2 * Copyright (C) 1995-2008 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 "Strength Reduction of Multiplications by Integer Constants" by Youfeng Wu.
28 * Implements Division and Modulo by Consts from "Hackers Delight",
36 #include "irgraph_t.h"
43 #include "dbginfo_t.h"
44 #include "iropt_dbg.h"
56 /** The params got from the factory in arch_dep_init(...). */
57 static const ir_settings_arch_dep_t *params = NULL;
59 /** The bit mask, which optimizations to apply. */
60 static arch_dep_opts_t opts;
62 void arch_dep_init(arch_dep_params_factory_t factory)
70 void arch_dep_set_opts(arch_dep_opts_t the_opts)
75 /** check, whether a mode allows a Mulh instruction. */
76 static int allow_Mulh(ir_mode *mode)
78 if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
80 return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
86 typedef struct instruction instruction;
88 insn_kind kind; /**< the instruction kind */
89 instruction *in[2]; /**< the ins */
90 unsigned shift_count; /**< shift count for LEA and SHIFT */
91 ir_node *irn; /**< the generated node for this instruction if any. */
92 int costs; /**< the costs for this instruction */
96 * The environment for the strength reduction of multiplications.
98 typedef struct _mul_env {
99 struct obstack obst; /**< an obstack for local space. */
100 ir_mode *mode; /**< the mode of the multiplication constant */
101 unsigned bits; /**< number of bits in the mode */
102 unsigned max_S; /**< the maximum LEA shift value. */
103 instruction *root; /**< the root of the instruction tree */
104 ir_node *op; /**< the operand that is multiplied */
105 ir_node *blk; /**< the block where the new graph is built */
106 dbg_info *dbg; /**< the debug info for the new graph. */
107 ir_mode *shf_mode; /**< the (unsigned) mode for the shift constants */
108 int fail; /**< set to 1 if the instruction sequence fails the constraints */
109 int n_shift; /**< maximum number of allowed shift instructions */
111 evaluate_costs_func evaluate; /**< the evaluate callback */
115 * Some kind of default evaluator. Return the cost of
118 static int default_evaluate(insn_kind kind, tarval *tv)
128 * emit a LEA (or an Add) instruction
130 static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift)
132 instruction *res = OALLOC(&env->obst, instruction);
133 res->kind = shift > 0 ? LEA : ADD;
136 res->shift_count = shift;
143 * emit a SHIFT (or an Add or a Zero) instruction
145 static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift)
147 instruction *res = OALLOC(&env->obst, instruction);
148 if (shift == env->bits) {
149 /* a 2^bits with bits resolution is a zero */
153 res->shift_count = 0;
154 } else if (shift != 1) {
158 res->shift_count = shift;
163 res->shift_count = 0;
171 * emit a SUB instruction
173 static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b)
175 instruction *res = OALLOC(&env->obst, instruction);
179 res->shift_count = 0;
186 * emit the ROOT instruction
188 static instruction *emit_ROOT(mul_env *env, ir_node *root_op)
190 instruction *res = OALLOC(&env->obst, instruction);
194 res->shift_count = 0;
202 * Returns the condensed representation of the tarval tv
204 static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr)
206 ir_mode *mode = get_tarval_mode(tv);
207 int bits = get_mode_size_bits(mode);
208 char *bitstr = get_tarval_bitpattern(tv);
210 unsigned char *R = obstack_alloc(&env->obst, bits);
213 for (i = 0; bitstr[i] != '\0'; ++i) {
214 if (bitstr[i] == '1') {
227 * Calculate the gain when using the generalized complementary technique
229 static int calculate_gain(unsigned char *R, int r)
235 /* the gain for r == 1 */
237 for (i = 2; i < r; ++i) {
238 /* calculate the gain for r from the gain for r-1 */
239 gain += 2 - R[i - 1];
241 if (gain > max_gain) {
250 * Calculates the condensed complement of a given (R,r) tuple
252 static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs)
254 unsigned char *value = obstack_alloc(&env->obst, env->bits);
258 memset(value, 0, env->bits);
261 for (i = 0; i < gain; ++i) {
266 /* negate and propagate 1 */
268 for (i = 0; i <= j; ++i) {
269 unsigned char v = !value[i];
275 /* condense it again */
278 for (i = 0; i <= j; ++i) {
291 * creates a tarval from a condensed representation.
