/*
- * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
* Implements "Strenght Reduction of Multiplications by Integer Constants" by Youfeng Wu.
* Implements Division and Modulo by Consts from "Hackers Delight",
*/
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
-
-#ifdef HAVE_STDLIB_H
-# include <stdlib.h>
-#endif
+#include "config.h"
+#include <stdlib.h>
#include <assert.h>
#include "irnode_t.h"
#include "irhooks.h"
#include "ircons.h"
#include "irarch.h"
+#include "irflag.h"
+#include "error.h"
#undef DEB
/** The bit mask, which optimizations to apply. */
static arch_dep_opts_t opts;
-/* we need this new pseudo op */
-static ir_op *op_Mulh = NULL;
-
-/**
- * construct a Mulh: Mulh(a,b) = (a * b) >> w, w is the with in bits of a, b
- */
-static ir_node *
-new_rd_Mulh (dbg_info *db, ir_graph *irg, ir_node *block,
- ir_node *op1, ir_node *op2, ir_mode *mode) {
- ir_node *in[2];
- ir_node *res;
-
- in[0] = op1;
- in[1] = op2;
- res = new_ir_node(db, irg, block, op_Mulh, mode, 2, in);
- res = optimize_node(res);
- IRN_VRFY_IRG(res, irg);
- return res;
-}
-
-ir_op *get_op_Mulh(void) { return op_Mulh; }
-
void arch_dep_init(arch_dep_params_factory_t factory) {
opts = arch_dep_none;
if (factory != NULL)
params = factory();
-
- if (! op_Mulh) {
- int mulh_opc = get_next_ir_opcode();
-
- /* create the Mulh operation */
- op_Mulh = new_ir_op(mulh_opc, "Mulh", op_pin_state_floats, irop_flag_commutative, oparity_binary, 0, 0, NULL);
- }
}
void arch_dep_set_opts(arch_dep_opts_t the_opts) {
struct instruction {
insn_kind kind; /**< the instruction kind */
instruction *in[2]; /**< the ins */
- int shift_count; /**< shift count for LEA and SHIFT */
+ unsigned shift_count; /**< shift count for LEA and SHIFT */
ir_node *irn; /**< the generated node for this instruction if any. */
int costs; /**< the costs for this instruction */
};
typedef struct _mul_env {
struct obstack obst; /**< an obstack for local space. */
ir_mode *mode; /**< the mode of the multiplication constant */
- int bits; /**< number of bits in the mode */
+ 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 */
} mul_env;
/**
- * Some kind of default evaluator.
+ * Some kind of default evaluator. Return the cost of
+ * instructions.
*/
static int default_evaluate(insn_kind kind, tarval *tv) {
+ (void) tv;
+
if (kind == MUL)
return 13;
return 1;
/**
* emit a LEA (or an Add) instruction
*/
-static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, int shift) {
+static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) {
instruction *res = obstack_alloc(&env->obst, sizeof(*res));
res->kind = shift > 0 ? LEA : ADD;
res->in[0] = a;
}
/**
- * emit a SHIFT (or an Add) instruction
+ * emit a SHIFT (or an Add or a Zero) instruction
*/
-static instruction *emit_SHIFT(mul_env *env, instruction *a, int shift) {
+static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) {
instruction *res = obstack_alloc(&env->obst, sizeof(*res));
- if (shift != 1) {
+ if (shift == env->bits) {
+ /* a 2^bits with bits resolution is a zero */
+ res->kind = ZERO;
+ res->in[0] = NULL;
+ res->in[1] = NULL;
+ res->shift_count = 0;
+ } else if (shift != 1) {
res->kind = SHIFT;
res->in[0] = a;
res->in[1] = NULL;
* Calculate the gain when using the generalized complementary technique
*/
static int calculate_gain(unsigned char *R, int r) {
- int max_gain = -1;
- int idx, i;
+ int max_gain = 0;
+ int idx = -1, i;
int gain;
/* the gain for r == 1 */
idx = i;
}
}
- if (max_gain > 0)
- return idx;
- return -1;
+ return idx;
}
/**
static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) {
instruction *ins, *ins2;
+ (void) N;
if (r == 1) {
return emit_SHIFT(env, env->root, R[0]);
} else {
assert(r == 2);
ins = env->root;
+ if (R[1] <= env->max_S) {
+ ins = emit_LEA(env, ins, ins, R[1]);
+ if (R[0] != 0) {
+ ins = emit_SHIFT(env, ins, R[0]);
+ }
+ return ins;
+ }
if (R[0] != 0) {
ins = emit_SHIFT(env, ins, R[0]);
}
- if (R[1] <= env->max_S)
- return emit_LEA(env, ins, ins, R[1]);
ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
return emit_LEA(env, ins, ins2, 0);
if (gain > 0) {
instruction *instr1, *instr2;
unsigned char *R1, *R2;
- int r1, r2, i, k;
+ int r1, r2, i, k, j;
R1 = complement_condensed(env, R, r, gain, &r1);
r2 = r - gain + 1;
}
R2[0] = k;
R2[1] = R[gain] - 1;
- for (i = gain; i < r; ++i) {
