Skip to content

GitLab

Commit ae91f67c authored by Michael Beck's avatar Michael Beck
Browse files

implement the Youfeng Wu algorithm for MulC 

[r15673]
parent 02eb94c0
Hide whitespace changes
Inline Side-by-side
include/libfirm/irarch.h
View file @ ae91f67c
......@@ -29,6 +29,28 @@
#include "firm_types.h"
/**
* The Multiplication replacement can consist of the following instructions.
*/
typedef enum instr {
LEA, /**< the LEA instruction */
SHIFT, /**< the SHIFT instruction */
SUB, /**< the SUB instruction */
ADD, /**< the ADD instruction */
MUL, /**< the original MUL instruction */
ROOT, /**< the ROOT value that is multiplied */
} insn_kind;
/**
* A Callback for evaluating the costs of an instruction.
*
* @param kind the instruction
* @param tv for MUL instruction, the multiplication constant
*
* @return the costs of this instruction
*/
typedef int (*evaluate_costs_func)(insn_kind kind, tarval *tv);
/**
* A parameter structure that drives the machine dependent Firm
* optimizations.
......@@ -40,6 +62,7 @@ struct ir_settings_arch_dep_t {
unsigned highest_shift_amount; /**< The highest shift amount you want to
tolerate. Muls which would require a higher
shift constant are left. */
evaluate_costs_func evaluate; /**< Evaluate the costs of a generated instruction. */
/* Div/Mod optimization */
unsigned allow_mulhs : 1; /**< Use the Mulhs operation for division by constant */
......
ir/ir/irarch.c
View file @ ae91f67c
......@@ -23,6 +23,9 @@
* @date 28.9.2004
* @author Sebastian Hack, Michael Beck
* @version $Id$
*
* 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"
......@@ -113,243 +116,486 @@ static int allow_Mulh(ir_mode *mode) {
return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
}
/* 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);
/**
* An instruction,
*/
typedef struct instruction instruction;
struct instruction {
insn_kind kind; /**< the instruction kind */
instruction *in[2]; /**< the ins */
int 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 */
};
/* 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;
/**
* The environment for the strength reduction of multiplications.
*/
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 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 */
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 */
int n_shift; /**< maximum number of allowed shift instructions */
evaluate_costs_func evaluate; /**< the evaluate callback */
} mul_env;
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;
/**
* Some kind of default evaluator.
*/
static int default_evaluate(insn_kind kind, tarval *tv) {
if (kind == MUL)
return 13;
return 1;
}
/* 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;
}
/**
* emit a LEA (or an Add) instruction
*/
static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, int shift) {
instruction *res = obstack_alloc(&env->obst, sizeof(*res));
res->kind = shift > 0 ? LEA : ADD;
res->in[0] = a;
res->in[1] = b;
res->shift_count = shift;
res->irn = NULL;
res->costs = -1;
return res;
}
if (tv != NULL) {
int maximum_shifts = params->maximum_shifts;
int also_use_subs = params->also_use_subs;
int highest_shift_amount = params->highest_shift_amount;
char *bitstr = get_tarval_bitpattern(tv);
char *p;
int i, last = 0;
int counter = 0;
int curr_bit;
int compr_len = 0;
char compr[MAX_BITSTR];
int singleton;
int end_of_group;
int shift_with_sub[MAX_BITSTR] = { 0 };
int shift_without_sub[MAX_BITSTR] = { 0 };
int shift_with_sub_pos = 0;
int shift_without_sub_pos = 0;
#if DEB
{
long val = get_tarval_long(tv);
fprintf(stderr, "Found mul with %ld(%lx) = ", val, val);
for(p = bitstr; *p != '\0'; p++)
printf("%c", *p);
printf("\n");
}
#endif
/**
* emit a SHIFT (or an Add) instruction
*/
static instruction *emit_SHIFT(mul_env *env, instruction *a, int shift) {
instruction *res = obstack_alloc(&env->obst, sizeof(*res));
if (shift != 1) {
res->kind = SHIFT;
res->in[0] = a;
res->in[1] = NULL;
res->shift_count = shift;
} else {
res->kind = ADD;
res->in[0] = a;
res->in[1] = a;
res->shift_count = 0;
}
res->irn = NULL;
res->costs = -1;
return res;
}
for (p = bitstr; *p != '\0'; p++) {
int bit = *p != '0';
/**
* emit a SUB instruction
*/
static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) {
instruction *res = obstack_alloc(&env->obst, sizeof(*res));
res->kind = SUB;
res->in[0] = a;
res->in[1] = b;
res->shift_count = 0;
res->irn = NULL;
res->costs = -1;
return res;
}
if (bit != last) {
/* The last was 1 we are now at 0 OR
* The last was 0 and we are now at 1 */
compr[compr_len++] = counter;
counter = 1;
} else
counter++;
/**