293 static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r)
299 tv = get_mode_one(env->mode);
301 for (i = 0; i < r; ++i) {
304 tarval *t = new_tarval_from_long(j, mode_Iu);
305 tv = tarval_shl(tv, t);
307 res = res ? tarval_add(res, tv) : tv;
313 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
316 * handle simple cases with up-to 2 bits set
318 static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N)
320 instruction *ins, *ins2;
324 return emit_SHIFT(env, env->root, R[0]);
329 if (R[1] <= env->max_S) {
330 ins = emit_LEA(env, ins, ins, R[1]);
332 ins = emit_SHIFT(env, ins, R[0]);
337 ins = emit_SHIFT(env, ins, R[0]);
340 ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
341 return emit_LEA(env, ins, ins2, 0);
346 * Main decompose driver.
348 static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N)
354 return decompose_simple_cases(env, R, r, N);
356 if (params->also_use_subs) {
357 gain = calculate_gain(R, r);
359 instruction *instr1, *instr2;
360 unsigned char *R1, *R2;
363 R1 = complement_condensed(env, R, r, gain, &r1);
365 R2 = obstack_alloc(&env->obst, r2);
368 for (i = 0; i < gain; ++i) {
375 /* Two identical bits: normalize */
380 for (i = gain + 1; i < r; ++i) {
384 instr1 = decompose_mul(env, R1, r1, NULL);
385 instr2 = decompose_mul(env, R2, r2, NULL);
386 return emit_SUB(env, instr2, instr1);
391 N = condensed_to_value(env, R, r);
393 for (i = env->max_S; i > 0; --i) {
394 tarval *div_res, *mod_res;
395 tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
397 div_res = tarval_divmod(N, tv, &mod_res);
398 if (mod_res == get_mode_null(env->mode)) {
402 Rs = value_to_condensed(env, div_res, &rs);
404 instruction *N1 = decompose_mul(env, Rs, rs, div_res);
405 return emit_LEA(env, N1, N1, i);
409 return basic_decompose_mul(env, R, r, N);
412 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
415 * basic decomposition routine
417 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N)
422 if (R[0] == 0) { /* Case 1 */
423 t = R[1] > IMAX(env->max_S, R[1]);
425 Ns = decompose_mul(env, &R[1], r - 1, N);
426 return emit_LEA(env, env->root, Ns, t);
427 } else if (R[0] <= env->max_S) { /* Case 2 */
430 Ns = decompose_mul(env, &R[1], r - 1, N);
431 return emit_LEA(env, Ns, env->root, t);
435 Ns = decompose_mul(env, R, r, N);
436 return emit_SHIFT(env, Ns, t);
441 * recursive build the graph form the instructions.
443 * @param env the environment
444 * @param inst the instruction
446 static ir_node *build_graph(mul_env *env, instruction *inst)
453 switch (inst->kind) {
455 l = build_graph(env, inst->in[0]);
456 r = build_graph(env, inst->in[1]);
457 c = new_Const_long(env->shf_mode, inst->shift_count);
458 r = new_rd_Shl(env->dbg, env->blk, r, c, env->mode);
459 return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
461 l = build_graph(env, inst->in[0]);
462 c = new_Const_long(env->shf_mode, inst->shift_count);
463 return inst->irn = new_rd_Shl(env->dbg, env->blk, l, c, env->mode);
465 l = build_graph(env, inst->in[0]);
466 r = build_graph(env, inst->in[1]);
467 return inst->irn = new_rd_Sub(env->dbg, env->blk, l, r, env->mode);
469 l = build_graph(env, inst->in[0]);
470 r = build_graph(env, inst->in[1]);
471 return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
473 return inst->irn = new_Const(get_mode_null(env->mode));
475 panic("Unsupported instruction kind");
480 * Calculate the costs for the given instruction sequence.
