);
+ step = fold_convert (type, step);
}
}
if (!base && !step)
continue;
- if (!operand_equal_p (base, cand->iv->base, 0))
- continue;
-
- if (zero_p (cand->iv->step))
- {
- if (zero_p (step))
- break;
- }
- else
- {
- if (step && operand_equal_p (step, cand->iv->step, 0))
- break;
- }
+ if (operand_equal_p (base, cand->iv->base, 0)
+ && operand_equal_p (step, cand->iv->step, 0))
+ break;
}
if (i == n_iv_cands (data))
EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi)
{
iv = ver_info (data, i)->iv;
- if (iv && iv->biv_p && !zero_p (iv->step))
+ if (iv && iv->biv_p && !integer_zerop (iv->step))
add_old_iv_candidates (data, iv);
}
}
{
const char *name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (obj));
x = gen_rtx_SYMBOL_REF (Pmode, name);
+ SET_SYMBOL_REF_DECL (x, obj);
+ x = gen_rtx_MEM (DECL_MODE (obj), x);
+ targetm.encode_section_info (obj, x, true);
}
else
- x = gen_raw_REG (Pmode, (*regno)++);
+ {
+ x = gen_raw_REG (Pmode, (*regno)++);
+ x = gen_rtx_MEM (DECL_MODE (obj), x);
+ }
- return gen_rtx_MEM (DECL_MODE (obj), x);
+ return x;
}
/* Prepares decl_rtl for variables referred in *EXPR_P. Callback for
{
tree obj = NULL_TREE;
rtx x = NULL_RTX;
- int *regno = data;
+ int *regno = (int *) data;
switch (TREE_CODE (*expr_p))
{
if (TREE_CODE (bot) != INTEGER_CST)
return false;
- p0 = double_int_sext (tree_to_double_int (bot), precision);
- p1 = double_int_sext (tree_to_double_int (top), precision);
+ p0 = double_int_sext (tree_to_double_int (top), precision);
+ p1 = double_int_sext (tree_to_double_int (bot), precision);
if (double_int_zero_p (p1))
return false;
*mul = double_int_sext (double_int_sdivmod (p0, p1, FLOOR_DIV_EXPR, &res),
}
}
-/* Sets COMB to CST. */
-
-static void
-aff_combination_const (struct affine_tree_combination *comb, tree type,
- unsigned HOST_WIDE_INT cst)
-{
- unsigned prec = TYPE_PRECISION (type);
-
- comb->type = type;
- comb->mask = (((unsigned HOST_WIDE_INT) 2 << (prec - 1)) - 1);
-
- comb->n = 0;
- comb->rest = NULL_TREE;
- comb->offset = cst & comb->mask;
-}
-
-/* Sets COMB to single element ELT. */
-
-static void
-aff_combination_elt (struct affine_tree_combination *comb, tree type, tree elt)
-{
- unsigned prec = TYPE_PRECISION (type);
-
- comb->type = type;
- comb->mask = (((unsigned HOST_WIDE_INT) 2 << (prec - 1)) - 1);
-
- comb->n = 1;
- comb->elts[0] = elt;
- comb->coefs[0] = 1;
- comb->rest = NULL_TREE;
- comb->offset = 0;
-}
-
-/* Scales COMB by SCALE. */
-
-static void
-aff_combination_scale (struct affine_tree_combination *comb,
- unsigned HOST_WIDE_INT scale)
-{
- unsigned i, j;
-
- if (scale == 1)
- return;
-
- if (scale == 0)
- {
- aff_combination_const (comb, comb->type, 0);
- return;
- }
-
- comb->offset = (scale * comb->offset) & comb->mask;
- for (i = 0, j = 0; i < comb->n; i++)
- {
- comb->coefs[j] = (scale * comb->coefs[i]) & comb->mask;
- comb->elts[j] = comb->elts[i];
- if (comb->coefs[j] != 0)
- j++;
- }
- comb->n = j;
-
- if (comb->rest)
- {
- if (comb->n < MAX_AFF_ELTS)
- {
- comb->coefs[comb->n] = scale;
- comb->elts[comb->n] = comb->rest;
- comb->rest = NULL_TREE;
- comb->n++;
- }
- else
- comb->rest = fold_build2 (MULT_EXPR, comb->type, comb->rest,
- build_int_cst_type (comb->type, scale));
- }
-}
-
-/* Adds ELT * SCALE to COMB. */
-
-static void
-aff_combination_add_elt (struct affine_tree_combination *comb, tree elt,
- unsigned HOST_WIDE_INT scale)
-{
- unsigned i;
-
- if (scale == 0)
- return;
-
- for (i = 0; i < comb->n; i++)
- if (operand_equal_p (comb->elts[i], elt, 0))
- {
- comb->coefs[i] = (comb->coefs[i] + scale) & comb->mask;
- if (comb->coefs[i])
- return;
-
- comb->n--;
- comb->coefs[i] = comb->coefs[comb->n];
- comb->elts[i] = comb->elts[comb->n];
-
- if (comb->rest)
- {
- gcc_assert (comb->n == MAX_AFF_ELTS - 1);
- comb->coefs[comb->n] = 1;
- comb->elts[comb->n] = comb->rest;
- comb->rest = NULL_TREE;
- comb->n++;
- }
- return;
- }
- if (comb->n < MAX_AFF_ELTS)
- {
- comb->coefs[comb->n] = scale;
- comb->elts[comb->n] = elt;
- comb->n++;
- return;
- }
-
- if (scale == 1)
- elt = fold_convert (comb->type, elt);
- else
- elt = fold_build2 (MULT_EXPR, comb->type,
- fold_convert (comb->type, elt),
- build_int_cst_type (comb->type, scale));
-
- if (comb->rest)
- comb->rest = fold_build2 (PLUS_EXPR, comb->type, comb->rest, elt);
- else
- comb->rest = elt;
-}
-
-/* Adds COMB2 to COMB1. */
-
-static void
-aff_combination_add (struct affine_tree_combination *comb1,
