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"
60 /* when we need verifying */
62 # define IRN_VRFY_IRG(res, irg)
64 # define IRN_VRFY_IRG(res, irg) irn_vrfy_irg(res, irg)
67 /** The params got from the factory in arch_dep_init(...). */
68 static const ir_settings_arch_dep_t *params = NULL;
70 /** The bit mask, which optimizations to apply. */
71 static arch_dep_opts_t opts;
73 void arch_dep_init(arch_dep_params_factory_t factory) {
80 void arch_dep_set_opts(arch_dep_opts_t the_opts) {
83 if (opts & arch_dep_mul_to_shift)
84 set_opt_arch_dep_running(1);
87 /** check, whether a mode allows a Mulh instruction. */
88 static int allow_Mulh(ir_mode *mode) {
89 if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
91 return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
97 typedef struct instruction instruction;
99 insn_kind kind; /**< the instruction kind */
100 instruction *in[2]; /**< the ins */
101 unsigned shift_count; /**< shift count for LEA and SHIFT */
102 ir_node *irn; /**< the generated node for this instruction if any. */
103 int costs; /**< the costs for this instruction */
107 * The environment for the strength reduction of multiplications.
109 typedef struct _mul_env {
110 struct obstack obst; /**< an obstack for local space. */
111 ir_mode *mode; /**< the mode of the multiplication constant */
112 unsigned bits; /**< number of bits in the mode */
113 unsigned max_S; /**< the maximum LEA shift value. */
114 instruction *root; /**< the root of the instruction tree */
115 ir_node *op; /**< the operand that is multiplied */
116 ir_node *blk; /**< the block where the new graph is built */
117 dbg_info *dbg; /**< the debug info for the new graph. */
118 ir_mode *shf_mode; /**< the (unsigned) mode for the shift constants */
119 int fail; /**< set to 1 if the instruction sequence fails the constraints */
120 int n_shift; /**< maximum number of allowed shift instructions */
122 evaluate_costs_func evaluate; /**< the evaluate callback */
126 * Some kind of default evaluator. Return the cost of
129 static int default_evaluate(insn_kind kind, tarval *tv) {
138 * emit a LEA (or an Add) instruction
140 static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) {
141 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
142 res->kind = shift > 0 ? LEA : ADD;
145 res->shift_count = shift;
152 * emit a SHIFT (or an Add or a Zero) instruction
154 static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) {
155 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
156 if (shift == env->bits) {
157 /* a 2^bits with bits resolution is a zero */
161 res->shift_count = 0;
162 } else if (shift != 1) {
166 res->shift_count = shift;
171 res->shift_count = 0;
179 * emit a SUB instruction
181 static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) {
182 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
186 res->shift_count = 0;
193 * emit the ROOT instruction
195 static instruction *emit_ROOT(mul_env *env, ir_node *root_op) {
196 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
200 res->shift_count = 0;
208 * Returns the condensed representation of the tarval tv
210 static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) {
211 ir_mode *mode = get_tarval_mode(tv);
212 int bits = get_mode_size_bits(mode);
213 char *bitstr = get_tarval_bitpattern(tv);
215 unsigned char *R = obstack_alloc(&env->obst, bits);
218 for (i = 0; bitstr[i] != '\0'; ++i) {
219 if (bitstr[i] == '1') {
232 * Calculate the gain when using the generalized complementary technique
234 static int calculate_gain(unsigned char *R, int r) {
239 /* the gain for r == 1 */
241 for (i = 2; i < r; ++i) {
242 /* calculate the gain for r from the gain for r-1 */
243 gain += 2 - R[i - 1];
245 if (gain > max_gain) {
256 * Calculates the condensed complement of a given (R,r) tuple
258 static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs) {
259 unsigned char *value = obstack_alloc(&env->obst, env->bits);
263 memset(value, 0, env->bits);
266 for (i = 0; i < gain; ++i) {
271 /* negate and propagate 1 */
273 for (i = 0; i <= j; ++i) {
274 unsigned char v = !value[i];
280 /* condense it again */
283 for (i = 0; i <= j; ++i) {
296 * creates a tarval from a condensed representation.