481 * Note that additional costs due to higher register pressure are NOT evaluated yet
483 static int evaluate_insn(mul_env *env, instruction *inst)
487 if (inst->costs >= 0) {
488 /* was already evaluated */
492 switch (inst->kind) {
496 costs = evaluate_insn(env, inst->in[0]);
497 costs += evaluate_insn(env, inst->in[1]);
498 costs += env->evaluate(inst->kind, NULL);
502 if (inst->shift_count > params->highest_shift_amount)
504 if (env->n_shift <= 0)
508 costs = evaluate_insn(env, inst->in[0]);
509 costs += env->evaluate(inst->kind, NULL);
513 inst->costs = costs = env->evaluate(inst->kind, NULL);
519 panic("Unsupported instruction kind");
523 * Evaluate the replacement instructions and build a new graph
524 * if faster than the Mul.
525 * Returns the root of the new graph then or irn otherwise.
527 * @param irn the Mul operation
528 * @param operand the multiplication operand
529 * @param tv the multiplication constant
531 * @return the new graph
533 static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv)
542 obstack_init(&env.obst);
543 env.mode = get_tarval_mode(tv);
544 env.bits = (unsigned)get_mode_size_bits(env.mode);
546 env.root = emit_ROOT(&env, operand);
548 env.n_shift = params->maximum_shifts;
549 env.evaluate = params->evaluate != NULL ? params->evaluate : default_evaluate;
551 R = value_to_condensed(&env, tv, &r);
552 inst = decompose_mul(&env, R, r, tv);
554 /* the paper suggests 70% here */
555 mul_costs = (env.evaluate(MUL, tv) * 7 + 5) / 10;
556 if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) {
558 env.blk = get_nodes_block(irn);
559 env.dbg = get_irn_dbg_info(irn);
560 env.shf_mode = find_unsigned_mode(env.mode);
561 if (env.shf_mode == NULL)
562 env.shf_mode = mode_Iu;
564 res = build_graph(&env, inst);
566 obstack_free(&env.obst, NULL);
570 /* Replace Muls with Shifts and Add/Subs. */
571 ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn)
575 ir_mode *mode = get_irn_mode(irn);
582 /* If the architecture dependent optimizations were not initialized
583 or this optimization was not enabled. */
584 if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
587 if (!is_Mul(irn) || !mode_is_int(mode))
590 /* we should never do the reverse transformations again
592 irg = get_irn_irg(irn);
593 set_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP);
595 left = get_binop_left(irn);
596 right = get_binop_right(irn);
600 /* Look, if one operand is a constant. */
601 if (is_Const(left)) {
602 tv = get_Const_tarval(left);
604 } else if (is_Const(right)) {
605 tv = get_Const_tarval(right);
610 res = do_decomposition(irn, operand, tv);
613 hook_arch_dep_replace_mul_with_shifts(irn);
622 * calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
624 static int tv_ld2(tarval *tv, int bits)
628 for (num = i = 0; i < bits; ++i) {
629 unsigned char v = get_tarval_sub_bits(tv, i);
634 for (j = 0; j < 8; ++j)
647 /* for shorter lines */
648 #define ABS(a) tarval_abs(a)
649 #define NEG(a) tarval_neg(a)
650 #define NOT(a) tarval_not(a)
651 #define SHL(a, b) tarval_shl(a, b)
652 #define SHR(a, b) tarval_shr(a, b)
653 #define ADD(a, b) tarval_add(a, b)
654 #define SUB(a, b) tarval_sub(a, b, NULL)
655 #define MUL(a, b) tarval_mul(a, b)
656 #define DIV(a, b) tarval_div(a, b)
657 #define MOD(a, b) tarval_mod(a, b)
