/*
- * 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 "Strenght 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"
-#ifdef HAVE_STDLIB_H
-# include <stdlib.h>
-#endif
-
+#include <stdlib.h>
#include <assert.h>
#include "irnode_t.h"
#include "iropt_t.h"
#include "ircons_t.h"
#include "irgmod.h"
-#include "irvrfy.h"
+#include "irverify.h"
#include "tv_t.h"
#include "dbginfo_t.h"
#include "iropt_dbg.h"
#include "ircons.h"
#include "irarch.h"
#include "irflag.h"
+#include "be.h"
#include "error.h"
-#undef DEB
-
-#define MAX_BITSTR 64
-
-/* when we need verifying */
-#ifdef NDEBUG
-# define IRN_VRFY_IRG(res, irg)
-#else
-# define IRN_VRFY_IRG(res, irg) irn_vrfy_irg(res, irg)
-#endif
-
-/** 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) {
+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;
return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
/**
* The environment for the strength reduction of multiplications.
*/
-typedef struct _mul_env {
+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, const ir_mode *mode, ir_tarval *tv)
+{
+ (void) mode;
(void) tv;
if (kind == MUL)
/**
* emit a LEA (or an Add) instruction
*/
-static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) {
- instruction *res = obstack_alloc(&env->obst, sizeof(*res));
+static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift)
+{
+ instruction *res = OALLOC(&env->obst, instruction);
res->kind = shift > 0 ? LEA : ADD;
res->in[0] = a;
res->in[1] = b;
/**
* emit a SHIFT (or an Add or a Zero) instruction
*/
-static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) {
- instruction *res = obstack_alloc(&env->obst, sizeof(*res));
+static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift)
+{
+ instruction *res = OALLOC(&env->obst, instruction);
if (shift == env->bits) {
/* a 2^bits with bits resolution is a zero */
res->kind = ZERO;
/**
* emit a SUB instruction
*/
-static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) {
- instruction *res = obstack_alloc(&env->obst, sizeof(*res));
+static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b)
+{
+ instruction *res = OALLOC(&env->obst, instruction);
res->kind = SUB;
res->in[0] = a;
res->in[1] = b;
/**
* emit the ROOT instruction
*/
-static instruction *emit_ROOT(mul_env *env, ir_node *root_op) {
- instruction *res = obstack_alloc(&env->obst, sizeof(*res));
+static instruction *emit_ROOT(mul_env *env, ir_node *root_op)
+{
+ instruction *res = OALLOC(&env->obst, instruction);
res->kind = ROOT;
res->in[0] = NULL;
res->in[1] = NULL;
/**
* 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);
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) {
/**
* Calculate the gain when using the generalized complementary technique
*/
-static int calculate_gain(unsigned char *R, int r) {
+static int calculate_gain(unsigned char *R, int r)
+{
int max_gain = 0;
int idx = -1, i;
int gain;
/**
* Calculates the condensed complement of a given (R,r) tuple
*/
-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);
+static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs)
+{
+ unsigned char *value = (unsigned char*)obstack_alloc(&env->obst, env->bits);
int i, l, j;
unsigned char c;
/**
* creates a tarval from a condensed representation.
*/
-static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) {
- tarval *res, *tv;
+static ir_tarval *condensed_to_value(mul_env *env, unsigned char *R, int r)
+{
+ 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;
(void) N;
/**
* 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;
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) {
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);
* @param env the environment
* @param inst the instruction
*/
-static ir_node *build_graph(mul_env *env, instruction *inst) {
+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_r_Const(current_ir_graph, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode));
- r = new_rd_Shl(env->dbg, current_ir_graph, env->blk, r, c, env->mode);
- return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
+ 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_r_Const(current_ir_graph, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode));
- return inst->irn = new_rd_Shl(env->dbg, current_ir_graph, env->blk, l, c, env->mode);
+ 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_Sub(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
+ return inst->irn = new_rd_Sub(env->dbg, env->blk, l, r, env->mode);
case ADD:
l = build_graph(env, inst->in[0]);
r = build_graph(env, inst->in[1]);
- return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
+ return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
case ZERO:
- return inst->irn = new_r_Const(current_ir_graph, env->mode, get_mode_null(env->mode));
+ return inst->irn = new_r_Const(irg, get_mode_null(env->mode));
default:
panic("Unsupported instruction kind");
- return NULL;
}
}
* Calculate the costs for the given instruction sequence.
