2 * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Machine dependent Firm optimizations.
24 * @author Sebastian Hack, Michael Beck
27 * Implements "Strenght Reduction of Multiplications by Integer Constants" by Youfeng Wu.
28 * Implements Division and Modulo by Consts from "Hackers Delight",
41 #include "irgraph_t.h"
48 #include "dbginfo_t.h"
49 #include "iropt_dbg.h"
59 /* when we need verifying */
61 # define IRN_VRFY_IRG(res, irg)
63 # define IRN_VRFY_IRG(res, irg) irn_vrfy_irg(res, irg)
66 /** The params got from the factory in arch_dep_init(...). */
67 static const ir_settings_arch_dep_t *params = NULL;
69 /** The bit mask, which optimizations to apply. */
70 static arch_dep_opts_t opts;
72 /* we need this new pseudo op */
73 static ir_op *op_Mulh = NULL;
76 * construct a Mulh: Mulh(a,b) = (a * b) >> w, w is the with in bits of a, b
79 new_rd_Mulh (dbg_info *db, ir_graph *irg, ir_node *block,
80 ir_node *op1, ir_node *op2, ir_mode *mode) {
86 res = new_ir_node(db, irg, block, op_Mulh, mode, 2, in);
87 res = optimize_node(res);
88 IRN_VRFY_IRG(res, irg);
92 ir_op *get_op_Mulh(void) { return op_Mulh; }
94 void arch_dep_init(arch_dep_params_factory_t factory) {
101 int mulh_opc = get_next_ir_opcode();
103 /* create the Mulh operation */
104 op_Mulh = new_ir_op(mulh_opc, "Mulh", op_pin_state_floats, irop_flag_commutative, oparity_binary, 0, 0, NULL);
108 void arch_dep_set_opts(arch_dep_opts_t the_opts) {
112 /** check, whether a mode allows a Mulh instruction. */
113 static int allow_Mulh(ir_mode *mode) {
114 if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
116 return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
122 typedef struct instruction instruction;
124 insn_kind kind; /**< the instruction kind */
125 instruction *in[2]; /**< the ins */
126 unsigned shift_count; /**< shift count for LEA and SHIFT */
127 ir_node *irn; /**< the generated node for this instruction if any. */
128 int costs; /**< the costs for this instruction */
132 * The environment for the strength reduction of multiplications.
134 typedef struct _mul_env {
135 struct obstack obst; /**< an obstack for local space. */
136 ir_mode *mode; /**< the mode of the multiplication constant */
137 int bits; /**< number of bits in the mode */
138 unsigned max_S; /**< the maximum LEA shift value. */
139 instruction *root; /**< the root of the instruction tree */
140 ir_node *op; /**< the operand that is multiplied */
141 ir_node *blk; /**< the block where the new graph is built */
142 dbg_info *dbg; /**< the debug info for the new graph. */
143 ir_mode *shf_mode; /**< the (unsigned) mode for the shift constants */
144 int fail; /**< set to 1 if the instruction sequence fails the constraints */
145 int n_shift; /**< maximum number of allowed shift instructions */
147 evaluate_costs_func evaluate; /**< the evaluate callback */
151 * Some kind of default evaluator.
