* @author Sebastian Hack, Michael Beck
* @version $Id$
*
- * Implements "Strength Reduction of Multiplications by Integer Constants" by Youfeng Wu.
+ * Implements "Strength Reduction of Multiplications by Integer Constants"
+ * by Youfeng Wu.
* Implements Division and Modulo by Consts from "Hackers Delight",
*/
#include "config.h"
#include "ircons.h"
#include "irarch.h"
#include "irflag.h"
+#include "be.h"
#include "error.h"
-#undef DEB
-
-#define MAX_BITSTR 64
-
-/** The params got from the factory in arch_dep_init(...). */
-static const ir_settings_arch_dep_t *params = NULL;
-
/** The bit mask, which optimizations to apply. */
static arch_dep_opts_t opts;
-void arch_dep_init(arch_dep_params_factory_t factory)
-{
- opts = arch_dep_none;
-
- if (factory != NULL)
- params = factory();
-}
-
void arch_dep_set_opts(arch_dep_opts_t the_opts)
{
opts = the_opts;
}
/** check, whether a mode allows a Mulh instruction. */
-static int allow_Mulh(ir_mode *mode)
+static int allow_Mulh(const ir_settings_arch_dep_t *params, ir_mode *mode)
{
if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
return 0;
*/
typedef struct mul_env {
struct obstack obst; /**< an obstack for local space. */
+ const ir_settings_arch_dep_t *params;
ir_mode *mode; /**< the mode of the multiplication constant */
unsigned bits; /**< number of bits in the mode */
unsigned max_S; /**< the maximum LEA shift value. */
instruction *root; /**< the root of the instruction tree */
ir_node *op; /**< the operand that is multiplied */
ir_node *blk; /**< the block where the new graph is built */
+ ir_graph *irg;
dbg_info *dbg; /**< the debug info for the new graph. */
ir_mode *shf_mode; /**< the (unsigned) mode for the shift constants */
int fail; /**< set to 1 if the instruction sequence fails the constraints */
* Some kind of default evaluator. Return the cost of
* instructions.
*/
-static int default_evaluate(insn_kind kind, tarval *tv)
+static int default_evaluate(insn_kind kind, ir_tarval *tv)
{
(void) tv;
/**
* Returns the condensed representation of the tarval tv
*/
-static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr)
+static unsigned char *value_to_condensed(mul_env *env, ir_tarval *tv, int *pr)
{
ir_mode *mode = get_tarval_mode(tv);
int bits = get_mode_size_bits(mode);
/**
* creates a tarval from a condensed representation.
*/
-static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r)
+static ir_tarval *condensed_to_value(mul_env *env, unsigned char *R, int r)
{
- tarval *res, *tv;
+ ir_tarval *res, *tv;
int i, j;
j = 0;
for (i = 0; i < r; ++i) {
j = R[i];
if (j) {
- tarval *t = new_tarval_from_long(j, mode_Iu);
+ ir_tarval *t = new_tarval_from_long(j, mode_Iu);
tv = tarval_shl(tv, t);
}
res = res ? tarval_add(res, tv) : tv;
}
/* forward */
-static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
+static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N);
/*
* handle simple cases with up-to 2 bits set
*/
-static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N)
+static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, ir_tarval *N)
{
instruction *ins, *ins2;
/**
* Main decompose driver.
*/
-static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N)
+static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N)
{
unsigned i;
int gain;
if (r <= 2)
return decompose_simple_cases(env, R, r, N);
- if (params->also_use_subs) {
+ if (env->params->also_use_subs) {
gain = calculate_gain(R, r);
if (gain > 0) {
instruction *instr1, *instr2;
N = condensed_to_value(env, R, r);
for (i = env->max_S; i > 0; --i) {
- tarval *div_res, *mod_res;
- tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
+ ir_tarval *div_res, *mod_res;
+ ir_tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
div_res = tarval_divmod(N, tv, &mod_res);
if (mod_res == get_mode_null(env->mode)) {
/**
* basic decomposition routine
*/
-static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N)
+static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N)
{
instruction *Ns;
unsigned t;
- if (R[0] == 0) { /* Case 1 */
+ if (R[0] == 0) { /* Case 1 */
t = R[1] > IMAX(env->max_S, R[1]);
R[1] -= t;
Ns = decompose_mul(env, &R[1], r - 1, N);
return emit_LEA(env, env->root, Ns, t);
- } else if (R[0] <= env->max_S) { /* Case 2 */
+ } else if (R[0] <= env->max_S) { /* Case 2 */
t = R[0];
R[1] += t;
Ns = decompose_mul(env, &R[1], r - 1, N);
static ir_node *build_graph(mul_env *env, instruction *inst)
{
ir_node *l, *r, *c;
+ ir_graph *irg = env->irg;
if (inst->irn)
return inst->irn;
case LEA:
l = build_graph(env, inst->in[0]);
r = build_graph(env, inst->in[1]);
- c = new_Const_long(env->shf_mode, inst->shift_count);
