/*
- * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
* @brief Machine dependent Firm optimizations.
* @date 28.9.2004
* @author Sebastian Hack, Michael Beck
- * @version $Id$
*
* Implements "Strength Reduction of Multiplications by Integer Constants"
* by Youfeng Wu.
* Some kind of default evaluator. Return the cost of
* instructions.
*/
-static int default_evaluate(insn_kind kind, ir_tarval *tv)
+static int default_evaluate(insn_kind kind, const ir_mode *mode, ir_tarval *tv)
{
+ (void) mode;
(void) tv;
if (kind == MUL)
int bits = get_mode_size_bits(mode);
char *bitstr = get_tarval_bitpattern(tv);
int i, l, r;
- unsigned char *R = obstack_alloc(&env->obst, bits);
+ unsigned char *R = (unsigned char*)obstack_alloc(&env->obst, bits);
l = r = 0;
for (i = 0; bitstr[i] != '\0'; ++i) {
*/
static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs)
{
- unsigned char *value = obstack_alloc(&env->obst, env->bits);
+ unsigned char *value = (unsigned char*)obstack_alloc(&env->obst, env->bits);
int i, l, j;
unsigned char c;
*/
static ir_tarval *condensed_to_value(mul_env *env, unsigned char *R, int r)
{
- ir_tarval *res, *tv;
- int i, j;
-
- j = 0;
- tv = get_mode_one(env->mode);
- res = NULL;
- for (i = 0; i < r; ++i) {
- j = R[i];
+ ir_tarval *tv = get_mode_one(env->mode);
+ ir_tarval *res = NULL;
+ for (int i = 0; i < r; ++i) {
+ int j = R[i];
if (j) {
ir_tarval *t = new_tarval_from_long(j, mode_Iu);
tv = tarval_shl(tv, t);
R1 = complement_condensed(env, R, r, gain, &r1);
r2 = r - gain + 1;
- R2 = obstack_alloc(&env->obst, r2);
+ R2 = (unsigned char*)obstack_alloc(&env->obst, r2);
k = 1;
for (i = 0; i < gain; ++i) {
case ADD:
costs = evaluate_insn(env, inst->in[0]);
costs += evaluate_insn(env, inst->in[1]);
- costs += env->evaluate(inst->kind, NULL);
+ costs += env->evaluate(inst->kind, env->mode, NULL);
inst->costs = costs;
return costs;
case SHIFT:
else
--env->n_shift;
costs = evaluate_insn(env, inst->in[0]);
- costs += env->evaluate(inst->kind, NULL);
+ costs += env->evaluate(inst->kind, env->mode, NULL);
inst->costs = costs;
return costs;
case ZERO:
- inst->costs = costs = env->evaluate(inst->kind, NULL);
+ inst->costs = costs = env->evaluate(inst->kind, env->mode, NULL);
return costs;
case MUL:
case ROOT:
inst = decompose_mul(&env, R, r, tv);
/* the paper suggests 70% here */
- mul_costs = (env.evaluate(MUL, tv) * 7 + 5) / 10;
+ mul_costs = (env.evaluate(MUL, env.mode, tv) * 7 + 5) / 10;
if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) {
env.op = operand;
env.blk = get_nodes_block(irn);
ir_tarval *tv;
const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
-
/* If the architecture dependent optimizations were not initialized
or this optimization was not enabled. */
if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
- return irn;
+ return res;
- if (!is_Mul(irn) || !mode_is_int(mode))
+ assert(is_Mul(irn));
+ if (!mode_is_int(mode))
return res;
/* we should never do the reverse transformations again
(like x+x -> 2*x) */
irg = get_irn_irg(irn);
- set_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP);
+ add_irg_constraints(irg, IR_GRAPH_CONSTRAINT_ARCH_DEP);
left = get_binop_left(irn);
right = get_binop_right(irn);
operand = left;
}
+ /* multiplications with 0 are a special case which we leave for
+ * equivalent_node_Mul because the code here can't handle them */
+ if (tv == get_mode_null(mode))
+ return res;
+
if (tv != NULL) {
res = do_decomposition(irn, operand, tv);
int bits = get_mode_size_bits(u_mode);
int p;