- struct affine_tree_combination *comb2)
-{
- unsigned i;
-
- comb1->offset = (comb1->offset + comb2->offset) & comb1->mask;
- for (i = 0; i < comb2->n; i++)
- aff_combination_add_elt (comb1, comb2->elts[i], comb2->coefs[i]);
- if (comb2->rest)
- aff_combination_add_elt (comb1, comb2->rest, 1);
-}
-
-/* Convert COMB to TYPE. */
-
-static void
-aff_combination_convert (tree type, struct affine_tree_combination *comb)
-{
- unsigned prec = TYPE_PRECISION (type);
- unsigned i;
-
- /* If the precision of both types is the same, it suffices to change the type
- of the whole combination -- the elements are allowed to have another type
- equivalent wrto STRIP_NOPS. */
- if (prec == TYPE_PRECISION (comb->type))
- {
- comb->type = type;
- return;
- }
-
- comb->mask = (((unsigned HOST_WIDE_INT) 2 << (prec - 1)) - 1);
- comb->offset = comb->offset & comb->mask;
-
- /* The type of the elements can be different from comb->type only as
- much as what STRIP_NOPS would remove. We can just directly cast
- to TYPE. */
- for (i = 0; i < comb->n; i++)
- comb->elts[i] = fold_convert (type, comb->elts[i]);
- if (comb->rest)
- comb->rest = fold_convert (type, comb->rest);
-
- comb->type = type;
-}
-
-/* Splits EXPR into an affine combination of parts. */
-
-static void
-tree_to_aff_combination (tree expr, tree type,
- struct affine_tree_combination *comb)
-{
- struct affine_tree_combination tmp;
- enum tree_code code;
- tree cst, core, toffset;
- HOST_WIDE_INT bitpos, bitsize;
- enum machine_mode mode;
- int unsignedp, volatilep;
-
- STRIP_NOPS (expr);
-
- code = TREE_CODE (expr);
- switch (code)
- {
- case INTEGER_CST:
- aff_combination_const (comb, type, int_cst_value (expr));
- return;
-
- case PLUS_EXPR:
- case MINUS_EXPR:
- tree_to_aff_combination (TREE_OPERAND (expr, 0), type, comb);
- tree_to_aff_combination (TREE_OPERAND (expr, 1), type, &tmp);
- if (code == MINUS_EXPR)
- aff_combination_scale (&tmp, -1);
- aff_combination_add (comb, &tmp);
- return;
-
- case MULT_EXPR:
- cst = TREE_OPERAND (expr, 1);
- if (TREE_CODE (cst) != INTEGER_CST)
- break;
- tree_to_aff_combination (TREE_OPERAND (expr, 0), type, comb);
- aff_combination_scale (comb, int_cst_value (cst));
- return;
-
- case NEGATE_EXPR:
- tree_to_aff_combination (TREE_OPERAND (expr, 0), type, comb);
- aff_combination_scale (comb, -1);
- return;
-
- case ADDR_EXPR:
- core = get_inner_reference (TREE_OPERAND (expr, 0), &bitsize, &bitpos,
- &toffset, &mode, &unsignedp, &volatilep,
- false);
- if (bitpos % BITS_PER_UNIT != 0)
- break;
- aff_combination_const (comb, type, bitpos / BITS_PER_UNIT);
- core = build_fold_addr_expr (core);
- if (TREE_CODE (core) == ADDR_EXPR)
- aff_combination_add_elt (comb, core, 1);
- else
- {
- tree_to_aff_combination (core, type, &tmp);
- aff_combination_add (comb, &tmp);
- }
- if (toffset)
- {
- tree_to_aff_combination (toffset, type, &tmp);
- aff_combination_add (comb, &tmp);
- }
- return;
-
- default:
- break;
- }
-
- aff_combination_elt (comb, type, expr);
-}
-
-/* Creates EXPR + ELT * SCALE in TYPE. MASK is the mask for width of TYPE. */
-
-static tree
-add_elt_to_tree (tree expr, tree type, tree elt, unsigned HOST_WIDE_INT scale,
- unsigned HOST_WIDE_INT mask)
-{
- enum tree_code code;
-
- scale &= mask;
- elt = fold_convert (type, elt);
-
- if (scale == 1)
- {
- if (!expr)
- return elt;
-
- return fold_build2 (PLUS_EXPR, type, expr, elt);
- }
-
- if (scale == mask)
- {
- if (!expr)
- return fold_build1 (NEGATE_EXPR, type, elt);
-
- return fold_build2 (MINUS_EXPR, type, expr, elt);
- }
-
- if (!expr)
- return fold_build2 (MULT_EXPR, type, elt,
- build_int_cst_type (type, scale));
-
- if ((scale | (mask >> 1)) == mask)
- {
- /* Scale is negative. */
- code = MINUS_EXPR;
- scale = (-scale) & mask;
- }
- else
- code = PLUS_EXPR;
-
- elt = fold_build2 (MULT_EXPR, type, elt,
- build_int_cst_type (type, scale));
- return fold_build2 (code, type, expr, elt);
-}
-
-/* Copies the tree elements of COMB to ensure that they are not shared. */
-
-static void
-unshare_aff_combination (struct affine_tree_combination *comb)
-{
- unsigned i;
-
- for (i = 0; i < comb->n; i++)
- comb->elts[i] = unshare_expr (comb->elts[i]);
- if (comb->rest)
- comb->rest = unshare_expr (comb->rest);
-}
-
-/* Makes tree from the affine combination COMB. */
-
-static tree
-aff_combination_to_tree (struct affine_tree_combination *comb)
-{
- tree type = comb->type;