+ c = new_r_Const_long(irg, env->shf_mode, inst->shift_count);
r = new_rd_Shl(env->dbg, env->blk, r, c, env->mode);
return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
case SHIFT:
l = build_graph(env, inst->in[0]);
- c = new_Const_long(env->shf_mode, inst->shift_count);
+ c = new_r_Const_long(irg, env->shf_mode, inst->shift_count);
return inst->irn = new_rd_Shl(env->dbg, env->blk, l, c, env->mode);
case SUB:
l = build_graph(env, inst->in[0]);
r = build_graph(env, inst->in[1]);
return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
case ZERO:
- return inst->irn = new_Const(get_mode_null(env->mode));
+ return inst->irn = new_r_Const(irg, get_mode_null(env->mode));
default:
panic("Unsupported instruction kind");
}
inst->costs = costs;
return costs;
case SHIFT:
- if (inst->shift_count > params->highest_shift_amount)
+ if (inst->shift_count > env->params->highest_shift_amount)
env->fail = 1;
if (env->n_shift <= 0)
env->fail = 1;
*
* @return the new graph
*/
-static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv)
+static ir_node *do_decomposition(ir_node *irn, ir_node *operand, ir_tarval *tv)
{
mul_env env;
instruction *inst;
int mul_costs;
obstack_init(&env.obst);
+ env.params = be_get_backend_param()->dep_param;
env.mode = get_tarval_mode(tv);
env.bits = (unsigned)get_mode_size_bits(env.mode);
env.max_S = 3;
env.root = emit_ROOT(&env, operand);
env.fail = 0;
- env.n_shift = params->maximum_shifts;
- env.evaluate = params->evaluate != NULL ? params->evaluate : default_evaluate;
+ env.n_shift = env.params->maximum_shifts;
+ env.evaluate = env.params->evaluate != NULL ? env.params->evaluate : default_evaluate;
+ env.irg = get_irn_irg(irn);
R = value_to_condensed(&env, tv, &r);
inst = decompose_mul(&env, R, r, tv);
/* Replace Muls with Shifts and Add/Subs. */
ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn)
{
- ir_graph *irg;
- ir_node *res = irn;
- ir_mode *mode = get_irn_mode(irn);
- ir_node *left;
- ir_node *right;
- ir_node *operand;
- tarval *tv;
+ ir_node *res = irn;
+ ir_mode *mode = get_irn_mode(irn);
+ ir_graph *irg;
+ ir_node *left;
+ ir_node *right;
+ ir_node *operand;
+ ir_tarval *tv;
+ const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
/* If the architecture dependent optimizations were not initialized
/**
* calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
*/
-static int tv_ld2(tarval *tv, int bits)
+static int tv_ld2(ir_tarval *tv, int bits)
{
int i, k = 0, num;
/** The result of a the magic() function. */
struct ms {
- tarval *M; /**< magic number */
+ ir_tarval *M; /**< magic number */
int s; /**< shift amount */
int need_add; /**< an additional add is needed */
int need_sub; /**< an additional sub is needed */
*
* see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
*/
-static struct ms magic(tarval *d)
+static struct ms magic(ir_tarval *d)
{
ir_mode *mode = get_tarval_mode(d);
ir_mode *u_mode = find_unsigned_mode(mode);
int bits = get_mode_size_bits(u_mode);
int p;
- tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
+ ir_tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
pn_Cmp d_cmp, M_cmp;
- tarval *bits_minus_1, *two_bits_1;
+ ir_tarval *bits_minus_1, *two_bits_1;
struct ms mag;
/** The result of the magicu() function. */
struct mu {
- tarval *M; /**< magic add constant */
+ ir_tarval *M; /**< magic add constant */
int s; /**< shift amount */
int need_add; /**< add indicator */
};
*
* see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned)
*/
-static struct mu magicu(tarval *d)
+static struct mu magicu(ir_tarval *d)
{
ir_mode *mode = get_tarval_mode(d);
int bits = get_mode_size_bits(mode);
int p;
- tarval *nc, *delta, *q1, *r1, *q2, *r2;
- tarval *bits_minus_1, *two_bits_1, *seven_ff;
+ ir_tarval *nc, *delta, *q1, *r1, *q2, *r2;
+ ir_tarval *bits_minus_1, *two_bits_1, *seven_ff;
struct mu magu;
*
* Note that 'div' might be a mod or DivMod operation as well
*/
-static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv)
+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);
return div;
if (mode_is_signed(mode)) {
+ ir_graph *irg = get_irn_irg(div);
struct ms mag = magic(tv);
/* generate the Mulh instruction */
- c = new_Const(mag.M);
+ c = new_r_Const(irg, mag.M);
q = new_rd_Mulh(dbg, block, n, c, mode);
/* do we need an Add or Sub */
/* Do we need the shift */
if (mag.s > 0) {
- c = new_Const_long(mode_Iu, mag.s);
+ c = new_r_Const_long(irg, mode_Iu, mag.s);
q = new_rd_Shrs(dbg, block, q, c, mode);
}
/* final */
- c = new_Const_long(mode_Iu, bits - 1);
+ c = new_r_Const_long(irg, mode_Iu, bits - 1);
t = new_rd_Shr(dbg, block, q, c, mode);