- R2[i] = R[i];
+ j = 2;
+ if (R2[1] == 0) {
+ /* Two identical bits: normalize */
+ ++R2[0];
+ --j;
+ --r2;
+ }
+ for (i = gain + 1; i < r; ++i) {
+ R2[j++] = R[i];
}
instr1 = decompose_mul(env, R1, r1, NULL);
return basic_decompose_mul(env, R, r, N);
}
+#define IMAX(a,b) ((a) > (b) ? (a) : (b))
+
/**
* basic decomposition routine
*/
unsigned t;
if (R[0] == 0) { /* Case 1 */
- t = R[1] > max(env->max_S, R[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);
case LEA:
l = build_graph(env, inst->in[0]);
r = build_graph(env, inst->in[1]);
- c = new_r_Const(current_ir_graph, env->blk, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode));
+ c = new_Const_long(env->shf_mode, inst->shift_count);
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);
case SHIFT:
l = build_graph(env, inst->in[0]);
- c = new_r_Const(current_ir_graph, env->blk, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode));
+ c = new_Const_long(env->shf_mode, inst->shift_count);
return inst->irn = new_rd_Shl(env->dbg, current_ir_graph, env->blk, l, c, env->mode);
case SUB:
l = build_graph(env, inst->in[0]);
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);
+ case ZERO:
+ return inst->irn = new_Const(get_mode_null(env->mode));
default:
- assert(0);
+ panic("Unsupported instruction kind");
return NULL;
}
}
costs += env->evaluate(inst->kind, NULL);
inst->costs = costs;
return costs;
- default:
- assert(0);
- return 0;
+ case ZERO:
+ inst->costs = costs = env->evaluate(inst->kind, NULL);
+ return costs;
+ case MUL:
+ case ROOT:
+ break;
}
+ panic("Unsupported instruction kind");
}
/**
* Evaluate the replacement instructions and build a new graph
* if faster than the Mul.
- * returns the root of the new graph then or irn otherwise.
+ * Returns the root of the new graph then or irn otherwise.
*
* @param irn the Mul operation
* @param operand the multiplication operand
obstack_init(&env.obst);
env.mode = get_tarval_mode(tv);
- env.bits = get_mode_size_bits(env.mode);
+ env.bits = (unsigned)get_mode_size_bits(env.mode);
env.max_S = 3;
env.root = emit_ROOT(&env, operand);
env.fail = 0;
inst = decompose_mul(&env, R, r, tv);
/* the paper suggests 70% here */
- mul_costs = (env.evaluate(MUL, tv) * 7) / 10;
+ mul_costs = (env.evaluate(MUL, tv) * 7 + 5) / 10;
if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) {
env.op = operand;
env.blk = get_nodes_block(irn);
if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
return irn;
- if (is_Mul(irn) && mode_is_int(mode)) {
- ir_node *block = get_nodes_block(irn);
- 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;
- }
+ 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;
+ }
- 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;
}
#define SHL(a, b) tarval_shl(a, b)
#define SHR(a, b) tarval_shr(a, b)
#define ADD(a, b) tarval_add(a, b)
-#define SUB(a, b) tarval_sub(a, b)
+#define SUB(a, b) tarval_sub(a, b, NULL)
#define MUL(a, b) tarval_mul(a, b)
#define DIV(a, b) tarval_div(a, b)
#define MOD(a, b) tarval_mod(a, b)
struct ms mag = magic(tv);
/* generate the Mulh instruction */
- c = new_r_Const(current_ir_graph, block, mode, mag.M);
+ c = new_Const(mag.M);
q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
/* do we need an Add or Sub */
/* Do we need the shift */
if (mag.s > 0) {
- c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
- q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
+ c = new_Const_long(mode_Iu, mag.s);
+ q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
}
/* final */
- c = new_r_Const_long(current_ir_graph, block, mode_Iu, bits-1);
+ c = new_Const_long(mode_Iu, bits - 1);
t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode);
ir_node *c;
/* generate the Mulh instruction */
- c = new_r_Const(current_ir_graph, block, mode, mag.M);
- q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
+ c = new_Const(mag.M);
+ q = new_rd_Mulh(dbg, current_ir_graph, 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);
- c = new_r_Const(current_ir_graph, block, mode_Iu, get_mode_one(mode_Iu));
+ c = new_Const(get_mode_one(mode_Iu));
t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode);
- c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s-1);
+ c = new_Const_long(mode_Iu, mag.s - 1);
q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
} else {
/* use the default scheme */