- tree expr = comb->rest;
- unsigned i;
- unsigned HOST_WIDE_INT off, sgn;
-
- if (comb->n == 0 && comb->offset == 0)
- {
- if (expr)
- {
- /* Handle the special case produced by get_computation_aff when
- the type does not fit in HOST_WIDE_INT. */
- return fold_convert (type, expr);
- }
- else
- return build_int_cst (type, 0);
- }
-
- gcc_assert (comb->n == MAX_AFF_ELTS || comb->rest == NULL_TREE);
-
- for (i = 0; i < comb->n; i++)
- expr = add_elt_to_tree (expr, type, comb->elts[i], comb->coefs[i],
- comb->mask);
-
- if ((comb->offset | (comb->mask >> 1)) == comb->mask)
- {
- /* Offset is negative. */
- off = (-comb->offset) & comb->mask;
- sgn = comb->mask;
- }
- else
- {
- off = comb->offset;
- sgn = 1;
- }
- return add_elt_to_tree (expr, type, build_int_cst_type (type, off), sgn,
- comb->mask);
-}
-
-/* Folds EXPR using the affine expressions framework. */
-
-static tree
-fold_affine_expr (tree expr)
-{
- tree type = TREE_TYPE (expr);
- struct affine_tree_combination comb;
-
- if (TYPE_PRECISION (type) > HOST_BITS_PER_WIDE_INT)
- return expr;
-
- tree_to_aff_combination (expr, type, &comb);
- return aff_combination_to_tree (&comb);
-}
-
/* If A is (TYPE) BA and B is (TYPE) BB, and the types of BA and BB have the
same precision that is at least as wide as the precision of TYPE, stores
BA to A and BB to B, and returns the type of BA. Otherwise, returns the
tree ubase = use->iv->base;
tree ustep = use->iv->step;
tree cbase = cand->iv->base;
- tree cstep = cand->iv->step;
+ tree cstep = cand->iv->step, cstep_common;
tree utype = TREE_TYPE (ubase), ctype = TREE_TYPE (cbase);
- tree common_type;
+ tree common_type, var;
tree uutype;
- tree expr, delta;
- tree ratio;
- unsigned HOST_WIDE_INT ustepi, cstepi;
- HOST_WIDE_INT ratioi;
- struct affine_tree_combination cbase_aff, expr_aff;
- tree cstep_orig = cstep, ustep_orig = ustep;
+ aff_tree cbase_aff, var_aff;
double_int rat;
if (TYPE_PRECISION (utype) > TYPE_PRECISION (ctype))
return false;
}
- expr = var_at_stmt (loop, cand, at);
-
- if (TREE_TYPE (expr) != ctype)
- {
- /* This may happen with the original ivs. */
- expr = fold_convert (ctype, expr);
- }
-
- if (TYPE_UNSIGNED (utype))
- uutype = utype;
- else
- {
- uutype = unsigned_type_for (utype);
- ubase = fold_convert (uutype, ubase);
- ustep = fold_convert (uutype, ustep);
- }
+ var = var_at_stmt (loop, cand, at);
+ uutype = unsigned_type_for (utype);
- if (uutype != ctype)
+ /* If the conversion is not noop, perform it. */
+ if (TYPE_PRECISION (utype) < TYPE_PRECISION (ctype))
{
- expr = fold_convert (uutype, expr);
- cbase = fold_convert (uutype, cbase);
cstep = fold_convert (uutype, cstep);
-
- /* If the conversion is not noop, we must take it into account when
- considering the value of the step. */
- if (TYPE_PRECISION (utype) < TYPE_PRECISION (ctype))
- cstep_orig = cstep;
- }
-
- if (cst_and_fits_in_hwi (cstep_orig)
- && cst_and_fits_in_hwi (ustep_orig))
- {
- ustepi = int_cst_value (ustep_orig);
- cstepi = int_cst_value (cstep_orig);
-
- if (!divide (TYPE_PRECISION (uutype), ustepi, cstepi, &ratioi))
- {
- /* TODO maybe consider case when ustep divides cstep and the ratio is
- a power of 2 (so that the division is fast to execute)? We would
- need to be much more careful with overflows etc. then. */
- return false;
- }
-
- ratio = build_int_cst_type (uutype, ratioi);
+ cbase = fold_convert (uutype, cbase);
+ var = fold_convert (uutype, var);
}
- else
- {
- if (!constant_multiple_of (ustep_orig, cstep_orig, &rat))
- return false;
- ratio = double_int_to_tree (uutype, rat);
- /* Ratioi is only used to detect special cases when the multiplicative
- factor is 1 or -1, so if rat does not fit to HOST_WIDE_INT, we may
- set it to 0. */
- if (double_int_fits_in_shwi_p (rat))
- ratioi = double_int_to_shwi (rat);
- else
- ratioi = 0;
- }
+ if (!constant_multiple_of (ustep, cstep, &rat))
+ return false;
/* In case both UBASE and CBASE are shortened to UUTYPE from some common
type, we achieve better folding by computing their difference in this
anyway. */
common_type = determine_common_wider_type (&ubase, &cbase);
- /* We may need to shift the value if we are after the increment. */
- if (stmt_after_increment (loop, cand, at))
- {
- if (uutype != common_type)
- cstep = fold_convert (common_type, cstep);
- cbase = fold_build2 (PLUS_EXPR, common_type, cbase, cstep);
- }
-
- /* use = ubase - ratio * cbase + ratio * var.