* Note that additional costs due to higher register pressure are NOT evaluated yet
*/
-static int evaluate_insn(mul_env *env, instruction *inst) {
+static int evaluate_insn(mul_env *env, instruction *inst)
+{
int costs;
if (inst->costs >= 0) {
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:
- 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;
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:
+ break;
}
panic("Unsupported instruction kind");
}
*
* @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;
unsigned char *R;
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);
/* 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);
}
/* Replace Muls with Shifts and Add/Subs. */
-ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) {
- ir_node *res = irn;
- ir_mode *mode = get_irn_mode(irn);
+ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn)
+{
+ 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
or this optimization was not enabled. */
if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
- return irn;
+ return res;
+
+ 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);
+ add_irg_constraints(irg, IR_GRAPH_CONSTRAINT_ARCH_DEP);
+
+ left = get_binop_left(irn);
+ right = get_binop_right(irn);
+ tv = NULL;
+ operand = NULL;
+
+ /* Look, if one operand is a constant. */
+ if (is_Const(left)) {
+ tv = get_Const_tarval(left);
+ operand = right;
+ } else if (is_Const(right)) {
+ tv = get_Const_tarval(right);
+ operand = left;
+ }
- set_arch_dep_running(1);
- {
- if (is_Mul(irn) && mode_is_int(mode)) {
- ir_node *left = get_binop_left(irn);
- ir_node *right = get_binop_right(irn);
- tarval *tv = NULL;
- ir_node *operand = NULL;
-
- /* Look, if one operand is a constant. */
- if (is_Const(left)) {
- tv = get_Const_tarval(left);
- operand = right;
- } else if (is_Const(right)) {
- tv = get_Const_tarval(right);
- 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);
+ if (tv != NULL) {
+ res = do_decomposition(irn, operand, tv);
- if (res != irn) {
- hook_arch_dep_replace_mul_with_shifts(irn);
- exchange(irn, res);
- }
- }
+ if (res != irn) {
+ hook_arch_dep_replace_mul_with_shifts(irn);
+ exchange(irn, res);
}
}
- //set_arch_dep_running(0);
return res;
}
/**
* 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;
for (num = i = 0; i < bits; ++i) {
/** 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 */
- pn_Cmp d_cmp, M_cmp;
+ ir_tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
+ ir_relation d_cmp, M_cmp;
- tarval *bits_minus_1, *two_bits_1;
+ ir_tarval *bits_minus_1, *two_bits_1;
struct ms mag;
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);
/** 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;
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, 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);
- 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))
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_r_Const(current_ir_graph, mode, mag.M);
- q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
+ 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 */
if (mag.need_add)
- q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
+ q = new_rd_Add(dbg, block, q, n, mode);
else if (mag.need_sub)
- q = new_rd_Sub(dbg, current_ir_graph, block, q, n, mode);
+ q = new_rd_Sub(dbg, block, q, n, mode);
/* Do we need the shift */
if (mag.s > 0) {
- c = new_r_Const_long(current_ir_graph, mode_Iu, mag.s);
- q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
+ c = new_r_Const_long(irg, mode_Iu, mag.s);
+ q = new_rd_Shrs(dbg, block, q, c, mode);
}
/* final */
- c = new_r_Const_long(current_ir_graph, mode_Iu, bits-1);
- t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
+ c = new_r_Const_long(irg, mode_Iu, bits - 1);
+ t = new_rd_Shr(dbg, block, q, c, mode);
- q = new_rd_Add(dbg, current_ir_graph, block, q, t, 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_r_Const(current_ir_graph, mode, mag.M);
- q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
+ 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, current_ir_graph, block, n, q, mode);
+ ir_node *t = new_rd_Sub(dbg, block, n, q, mode);
- c = new_r_Const(current_ir_graph, mode_Iu, get_mode_one(mode_Iu));
- t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
+ c = new_r_Const(irg, get_mode_one(mode_Iu));
+ t = new_rd_Shr(dbg, block, t, c, mode);
- t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode);
+ t = new_rd_Add(dbg, block, t, q, mode);
- c = new_r_Const_long(current_ir_graph, mode_Iu, mag.s-1);
- q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
+ 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, current_ir_graph, block, q, n, mode);
+ q = new_rd_Add(dbg, block, q, n, mode);
}
} else if (mag.s > 0) { /* default scheme, shift needed */
- c = new_r_Const_long(current_ir_graph, mode_Iu, mag.s);
- q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
+ c = new_r_Const_long(irg, mode_Iu, mag.s);
+ q = new_rd_Shr(dbg, block, q, c, mode);
}
}
return q;
}
/* Replace Divs with Shifts and Add/Subs and Mulh. */
-ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
+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
- 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;
- tarval *tv, *ntv;
- dbg_info *dbg;
- int n, bits;
- int k, n_flag;
+ 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 */
- if (mode_is_signed(mode)) {
- ir_node *k_node;
- ir_node *curr = left;
+ 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);
+ }
- /* create the correction code for signed values only if there might be a remainder */
- if (! is_Div_remainderless(irn)) {
- if (k != 1) {
- k_node = new_r_Const_long(current_ir_graph, mode_Iu, k - 1);
- curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
- }
+ if (k < 0) {
+ n_flag = 0;
+ k = tv_ld2(tv, n);
+ }
- k_node = new_r_Const_long(current_ir_graph, mode_Iu, bits - k);
- curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
+ 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;
- curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
- } else {
- k_node = 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(current_ir_graph, mode_Iu, k);
- res = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode);
-
- if (n_flag) { /* negate the result */
- ir_node *k_node;
+ 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.