* emit the ROOT instruction
*/
static instruction *emit_ROOT(mul_env *env, ir_node *root_op) {
instruction *res = obstack_alloc(&env->obst, sizeof(*res));
res->kind = ROOT;
res->in[0] = NULL;
res->in[1] = NULL;
res->shift_count = 0;
res->irn = root_op;
res->costs = 0;
return res;
}
last = bit;
}
compr[compr_len++] = counter;
#ifdef DEB
{
const char *prefix = "";
for(i = 0; i < compr_len; i++, prefix = ",")
fprintf(stderr, "%s%d", prefix, compr[i]);
fprintf("\n");
}
#endif
/* Go over all recorded one groups. */
curr_bit = compr[0];
/**
* Returns the condensed representation of the tarval tv
*/
static unsigned char *value_to_condensed(mul_env *env, 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);
l = r = 0;
for (i = 0; bitstr[i] != '\0'; ++i) {
if (bitstr[i] == '1') {
R[r] = i - l;
l = i;
++r;
}
}
free(bitstr);
for(i = 1; i < compr_len; i = end_of_group + 2) {
int j, zeros_in_group, ones_in_group;
*pr = r;
return R;
}
ones_in_group = compr[i];
zeros_in_group = 0;
/**
* 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 gain;
/* the gain for r == 1 */
gain = 2 - 3 - R[0];
for (i = 2; i < r; ++i) {
/* calculate the gain for r from the gain for r-1 */
gain += 2 - R[i - 1];
if (gain > max_gain) {
max_gain = gain;
idx = i;
}
}
if (max_gain > 0)
return idx;
return -1;
}
/* Scan for singular 0s in a sequence. */
for(j = i + 1; j < compr_len && compr[j] == 1; j += 2) {
zeros_in_group += 1;
ones_in_group += (j + 1 < compr_len ? compr[j + 1] : 0);
}
end_of_group = j - 1;
/**
* 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);
int i, l, j;
unsigned char c;
if(zeros_in_group >= ones_in_group - 1)
end_of_group = i;
memset(value, 0, env->bits);
#ifdef DEB
fprintf(stderr, " i:%d, eg:%d\n", i, end_of_group);
#endif
j = 0;
for (i = 0; i < gain; ++i) {
j += R[i];
value[j] = 1;
}
singleton = compr[i] == 1 && i == end_of_group;
for(j = i; j <= end_of_group; j += 2) {
int curr_ones = compr[j];
int biased_curr_bit = curr_bit + 1;
int k;
/* negate and propagate 1 */
c = 1;
for (i = 0; i <= j; ++i) {
unsigned char v = !value[i];
#ifdef DEB
fprintf(stderr, " j:%d, ones:%d\n", j, curr_ones);
#endif
value[i] = v ^ c;
c = v & c;
}
/* If this ones group is a singleton group (it has no
singleton zeros inside. */
if(singleton)
shift_with_sub[shift_with_sub_pos++] = biased_curr_bit;
else if(j == i)
shift_with_sub[shift_with_sub_pos++] = -biased_curr_bit;
/* condense it again */
l = r = 0;
R = value;
for (i = 0; i <= j; ++i) {
if (value[i] == 1) {
R[r] = i - l;
l = i;
++r;
}
}
for(k = 0; k < curr_ones; k++)
shift_without_sub[shift_without_sub_pos++] = biased_curr_bit + k;
*prs = r;
return R;
}
curr_bit += curr_ones;
biased_curr_bit = curr_bit + 1;
/**
* creates a tarval from a condensed representation.
*/
static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) {
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];
if (j) {
tarval *t = new_tarval_from_long(j, mode_Iu);
tv = tarval_shl(tv, t);
}
res = res ? tarval_add(res, tv) : tv;
}
return res;
}
if(!singleton && j == end_of_group)
shift_with_sub[shift_with_sub_pos++] = biased_curr_bit;
else if(j != end_of_group)
shift_with_sub[shift_with_sub_pos++] = -biased_curr_bit;
/* forward */
static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
curr_bit += compr[j + 1];
}
}
/*
* handle simple cases with up-to 2 bits set
*/
static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) {
instruction *ins, *ins2;
{
int *shifts = shift_with_sub;
int n = shift_with_sub_pos;
int highest_shift_wide = 0;
int highest_shift_seq = 0;
int last_shift = 0;
/* If we may not use subs, or we can achieve the same with adds,
prefer adds. */
if(!also_use_subs || shift_with_sub_pos >= shift_without_sub_pos) {
shifts = shift_without_sub;
n = shift_without_sub_pos;
}
if (r == 1) {
return emit_SHIFT(env, env->root, R[0]);
} else {
assert(r == 2);
/* If the number of needed shifts exceeds the given maximum,
use the Mul and exit. */
if(n > maximum_shifts) {
#ifdef DEB
fprintf(stderr, "Only allowed %d shifts, but %d are needed\n",
maximum_shifts, n);
#endif
goto end;
}
ins = env->root;
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]);
/* Compute the highest shift needed for both, the
sequential and wide representations. */
for(i = 0; i < n; i++) {
int curr = abs(shifts[i]);
int curr_seq = curr - last;
ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
return emit_LEA(env, ins, ins2, 0);
}
}
highest_shift_wide = curr > highest_shift_wide ? curr : highest_shift_wide;
highest_shift_seq = curr_seq > highest_shift_seq ? curr_seq : highest_shift_seq;
/**
* Main decompose driver.