-
- In general case ubase + ratio * (var - cbase) could be better (one less
- multiplication), but often it is possible to eliminate redundant parts
- of computations from (ubase - ratio * cbase) term, and if it does not
- happen, fold is able to apply the distributive law to obtain this form
- anyway. */
+ /* use = ubase - ratio * cbase + ratio * var. */
+ tree_to_aff_combination (ubase, common_type, aff);
+ tree_to_aff_combination (cbase, common_type, &cbase_aff);
+ tree_to_aff_combination (var, uutype, &var_aff);
- if (TYPE_PRECISION (common_type) > HOST_BITS_PER_WIDE_INT)
+ /* We need to shift the value if we are after the increment. */
+ if (stmt_after_increment (loop, cand, at))
{
- /* Let's compute in trees and just return the result in AFF. This case
- should not be very common, and fold itself is not that bad either,
- so making the aff. functions more complicated to handle this case
- is not that urgent. */
- if (ratioi == 1)
- {
- delta = fold_build2 (MINUS_EXPR, common_type, ubase, cbase);
- if (uutype != common_type)
- delta = fold_convert (uutype, delta);
- expr = fold_build2 (PLUS_EXPR, uutype, expr, delta);
- }
- else if (ratioi == -1)
- {
- delta = fold_build2 (PLUS_EXPR, common_type, ubase, cbase);
- if (uutype != common_type)
- delta = fold_convert (uutype, delta);
- expr = fold_build2 (MINUS_EXPR, uutype, delta, expr);
- }
+ aff_tree cstep_aff;
+
+ if (common_type != uutype)
+ cstep_common = fold_convert (common_type, cstep);
else
- {
- delta = fold_build2 (MULT_EXPR, common_type, cbase, ratio);
- delta = fold_build2 (MINUS_EXPR, common_type, ubase, delta);
- if (uutype != common_type)
- delta = fold_convert (uutype, delta);
- expr = fold_build2 (MULT_EXPR, uutype, ratio, expr);
- expr = fold_build2 (PLUS_EXPR, uutype, delta, expr);
- }
+ cstep_common = cstep;
- aff->type = uutype;
- aff->n = 0;
- aff->offset = 0;
- aff->mask = 0;
- aff->rest = expr;
- return true;
+ tree_to_aff_combination (cstep_common, common_type, &cstep_aff);
+ aff_combination_add (&cbase_aff, &cstep_aff);
}
- /* If we got here, the types fits in HOST_WIDE_INT, thus it must be
- possible to compute ratioi. */
- gcc_assert (ratioi);
-
- tree_to_aff_combination (ubase, common_type, aff);
- tree_to_aff_combination (cbase, common_type, &cbase_aff);
- tree_to_aff_combination (expr, uutype, &expr_aff);
- aff_combination_scale (&cbase_aff, -ratioi);
- aff_combination_scale (&expr_aff, ratioi);
+ aff_combination_scale (&cbase_aff, double_int_neg (rat));
aff_combination_add (aff, &cbase_aff);
if (common_type != uutype)
- aff_combination_convert (uutype, aff);
- aff_combination_add (aff, &expr_aff);
+ aff_combination_convert (aff, uutype);
+
+ aff_combination_scale (&var_aff, rat);
+ aff_combination_add (aff, &var_aff);
return true;
}
get_computation_at (struct loop *loop,
struct iv_use *use, struct iv_cand *cand, tree at)
{
- struct affine_tree_combination aff;
+ aff_tree aff;
tree type = TREE_TYPE (use->iv->base);
if (!get_computation_aff (loop, use, cand, at, &aff))
static hashval_t
mbc_entry_hash (const void *entry)
{
- const struct mbc_entry *e = entry;
+ const struct mbc_entry *e = (const struct mbc_entry *) entry;
return 57 * (hashval_t) e->mode + (hashval_t) (e->cst % 877);
}
static int
mbc_entry_eq (const void *entry1, const void *entry2)
{
- const struct mbc_entry *e1 = entry1;
- const struct mbc_entry *e2 = entry2;
+ const struct mbc_entry *e1 = (const struct mbc_entry *) entry1;
+ const struct mbc_entry *e2 = (const struct mbc_entry *) entry2;
return (e1->mode == e2->mode
&& e1->cst == e2->cst);
return cost;
}
-/* Returns true if multiplying by RATIO is allowed in address. */
+/* Returns true if multiplying by RATIO is allowed in an address. Test the
+ validity for a memory reference accessing memory of mode MODE. */
bool
-multiplier_allowed_in_address_p (HOST_WIDE_INT ratio)
+multiplier_allowed_in_address_p (HOST_WIDE_INT ratio, enum machine_mode mode)
{
#define MAX_RATIO 128
- static sbitmap valid_mult;
+ static sbitmap valid_mult[MAX_MACHINE_MODE];
- if (!valid_mult)
+ if (!valid_mult[mode])
{
rtx reg1 = gen_raw_REG (Pmode, LAST_VIRTUAL_REGISTER + 1);
rtx addr;
HOST_WIDE_INT i;
- valid_mult = sbitmap_alloc (2 * MAX_RATIO + 1);
- sbitmap_zero (valid_mult);
+ valid_mult[mode] = sbitmap_alloc (2 * MAX_RATIO + 1);
+ sbitmap_zero (valid_mult[mode]);
addr = gen_rtx_fmt_ee (MULT, Pmode, reg1, NULL_RTX);
for (i = -MAX_RATIO; i <= MAX_RATIO; i++)
{
XEXP (addr, 1) = gen_int_mode (i, Pmode);
- if (memory_address_p (Pmode, addr))
- SET_BIT (valid_mult, i + MAX_RATIO);
+ if (memory_address_p (mode, addr))
+ SET_BIT (valid_mult[mode], i + MAX_RATIO);
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, " allowed multipliers:");
for (i = -MAX_RATIO; i <= MAX_RATIO; i++)
- if (TEST_BIT (valid_mult, i + MAX_RATIO))
+ if (TEST_BIT (valid_mult[mode], i + MAX_RATIO))
fprintf (dump_file, " %d", (int) i);
fprintf (dump_file, "\n");
fprintf (dump_file, "\n");
if (ratio > MAX_RATIO || ratio < -MAX_RATIO)
return false;
- return TEST_BIT (valid_mult, ratio + MAX_RATIO);
+ return TEST_BIT (valid_mult[mode], ratio + MAX_RATIO);
}
/* Returns cost of address in shape symbol + var + OFFSET + RATIO * index.
If SYMBOL_PRESENT is false, symbol is omitted. If VAR_PRESENT is false,
- variable is omitted. The created memory accesses MODE.
-
+ variable is omitted. Compute the cost for a memory reference that accesses
+ a memory location of mode MEM_MODE.