ir_tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
- pn_Cmp d_cmp, M_cmp;
+ ir_relation d_cmp, M_cmp;
ir_tarval *bits_minus_1, *two_bits_1;
q1 = ADD(q1, q1); /* Update q1 = 2^p/|nc| */
r1 = ADD(r1, r1); /* Update r1 = rem(2^p, |nc|) */
- if (CMP(r1, anc) & pn_Cmp_Ge) {
+ if (CMP(r1, anc) & ir_relation_greater_equal) {
q1 = ADD(q1, ONE(u_mode));
r1 = SUB(r1, anc);
}
q2 = ADD(q2, q2); /* Update q2 = 2^p/|d| */
r2 = ADD(r2, r2); /* Update r2 = rem(2^p, |d|) */
- if (CMP(r2, ad) & pn_Cmp_Ge) {
+ if (CMP(r2, ad) & ir_relation_greater_equal) {
q2 = ADD(q2, ONE(u_mode));
r2 = SUB(r2, ad);
}
delta = SUB(ad, r2);
- } while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq));
+ } while (CMP(q1, delta) & ir_relation_less || (CMP(q1, delta) & ir_relation_equal && CMP(r1, ZERO(u_mode)) & ir_relation_equal));
d_cmp = CMP(d, ZERO(mode));
- if (d_cmp & pn_Cmp_Ge)
+ if (d_cmp & ir_relation_greater_equal)
mag.M = ADD(CNV(q2, mode), ONE(mode));
else
mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode)));
mag.s = p - bits;
/* need an add if d > 0 && M < 0 */
- mag.need_add = d_cmp & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt;
+ mag.need_add = d_cmp & ir_relation_greater && M_cmp & ir_relation_less;
/* need a sub if d < 0 && M > 0 */
- mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt;
+ mag.need_sub = d_cmp & ir_relation_less && M_cmp & ir_relation_greater;
tarval_set_integer_overflow_mode(rem);
do {
++p;
- if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
+ if (CMP(r1, SUB(nc, r1)) & ir_relation_greater_equal) {
q1 = ADD(ADD(q1, q1), ONE(mode));
r1 = SUB(ADD(r1, r1), nc);
}
r1 = ADD(r1, r1);
}
- if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
- if (CMP(q2, seven_ff) & pn_Cmp_Ge)
+ if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & ir_relation_greater_equal) {
+ if (CMP(q2, seven_ff) & ir_relation_greater_equal)
magu.need_add = 1;
q2 = ADD(ADD(q2, q2), ONE(mode));
r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
}
else {
- if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
+ if (CMP(q2, two_bits_1) & ir_relation_greater_equal)
magu.need_add = 1;
q2 = ADD(q2, q2);
}
delta = SUB(SUB(d, ONE(mode)), r2);
} while (p < 2*bits &&
- (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
+ (CMP(q1, delta) & ir_relation_less || (CMP(q1, delta) & ir_relation_equal && CMP(r1, ZERO(mode)) & ir_relation_equal)));
magu.M = ADD(q2, ONE(mode)); /* Magic number */
magu.s = p - bits; /* and shift amount */
/**
* Build the Mulh replacement code for n / tv.
*
- * Note that 'div' might be a mod or DivMod operation as well
+ * Note that 'div' might be a Mod operation as well
*/
static ir_node *replace_div_by_mulh(ir_node *div, ir_tarval *tv)
{
dbg_info *dbg = get_irn_dbg_info(div);
ir_node *n = get_binop_left(div);
- ir_node *block = get_irn_n(div, -1);
+ ir_node *block = get_nodes_block(div);
ir_mode *mode = get_irn_mode(n);
int bits = get_mode_size_bits(mode);
- ir_node *q, *t, *c;
+ ir_node *q;
/* Beware: do not transform bad code */
if (is_Bad(n) || is_Bad(block))
struct ms mag = magic(tv);
/* generate the Mulh instruction */
- c = new_r_Const(irg, mag.M);
+ ir_node *c = new_r_Const(irg, mag.M);
+ ir_node *t;
q = new_rd_Mulh(dbg, block, n, c, mode);
/* do we need an Add or Sub */
q = new_rd_Add(dbg, block, q, t, mode);
} else {
struct mu mag = magicu(tv);
- ir_node *c;
ir_graph *irg = get_irn_irg(div);
/* generate the Mulh instruction */
- c = new_r_Const(irg, mag.M);
+ ir_node *c = new_r_Const(irg, mag.M);