298 static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) {
303 tv = get_mode_one(env->mode);
305 for (i = 0; i < r; ++i) {
308 tarval *t = new_tarval_from_long(j, mode_Iu);
309 tv = tarval_shl(tv, t);
311 res = res ? tarval_add(res, tv) : tv;
317 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
320 * handle simple cases with up-to 2 bits set
322 static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) {
323 instruction *ins, *ins2;
327 return emit_SHIFT(env, env->root, R[0]);
333 ins = emit_SHIFT(env, ins, R[0]);
335 if (R[1] <= env->max_S)
336 return emit_LEA(env, ins, ins, R[1]);
338 ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
339 return emit_LEA(env, ins, ins2, 0);
344 * Main decompose driver.
346 static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
351 return decompose_simple_cases(env, R, r, N);
353 if (params->also_use_subs) {
354 gain = calculate_gain(R, r);
356 instruction *instr1, *instr2;
357 unsigned char *R1, *R2;
360 R1 = complement_condensed(env, R, r, gain, &r1);
362 R2 = obstack_alloc(&env->obst, r2);
365 for (i = 0; i < gain; ++i) {
372 /* Two identical bits: normalize */
377 for (i = gain + 1; i < r; ++i) {
381 instr1 = decompose_mul(env, R1, r1, NULL);
382 instr2 = decompose_mul(env, R2, r2, NULL);
383 return emit_SUB(env, instr2, instr1);
388 N = condensed_to_value(env, R, r);
390 for (i = env->max_S; i > 0; --i) {
391 tarval *div_res, *mod_res;
392 tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
394 div_res = tarval_divmod(N, tv, &mod_res);
395 if (mod_res == get_mode_null(env->mode)) {
399 Rs = value_to_condensed(env, div_res, &rs);
401 instruction *N1 = decompose_mul(env, Rs, rs, div_res);
402 return emit_LEA(env, N1, N1, i);
406 return basic_decompose_mul(env, R, r, N);
409 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
412 * basic decomposition routine
414 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
418 if (R[0] == 0) { /* Case 1 */
419 t = R[1] > IMAX(env->max_S, R[1]);
421 Ns = decompose_mul(env, &R[1], r - 1, N);
422 return emit_LEA(env, env->root, Ns, t);
423 } else if (R[0] <= env->max_S) { /* Case 2 */
426 Ns = decompose_mul(env, &R[1], r - 1, N);
427 return emit_LEA(env, Ns, env->root, t);
431 Ns = decompose_mul(env, R, r, N);
432 return emit_SHIFT(env, Ns, t);
437 * recursive build the graph form the instructions.
439 * @param env the environment
440 * @param inst the instruction
442 static ir_node *build_graph(mul_env *env, instruction *inst) {
448 switch (inst->kind) {
450 l = build_graph(env, inst->in[0]);
451 r = build_graph(env, inst->in[1]);
452 c = new_r_Const(current_ir_graph, env->blk, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode));
453 r = new_rd_Shl(env->dbg, current_ir_graph, env->blk, r, c, env->mode);
454 return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
456 l = build_graph(env, inst->in[0]);
457 c = new_r_Const(current_ir_graph, env->blk, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode));
458 return inst->irn = new_rd_Shl(env->dbg, current_ir_graph, env->blk, l, c, env->mode);
460 l = build_graph(env, inst->in[0]);
461 r = build_graph(env, inst->in[1]);
462 return inst->irn = new_rd_Sub(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
464 l = build_graph(env, inst->in[0]);
465 r = build_graph(env, inst->in[1]);
466 return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
468 return inst->irn = new_r_Const(current_ir_graph, env->blk, env->mode, get_mode_null(env->mode));
476 * Calculate the costs for the given instruction sequence.