+
TODO -- there must be some better way. This all is quite crude. */
static unsigned
get_address_cost (bool symbol_present, bool var_present,
- unsigned HOST_WIDE_INT offset, HOST_WIDE_INT ratio)
+ unsigned HOST_WIDE_INT offset, HOST_WIDE_INT ratio,
+ enum machine_mode mem_mode)
{
- static bool initialized = false;
- static HOST_WIDE_INT rat, off;
- static HOST_WIDE_INT min_offset, max_offset;
- static unsigned costs[2][2][2][2];
+ static bool initialized[MAX_MACHINE_MODE];
+ static HOST_WIDE_INT rat[MAX_MACHINE_MODE], off[MAX_MACHINE_MODE];
+ static HOST_WIDE_INT min_offset[MAX_MACHINE_MODE], max_offset[MAX_MACHINE_MODE];
+ static unsigned costs[MAX_MACHINE_MODE][2][2][2][2];
unsigned cost, acost;
bool offset_p, ratio_p;
HOST_WIDE_INT s_offset;
unsigned HOST_WIDE_INT mask;
unsigned bits;
- if (!initialized)
+ if (!initialized[mem_mode])
{
HOST_WIDE_INT i;
+ HOST_WIDE_INT start = BIGGEST_ALIGNMENT / BITS_PER_UNIT;
int old_cse_not_expected;
unsigned sym_p, var_p, off_p, rat_p, add_c;
rtx seq, addr, base;
rtx reg0, reg1;
- initialized = true;
+ initialized[mem_mode] = true;
reg1 = gen_raw_REG (Pmode, LAST_VIRTUAL_REGISTER + 1);
addr = gen_rtx_fmt_ee (PLUS, Pmode, reg1, NULL_RTX);
- for (i = 1; i <= 1 << 20; i <<= 1)
+ for (i = start; i <= 1 << 20; i <<= 1)
{
XEXP (addr, 1) = gen_int_mode (i, Pmode);
- if (!memory_address_p (Pmode, addr))
+ if (!memory_address_p (mem_mode, addr))
break;
}
- max_offset = i >> 1;
- off = max_offset;
+ max_offset[mem_mode] = i == start ? 0 : i >> 1;
+ off[mem_mode] = max_offset[mem_mode];
- for (i = 1; i <= 1 << 20; i <<= 1)
+ for (i = start; i <= 1 << 20; i <<= 1)
{
XEXP (addr, 1) = gen_int_mode (-i, Pmode);
- if (!memory_address_p (Pmode, addr))
+ if (!memory_address_p (mem_mode, addr))
break;
}
- min_offset = -(i >> 1);
+ min_offset[mem_mode] = i == start ? 0 : -(i >> 1);
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "get_address_cost:\n");
- fprintf (dump_file, " min offset %d\n", (int) min_offset);
- fprintf (dump_file, " max offset %d\n", (int) max_offset);
+ fprintf (dump_file, " min offset %s %d\n",
+ GET_MODE_NAME (mem_mode),
+ (int) min_offset[mem_mode]);
+ fprintf (dump_file, " max offset %s %d\n",
+ GET_MODE_NAME (mem_mode),
+ (int) max_offset[mem_mode]);
}
- rat = 1;
+ rat[mem_mode] = 1;
for (i = 2; i <= MAX_RATIO; i++)
- if (multiplier_allowed_in_address_p (i))
+ if (multiplier_allowed_in_address_p (i, mem_mode))
{
- rat = i;
+ rat[mem_mode] = i;
break;
}
addr = reg0;
if (rat_p)
- addr = gen_rtx_fmt_ee (MULT, Pmode, addr, gen_int_mode (rat, Pmode));
+ addr = gen_rtx_fmt_ee (MULT, Pmode, addr,
+ gen_int_mode (rat[mem_mode], Pmode));
if (var_p)
addr = gen_rtx_fmt_ee (PLUS, Pmode, addr, reg1);
if (sym_p)
{
base = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (""));
+ /* ??? We can run into trouble with some backends by presenting
+ it with symbols which havn't been properly passed through
+ targetm.encode_section_info. By setting the local bit, we
+ enhance the probability of things working. */
+ SYMBOL_REF_FLAGS (base) = SYMBOL_FLAG_LOCAL;
+
if (off_p)
base = gen_rtx_fmt_e (CONST, Pmode,
gen_rtx_fmt_ee (PLUS, Pmode,
base,
- gen_int_mode (off, Pmode)));
+ gen_int_mode (off[mem_mode],
+ Pmode)));
}
else if (off_p)
- base = gen_int_mode (off, Pmode);
+ base = gen_int_mode (off[mem_mode], Pmode);
else
base = NULL_RTX;
follow. */
old_cse_not_expected = cse_not_expected;
cse_not_expected = true;
- addr = memory_address (Pmode, addr);
+ addr = memory_address (mem_mode, addr);
cse_not_expected = old_cse_not_expected;
seq = get_insns ();
end_sequence ();
acost = seq_cost (seq);
- acost += address_cost (addr, Pmode);
+ acost += address_cost (addr, mem_mode);
if (!acost)
acost = 1;
- costs[sym_p][var_p][off_p][rat_p] = acost;
+ costs[mem_mode][sym_p][var_p][off_p][rat_p] = acost;
}
/* On some targets, it is quite expensive to load symbol to a register,
off_p = (i >> 1) & 1;
rat_p = (i >> 2) & 1;
- acost = costs[0][1][off_p][rat_p] + 1;
+ acost = costs[mem_mode][0][1][off_p][rat_p] + 1;
if (var_p)
acost += add_c;
- if (acost < costs[1][var_p][off_p][rat_p])
- costs[1][var_p][off_p][rat_p] = acost;
+ if (acost < costs[mem_mode][1][var_p][off_p][rat_p])
+ costs[mem_mode][1][var_p][off_p][rat_p] = acost;
}
if (dump_file && (dump_flags & TDF_DETAILS))
if (rat_p)
fprintf (dump_file, "rat * ");
- acost = costs[sym_p][var_p][off_p][rat_p];
+ acost = costs[mem_mode][sym_p][var_p][off_p][rat_p];
fprintf (dump_file, "index costs %d\n", acost);
}
fprintf (dump_file, "\n");
cost = 0;
offset_p = (s_offset != 0
- && min_offset <= s_offset && s_offset <= max_offset);
+ && min_offset[mem_mode] <= s_offset
+ && s_offset <= max_offset[mem_mode]);
ratio_p = (ratio != 1
- && multiplier_allowed_in_address_p (ratio));
+ && multiplier_allowed_in_address_p (ratio, mem_mode));
if (ratio != 1 && !ratio_p)
cost += multiply_by_cost (ratio, Pmode);
if (s_offset && !offset_p && !symbol_present)
- {
- cost += add_cost (Pmode);
- var_present = true;