*/
static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
unsigned i;
int gain;
if (r <= 2)
return decompose_simple_cases(env, R, r, N);
if (params->also_use_subs) {
gain = calculate_gain(R, r);
if (gain > 0) {
instruction *instr1, *instr2;
unsigned char *R1, *R2;
int r1, r2, i, k;
R1 = complement_condensed(env, R, r, gain, &r1);
r2 = r - gain + 1;
R2 = obstack_alloc(&env->obst, r2);
k = 1;
for (i = 0; i < gain; ++i) {
k += R[i];
}
R2[0] = k;
R2[1] = R[gain] - 1;
for (i = gain; i < r; ++i) {
R2[i] = R[i];
}
last_shift = curr;
}
instr1 = decompose_mul(env, R1, r1, NULL);
instr2 = decompose_mul(env, R2, r2, NULL);
return emit_SUB(env, instr2, instr1);
}
}
/* If the highest shift amount is greater than the given limit,
give back the Mul */
if(highest_shift_seq > highest_shift_amount) {
#ifdef DEB
fprintf(stderr, "Shift argument %d exceeds maximum %d\n",
highest_shift_seq, highest_shift_amount);
#endif
goto end;
}
if (N == NULL)
N = condensed_to_value(env, R, r);
/* If we have subs, we cannot do sequential. */
if(1 /* also_use_subs */) {
if(n > 0) {
ir_node *curr = NULL;
for (i = env->max_S; i > 0; --i) {
tarval *div_res, *mod_res;
tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
i = n - 1;
div_res = tarval_divmod(N, tv, &mod_res);
if (mod_res == get_mode_null(env->mode)) {
unsigned char *Rs;
int rs;
do {
int curr_shift = shifts[i];
int sub = curr_shift < 0;
int amount = abs(curr_shift) - 1;
ir_node *aux = operand;
Rs = value_to_condensed(env, div_res, &rs);
if (rs < r) {
instruction *N1 = decompose_mul(env, Rs, rs, div_res);
return emit_LEA(env, N1, N1, i);
}
}
}
return basic_decompose_mul(env, R, r, N);
}
assert(amount >= 0 && "What is a negative shift??");
/**
* basic decomposition routine
*/
static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
instruction *Ns;
unsigned t;
if (R[0] == 0) { /* Case 1 */
t = R[1] > max(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 */
t = R[0];
R[1] += t;
Ns = decompose_mul(env, &R[1], r - 1, N);
return emit_LEA(env, Ns, env->root, t);
} else {
t = R[0];
R[0] = 0;
Ns = decompose_mul(env, R, r, N);
return emit_SHIFT(env, Ns, t);
}
}
if (amount != 0) {
ir_node *cnst = new_r_Const_long(current_ir_graph, block, mode_Iu, amount);
aux = new_r_Shl(current_ir_graph, block, operand, cnst, mode);
}
/**
* recursive build the graph form the instructions.
*
* @param env the environment
* @param inst the instruction
*/
static ir_node *build_graph(mul_env *env, instruction *inst) {
ir_node *l, *r, *c;
if (inst->irn)
return inst->irn;
switch (inst->kind) {
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));
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));
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]);
r = build_graph(env, inst->in[1]);
return inst->irn = new_rd_Sub(env->dbg, current_ir_graph, 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);
default:
assert(0);
return NULL;
}
}
if (curr) {
if (sub)
curr = new_r_Sub(current_ir_graph, block, curr, aux, mode);
else
curr = new_r_Add(current_ir_graph, block, curr, aux, mode);
} else
curr = aux;
/**
* 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) {
int costs;
} while(--i >= 0);
if (inst->costs >= 0) {
/* was already evaluated */
return 0;
}
res = curr;
}
}
switch (inst->kind) {
case LEA:
case SUB:
case ADD:
costs = evaluate_insn(env, inst->in[0]);
costs += evaluate_insn(env, inst->in[1]);
costs += env->evaluate(inst->kind, NULL);
inst->costs = costs;
return costs;
case SHIFT:
if (inst->shift_count > params->highest_shift_amount)