477 * Note that additional costs due to higher register pressure are NOT evaluated yet
479 static int evaluate_insn(mul_env *env, instruction *inst) {
482 if (inst->costs >= 0) {
483 /* was already evaluated */
487 switch (inst->kind) {
491 costs = evaluate_insn(env, inst->in[0]);
492 costs += evaluate_insn(env, inst->in[1]);
493 costs += env->evaluate(inst->kind, NULL);
497 if (inst->shift_count > params->highest_shift_amount)
499 if (env->n_shift <= 0)
503 costs = evaluate_insn(env, inst->in[0]);
504 costs += env->evaluate(inst->kind, NULL);
508 inst->costs = costs = env->evaluate(inst->kind, NULL);
517 * Evaluate the replacement instructions and build a new graph
518 * if faster than the Mul.
519 * returns the root of the new graph then or irn otherwise.
521 * @param irn the Mul operation
522 * @param operand the multiplication operand
523 * @param tv the multiplication constant
525 * @return the new graph
527 static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) {
535 obstack_init(&env.obst);
536 env.mode = get_tarval_mode(tv);
537 env.bits = (unsigned)get_mode_size_bits(env.mode);
539 env.root = emit_ROOT(&env, operand);
541 env.n_shift = params->maximum_shifts;
542 env.evaluate = params->evaluate != NULL ? params->evaluate : default_evaluate;
544 R = value_to_condensed(&env, tv, &r);
545 inst = decompose_mul(&env, R, r, tv);
547 /* the paper suggests 70% here */
548 mul_costs = (env.evaluate(MUL, tv) * 7) / 10;
549 if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) {
551 env.blk = get_nodes_block(irn);
552 env.dbg = get_irn_dbg_info(irn);
553 env.shf_mode = find_unsigned_mode(env.mode);
554 if (env.shf_mode == NULL)
555 env.shf_mode = mode_Iu;
557 res = build_graph(&env, inst);
559 obstack_free(&env.obst, NULL);
563 /* Replace Muls with Shifts and Add/Subs. */
564 ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) {
566 ir_mode *mode = get_irn_mode(irn);
568 /* If the architecture dependent optimizations were not initialized
569 or this optimization was not enabled. */
570 if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
573 if (is_Mul(irn) && mode_is_int(mode)) {
574 ir_node *left = get_binop_left(irn);
575 ir_node *right = get_binop_right(irn);
577 ir_node *operand = NULL;
579 /* Look, if one operand is a constant. */
580 if (is_Const(left)) {
581 tv = get_Const_tarval(left);
583 } else if (is_Const(right)) {
584 tv = get_Const_tarval(right);
589 res = do_decomposition(irn, operand, tv);
592 hook_arch_dep_replace_mul_with_shifts(irn);
602 * calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
604 static int tv_ld2(tarval *tv, int bits) {
607 for (num = i = 0; i < bits; ++i) {
608 unsigned char v = get_tarval_sub_bits(tv, i);
613 for (j = 0; j < 8; ++j)
626 /* for shorter lines */
627 #define ABS(a) tarval_abs(a)
628 #define NEG(a) tarval_neg(a)
629 #define NOT(a) tarval_not(a)
630 #define SHL(a, b) tarval_shl(a, b)
631 #define SHR(a, b) tarval_shr(a, b)
632 #define ADD(a, b) tarval_add(a, b)
633 #define SUB(a, b) tarval_sub(a, b)
634 #define MUL(a, b) tarval_mul(a, b)
635 #define DIV(a, b) tarval_div(a, b)
636 #define MOD(a, b) tarval_mod(a, b)
637 #define CMP(a, b) tarval_cmp(a, b)
638 #define CNV(a, m) tarval_convert_to(a, m)
639 #define ONE(m) get_mode_one(m)
640 #define ZERO(m) get_mode_null(m)
642 /** The result of a the magic() function. */
644 tarval *M; /**< magic number */
645 int s; /**< shift amount */
646 int need_add; /**< an additional add is needed */
647 int need_sub; /**< an additional sub is needed */
651 * Signed division by constant d: calculate the Magic multiplier M and the shift amount s
653 * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
655 static struct ms magic(tarval *d) {
656 ir_mode *mode = get_tarval_mode(d);