153 static int default_evaluate(insn_kind kind, tarval *tv) {
160 * emit a LEA (or an Add) instruction
162 static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) {
163 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
164 res->kind = shift > 0 ? LEA : ADD;
167 res->shift_count = shift;
174 * emit a SHIFT (or an Add or a Zero) instruction
176 static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) {
177 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
178 if (shift == env->bits) {
179 /* a 2^bits with bits resolution is a zero */
183 res->shift_count = 0;
184 } else if (shift != 1) {
188 res->shift_count = shift;
193 res->shift_count = 0;
201 * emit a SUB instruction
203 static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) {
204 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
208 res->shift_count = 0;
215 * emit the ROOT instruction
217 static instruction *emit_ROOT(mul_env *env, ir_node *root_op) {
218 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
222 res->shift_count = 0;
230 * Returns the condensed representation of the tarval tv
232 static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) {
233 ir_mode *mode = get_tarval_mode(tv);
234 int bits = get_mode_size_bits(mode);
235 char *bitstr = get_tarval_bitpattern(tv);
237 unsigned char *R = obstack_alloc(&env->obst, bits);
240 for (i = 0; bitstr[i] != '\0'; ++i) {
241 if (bitstr[i] == '1') {
254 * Calculate the gain when using the generalized complementary technique
256 static int calculate_gain(unsigned char *R, int r) {
261 /* the gain for r == 1 */
263 for (i = 2; i < r; ++i) {
264 /* calculate the gain for r from the gain for r-1 */
265 gain += 2 - R[i - 1];
267 if (gain > max_gain) {
278 * Calculates the condensed complement of a given (R,r) tuple
280 static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs) {
281 unsigned char *value = obstack_alloc(&env->obst, env->bits);
285 memset(value, 0, env->bits);
288 for (i = 0; i < gain; ++i) {
293 /* negate and propagate 1 */
295 for (i = 0; i <= j; ++i) {
296 unsigned char v = !value[i];
302 /* condense it again */
305 for (i = 0; i <= j; ++i) {
318 * creates a tarval from a condensed representation.
320 static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) {
325 tv = get_mode_one(env->mode);
327 for (i = 0; i < r; ++i) {
330 tarval *t = new_tarval_from_long(j, mode_Iu);
331 tv = tarval_shl(tv, t);
333 res = res ? tarval_add(res, tv) : tv;
339 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
342 * handle simple cases with up-to 2 bits set
344 static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) {
345 instruction *ins, *ins2;
348 return emit_SHIFT(env, env->root, R[0]);
354 ins = emit_SHIFT(env, ins, R[0]);
356 if (R[1] <= env->max_S)
357 return emit_LEA(env, ins, ins, R[1]);
359 ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
360 return emit_LEA(env, ins, ins2, 0);
365 * Main decompose driver.
367 static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
372 return decompose_simple_cases(env, R, r, N);
374 if (params->also_use_subs) {
375 gain = calculate_gain(R, r);
377 instruction *instr1, *instr2;
378 unsigned char *R1, *R2;
381 R1 = complement_condensed(env, R, r, gain, &r1);
383 R2 = obstack_alloc(&env->obst, r2);
386 for (i = 0; i < gain; ++i) {
393 /* Two identical bits: normalize */
398 for (i = gain + 1; i < r; ++i) {
402 instr1 = decompose_mul(env, R1, r1, NULL);
403 instr2 = decompose_mul(env, R2, r2, NULL);
404 return emit_SUB(env, instr2, instr1);
409 N = condensed_to_value(env, R, r);
411 for (i = env->max_S; i > 0; --i) {
412 tarval *div_res, *mod_res;
413 tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
415 div_res = tarval_divmod(N, tv, &mod_res);
416 if (mod_res == get_mode_null(env->mode)) {
420 Rs = value_to_condensed(env, div_res, &rs);
422 instruction *N1 = decompose_mul(env, Rs, rs, div_res);
423 return emit_LEA(env, N1, N1, i);
427 return basic_decompose_mul(env, R, r, N);
430 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
433 * basic decomposition routine
435 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
439 if (R[0] == 0) { /* Case 1 */
440 t = R[1] > IMAX(env->max_S, R[1]);
442 Ns = decompose_mul(env, &R[1], r - 1, N);
443 return emit_LEA(env, env->root, Ns, t);
444 } else if (R[0] <= env->max_S) { /* Case 2 */
447 Ns = decompose_mul(env, &R[1], r - 1, N);
448 return emit_LEA(env, Ns, env->root, t);
452 Ns = decompose_mul(env, R, r, N);
453 return emit_SHIFT(env, Ns, t);
458 * recursive build the graph form the instructions.