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_Const(mag.M);
+ c = new_r_Const(irg, mag.M);
q = new_rd_Mulh(dbg, block, n, c, mode);
if (mag.need_add) {
/* use the GM scheme */
t = new_rd_Sub(dbg, block, n, q, mode);
- c = new_Const(get_mode_one(mode_Iu));
+ c = new_r_Const(irg, get_mode_one(mode_Iu));
t = new_rd_Shr(dbg, block, t, c, mode);
t = new_rd_Add(dbg, block, t, q, mode);
- c = new_Const_long(mode_Iu, mag.s - 1);
+ c = new_r_Const_long(irg, mode_Iu, mag.s - 1);
q = new_rd_Shr(dbg, block, t, c, mode);
} else {
/* use the default scheme */
q = new_rd_Add(dbg, block, q, n, mode);
}
} else if (mag.s > 0) { /* default scheme, shift needed */
- c = new_Const_long(mode_Iu, mag.s);
+ c = new_r_Const_long(irg, mode_Iu, mag.s);
q = new_rd_Shr(dbg, block, q, c, mode);
}
}
/* Replace Divs with Shifts and Add/Subs and Mulh. */
ir_node *arch_dep_replace_div_by_const(ir_node *irn)
{
+ const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
ir_node *res = irn;
/* If the architecture dependent optimizations were not initialized
ir_node *c = get_Div_right(irn);
ir_node *block, *left;
ir_mode *mode;
- tarval *tv, *ntv;
+ ir_tarval *tv, *ntv;
dbg_info *dbg;
int n, bits;
int k;
}
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_Const_long(mode_Iu, 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_Const_long(mode_Iu, bits - k);
+ 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);
k_node = left;
}
- k_node = new_Const_long(mode_Iu, k);
+ k_node = new_r_Const_long(irg, mode_Iu, k);
res = new_rd_Shrs(dbg, block, curr, k_node, mode);
if (n_flag) { /* negate the result */
ir_node *k_node;
- k_node = new_Const(get_mode_null(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_Const_long(mode_Iu, k);
+ k_node = new_r_Const_long(irg, mode_Iu, k);
res = new_rd_Shr(dbg, block, left, k_node, mode);
}
} else {
/* other constant */
- if (allow_Mulh(mode))
+ if (allow_Mulh(params, mode))
res = replace_div_by_mulh(irn, tv);
}
}
/* Replace Mods with Shifts and Add/Subs and Mulh. */
ir_node *arch_dep_replace_mod_by_const(ir_node *irn)
{
+ const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
ir_node *res = irn;
/* If the architecture dependent optimizations were not initialized
ir_node *c = get_Mod_right(irn);
ir_node *block, *left;
ir_mode *mode;
- tarval *tv, *ntv;
+ ir_tarval *tv, *ntv;
dbg_info *dbg;
int n, bits;
int k;
}
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
*/
ir_node *curr = left;
if (k != 1) {
- k_node = new_Const_long(mode_Iu, 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_Const_long(mode_Iu, bits - k);
+ 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);
- k_node = new_Const_long(mode, (-1) << k);
+ k_node = new_r_Const_long(irg, mode, (-1) << k);
curr = new_rd_And(dbg, block, curr, k_node, mode);
res = new_rd_Sub(dbg, block, left, curr, mode);
} else { /* unsigned case */
ir_node *k_node;
- k_node = new_Const_long(mode, (1 << k) - 1);
+ k_node = new_r_Const_long(irg, mode, (1 << k) - 1);
res = new_rd_And(dbg, block, left, k_node, mode);
}
} else {
/* other constant */
- if (allow_Mulh(mode)) {
+ if (allow_Mulh(params, mode)) {
res = replace_div_by_mulh(irn, tv);
res = new_rd_Mul(dbg, block, res, c, mode);
/* 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
ir_node *c = get_DivMod_right(irn);
ir_node *block, *left;
ir_mode *mode;
- tarval *tv, *ntv;
+ ir_tarval *tv, *ntv;
dbg_info *dbg;
int n, bits;
int k;
}
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_Const_long(mode_Iu, 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_Const_long(mode_Iu, bits - k);
+ 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_Const_long(mode_Iu, k);
+ 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;
-
- k_node = new_Const(get_mode_null(mode));
+ 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_Const_long(mode, (-1) << k);
+ 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;
-
- k_node = new_Const_long(mode_Iu, k);
+ 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_Const_long(mode, (1 << k) - 1);
+ 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(mode)) {
+ if (allow_Mulh(params, mode)) {
ir_node *t;
*div = replace_div_by_mulh(irn, tv);
if (*div)
hook_arch_dep_replace_division_by_const(irn);
}
-
-
-static const ir_settings_arch_dep_t default_params = {
- 1, /* also use subs */
- 4, /* maximum shifts */
- 31, /* maximum shift amount */
- default_evaluate, /* default evaluator */
-
- 0, /* allow Mulhs */
- 0, /* allow Mulus */
- 32 /* Mulh allowed up to 32 bit */
-};
-
-/* A default parameter factory for testing purposes. */
-const ir_settings_arch_dep_t *arch_dep_default_factory(void)
-{
- return &default_params;
-}