q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
}
} else if (mag.s > 0) { /* default scheme, shift needed */
- c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
+ c = new_Const_long(mode_Iu, mag.s);
q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
}
}
if (params == NULL || (opts & arch_dep_div_by_const) == 0)
return irn;
- if (get_irn_opcode(irn) == iro_Div) {
+ if (is_Div(irn)) {
ir_node *c = get_Div_right(irn);
ir_node *block, *left;
ir_mode *mode;
int n, bits;
int k, n_flag;
- if (get_irn_op(c) != op_Const)
+ if (! is_Const(c))
return irn;
tv = get_Const_tarval(c);
/* check for division by zero */
- if (classify_tarval(tv) == TV_CLASSIFY_NULL)
+ if (tarval_is_null(tv))
return irn;
left = get_Div_left(irn);
ir_node *k_node;
ir_node *curr = left;
- if (k != 1) {
- k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
- curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
- }
+ /* 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);
+ curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
+ }
- k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
- curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
+ k_node = new_Const_long(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);
+ curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
+ } else {
+ k_node = left;
+ }
- k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
+ k_node = new_Const_long(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(current_ir_graph, block, mode, get_mode_null(mode));
+ k_node = new_Const(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(current_ir_graph, block, mode_Iu, k);
+ k_node = new_Const_long(mode_Iu, k);
res = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
}
} else {
if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
return irn;
- if (get_irn_opcode(irn) == iro_Mod) {
+ if (is_Mod(irn)) {
ir_node *c = get_Mod_right(irn);
ir_node *block, *left;
ir_mode *mode;
int n, bits;
int k;
- if (get_irn_op(c) != op_Const)
+ if (! is_Const(c))
return irn;
tv = get_Const_tarval(c);
/* check for division by zero */
- if (classify_tarval(tv) == TV_CLASSIFY_NULL)
+ if (tarval_is_null(tv))
return irn;
left = get_Mod_left(irn);
ir_node *curr = left;
if (k != 1) {
- k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
+ k_node = new_Const_long(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, block, mode_Iu, bits - k);
+ k_node = new_Const_long(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);
- k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
+ k_node = new_Const_long(mode, (-1) << k);
curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
res = 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, block, mode, (1 << k) - 1);
+ k_node = new_Const_long(mode, (1 << k) - 1);
res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
}
} else {
((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
return;
- if (get_irn_opcode(irn) == iro_DivMod) {
+ if (is_DivMod(irn)) {
ir_node *c = get_DivMod_right(irn);
ir_node *block, *left;
ir_mode *mode;
int n, bits;
int k, n_flag;
- if (get_irn_op(c) != op_Const)
+ if (! is_Const(c))
return;
tv = get_Const_tarval(c);
/* check for division by zero */
- if (classify_tarval(tv) == TV_CLASSIFY_NULL)
+ if (tarval_is_null(tv))
return;
left = get_DivMod_left(irn);
ir_node *curr = left;
if (k != 1) {
- k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
+ k_node = new_Const_long(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, block, mode_Iu, bits - k);
+ k_node = new_Const_long(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, block, mode_Iu, k);
+ c_k = new_Const_long(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, block, mode, get_mode_null(mode));
+ k_node = new_Const(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, block, mode, (-1) << k);
+ k_node = new_Const_long(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, block, mode_Iu, k);
+ k_node = new_Const_long(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, block, mode, (1 << k) - 1);
+ k_node = new_Const_long(mode, (1 << k) - 1);
*mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
}
} else {
1, /* also use subs */
4, /* maximum shifts */
31, /* maximum shift amount */
+ default_evaluate, /* default evaluator */
0, /* allow Mulhs */
0, /* allow Mulus */