460 * @param env the environment
461 * @param inst the instruction
463 static ir_node *build_graph(mul_env *env, instruction *inst) {
469 switch (inst->kind) {
471 l = build_graph(env, inst->in[0]);
472 r = build_graph(env, inst->in[1]);
473 c = new_r_Const(current_ir_graph, env->blk, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode));
474 r = new_rd_Shl(env->dbg, current_ir_graph, env->blk, r, c, env->mode);
475 return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
477 l = build_graph(env, inst->in[0]);
478 c = new_r_Const(current_ir_graph, env->blk, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode));
479 return inst->irn = new_rd_Shl(env->dbg, current_ir_graph, env->blk, l, c, env->mode);
481 l = build_graph(env, inst->in[0]);
482 r = build_graph(env, inst->in[1]);
483 return inst->irn = new_rd_Sub(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
485 l = build_graph(env, inst->in[0]);
486 r = build_graph(env, inst->in[1]);
487 return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
489 return inst->irn = new_r_Const(current_ir_graph, env->blk, env->mode, get_mode_null(env->mode));
497 * Calculate the costs for the given instruction sequence.
498 * Note that additional costs due to higher register pressure are NOT evaluated yet
500 static int evaluate_insn(mul_env *env, instruction *inst) {
503 if (inst->costs >= 0) {
504 /* was already evaluated */
508 switch (inst->kind) {
512 costs = evaluate_insn(env, inst->in[0]);
513 costs += evaluate_insn(env, inst->in[1]);
514 costs += env->evaluate(inst->kind, NULL);
518 if (inst->shift_count > params->highest_shift_amount)
520 if (env->n_shift <= 0)
524 costs = evaluate_insn(env, inst->in[0]);
525 costs += env->evaluate(inst->kind, NULL);
529 inst->costs = costs = env->evaluate(inst->kind, NULL);
538 * Evaluate the replacement instructions and build a new graph
539 * if faster than the Mul.
540 * returns the root of the new graph then or irn otherwise.
542 * @param irn the Mul operation
543 * @param operand the multiplication operand
544 * @param tv the multiplication constant
546 * @return the new graph
548 static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) {
556 obstack_init(&env.obst);
557 env.mode = get_tarval_mode(tv);
558 env.bits = get_mode_size_bits(env.mode);
560 env.root = emit_ROOT(&env, operand);
562 env.n_shift = params->maximum_shifts;
563 env.evaluate = params->evaluate != NULL ? params->evaluate : default_evaluate;
565 R = value_to_condensed(&env, tv, &r);
566 inst = decompose_mul(&env, R, r, tv);
568 /* the paper suggests 70% here */
569 mul_costs = (env.evaluate(MUL, tv) * 7) / 10;
570 if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) {
572 env.blk = get_nodes_block(irn);
573 env.dbg = get_irn_dbg_info(irn);
574 env.shf_mode = find_unsigned_mode(env.mode);
575 if (env.shf_mode == NULL)
576 env.shf_mode = mode_Iu;
578 res = build_graph(&env, inst);
580 obstack_free(&env.obst, NULL);
584 /* Replace Muls with Shifts and Add/Subs. */
585 ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) {
587 ir_mode *mode = get_irn_mode(irn);
589 /* If the architecture dependent optimizations were not initialized
590 or this optimization was not enabled. */
591 if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
594 if (is_Mul(irn) && mode_is_int(mode)) {
595 ir_node *block = get_nodes_block(irn);
596 ir_node *left = get_binop_left(irn);
597 ir_node *right = get_binop_right(irn);
599 ir_node *operand = NULL;
601 /* Look, if one operand is a constant. */
602 if (is_Const(left)) {
603 tv = get_Const_tarval(left);
605 } else if (is_Const(right)) {
606 tv = get_Const_tarval(right);
611 res = do_decomposition(irn, operand, tv);
614 hook_arch_dep_replace_mul_with_shifts(irn);
624 * calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
626 static int tv_ld2(tarval *tv, int bits) {
629 for (num = i = 0; i < bits; ++i) {
630 unsigned char v = get_tarval_sub_bits(tv, i);
635 for (j = 0; j < 8; ++j)
648 /* for shorter lines */
649 #define ABS(a) tarval_abs(a)
650 #define NEG(a) tarval_neg(a)
651 #define NOT(a) tarval_not(a)
652 #define SHL(a, b) tarval_shl(a, b)
653 #define SHR(a, b) tarval_shr(a, b)
654 #define ADD(a, b) tarval_add(a, b)
655 #define SUB(a, b) tarval_sub(a, b)
656 #define MUL(a, b) tarval_mul(a, b)
657 #define DIV(a, b) tarval_div(a, b)
658 #define MOD(a, b) tarval_mod(a, b)
659 #define CMP(a, b) tarval_cmp(a, b)
