X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firarch.c;h=f51c9b40aa020393870563b90355ffbd144efc97;hb=9d3c8631459f431c313160dab5778e8a7b88dd92;hp=3e60f090b736cd324aa5d4e4dac92c46f94b532a;hpb=a8372ea8e05bc77b5be9b65411ea2acf2b586cc3;p=libfirm diff --git a/ir/ir/irarch.c b/ir/ir/irarch.c index 3e60f090b..f51c9b40a 100644 --- a/ir/ir/irarch.c +++ b/ir/ir/irarch.c @@ -1,11 +1,34 @@ - /** - * @file irarch.c - * @date 28.9.2004 - * @author Sebastian Hack - * @brief Machine dependent firm optimizations. - * - * $Id$ - */ +/* + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. + * + * This file is part of libFirm. + * + * This file may be distributed and/or modified under the terms of the + * GNU General Public License version 2 as published by the Free Software + * Foundation and appearing in the file LICENSE.GPL included in the + * packaging of this file. + * + * Licensees holding valid libFirm Professional Edition licenses may use + * this file in accordance with the libFirm Commercial License. + * Agreement provided with the Software. + * + * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE + * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR + * PURPOSE. + */ + +/** + * @file + * @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. + * Implements Division and Modulo by Consts from "Hackers Delight", + */ +#include "config.h" #include #include @@ -16,294 +39,1159 @@ #include "iropt_t.h" #include "ircons_t.h" #include "irgmod.h" -#include "irvrfy.h" -#include "tv.h" +#include "irverify.h" +#include "tv_t.h" #include "dbginfo_t.h" #include "iropt_dbg.h" #include "irflag_t.h" -#include "firmstat.h" +#include "irhooks.h" #include "ircons.h" #include "irarch.h" +#include "irflag.h" +#include "be.h" +#include "error.h" +/** The bit mask, which optimizations to apply. */ +static arch_dep_opts_t opts; -#undef DEB +void arch_dep_set_opts(arch_dep_opts_t the_opts) +{ + opts = the_opts; +} -#define MAX_BITSTR 64 +/** check, whether a mode allows a Mulh instruction. */ +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 params got from the factopry in arch_dep_init(...). */ -static const arch_dep_params_t *params = NULL; +/** + * An instruction, + */ +typedef struct instruction instruction; +struct instruction { + insn_kind kind; /**< the instruction kind */ + instruction *in[2]; /**< the ins */ + 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 */ +}; -/** The bit mask, which optimizations to apply. */ -static arch_dep_opts_t opts; +/** + * The environment for the strength reduction of multiplications. + */ +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 */ + int n_shift; /**< maximum number of allowed shift instructions */ -void arch_dep_init(arch_dep_params_factory_t factory) + evaluate_costs_func evaluate; /**< the evaluate callback */ +} mul_env; + +/** + * Some kind of default evaluator. Return the cost of + * instructions. + */ +static int default_evaluate(insn_kind kind, ir_tarval *tv) { - opts = arch_dep_none; + (void) tv; - if(factory != NULL) - params = factory(); + if (kind == MUL) + return 13; + return 1; } -void arch_dep_set_opts(arch_dep_opts_t the_opts) { - opts = the_opts; +/** + * emit a LEA (or an Add) instruction + */ +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; + res->shift_count = shift; + res->irn = NULL; + res->costs = -1; + return res; } -ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) +/** + * emit a SHIFT (or an Add or a Zero) instruction + */ +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; + 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; + 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; +} + +/** + * emit a SUB instruction + */ +static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) { - ir_node *block = get_nodes_block(irn); - ir_node *res = irn; - ir_node *operand = NULL; - ir_node *left, *right; - ir_mode *mode = get_irn_mode(irn); - tarval *tv = NULL; + instruction *res = OALLOC(&env->obst, instruction); + res->kind = SUB; + res->in[0] = a; + res->in[1] = b; + res->shift_count = 0; + res->irn = NULL; + res->costs = -1; + return res; +} - /* 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; +/** + * emit the ROOT instruction + */ +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; + res->shift_count = 0; + res->irn = root_op; + res->costs = 0; + return res; +} + + +/** + * Returns the condensed representation of the tarval tv + */ +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); + + l = r = 0; + for (i = 0; bitstr[i] != '\0'; ++i) { + if (bitstr[i] == '1') { + R[r] = i - l; + l = i; + ++r; + } + } + free(bitstr); - if(is_ir_node(irn) - && get_irn_opcode(irn) == iro_Mul - && mode_is_int(mode)) { + *pr = r; + return R; +} + +/** + * Calculate the gain when using the generalized complementary technique + */ +static int calculate_gain(unsigned char *R, int r) +{ + int max_gain = 0; + int idx = -1, 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; + } + } + return idx; +} + +/** + * 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; + + memset(value, 0, env->bits); + + j = 0; + for (i = 0; i < gain; ++i) { + j += R[i]; + value[j] = 1; + } + + /* negate and propagate 1 */ + c = 1; + for (i = 0; i <= j; ++i) { + unsigned char v = !value[i]; + + value[i] = v ^ c; + c = v & c; + } + + /* 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; + } + } - left = get_binop_left(irn); - right = get_binop_right(irn); + *prs = r; + return R; +} + +/** + * creates a tarval from a condensed representation. + */ +static ir_tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) +{ + ir_tarval *res, *tv; + int i, j; - /* Look, if one operand is a constant. */ - if(get_irn_opcode(left) == iro_Const) { - tv = get_Const_tarval(left); - operand = right; - } else if(get_irn_opcode(right) == iro_Const) { - tv = get_Const_tarval(right); - operand = left; + j = 0; + tv = get_mode_one(env->mode); + res = NULL; + for (i = 0; i < r; ++i) { + j = R[i]; + if (j) { + ir_tarval *t = new_tarval_from_long(j, mode_Iu); + tv = tarval_shl(tv, t); } + res = res ? tarval_add(res, tv) : tv; + } + return res; +} + +/* forward */ +static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N); - 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 - { - int val = (int) get_tarval_long(tv); - fprintf(stderr, "Found mul with %d(%x) = ", val, val); - for(p = bitstr; *p != '\0'; p++) - printf("%c", *p); - printf("\n"); +/* + * handle simple cases with up-to 2 bits set + */ +static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, ir_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]); } -#endif - - for(p = bitstr; *p != '\0'; p++) { - int bit = *p != '0'; - - switch(bit - last) { - case -1: // The last was 1 we are now at 0 - case 1: // The last was 0 and we are now at 1 - compr[compr_len++] = counter; - counter = 1; - break; - default: - counter++; - } + return ins; + } + if (R[0] != 0) { + ins = emit_SHIFT(env, ins, R[0]); + } + + ins2 = emit_SHIFT(env, env->root, R[0] + R[1]); + return emit_LEA(env, ins, ins2, 0); + } +} + +/** + * Main decompose driver. + */ +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 (env->params->also_use_subs) { + gain = calculate_gain(R, r); + if (gain > 0) { + instruction *instr1, *instr2; + unsigned char *R1, *R2; + int r1, r2, i, k, j; - last = bit; + 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; + 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]; } - compr[compr_len++] = counter; + instr1 = decompose_mul(env, R1, r1, NULL); + instr2 = decompose_mul(env, R2, r2, NULL); + return emit_SUB(env, instr2, instr1); + } + } + + if (N == NULL) + N = condensed_to_value(env, R, r); + + for (i = env->max_S; i > 0; --i) { + 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)) { + unsigned char *Rs; + int rs; -#ifdef DEF - { - const char *prefix = ""; - for(i = 0; i < compr_len; i++, prefix = ",") - fprintf(stderr, "%s%d", prefix, compr[i]); - fprintf("\n"); + 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); } -#endif + } + } + return basic_decompose_mul(env, R, r, N); +} - // Go over all recorded one groups. - curr_bit = compr[0]; +#define IMAX(a,b) ((a) > (b) ? (a) : (b)) - for(i = 1; i < compr_len; i = end_of_group + 2) { - int j, zeros_in_group, ones_in_group; +/** + * basic decomposition routine + */ +static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N) +{ + instruction *Ns; + unsigned t; - ones_in_group = compr[i]; - zeros_in_group = 0; + 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 */ + 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); + } +} - // 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; +/** + * 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; + ir_graph *irg = env->irg; - if(zeros_in_group >= ones_in_group - 1) - end_of_group = i; + if (inst->irn) + return inst->irn; -#ifdef DEB - fprintf(stderr, " i:%d, eg:%d\n", i, end_of_group); -#endif + switch (inst->kind) { + case LEA: + l = build_graph(env, inst->in[0]); + r = build_graph(env, inst->in[1]); + 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_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, 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, env->blk, l, r, env->mode); + case ZERO: + return inst->irn = new_r_Const(irg, get_mode_null(env->mode)); + default: + panic("Unsupported instruction kind"); + } +} - 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; +/** + * 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; -#ifdef DEB - fprintf(stderr, " j:%d, ones:%d\n", j, curr_ones); -#endif + if (inst->costs >= 0) { + /* was already evaluated */ + return 0; + } - // 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; + 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 > 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); + inst->costs = costs; + return costs; + case ZERO: + inst->costs = costs = env->evaluate(inst->kind, NULL); + return costs; + case MUL: + case ROOT: + break; + } + panic("Unsupported instruction kind"); +} - for(k = 0; k < curr_ones; k++) - shift_without_sub[shift_without_sub_pos++] = biased_curr_bit + k; +/** + * 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. + * + * @param irn the Mul operation + * @param operand the multiplication operand + * @param tv the multiplication constant + * + * @return the new graph + */ +static ir_node *do_decomposition(ir_node *irn, ir_node *operand, ir_tarval *tv) +{ + mul_env env; + instruction *inst; + unsigned char *R; + int r; + ir_node *res = irn; + int mul_costs; - curr_bit += curr_ones; - biased_curr_bit = curr_bit + 1; + 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 = env.params->maximum_shifts; + env.evaluate = env.params->evaluate != NULL ? env.params->evaluate : default_evaluate; + env.irg = get_irn_irg(irn); - 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; + R = value_to_condensed(&env, tv, &r); + inst = decompose_mul(&env, R, r, tv); - curr_bit += compr[j + 1]; - } + /* the paper suggests 70% here */ + 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); + env.dbg = get_irn_dbg_info(irn); + env.shf_mode = find_unsigned_mode(env.mode); + if (env.shf_mode == NULL) + env.shf_mode = mode_Iu; - } + res = build_graph(&env, inst); + } + obstack_free(&env.obst, NULL); + return res; +} - { - 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 achive 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; - } +/* 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_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 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 - return irn; - } - /* 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; + /* 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; - highest_shift_wide = curr > highest_shift_wide ? curr - : highest_shift_wide; - highest_shift_seq = curr_seq > highest_shift_seq ? curr_seq - : highest_shift_seq; + if (!is_Mul(irn) || !mode_is_int(mode)) + return res; - last_shift = curr; - } + /* 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); + + 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; + } + + if (tv != NULL) { + res = do_decomposition(irn, operand, tv); + + if (res != irn) { + hook_arch_dep_replace_mul_with_shifts(irn); + exchange(irn, res); + } + } + + return res; +} + +/** + * calculated the ld2 of a tarval if tarval is 2^n, else returns -1. + */ +static int tv_ld2(ir_tarval *tv, int bits) +{ + int i, k = 0, num; + + for (num = i = 0; i < bits; ++i) { + unsigned char v = get_tarval_sub_bits(tv, i); + + if (v) { + int j; - /* 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 - return irn; + for (j = 0; j < 8; ++j) + if ((1 << j) & v) { + ++num; + k = 8 * i + j; } + } + } + if (num == 1) + return k; + return -1; +} + - /* If we have subs, we