q = new_rd_Mulh(dbg, block, n, c, mode);
if (mag.need_add) {
if (mag.s > 0) {
/* use the GM scheme */
- t = new_rd_Sub(dbg, block, n, q, mode);
+ ir_node *t = new_rd_Sub(dbg, block, n, q, mode);
c = new_r_Const(irg, get_mode_one(mode_Iu));
t = new_rd_Shr(dbg, block, t, c, mode);
ir_node *res = irn;
/* If the architecture dependent optimizations were not initialized
- or this optimization was not enabled. */
+ or this optimization was not enabled. */
if (params == NULL || (opts & arch_dep_div_by_const) == 0)
return irn;
- if (is_Div(irn)) {
- ir_node *c = get_Div_right(irn);
- ir_node *block, *left;
- ir_mode *mode;
- ir_tarval *tv, *ntv;
- dbg_info *dbg;
- int n, bits;
- int k;
- int n_flag = 0;
+ if (!is_Div(irn))
+ return irn;
- if (! is_Const(c))
- return irn;
+ ir_node *c = get_Div_right(irn);
+ ir_node *block, *left;
+ ir_mode *mode;
+ ir_tarval *tv, *ntv;
+ dbg_info *dbg;
+ int n, bits;
+ int k;
+ int n_flag = 0;
- tv = get_Const_tarval(c);
+ if (! is_Const(c))
+ return irn;
- /* check for division by zero */
- if (tarval_is_null(tv))
- return irn;
+ tv = get_Const_tarval(c);
- left = get_Div_left(irn);
- mode = get_irn_mode(left);
- block = get_irn_n(irn, -1);
- dbg = get_irn_dbg_info(irn);
+ /* check for division by zero */
+ if (tarval_is_null(tv))
+ return irn;
- bits = get_mode_size_bits(mode);
- n = (bits + 7) / 8;
+ left = get_Div_left(irn);
+ mode = get_irn_mode(left);
- k = -1;
- if (mode_is_signed(mode)) {
- /* for signed divisions, the algorithm works for a / -2^k by negating the result */
- ntv = tarval_neg(tv);
- n_flag = 1;
- k = tv_ld2(ntv, n);
- }
+ /* can only handle integer Div's */
+ if (!mode_is_int(mode))
+ return irn;
- if (k < 0) {
- n_flag = 0;
- k = tv_ld2(tv, n);
- }
+ block = get_nodes_block(irn);
+ dbg = get_irn_dbg_info(irn);
- if (k >= 0) { /* division by 2^k or -2^k */
- ir_graph *irg = get_irn_irg(irn);
- if (mode_is_signed(mode)) {
- ir_node *k_node;
- ir_node *curr = left;
+ bits = get_mode_size_bits(mode);
+ n = (bits + 7) / 8;
- /* create the correction code for signed values only if there might be a remainder */
- if (! get_Div_no_remainder(irn)) {
- if (k != 1) {
- k_node = new_r_Const_long(irg, mode_Iu, k - 1);
- curr = new_rd_Shrs(dbg, block, left, k_node, mode);
- }
+ k = -1;
+ if (mode_is_signed(mode)) {
+ /* for signed divisions, the algorithm works for a / -2^k by negating the result */
+ ntv = tarval_neg(tv);
+ n_flag = 1;
+ k = tv_ld2(ntv, n);
+ }
- k_node = new_r_Const_long(irg, mode_Iu, bits - k);
- curr = new_rd_Shr(dbg, block, curr, k_node, mode);
+ if (k < 0) {
+ n_flag = 0;
+ k = tv_ld2(tv, n);
+ }
- curr = new_rd_Add(dbg, block, left, curr, mode);
- } else {
- k_node = left;
+ if (k > 0) { /* division by 2^k or -2^k */
+ ir_graph *irg = get_irn_irg(irn);
+ if (mode_is_signed(mode)) {
+ ir_node *k_node;
+ ir_node *curr = left;
+
+ /* create the correction code for signed values only if there might be a remainder */
+ if (! get_Div_no_remainder(irn)) {
+ if (k != 1) {
+ k_node = new_r_Const_long(irg, mode_Iu, k - 1);
+ curr = new_rd_Shrs(dbg, block, left, k_node, mode);
}
- k_node = new_r_Const_long(irg, mode_Iu, k);
- res = new_rd_Shrs(dbg, block, curr, k_node, mode);
+ k_node = new_r_Const_long(irg, mode_Iu, bits - k);
+ curr = new_rd_Shr(dbg, block, curr, k_node, mode);
+ /* curr is now 2^(k-1) in case left < 0
+ * or 0 in case left >= 0
+ *
+ * For an example, where this fixup is necessary consider -3 / 2,
+ * which should compute to -1,
+ * but simply shifting right by one computes -2.