- }
+ cost += add_cost (Pmode);
- acost = costs[symbol_present][var_present][offset_p][ratio_p];
+ acost = costs[mem_mode][symbol_present][var_present][offset_p][ratio_p];
return cost + acost;
}
if (!costs_initialized)
{
- tree var = create_tmp_var_raw (integer_type_node, "test_var");
- rtx x = gen_rtx_MEM (DECL_MODE (var),
- gen_rtx_SYMBOL_REF (Pmode, "test_var"));
- tree addr;
tree type = build_pointer_type (integer_type_node);
+ tree var, addr;
+ rtx x;
+
+ var = create_tmp_var_raw (integer_type_node, "test_var");
+ TREE_STATIC (var) = 1;
+ x = produce_memory_decl_rtl (var, NULL);
+ SET_DECL_RTL (var, x);
integer_cost = computation_cost (build_int_cst (integer_type_node,
2000));
- SET_DECL_RTL (var, x);
- TREE_STATIC (var) = 1;
addr = build1 (ADDR_EXPR, type, var);
symbol_cost = computation_cost (addr) + 1;
address_cost
- = computation_cost (build2 (PLUS_EXPR, type,
+ = computation_cost (build2 (POINTER_PLUS_EXPR, type,
addr,
- build_int_cst (type, 2000))) + 1;
+ build_int_cst (sizetype, 2000))) + 1;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "force_expr_to_var_cost:\n");
switch (TREE_CODE (expr))
{
+ case POINTER_PLUS_EXPR:
case PLUS_EXPR:
case MINUS_EXPR:
case MULT_EXPR:
mode = TYPE_MODE (TREE_TYPE (expr));
switch (TREE_CODE (expr))
{
+ case POINTER_PLUS_EXPR:
case PLUS_EXPR:
case MINUS_EXPR:
cost = add_cost (mode);
return 0;
}
*var_present = true;
- if (zero_p (e2))
+ if (integer_zerop (e2))
return force_var_cost (data, e1, depends_on);
- if (zero_p (e1))
+ if (integer_zerop (e1))
{
cost = force_var_cost (data, e2, depends_on);
cost += multiply_by_cost (-1, mode);
tree ubase = use->iv->base, ustep = use->iv->step;
tree cbase, cstep;
tree utype = TREE_TYPE (ubase), ctype;
- unsigned HOST_WIDE_INT ustepi, cstepi, offset = 0;
+ unsigned HOST_WIDE_INT cstepi, offset = 0;
HOST_WIDE_INT ratio, aratio;
bool var_present, symbol_present;
unsigned cost = 0, n_sums;
+ double_int rat;
*depends_on = NULL;
else
cstepi = 0;
- if (cst_and_fits_in_hwi (ustep)
- && cst_and_fits_in_hwi (cstep))
- {
- ustepi = int_cst_value (ustep);
-
- if (!divide (TYPE_PRECISION (utype), ustepi, cstepi, &ratio))
- return INFTY;
- }
- else
- {
- double_int rat;
-
- if (!constant_multiple_of (ustep, cstep, &rat))
- return INFTY;
+ if (!constant_multiple_of (ustep, cstep, &rat))
+ return INFTY;
- if (double_int_fits_in_shwi_p (rat))
- ratio = double_int_to_shwi (rat);
- else
- return INFTY;
- }
+ if (double_int_fits_in_shwi_p (rat))
+ ratio = double_int_to_shwi (rat);
+ else
+ return INFTY;
/* use = ubase + ratio * (var - cbase). If either cbase is a constant
or ratio == 1, it is better to handle this like
(symbol/var/const parts may be omitted). If we are looking for an address,
find the cost of addressing this. */
if (address_p)
- return cost + get_address_cost (symbol_present, var_present, offset, ratio);
+ return cost + get_address_cost (symbol_present, var_present, offset, ratio,
+ TYPE_MODE (TREE_TYPE (*use->op_p)));
/* Otherwise estimate the costs for computing the expression. */
aratio = ratio > 0 ? ratio : -ratio;
return cost != INFTY;
}
-/* Computes value of induction variable IV in iteration NITER. */
+/* Computes value of candidate CAND at position AT in iteration NITER, and
+ stores it to VAL. */
-static tree
-iv_value (struct iv *iv, tree niter)
+static void
+cand_value_at (struct loop *loop, struct iv_cand *cand, tree at, tree niter,
+ aff_tree *val)
{
- tree val;
+ aff_tree step, delta, nit;
+ struct iv *iv = cand->iv;
tree type = TREE_TYPE (iv->base);
- niter = fold_convert (type, niter);
- val = fold_build2 (MULT_EXPR, type, iv->step, niter);
-
- return fold_build2 (PLUS_EXPR, type, iv->base, val);
-}
-
-/* Computes value of candidate CAND at position AT in iteration NITER. */
-
-static tree
-cand_value_at (struct loop *loop, struct iv_cand *cand, tree at, tree niter)
-{
- tree val = iv_value (cand->iv, niter);
- tree type = TREE_TYPE (cand->iv->base);
-
+ tree_to_aff_combination (iv->step, type, &step);
+ tree_to_aff_combination (niter, TREE_TYPE (niter), &nit);
+ aff_combination_convert (&nit, type);
+ aff_combination_mult (&nit, &step, &delta);
if (stmt_after_increment (loop, cand, at))
- val = fold_build2 (PLUS_EXPR, type, val, cand->iv->step);
+ aff_combination_add (&delta, &step);
- return val;
+ tree_to_aff_combination (iv->base, type, val);
+ aff_combination_add (val, &delta);
}
/* Returns period of induction variable iv. */
tree nit, nit_type;
tree wider_type, period, per_type;
struct loop *loop = data->current_loop;
+ aff_tree bnd;
if (TREE_CODE (cand->iv->step) != INTEGER_CST)
return false;
fold_convert (wider_type, nit))))
return false;
- *bound = fold_affine_expr (cand_value_at (loop, cand, use->stmt, nit));
+ cand_value_at (loop, cand, use->stmt, nit, &bnd);
+ *bound = aff_combination_to_tree (&bnd);
return true;
}
determine_use_iv_cost_condition (struct ivopts_data *data,
struct iv_use *use, struct iv_cand *cand)
{