658 #define CMP(a, b) tarval_cmp(a, b)
659 #define CNV(a, m) tarval_convert_to(a, m)
660 #define ONE(m) get_mode_one(m)
661 #define ZERO(m) get_mode_null(m)
663 /** The result of a the magic() function. */
665 tarval *M; /**< magic number */
666 int s; /**< shift amount */
667 int need_add; /**< an additional add is needed */
668 int need_sub; /**< an additional sub is needed */
672 * Signed division by constant d: calculate the Magic multiplier M and the shift amount s
674 * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
676 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)
764 ir_mode *mode = get_tarval_mode(d);
765 int bits = get_mode_size_bits(mode);
767 tarval *nc, *delta, *q1, *r1, *q2, *r2;
768 tarval *bits_minus_1, *two_bits_1, *seven_ff;
772 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
774 /* we need overflow mode to work correctly */
775 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
777 bits_minus_1 = new_tarval_from_long(bits - 1, mode);
778 two_bits_1 = SHL(get_mode_one(mode), bits_minus_1);
779 seven_ff = SUB(two_bits_1, ONE(mode));
781 magu.need_add = 0; /* initialize the add indicator */
782 nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d));
783 p = bits - 1; /* Init: p */
784 q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */
785 r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */
786 q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */
787 r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */
791 if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
792 q1 = ADD(ADD(q1, q1), ONE(mode));
793 r1 = SUB(ADD(r1, r1), nc);
800 if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
801 if (CMP(q2, seven_ff) & pn_Cmp_Ge)
804 q2 = ADD(ADD(q2, q2), ONE(mode));
805 r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
808 if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
812 r2 = ADD(ADD(r2, r2), ONE(mode));
814 delta = SUB(SUB(d, ONE(mode)), r2);
815 } while (p < 2*bits &&
816 (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
818 magu.M = ADD(q2, ONE(mode)); /* Magic number */
819 magu.s = p - bits; /* and shift amount */
821 tarval_set_integer_overflow_mode(rem);
827 * Build the Mulh replacement code for n / tv.
829 * Note that 'div' might be a mod or DivMod operation as well
831 static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv)
833 dbg_info *dbg = get_irn_dbg_info(div);
834 ir_node *n = get_binop_left(div);
835 ir_node *block = get_irn_n(div, -1);
836 ir_mode *mode = get_irn_mode(n);
837 int bits = get_mode_size_bits(mode);
840 /* Beware: do not transform bad code */
841 if (is_Bad(n) || is_Bad(block))
844 if (mode_is_signed(mode)) {
845 struct ms mag = magic(tv);
847 /* generate the Mulh instruction */
848 c = new_Const(mag.M);
849 q = new_rd_Mulh(dbg, block, n, c, mode);
851 /* do we need an Add or Sub */
853 q = new_rd_Add(dbg, block, q, n, mode);
854 else if (mag.need_sub)
855 q = new_rd_Sub(dbg, block, q, n, mode);
857 /* Do we need the shift */
859 c = new_Const_long(mode_Iu, mag.s);
860 q = new_rd_Shrs(dbg, block, q, c, mode);
864 c = new_Const_long(mode_Iu, bits - 1);
865 t = new_rd_Shr(dbg, block, q, c, mode);
867 q = new_rd_Add(dbg, block, q, t, mode);
869 struct mu mag = magicu(tv);
872 /* generate the Mulh instruction */
873 c = new_Const(mag.M);