660 #define CNV(a, m) tarval_convert_to(a, m)
661 #define ONE(m) get_mode_one(m)
662 #define ZERO(m) get_mode_null(m)
664 /** The result of a the magic() function. */
666 tarval *M; /**< magic number */
667 int s; /**< shift amount */
668 int need_add; /**< an additional add is needed */
669 int need_sub; /**< an additional sub is needed */
673 * Signed division by constant d: calculate the Magic multiplier M and the shift amount s
675 * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
677 static struct ms magic(tarval *d) {
678 ir_mode *mode = get_tarval_mode(d);
679 ir_mode *u_mode = find_unsigned_mode(mode);
680 int bits = get_mode_size_bits(u_mode);
682 tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
685 tarval *bits_minus_1, *two_bits_1;
689 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
691 /* we need overflow mode to work correctly */
692 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
695 bits_minus_1 = new_tarval_from_long(bits - 1, u_mode);
696 two_bits_1 = SHL(get_mode_one(u_mode), bits_minus_1);
698 ad = CNV(ABS(d), u_mode);
699 t = ADD(two_bits_1, SHR(CNV(d, u_mode), bits_minus_1));
700 anc = SUB(SUB(t, ONE(u_mode)), MOD(t, ad)); /* Absolute value of nc */
701 p = bits - 1; /* Init: p */
702 q1 = DIV(two_bits_1, anc); /* Init: q1 = 2^p/|nc| */
703 r1 = SUB(two_bits_1, MUL(q1, anc)); /* Init: r1 = rem(2^p, |nc|) */
704 q2 = DIV(two_bits_1, ad); /* Init: q2 = 2^p/|d| */
705 r2 = SUB(two_bits_1, MUL(q2, ad)); /* Init: r2 = rem(2^p, |d|) */
709 q1 = ADD(q1, q1); /* Update q1 = 2^p/|nc| */
710 r1 = ADD(r1, r1); /* Update r1 = rem(2^p, |nc|) */
712 if (CMP(r1, anc) & pn_Cmp_Ge) {
713 q1 = ADD(q1, ONE(u_mode));
717 q2 = ADD(q2, q2); /* Update q2 = 2^p/|d| */
718 r2 = ADD(r2, r2); /* Update r2 = rem(2^p, |d|) */
720 if (CMP(r2, ad) & pn_Cmp_Ge) {
721 q2 = ADD(q2, ONE(u_mode));
726 } while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq));
728 d_cmp = CMP(d, ZERO(mode));
730 if (d_cmp & pn_Cmp_Ge)
731 mag.M = ADD(CNV(q2, mode), ONE(mode));
733 mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode)));
735 M_cmp = CMP(mag.M, ZERO(mode));
739 /* need an add if d > 0 && M < 0 */
740 mag.need_add = d_cmp & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt;
742 /* need a sub if d < 0 && M > 0 */
743 mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt;
745 tarval_set_integer_overflow_mode(rem);
750 /** The result of the magicu() function. */
752 tarval *M; /**< magic add constant */
753 int s; /**< shift amount */
754 int need_add; /**< add indicator */
758 * Unsigned division by constant d: calculate the Magic multiplier M and the shift amount s
760 * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned)
762 static struct mu magicu(tarval *d) {
763 ir_mode *mode = get_tarval_mode(d);
764 int bits = get_mode_size_bits(mode);
766 tarval *nc, *delta, *q1, *r1, *q2, *r2;
767 tarval *bits_minus_1, *two_bits_1, *seven_ff;
771 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
773 /* we need overflow mode to work correctly */
774 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
776 bits_minus_1 = new_tarval_from_long(bits - 1, mode);
777 two_bits_1 = SHL(get_mode_one(mode), bits_minus_1);
778 seven_ff = SUB(two_bits_1, ONE(mode));
780 magu.need_add = 0; /* initialize the add indicator */
781 nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d));
782 p = bits - 1; /* Init: p */
783 q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */
784 r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */
785 q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */
786 r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */
790 if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
791 q1 = ADD(ADD(q1, q1), ONE(mode));
792 r1 = SUB(ADD(r1, r1), nc);
799 if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
800 if (CMP(q2, seven_ff) & pn_Cmp_Ge)
803 q2 = ADD(ADD(q2, q2), ONE(mode));
804 r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
807 if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
811 r2 = ADD(ADD(r2, r2), ONE(mode));
813 delta = SUB(SUB(d, ONE(mode)), r2);
814 } while (p < 2*bits &&
815 (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
817 magu.M = ADD(q2, ONE(mode)); /* Magic number */
818 magu.s = p - bits; /* and shift amount */
820 tarval_set_integer_overflow_mode(rem);
826 * Build the Mulh replacement code for n / tv.