- tree bound = NULL_TREE, op, cond;
- bitmap depends_on = NULL;
- unsigned cost;
+ tree bound = NULL_TREE;
+ struct iv *cmp_iv;
+ bitmap depends_on_elim = NULL, depends_on_express = NULL, depends_on;
+ unsigned elim_cost, express_cost, cost;
+ bool ok;
/* Only consider real candidates. */
if (!cand->iv)
return false;
}
+ /* Try iv elimination. */
if (may_eliminate_iv (data, use, cand, &bound))
- {
- cost = force_var_cost (data, bound, &depends_on);
+ elim_cost = force_var_cost (data, bound, &depends_on_elim);
+ else
+ elim_cost = INFTY;
- set_use_iv_cost (data, use, cand, cost, depends_on, bound);
- return cost != INFTY;
- }
+ /* Try expressing the original giv. If it is compared with an invariant,
+ note that we cannot get rid of it. */
+ ok = extract_cond_operands (data, use->op_p, NULL, NULL, NULL, &cmp_iv);
+ gcc_assert (ok);
- /* The induction variable elimination failed; just express the original
- giv. If it is compared with an invariant, note that we cannot get
- rid of it. */
- cost = get_computation_cost (data, use, cand, false, &depends_on);
+ express_cost = get_computation_cost (data, use, cand, false,
+ &depends_on_express);
+ fd_ivopts_data = data;
+ walk_tree (&cmp_iv->base, find_depends, &depends_on_express, NULL);
- cond = *use->op_p;
- if (TREE_CODE (cond) != SSA_NAME)
+ /* Choose the better approach. */
+ if (elim_cost < express_cost)
{
- op = TREE_OPERAND (cond, 0);
- if (TREE_CODE (op) == SSA_NAME && !zero_p (get_iv (data, op)->step))
- op = TREE_OPERAND (cond, 1);
- if (TREE_CODE (op) == SSA_NAME)
- {
- op = get_iv (data, op)->base;
- fd_ivopts_data = data;
- walk_tree (&op, find_depends, &depends_on, NULL);
- }
+ cost = elim_cost;
+ depends_on = depends_on_elim;
+ depends_on_elim = NULL;
+ }
+ else
+ {
+ cost = express_cost;
+ depends_on = depends_on_express;
+ depends_on_express = NULL;
+ bound = NULL_TREE;
}
- set_use_iv_cost (data, use, cand, cost, depends_on, NULL);
+ set_use_iv_cost (data, use, cand, cost, depends_on, bound);
+
+ if (depends_on_elim)
+ BITMAP_FREE (depends_on_elim);
+ if (depends_on_express)
+ BITMAP_FREE (depends_on_express);
+
return cost != INFTY;
}
static unsigned
ivopts_global_cost_for_size (struct ivopts_data *data, unsigned size)
{
- return global_cost_for_size (size, data->regs_used, n_iv_uses (data));
+ /* We add size to the cost, so that we prefer eliminating ivs
+ if possible. */
+ return size + estimate_reg_pressure_cost (size, data->regs_used);
}
/* For each size of the induction variable set determine the penalty. */
{
fprintf (dump_file, "Global costs:\n");
fprintf (dump_file, " target_avail_regs %d\n", target_avail_regs);
- fprintf (dump_file, " target_small_cost %d\n", target_small_cost);
- fprintf (dump_file, " target_pres_cost %d\n", target_pres_cost);
+ fprintf (dump_file, " target_reg_cost %d\n", target_reg_cost);
fprintf (dump_file, " target_spill_cost %d\n", target_spill_cost);
}
{
if (TREE_CODE (stmt) == PHI_NODE)
{
- if (!including_defined_name)
- {
- /* Prevent the ssa name defined by the statement from being removed. */
- SET_PHI_RESULT (stmt, NULL);
- }
- remove_phi_node (stmt, NULL_TREE);
+ remove_phi_node (stmt, NULL_TREE, including_defined_name);
}
else
{
block_stmt_iterator bsi = bsi_for_stmt (stmt);
bsi_remove (&bsi, true);
+ release_defs (stmt);
}
}
{
tree comp;
tree op, stmts, tgt, ass;
- block_stmt_iterator bsi, pbsi;
+ block_stmt_iterator bsi;
/* An important special case -- if we are asked to express value of
the original iv by itself, just exit; there is no need to
tree step, ctype, utype;
enum tree_code incr_code = PLUS_EXPR;
- gcc_assert (TREE_CODE (use->stmt) == MODIFY_EXPR);
- gcc_assert (TREE_OPERAND (use->stmt, 0) == cand->var_after);
+ gcc_assert (TREE_CODE (use->stmt) == GIMPLE_MODIFY_STMT);
+ gcc_assert (GIMPLE_STMT_OPERAND (use->stmt, 0) == cand->var_after);
step = cand->iv->step;
ctype = TREE_TYPE (step);
computations in the loop -- otherwise, the computation
we rely upon may be removed in remove_unused_ivs,
thus leading to ICE. */
- op = TREE_OPERAND (use->stmt, 1);
+ op = GIMPLE_STMT_OPERAND (use->stmt, 1);
if (TREE_CODE (op) == PLUS_EXPR
|| TREE_CODE (op) == MINUS_EXPR)
{
unshare_expr (step)));
}
else
- comp = get_computation (data->current_loop, use, cand);
+ {
+ comp = get_computation (data->current_loop, use, cand);
+ gcc_assert (comp != NULL_TREE);
+ }
switch (TREE_CODE (use->stmt))
{
if (name_info (data, tgt)->preserve_biv)
return;
- pbsi = bsi = bsi_start (bb_for_stmt (use->stmt));
- while (!bsi_end_p (pbsi)
- && TREE_CODE (bsi_stmt (pbsi)) == LABEL_EXPR)
- {
- bsi = pbsi;
- bsi_next (&pbsi);
- }
+ bsi = bsi_after_labels (bb_for_stmt (use->stmt));
break;
- case MODIFY_EXPR:
- tgt = TREE_OPERAND (use->stmt, 0);
+ case GIMPLE_MODIFY_STMT:
+ tgt = GIMPLE_STMT_OPERAND (use->stmt, 0);
bsi = bsi_for_stmt (use->stmt);
break;
}
op = force_gimple_operand (comp, &stmts, false, SSA_NAME_VAR (tgt));
+ if (stmts)
+ bsi_insert_before (&bsi, stmts, BSI_SAME_STMT);