874 q = new_rd_Mulh(dbg, block, n, c, mode);
878 /* use the GM scheme */
879 t = new_rd_Sub(dbg, block, n, q, mode);
881 c = new_Const(get_mode_one(mode_Iu));
882 t = new_rd_Shr(dbg, block, t, c, mode);
884 t = new_rd_Add(dbg, block, t, q, mode);
886 c = new_Const_long(mode_Iu, mag.s - 1);
887 q = new_rd_Shr(dbg, block, t, c, mode);
889 /* use the default scheme */
890 q = new_rd_Add(dbg, block, q, n, mode);
892 } else if (mag.s > 0) { /* default scheme, shift needed */
893 c = new_Const_long(mode_Iu, mag.s);
894 q = new_rd_Shr(dbg, block, q, c, mode);
900 /* Replace Divs with Shifts and Add/Subs and Mulh. */
901 ir_node *arch_dep_replace_div_by_const(ir_node *irn)
905 /* If the architecture dependent optimizations were not initialized
906 or this optimization was not enabled. */
907 if (params == NULL || (opts & arch_dep_div_by_const) == 0)
911 ir_node *c = get_Div_right(irn);
912 ir_node *block, *left;
923 tv = get_Const_tarval(c);
925 /* check for division by zero */
926 if (tarval_is_null(tv))
929 left = get_Div_left(irn);
930 mode = get_irn_mode(left);
931 block = get_irn_n(irn, -1);
932 dbg = get_irn_dbg_info(irn);
934 bits = get_mode_size_bits(mode);
938 if (mode_is_signed(mode)) {
939 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
940 ntv = tarval_neg(tv);
950 if (k >= 0) { /* division by 2^k or -2^k */
951 if (mode_is_signed(mode)) {
953 ir_node *curr = left;
955 /* create the correction code for signed values only if there might be a remainder */
956 if (! get_Div_no_remainder(irn)) {
958 k_node = new_Const_long(mode_Iu, k - 1);
959 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
962 k_node = new_Const_long(mode_Iu, bits - k);
963 curr = new_rd_Shr(dbg, block, curr, k_node, mode);
965 curr = new_rd_Add(dbg, block, left, curr, mode);
970 k_node = new_Const_long(mode_Iu, k);
971 res = new_rd_Shrs(dbg, block, curr, k_node, mode);
973 if (n_flag) { /* negate the result */
976 k_node = new_Const(get_mode_null(mode));
977 res = new_rd_Sub(dbg, block, k_node, res, mode);
979 } else { /* unsigned case */
982 k_node = new_Const_long(mode_Iu, k);
983 res = new_rd_Shr(dbg, block, left, k_node, mode);
987 if (allow_Mulh(mode))
988 res = replace_div_by_mulh(irn, tv);
993 hook_arch_dep_replace_division_by_const(irn);
998 /* Replace Mods with Shifts and Add/Subs and Mulh. */
999 ir_node *arch_dep_replace_mod_by_const(ir_node *irn)
1003 /* If the architecture dependent optimizations were not initialized
1004 or this optimization was not enabled. */
1005 if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
1009 ir_node *c = get_Mod_right(irn);
1010 ir_node *block, *left;
1020 tv = get_Const_tarval(c);
1022 /* check for division by zero */
1023 if (tarval_is_null(tv))
1026 left = get_Mod_left(irn);
1027 mode = get_irn_mode(left);
1028 block = get_irn_n(irn, -1);
1029 dbg = get_irn_dbg_info(irn);
1030 bits = get_mode_size_bits(mode);
1034 if (mode_is_signed(mode)) {
1035 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1036 ntv = tarval_neg(tv);
1045 /* division by 2^k or -2^k:
1046 * we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k
1048 if (mode_is_signed(mode)) {
1050 ir_node *curr = left;
1053 k_node = new_Const_long(mode_Iu, k - 1);
1054 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
1057 k_node = new_Const_long(mode_Iu, bits - k);