828 * Note that 'div' might be a mod or DivMod operation as well
830 static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
831 dbg_info *dbg = get_irn_dbg_info(div);
832 ir_node *n = get_binop_left(div);
833 ir_node *block = get_irn_n(div, -1);
834 ir_mode *mode = get_irn_mode(n);
835 int bits = get_mode_size_bits(mode);
838 /* Beware: do not transform bad code */
839 if (is_Bad(n) || is_Bad(block))
842 if (mode_is_signed(mode)) {
843 struct ms mag = magic(tv);
845 /* generate the Mulh instruction */
846 c = new_r_Const(current_ir_graph, block, mode, mag.M);
847 q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
849 /* do we need an Add or Sub */
851 q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
852 else if (mag.need_sub)
853 q = new_rd_Sub(dbg, current_ir_graph, block, q, n, mode);
855 /* Do we need the shift */
857 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
858 q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
862 c = new_r_Const_long(current_ir_graph, block, mode_Iu, bits-1);
863 t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
865 q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode);
867 struct mu mag = magicu(tv);
870 /* generate the Mulh instruction */
871 c = new_r_Const(current_ir_graph, block, mode, mag.M);
872 q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
876 /* use the GM scheme */
877 t = new_rd_Sub(dbg, current_ir_graph, block, n, q, mode);
879 c = new_r_Const(current_ir_graph, block, mode_Iu, get_mode_one(mode_Iu));
880 t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
882 t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode);
884 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s-1);
885 q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
887 /* use the default scheme */
888 q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
890 } else if (mag.s > 0) { /* default scheme, shift needed */
891 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
892 q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
898 /* Replace Divs with Shifts and Add/Subs and Mulh. */
899 ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
902 /* If the architecture dependent optimizations were not initialized
903 or this optimization was not enabled. */
904 if (params == NULL || (opts & arch_dep_div_by_const) == 0)
907 if (get_irn_opcode(irn) == iro_Div) {
908 ir_node *c = get_Div_right(irn);
909 ir_node *block, *left;
916 if (get_irn_op(c) != op_Const)
919 tv = get_Const_tarval(c);
921 /* check for division by zero */
922 if (classify_tarval(tv) == TV_CLASSIFY_NULL)
925 left = get_Div_left(irn);
926 mode = get_irn_mode(left);
927 block = get_irn_n(irn, -1);
928 dbg = get_irn_dbg_info(irn);
930 bits = get_mode_size_bits(mode);
934 if (mode_is_signed(mode)) {
935 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
936 ntv = tarval_neg(tv);
946 if (k >= 0) { /* division by 2^k or -2^k */
947 if (mode_is_signed(mode)) {
949 ir_node *curr = left;
952 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
953 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
956 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
957 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
959 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
961 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
962 res = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode);
964 if (n_flag) { /* negate the result */
967 k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode));
968 res = new_rd_Sub(dbg, current_ir_graph, block, k_node, res, mode);
970 } else { /* unsigned case */
973 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
974 res = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
978 if (allow_Mulh(mode))
979 res = replace_div_by_mulh(irn, tv);
984 hook_arch_dep_replace_division_by_const(irn);
989 /* Replace Mods with Shifts and Add/Subs and Mulh. */
990 ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
993 /* If the architecture dependent optimizations were not initialized
994 or this optimization was not enabled. */
995 if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
998 if (get_irn_opcode(irn) == iro_Mod) {
999 ir_node *c = get_Mod_right(irn);
1000 ir_node *block, *left;
1007 if (get_irn_op(c) != op_Const)
1010 tv = get_Const_tarval(c);
1012 /* check for division by zero */