if (TREE_CODE (use->stmt) == PHI_NODE)
{
- if (stmts)
- bsi_insert_after (&bsi, stmts, BSI_CONTINUE_LINKING);
- ass = build2 (MODIFY_EXPR, TREE_TYPE (tgt), tgt, op);
- bsi_insert_after (&bsi, ass, BSI_NEW_STMT);
+ ass = build_gimple_modify_stmt (tgt, op);
+ bsi_insert_before (&bsi, ass, BSI_SAME_STMT);
remove_statement (use->stmt, false);
SSA_NAME_DEF_STMT (tgt) = ass;
}
else
- {
- if (stmts)
- bsi_insert_before (&bsi, stmts, BSI_SAME_STMT);
- TREE_OPERAND (use->stmt, 1) = op;
- }
+ GIMPLE_STMT_OPERAND (use->stmt, 1) = op;
}
/* Replaces ssa name in index IDX by its basic variable. Callback for
}
var = SSA_NAME_VAR (var);
- tag = var_ann (var)->symbol_mem_tag;
+ tag = symbol_mem_tag (var);
gcc_assert (tag != NULL_TREE);
return tag;
}
if (!DECL_P (var))
return NULL_TREE;
- tag = var_ann (var)->symbol_mem_tag;
+ tag = symbol_mem_tag (var);
if (tag)
return tag;
rewrite_use_address (struct ivopts_data *data,
struct iv_use *use, struct iv_cand *cand)
{
- struct affine_tree_combination aff;
+ aff_tree aff;
block_stmt_iterator bsi = bsi_for_stmt (use->stmt);
tree ref;
+ bool ok;
- get_computation_aff (data->current_loop, use, cand, use->stmt, &aff);
+ ok = get_computation_aff (data->current_loop, use, cand, use->stmt, &aff);
+ gcc_assert (ok);
unshare_aff_combination (&aff);
ref = create_mem_ref (&bsi, TREE_TYPE (*use->op_p), &aff);
rewrite_use_compare (struct ivopts_data *data,
struct iv_use *use, struct iv_cand *cand)
{
- tree comp;
- tree *op_p, cond, op, stmts, bound;
+ tree comp, *var_p, op, bound;
block_stmt_iterator bsi = bsi_for_stmt (use->stmt);
enum tree_code compare;
struct cost_pair *cp = get_use_iv_cost (data, use, cand);
-
+ bool ok;
+
bound = cp->value;
if (bound)
{
tree var_type = TREE_TYPE (var);
compare = iv_elimination_compare (data, use);
- bound = fold_convert (var_type, bound);
- op = force_gimple_operand (unshare_expr (bound), &stmts,
- true, NULL_TREE);
-
- if (stmts)
- bsi_insert_before (&bsi, stmts, BSI_SAME_STMT);
+ bound = unshare_expr (fold_convert (var_type, bound));
+ op = force_gimple_operand_bsi (&bsi, bound, true, NULL_TREE);
*use->op_p = build2 (compare, boolean_type_node, var, op);
- update_stmt (use->stmt);
return;
}
/* The induction variable elimination failed; just express the original
giv. */
comp = get_computation (data->current_loop, use, cand);
+ gcc_assert (comp != NULL_TREE);
- cond = *use->op_p;
- op_p = &TREE_OPERAND (cond, 0);
- if (TREE_CODE (*op_p) != SSA_NAME
- || zero_p (get_iv (data, *op_p)->step))
- op_p = &TREE_OPERAND (cond, 1);
+ ok = extract_cond_operands (data, use->op_p, &var_p, NULL, NULL, NULL);
+ gcc_assert (ok);
- op = force_gimple_operand (comp, &stmts, true, SSA_NAME_VAR (*op_p));
- if (stmts)
- bsi_insert_before (&bsi, stmts, BSI_SAME_STMT);
-
- *op_p = op;
+ *var_p = force_gimple_operand_bsi (&bsi, comp, true, SSA_NAME_VAR (*var_p));
}
/* Rewrites USE using candidate CAND. */
static void
-rewrite_use (struct ivopts_data *data,
- struct iv_use *use, struct iv_cand *cand)
+rewrite_use (struct ivopts_data *data, struct iv_use *use, struct iv_cand *cand)
{
+ push_stmt_changes (&use->stmt);
+
switch (use->type)
{
case USE_NONLINEAR_EXPR:
default:
gcc_unreachable ();
}
- mark_new_vars_to_rename (use->stmt);
+
+ pop_stmt_changes (&use->stmt);
}
/* Rewrite the uses using the selected induction variables. */
info = ver_info (data, j);
if (info->iv
- && !zero_p (info->iv->step)
+ && !integer_zerop (info->iv->step)
&& !info->inv_id
&& !info->iv->have_use_for
&& !info->preserve_biv)
bitmap_iterator bi;
tree obj;
- htab_empty (data->niters);
+ if (data->niters)
+ {
+ pointer_map_destroy (data->niters);
+ data->niters = NULL;
+ }
EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi)
{
free (data->version_info);
BITMAP_FREE (data->relevant);
BITMAP_FREE (data->important_candidates);
- htab_delete (data->niters);
VEC_free (tree, heap, decl_rtl_to_reset);
VEC_free (iv_use_p, heap, data->iv_uses);
struct iv_ca *iv_ca;
edge exit;
+ gcc_assert (!data->niters);
data->current_loop = loop;
if (dump_file && (dump_flags & TDF_DETAILS))
return changed;
}
-/* Main entry point. Optimizes induction variables in LOOPS. */
+/* Main entry point. Optimizes induction variables in loops. */
void
-tree_ssa_iv_optimize (struct loops *loops)
+tree_ssa_iv_optimize (void)
{
struct loop *loop;
struct ivopts_data data;
+ loop_iterator li;
tree_ssa_iv_optimize_init (&data);
/* Optimize the loops starting with the innermost ones. */
- loop = loops->tree_root;
- while (loop->inner)
- loop = loop->inner;
-
- /* Scan the loops, inner ones first. */
- while (loop != loops->tree_root)
+ FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
{
if (dump_file && (dump_flags & TDF_DETAILS))
flow_loop_dump (loop, dump_file, NULL, 1);
tree_ssa_iv_optimize_loop (&data, loop);
-
- if (loop->next)
- {
- loop = loop->next;
- while (loop->inner)
- loop = loop->inner;
- }
- else
- loop = loop->outer;
}
tree_ssa_iv_optimize_finalize (&data);
OSDN Git Service