1058 curr = new_rd_Shr(dbg, block, curr, k_node, mode);
1060 curr = new_rd_Add(dbg, block, left, curr, mode);
1062 k_node = new_Const_long(mode, (-1) << k);
1063 curr = new_rd_And(dbg, block, curr, k_node, mode);
1065 res = new_rd_Sub(dbg, block, left, curr, mode);
1066 } else { /* unsigned case */
1069 k_node = new_Const_long(mode, (1 << k) - 1);
1070 res = new_rd_And(dbg, block, left, k_node, mode);
1073 /* other constant */
1074 if (allow_Mulh(mode)) {
1075 res = replace_div_by_mulh(irn, tv);
1077 res = new_rd_Mul(dbg, block, res, c, mode);
1079 /* res = arch_dep_mul_to_shift(res); */
1081 res = new_rd_Sub(dbg, block, left, res, mode);
1087 hook_arch_dep_replace_division_by_const(irn);
1092 /* Replace DivMods with Shifts and Add/Subs and Mulh. */
1093 void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn)
1097 /* If the architecture dependent optimizations were not initialized
1098 or this optimization was not enabled. */
1099 if (params == NULL ||
1100 ((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
1103 if (is_DivMod(irn)) {
1104 ir_node *c = get_DivMod_right(irn);
1105 ir_node *block, *left;
1116 tv = get_Const_tarval(c);
1118 /* check for division by zero */
1119 if (tarval_is_null(tv))
1122 left = get_DivMod_left(irn);
1123 mode = get_irn_mode(left);
1124 block = get_irn_n(irn, -1);
1125 dbg = get_irn_dbg_info(irn);
1127 bits = get_mode_size_bits(mode);
1131 if (mode_is_signed(mode)) {
1132 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1133 ntv = tarval_neg(tv);
1143 if (k >= 0) { /* division by 2^k or -2^k */
1144 if (mode_is_signed(mode)) {
1145 ir_node *k_node, *c_k;
1146 ir_node *curr = left;
1149 k_node = new_Const_long(mode_Iu, k - 1);
1150 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
1153 k_node = new_Const_long(mode_Iu, bits - k);
1154 curr = new_rd_Shr(dbg, block, curr, k_node, mode);
1156 curr = new_rd_Add(dbg, block, left, curr, mode);
1158 c_k = new_Const_long(mode_Iu, k);
1160 *div = new_rd_Shrs(dbg, block, curr, c_k, mode);
1162 if (n_flag) { /* negate the div result */
1165 k_node = new_Const(get_mode_null(mode));
1166 *div = new_rd_Sub(dbg, block, k_node, *div, mode);
1169 k_node = new_Const_long(mode, (-1) << k);
1170 curr = new_rd_And(dbg, block, curr, k_node, mode);
1172 *mod = new_rd_Sub(dbg, block, left, curr, mode);
1173 } else { /* unsigned case */
1176 k_node = new_Const_long(mode_Iu, k);
1177 *div = new_rd_Shr(dbg, block, left, k_node, mode);
1179 k_node = new_Const_long(mode, (1 << k) - 1);
1180 *mod = new_rd_And(dbg, block, left, k_node, mode);
1183 /* other constant */
1184 if (allow_Mulh(mode)) {
1187 *div = replace_div_by_mulh(irn, tv);
1189 t = new_rd_Mul(dbg, block, *div, c, mode);
1191 /* t = arch_dep_mul_to_shift(t); */
1193 *mod = new_rd_Sub(dbg, block, left, t, mode);
1199 hook_arch_dep_replace_division_by_const(irn);
1203 static const ir_settings_arch_dep_t default_params = {
1204 1, /* also use subs */
1205 4, /* maximum shifts */
1206 31, /* maximum shift amount */
1207 default_evaluate, /* default evaluator */
1209 0, /* allow Mulhs */
1210 0, /* allow Mulus */
1211 32 /* Mulh allowed up to 32 bit */
1214 /* A default parameter factory for testing purposes. */
1215 const ir_settings_arch_dep_t *arch_dep_default_factory(void)
1217 return &default_params;