657 ir_mode *u_mode = find_unsigned_mode(mode);
658 int bits = get_mode_size_bits(u_mode);
660 tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
663 tarval *bits_minus_1, *two_bits_1;
667 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
669 /* we need overflow mode to work correctly */
670 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
673 bits_minus_1 = new_tarval_from_long(bits - 1, u_mode);
674 two_bits_1 = SHL(get_mode_one(u_mode), bits_minus_1);
676 ad = CNV(ABS(d), u_mode);
677 t = ADD(two_bits_1, SHR(CNV(d, u_mode), bits_minus_1));
678 anc = SUB(SUB(t, ONE(u_mode)), MOD(t, ad)); /* Absolute value of nc */
679 p = bits - 1; /* Init: p */
680 q1 = DIV(two_bits_1, anc); /* Init: q1 = 2^p/|nc| */
681 r1 = SUB(two_bits_1, MUL(q1, anc)); /* Init: r1 = rem(2^p, |nc|) */
682 q2 = DIV(two_bits_1, ad); /* Init: q2 = 2^p/|d| */
683 r2 = SUB(two_bits_1, MUL(q2, ad)); /* Init: r2 = rem(2^p, |d|) */
687 q1 = ADD(q1, q1); /* Update q1 = 2^p/|nc| */
688 r1 = ADD(r1, r1); /* Update r1 = rem(2^p, |nc|) */
690 if (CMP(r1, anc) & pn_Cmp_Ge) {
691 q1 = ADD(q1, ONE(u_mode));
695 q2 = ADD(q2, q2); /* Update q2 = 2^p/|d| */
696 r2 = ADD(r2, r2); /* Update r2 = rem(2^p, |d|) */
698 if (CMP(r2, ad) & pn_Cmp_Ge) {
699 q2 = ADD(q2, ONE(u_mode));
704 } while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq));
706 d_cmp = CMP(d, ZERO(mode));
708 if (d_cmp & pn_Cmp_Ge)
709 mag.M = ADD(CNV(q2, mode), ONE(mode));
711 mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode)));
713 M_cmp = CMP(mag.M, ZERO(mode));
717 /* need an add if d > 0 && M < 0 */
718 mag.need_add = d_cmp & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt;
720 /* need a sub if d < 0 && M > 0 */
721 mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt;
723 tarval_set_integer_overflow_mode(rem);
728 /** The result of the magicu() function. */
730 tarval *M; /**< magic add constant */
731 int s; /**< shift amount */
732 int need_add; /**< add indicator */
736 * Unsigned division by constant d: calculate the Magic multiplier M and the shift amount s
738 * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned)
740 static struct mu magicu(tarval *d) {
741 ir_mode *mode = get_tarval_mode(d);
742 int bits = get_mode_size_bits(mode);
744 tarval *nc, *delta, *q1, *r1, *q2, *r2;
745 tarval *bits_minus_1, *two_bits_1, *seven_ff;
749 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
751 /* we need overflow mode to work correctly */
752 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
754 bits_minus_1 = new_tarval_from_long(bits - 1, mode);
755 two_bits_1 = SHL(get_mode_one(mode), bits_minus_1);
756 seven_ff = SUB(two_bits_1, ONE(mode));
758 magu.need_add = 0; /* initialize the add indicator */
759 nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d));
760 p = bits - 1; /* Init: p */
761 q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */
762 r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */
763 q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */
764 r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */
768 if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
769 q1 = ADD(ADD(q1, q1), ONE(mode));
770 r1 = SUB(ADD(r1, r1), nc);
777 if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
778 if (CMP(q2, seven_ff) & pn_Cmp_Ge)
781 q2 = ADD(ADD(q2, q2), ONE(mode));
782 r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
785 if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
789 r2 = ADD(ADD(r2, r2), ONE(mode));
791 delta = SUB(SUB(d, ONE(mode)), r2);
792 } while (p < 2*bits &&
793 (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
795 magu.M = ADD(q2, ONE(mode)); /* Magic number */
796 magu.s = p - bits; /* and shift amount */
798 tarval_set_integer_overflow_mode(rem);
804 * Build the Mulh replacement code for n / tv.