cannot do sequential. */ - if(1 /* also_use_subs */) { - if(n > 0) { - ir_node *curr = NULL; +/* for shorter lines */ +#define ABS(a) tarval_abs(a) +#define NEG(a) tarval_neg(a) +#define NOT(a) tarval_not(a) +#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, 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) +#define CMP(a, b) tarval_cmp(a, b) +#define CNV(a, m) tarval_convert_to(a, m) +#define ONE(m) get_mode_one(m) +#define ZERO(m) get_mode_null(m) - i = n - 1; +/** The result of a the magic() function. */ +struct ms { + 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 */ +}; - do { - int curr_shift = shifts[i]; - int sub = curr_shift < 0; - int amount = abs(curr_shift) - 1; - ir_node *aux = operand; +/** + * Signed division by constant d: calculate the Magic multiplier M and the shift amount s + * + * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler + */ +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; + ir_tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */ + pn_Cmp d_cmp, M_cmp; + ir_tarval *bits_minus_1, *two_bits_1; - assert(amount >= 0 && "What is a negative shift??"); + struct ms mag; - if(amount != 0) { - tarval *shift_amount = new_tarval_from_long(amount, mode_Iu); - ir_node *cnst = new_r_Const(current_ir_graph, block, mode_Iu, shift_amount); - aux = new_r_Shl(current_ir_graph, block, operand, cnst, mode); - } + tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode(); - 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; + /* we need overflow mode to work correctly */ + tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP); + + /* 2^(bits-1) */ + bits_minus_1 = new_tarval_from_long(bits - 1, u_mode); + two_bits_1 = SHL(get_mode_one(u_mode), bits_minus_1); + + ad = CNV(ABS(d), u_mode); + t = ADD(two_bits_1, SHR(CNV(d, u_mode), bits_minus_1)); + anc = SUB(SUB(t, ONE(u_mode)), MOD(t, ad)); /* Absolute value of nc */ + p = bits - 1; /* Init: p */ + q1 = DIV(two_bits_1, anc); /* Init: q1 = 2^p/|nc| */ + r1 = SUB(two_bits_1, MUL(q1, anc)); /* Init: r1 = rem(2^p, |nc|) */ + q2 = DIV(two_bits_1, ad); /* Init: q2 = 2^p/|d| */ + r2 = SUB(two_bits_1, MUL(q2, ad)); /* Init: r2 = rem(2^p, |d|) */ + + do { + ++p; + 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) { + 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) { + q2 = ADD(q2, ONE(u_mode)); + r2 = SUB(r2, ad); + } - } while(--i >= 0); + 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)); - res = curr; + d_cmp = CMP(d, ZERO(mode)); + + if (d_cmp & pn_Cmp_Ge) + mag.M = ADD(CNV(q2, mode), ONE(mode)); + else + mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode))); + + M_cmp = CMP(mag.M, ZERO(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; + + /* need a sub if d < 0 && M > 0 */ + mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt; + + tarval_set_integer_overflow_mode(rem); + + return mag; +} + +/** The result of the magicu() function. */ +struct mu { + ir_tarval *M; /**< magic add constant */ + int s; /**< shift amount */ + int need_add; /**< add indicator */ +}; + +/** + * Unsigned division by constant d: calculate the Magic multiplier M and the shift amount s + * + * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned) + */ +static struct mu magicu(ir_tarval *d) +{ + ir_mode *mode = get_tarval_mode(d); + int bits = get_mode_size_bits(mode); + int p; + ir_tarval *nc, *delta, *q1, *r1, *q2, *r2; + ir_tarval *bits_minus_1, *two_bits_1, *seven_ff; + + struct mu magu; + + tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode(); + + /* we need overflow mode to work correctly */ + tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP); + + bits_minus_1 = new_tarval_from_long(bits - 1, mode); + two_bits_1 = SHL(get_mode_one(mode), bits_minus_1); + seven_ff = SUB(two_bits_1, ONE(mode)); + + magu.need_add = 0; /* initialize the add indicator */ + nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d)); + p = bits - 1; /* Init: p */ + q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */ + r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */ + q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */ + r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */ + + do { + ++p; + if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) { + q1 = ADD(ADD(q1, q1), ONE(mode)); + r1 = SUB(ADD(r1, r1), nc); + } + else { + q1 = ADD(q1, q1); + r1 = ADD(r1, r1); + } + + if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) { + if (CMP(q2, seven_ff) & pn_Cmp_Ge) + 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) + magu.need_add = 1; + + q2 = ADD(q2, q2); + r2 = ADD(ADD(r2, r2), ONE(mode)); + } + 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))); + + magu.M = ADD(q2, ONE(mode)); /* Magic number */ + magu.s = p - bits; /* and shift amount */ + + tarval_set_integer_overflow_mode(rem); + + return magu; +} + +/** + * Build the Mulh replacement code for n / tv. + * + * Note that 'div' might be a mod or DivMod 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_mode *mode = get_irn_mode(n); + int bits = get_mode_size_bits(mode); + ir_node *q, *t, *c; + + /* 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(irg, mag.M); + 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, block, q, n, mode); + else if (mag.need_sub) + q = new_rd_Sub(dbg, block, q, n, mode); + + /* Do we need the shift */ + if (mag.s > 0) { + 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(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_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); + + 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_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_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) +{ + 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. */ + 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_Const(c)) + return irn; + + tv = get_Const_tarval(c); + + /* check for division by zero */ + if (tarval_is_null(tv)) + return irn; + + left = get_Div_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; + 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, bits - k); + curr = new_rd_Shr(dbg, block, curr, k_node, mode); + + curr = new_rd_Add(dbg, block, left, curr, mode); + } else { + k_node = left; } -#ifdef DEB - { - const char *prefix = ""; - for(i = 0; i < n; i++) { - fprintf(stderr, "%s%d", prefix, shifts[i]); - prefix = ", "; - } - fprintf(stderr, "\n"); + 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_r_Const(irg, get_mode_null(mode)); + res = new_rd_Sub(dbg, block, k_node, res, mode); } -#endif + } 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 { + /* other constant */ + if (allow_Mulh(params, mode)) + res = replace_div_by_mulh(irn, tv); + } + } + + if (res != irn) + hook_arch_dep_replace_division_by_const(irn); + + return res; +} + +/* 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 + or this optimization was not enabled. */ + if (params == NULL || (opts & arch_dep_mod_by_const) == 0) + return irn; + + if (is_Mod(irn)) { + ir_node *c = get_Mod_right(irn); + ir_node *block, *left; + ir_mode *mode; + ir_tarval *tv, *ntv; + dbg_info *dbg; + int n, bits; + int k; + + if (! is_Const(c)) + return irn; + + tv = get_Const_tarval(c); + + /* check for division by zero */ + if (tarval_is_null(tv)) + return irn; - if(bitstr) - free(bitstr); + left = get_Mod_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); + k = tv_ld2(ntv, n); + } + + if (k < 0) { + k = tv_ld2(tv, n); } + 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 + */ + if (mode_is_signed(mode)) { + ir_node *k_node; + 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); + + 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_r_Const_long(irg, mode, (1 << k) - 1); + res = new_rd_And(dbg, block, left, k_node, mode); + } + } else { + /* other constant */ + if (allow_Mulh(params, mode)) { + res = replace_div_by_mulh(irn, tv); + + res = new_rd_Mul(dbg, block, res, c, mode); + + /* res = arch_dep_mul_to_shift(res); */ + + res = new_rd_Sub(dbg, block, left, res, mode); + } + } } + if (res != irn) + hook_arch_dep_replace_division_by_const(irn); + 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; -static const arch_dep_params_t default_params = { - 1, /* also use subs */ - 4, /* maximum shifts */ - 31 /* maximum shift amount */ -}; + /* 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); + } + } + } -const arch_dep_params_t *arch_dep_default_factory(void) { - return &default_params; + if (*div) + hook_arch_dep_replace_division_by_const(irn); }