1013 if (classify_tarval(tv) == TV_CLASSIFY_NULL)
1016 left = get_Mod_left(irn);
1017 mode = get_irn_mode(left);
1018 block = get_irn_n(irn, -1);
1019 dbg = get_irn_dbg_info(irn);
1020 bits = get_mode_size_bits(mode);
1024 if (mode_is_signed(mode)) {
1025 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1026 ntv = tarval_neg(tv);
1035 /* division by 2^k or -2^k:
1036 * we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k
1038 if (mode_is_signed(mode)) {
1040 ir_node *curr = left;
1043 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
1044 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
1047 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
1048 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
1050 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
1052 k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
1053 curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
1055 res = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
1056 } else { /* unsigned case */
1059 k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1);
1060 res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
1063 /* other constant */
1064 if (allow_Mulh(mode)) {
1065 res = replace_div_by_mulh(irn, tv);
1067 res = new_rd_Mul(dbg, current_ir_graph, block, res, c, mode);
1069 /* res = arch_dep_mul_to_shift(res); */
1071 res = new_rd_Sub(dbg, current_ir_graph, block, left, res, mode);
1077 hook_arch_dep_replace_division_by_const(irn);
1082 /* Replace DivMods with Shifts and Add/Subs and Mulh. */
1083 void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn) {
1086 /* If the architecture dependent optimizations were not initialized
1087 or this optimization was not enabled. */
1088 if (params == NULL ||
1089 ((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
1092 if (get_irn_opcode(irn) == iro_DivMod) {
1093 ir_node *c = get_DivMod_right(irn);
1094 ir_node *block, *left;
1101 if (get_irn_op(c) != op_Const)
1104 tv = get_Const_tarval(c);
1106 /* check for division by zero */
1107 if (classify_tarval(tv) == TV_CLASSIFY_NULL)
1110 left = get_DivMod_left(irn);
1111 mode = get_irn_mode(left);
1112 block = get_irn_n(irn, -1);
1113 dbg = get_irn_dbg_info(irn);
1115 bits = get_mode_size_bits(mode);
1119 if (mode_is_signed(mode)) {
1120 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1121 ntv = tarval_neg(tv);
1131 if (k >= 0) { /* division by 2^k or -2^k */
1132 if (mode_is_signed(mode)) {
1133 ir_node *k_node, *c_k;
1134 ir_node *curr = left;
1137 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
1138 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
1141 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
1142 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
1144 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
1146 c_k = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
1148 *div = new_rd_Shrs(dbg, current_ir_graph, block, curr, c_k, mode);
1150 if (n_flag) { /* negate the div result */
1153 k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode));
1154 *div = new_rd_Sub(dbg, current_ir_graph, block, k_node, *div, mode);
1157 k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
1158 curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
1160 *mod = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
1161 } else { /* unsigned case */
1164 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
1165 *div = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
1167 k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1);
1168 *mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
1171 /* other constant */
1172 if (allow_Mulh(mode)) {
1175 *div = replace_div_by_mulh(irn, tv);
1177 t = new_rd_Mul(dbg, current_ir_graph, block, *div, c, mode);
1179 /* t = arch_dep_mul_to_shift(t); */
1181 *mod = new_rd_Sub(dbg, current_ir_graph, block, left, t, mode);
1187 hook_arch_dep_replace_division_by_const(irn);
1191 static const ir_settings_arch_dep_t default_params = {
1192 1, /* also use subs */
1193 4, /* maximum shifts */
1194 31, /* maximum shift amount */
1196 0, /* allow Mulhs */
1197 0, /* allow Mulus */
1198 32 /* Mulh allowed up to 32 bit */
1201 /* A default parameter factory for testing purposes. */
1202 const ir_settings_arch_dep_t *arch_dep_default_factory(void) {
1203 return &default_params;