806 * Note that 'div' might be a mod or DivMod operation as well
808 static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
809 dbg_info *dbg = get_irn_dbg_info(div);
810 ir_node *n = get_binop_left(div);
811 ir_node *block = get_irn_n(div, -1);
812 ir_mode *mode = get_irn_mode(n);
813 int bits = get_mode_size_bits(mode);
816 /* Beware: do not transform bad code */
817 if (is_Bad(n) || is_Bad(block))
820 if (mode_is_signed(mode)) {
821 struct ms mag = magic(tv);
823 /* generate the Mulh instruction */
824 c = new_r_Const(current_ir_graph, block, mode, mag.M);
825 q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
827 /* do we need an Add or Sub */
829 q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
830 else if (mag.need_sub)
831 q = new_rd_Sub(dbg, current_ir_graph, block, q, n, mode);
833 /* Do we need the shift */
835 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
836 q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
840 c = new_r_Const_long(current_ir_graph, block, mode_Iu, bits-1);
841 t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
843 q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode);
845 struct mu mag = magicu(tv);
848 /* generate the Mulh instruction */
849 c = new_r_Const(current_ir_graph, block, mode, mag.M);
850 q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
854 /* use the GM scheme */
855 t = new_rd_Sub(dbg, current_ir_graph, block, n, q, mode);
857 c = new_r_Const(current_ir_graph, block, mode_Iu, get_mode_one(mode_Iu));
858 t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
860 t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode);
862 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s-1);
863 q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
865 /* use the default scheme */
866 q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
868 } else if (mag.s > 0) { /* default scheme, shift needed */
869 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
870 q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
876 /* Replace Divs with Shifts and Add/Subs and Mulh. */
877 ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
880 /* If the architecture dependent optimizations were not initialized
881 or this optimization was not enabled. */
882 if (params == NULL || (opts & arch_dep_div_by_const) == 0)
885 if (get_irn_opcode(irn) == iro_Div) {
886 ir_node *c = get_Div_right(irn);
887 ir_node *block, *left;
894 if (get_irn_op(c) != op_Const)
897 tv = get_Const_tarval(c);
899 /* check for division by zero */
900 if (classify_tarval(tv) == TV_CLASSIFY_NULL)
903 left = get_Div_left(irn);
904 mode = get_irn_mode(left);
905 block = get_irn_n(irn, -1);
906 dbg = get_irn_dbg_info(irn);
908 bits = get_mode_size_bits(mode);
912 if (mode_is_signed(mode)) {
913 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
914 ntv = tarval_neg(tv);
924 if (k >= 0) { /* division by 2^k or -2^k */
925 if (mode_is_signed(mode)) {
927 ir_node *curr = left;
930 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
931 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
934 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
935 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
937 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
939 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
940 res = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode);
942 if (n_flag) { /* negate the result */
945 k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode));
946 res = new_rd_Sub(dbg, current_ir_graph, block, k_node, res, mode);
948 } else { /* unsigned case */
951 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
952 res = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
956 if (allow_Mulh(mode))
957 res = replace_div_by_mulh(irn, tv);
962 hook_arch_dep_replace_division_by_const(irn);
967 /* Replace Mods with Shifts and Add/Subs and Mulh. */
968 ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
971 /* If the architecture dependent optimizations were not initialized
972 or this optimization was not enabled. */
973 if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
976 if (get_irn_opcode(irn) == iro_Mod) {
977 ir_node *c = get_Mod_right(irn);
978 ir_node *block, *left;
985 if (get_irn_op(c) != op_Const)
988 tv = get_Const_tarval(c);
990 /* check for division by zero */