+ */
- if (n_flag) { /* negate the result */
- ir_node *k_node;
+ curr = new_rd_Add(dbg, block, left, curr, mode);
+ }
- k_node = new_r_Const(irg, get_mode_null(mode));
- res = new_rd_Sub(dbg, block, k_node, res, mode);
- }
- } else { /* unsigned case */
- ir_node *k_node;
+ k_node = new_r_Const_long(irg, mode_Iu, k);
+ res = new_rd_Shrs(dbg, block, curr, k_node, mode);
- k_node = new_r_Const_long(irg, mode_Iu, k);
- res = new_rd_Shr(dbg, block, left, k_node, mode);
+ if (n_flag) { /* negate the result */
+ k_node = new_r_Const(irg, get_mode_null(mode));
+ res = new_rd_Sub(dbg, block, k_node, res, mode);
}
- } else {
- /* other constant */
- if (allow_Mulh(params, mode))
- res = replace_div_by_mulh(irn, tv);
+ } else { /* unsigned case */
+ ir_node *k_node;
+
+ k_node = new_r_Const_long(irg, mode_Iu, k);
+ res = new_rd_Shr(dbg, block, left, k_node, mode);
}
+ } else if (k != 0) {
+ /* other constant */
+ if (allow_Mulh(params, mode))
+ res = replace_div_by_mulh(irn, tv);
+ } else { /* k == 0 i.e. division by 1 */
+ res = left;
}
if (res != irn)
left = get_Mod_left(irn);
mode = get_irn_mode(left);
- block = get_irn_n(irn, -1);
+ block = get_nodes_block(irn);
dbg = get_irn_dbg_info(irn);
bits = get_mode_size_bits(mode);
n = (bits + 7) / 8;
k = tv_ld2(tv, n);
}
- if (k >= 0) {
+ /* k == 0 i.e. modulo by 1 */
+ if (k == 0) {
+ ir_graph *irg = get_irn_irg(irn);
+
+ res = new_r_Const(irg, get_mode_null(mode));
+ }
+ else if (k > 0) {
ir_graph *irg = get_irn_irg(irn);
/* division by 2^k or -2^k:
* we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k
return res;
}
-
-/* Replace DivMods with Shifts and Add/Subs and Mulh. */
-void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn)
-{
- const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
- *div = *mod = NULL;
-
- /* If the architecture dependent optimizations were not initialized
- or this optimization was not enabled. */
- if (params == NULL ||
- ((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
- return;
-
- if (is_DivMod(irn)) {
- ir_node *c = get_DivMod_right(irn);
- ir_node *block, *left;
- ir_mode *mode;
- ir_tarval *tv, *ntv;
- dbg_info *dbg;
- int n, bits;
- int k;
- int n_flag = 0;
-
- if (! is_Const(c))
- return;
-
- tv = get_Const_tarval(c);
-
- /* check for division by zero */
- if (tarval_is_null(tv))
- return;
-
- left = get_DivMod_left(irn);
- mode = get_irn_mode(left);
- block = get_irn_n(irn, -1);
- dbg = get_irn_dbg_info(irn);
-
- bits = get_mode_size_bits(mode);
- n = (bits + 7) / 8;
-
- k = -1;
- if (mode_is_signed(mode)) {
- /* for signed divisions, the algorithm works for a / -2^k by negating the result */
- ntv = tarval_neg(tv);
- n_flag = 1;
- k = tv_ld2(ntv, n);
- }
-
- if (k < 0) {
- n_flag = 0;
- k = tv_ld2(tv, n);
- }
-
- if (k >= 0) { /* division by 2^k or -2^k */
- ir_graph *irg = get_irn_irg(irn);
- if (mode_is_signed(mode)) {
- ir_node *k_node, *c_k;
- ir_node *curr = left;
-
- if (k != 1) {
- k_node = new_r_Const_long(irg, mode_Iu, k - 1);
- curr = new_rd_Shrs(dbg, block, left, k_node, mode);
- }
-
- k_node = new_r_Const_long(irg, mode_Iu, bits - k);
- curr = new_rd_Shr(dbg, block, curr, k_node, mode);
-
- curr = new_rd_Add(dbg, block, left, curr, mode);
-
- c_k = new_r_Const_long(irg, mode_Iu, k);
-
- *div = new_rd_Shrs(dbg, block, curr, c_k, mode);
-
- if (n_flag) { /* negate the div result */
- ir_node *k_node = new_r_Const(irg, get_mode_null(mode));
- *div = new_rd_Sub(dbg, block, k_node, *div, mode);
- }
-
- k_node = new_r_Const_long(irg, mode, (-1) << k);
- curr = new_rd_And(dbg, block, curr, k_node, mode);
-
- *mod = new_rd_Sub(dbg, block, left, curr, mode);
- } else { /* unsigned case */
- ir_node *k_node = new_r_Const_long(irg, mode_Iu, k);
- *div = new_rd_Shr(dbg, block, left, k_node, mode);
-
- k_node = new_r_Const_long(irg, mode, (1 << k) - 1);
- *mod = new_rd_And(dbg, block, left, k_node, mode);
- }
- } else {
- /* other constant */
- if (allow_Mulh(params, mode)) {
- ir_node *t;
-
- *div = replace_div_by_mulh(irn, tv);
-
- t = new_rd_Mul(dbg, block, *div, c, mode);
-
- /* t = arch_dep_mul_to_shift(t); */
-
- *mod = new_rd_Sub(dbg, block, left, t, mode);
- }
- }
- }
-
- if (*div)
- hook_arch_dep_replace_division_by_const(irn);
-}