+ */
+
+ curr = new_rd_Add(dbg, block, left, curr, mode);
+ }
- k_node = new_r_Const(current_ir_graph, mode, get_mode_null(mode));
- res = new_rd_Sub(dbg, current_ir_graph, 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(current_ir_graph, mode_Iu, k);
- res = new_rd_Shr(dbg, current_ir_graph, 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(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)
}
/* Replace Mods with Shifts and Add/Subs and Mulh. */
-ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
+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;
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
*/
ir_node *curr = left;
if (k != 1) {
- k_node = new_r_Const_long(current_ir_graph, mode_Iu, k - 1);
- curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
+ 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(current_ir_graph, mode_Iu, bits - k);
- curr = new_rd_Shr(dbg, current_ir_graph, 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 = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
+ curr = new_rd_Add(dbg, block, left, curr, mode);
- k_node = new_r_Const_long(current_ir_graph, mode, (-1) << k);
- curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
+ 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, current_ir_graph, block, left, curr, mode);
+ res = new_rd_Sub(dbg, block, left, curr, mode);
} else { /* unsigned case */
ir_node *k_node;
- k_node = new_r_Const_long(current_ir_graph, mode, (1 << k) - 1);
- res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
+ 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, current_ir_graph, block, res, c, mode);
+ res = new_rd_Mul(dbg, block, res, c, mode);
/* res = arch_dep_mul_to_shift(res); */
- res = new_rd_Sub(dbg, current_ir_graph, block, left, res, mode);
+ res = new_rd_Sub(dbg, block, left, res, mode);
}
}
}
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) {
- *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;
- tarval *tv, *ntv;
- dbg_info *dbg;
- int n, bits;
- int k, n_flag;
-
- 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 */
- if (mode_is_signed(mode)) {
- ir_node *k_node, *c_k;
- ir_node *curr = left;
-
- if (k != 1) {
- k_node = new_r_Const_long(current_ir_graph, mode_Iu, k - 1);
- curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
- }
-
- k_node = new_r_Const_long(current_ir_graph, mode_Iu, bits - k);
- curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
-
- curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
-
- c_k = new_r_Const_long(current_ir_graph, mode_Iu, k);
-
- *div = new_rd_Shrs(dbg, current_ir_graph, block, curr, c_k, mode);
-
- if (n_flag) { /* negate the div result */
- ir_node *k_node;
-
- k_node = new_r_Const(current_ir_graph, mode, get_mode_null(mode));
- *div = new_rd_Sub(dbg, current_ir_graph, block, k_node, *div, mode);
- }
-
- k_node = new_r_Const_long(current_ir_graph, mode, (-1) << k);
- curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
-
- *mod = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
- } else { /* unsigned case */
- ir_node *k_node;
-
- k_node = new_r_Const_long(current_ir_graph, mode_Iu, k);
- *div = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
-
- k_node = new_r_Const_long(current_ir_graph, mode, (1 << k) - 1);
- *mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
- }
- } else {
- /* other constant */
- if (allow_Mulh(mode)) {
- ir_node *t;
-
- *div = replace_div_by_mulh(irn, tv);
-
- t = new_rd_Mul(dbg, current_ir_graph, block, *div, c, mode);
-
- /* t = arch_dep_mul_to_shift(t); */
-
- *mod = new_rd_Sub(dbg, current_ir_graph, block, left, t, mode);
- }
- }
- }
-
- 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;
-}