991 if (classify_tarval(tv) == TV_CLASSIFY_NULL)
994 left = get_Mod_left(irn);
995 mode = get_irn_mode(left);
996 block = get_irn_n(irn, -1);
997 dbg = get_irn_dbg_info(irn);
998 bits = get_mode_size_bits(mode);
1002 if (mode_is_signed(mode)) {
1003 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1004 ntv = tarval_neg(tv);
1013 /* division by 2^k or -2^k:
1014 * we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k
1016 if (mode_is_signed(mode)) {
1018 ir_node *curr = left;
1021 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
1022 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
1025 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
1026 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
1028 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
1030 k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
1031 curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
1033 res = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
1034 } else { /* unsigned case */
1037 k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1);
1038 res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
1041 /* other constant */
1042 if (allow_Mulh(mode)) {
1043 res = replace_div_by_mulh(irn, tv);
1045 res = new_rd_Mul(dbg, current_ir_graph, block, res, c, mode);
1047 /* res = arch_dep_mul_to_shift(res); */
1049 res = new_rd_Sub(dbg, current_ir_graph, block, left, res, mode);
1055 hook_arch_dep_replace_division_by_const(irn);
1060 /* Replace DivMods with Shifts and Add/Subs and Mulh. */
1061 void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn) {
1064 /* If the architecture dependent optimizations were not initialized
1065 or this optimization was not enabled. */
1066 if (params == NULL ||
1067 ((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
1070 if (get_irn_opcode(irn) == iro_DivMod) {
1071 ir_node *c = get_DivMod_right(irn);
1072 ir_node *block, *left;
1079 if (get_irn_op(c) != op_Const)
1082 tv = get_Const_tarval(c);
1084 /* check for division by zero */
1085 if (classify_tarval(tv) == TV_CLASSIFY_NULL)
1088 left = get_DivMod_left(irn);
1089 mode = get_irn_mode(left);
1090 block = get_irn_n(irn, -1);
1091 dbg = get_irn_dbg_info(irn);
1093 bits = get_mode_size_bits(mode);
1097 if (mode_is_signed(mode)) {
1098 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1099 ntv = tarval_neg(tv);
1109 if (k >= 0) { /* division by 2^k or -2^k */
1110 if (mode_is_signed(mode)) {
1111 ir_node *k_node, *c_k;
1112 ir_node *curr = left;
1115 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
1116 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
1119 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
1120 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
1122 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
1124 c_k = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
1126 *div = new_rd_Shrs(dbg, current_ir_graph, block, curr, c_k, mode);
1128 if (n_flag) { /* negate the div result */
1131 k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode));
1132 *div = new_rd_Sub(dbg, current_ir_graph, block, k_node, *div, mode);
1135 k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
1136 curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
1138 *mod = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
1139 } else { /* unsigned case */
1142 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
1143 *div = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
1145 k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1);
1146 *mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
1149 /* other constant */
1150 if (allow_Mulh(mode)) {
1153 *div = replace_div_by_mulh(irn, tv);
1155 t = new_rd_Mul(dbg, current_ir_graph, block, *div, c, mode);
1157 /* t = arch_dep_mul_to_shift(t); */
1159 *mod = new_rd_Sub(dbg, current_ir_graph, block, left, t, mode);
1165 hook_arch_dep_replace_division_by_const(irn);
1169 static const ir_settings_arch_dep_t default_params = {
1170 1, /* also use subs */
1171 4, /* maximum shifts */
1172 31, /* maximum shift amount */
1173 default_evaluate, /* default evaluator */
1175 0, /* allow Mulhs */
1176 0, /* allow Mulus */
1177 32 /* Mulh allowed up to 32 bit */
1180 /* A default parameter factory for testing purposes. */
1181 const ir_settings_arch_dep_t *arch_dep_default_factory(void) {
1182 return &default_params;