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 "Strenght 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 /* when we need verifying */
58 # define IRN_VRFY_IRG(res, irg)
60 # define IRN_VRFY_IRG(res, irg) irn_vrfy_irg(res, irg)
63 /** The params got from the factory in arch_dep_init(...). */
64 static const ir_settings_arch_dep_t *params = NULL;
66 /** The bit mask, which optimizations to apply. */
67 static arch_dep_opts_t opts;
69 void arch_dep_init(arch_dep_params_factory_t factory) {
76 void arch_dep_set_opts(arch_dep_opts_t the_opts) {
80 /** check, whether a mode allows a Mulh instruction. */
81 static int allow_Mulh(ir_mode *mode) {
82 if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
84 return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
90 typedef struct instruction instruction;
92 insn_kind kind; /**< the instruction kind */
93 instruction *in[2]; /**< the ins */
94 unsigned shift_count; /**< shift count for LEA and SHIFT */
95 ir_node *irn; /**< the generated node for this instruction if any. */
96 int costs; /**< the costs for this instruction */
100 * The environment for the strength reduction of multiplications.
102 typedef struct _mul_env {
103 struct obstack obst; /**< an obstack for local space. */
104 ir_mode *mode; /**< the mode of the multiplication constant */
105 unsigned bits; /**< number of bits in the mode */
106 unsigned max_S; /**< the maximum LEA shift value. */
107 instruction *root; /**< the root of the instruction tree */
108 ir_node *op; /**< the operand that is multiplied */
109 ir_node *blk; /**< the block where the new graph is built */
110 dbg_info *dbg; /**< the debug info for the new graph. */
111 ir_mode *shf_mode; /**< the (unsigned) mode for the shift constants */
112 int fail; /**< set to 1 if the instruction sequence fails the constraints */
113 int n_shift; /**< maximum number of allowed shift instructions */
115 evaluate_costs_func evaluate; /**< the evaluate callback */
119 * Some kind of default evaluator. Return the cost of
122 static int default_evaluate(insn_kind kind, tarval *tv) {
131 * emit a LEA (or an Add) instruction
133 static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) {
134 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
135 res->kind = shift > 0 ? LEA : ADD;
138 res->shift_count = shift;
145 * emit a SHIFT (or an Add or a Zero) instruction
147 static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) {
148 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
149 if (shift == env->bits) {
150 /* a 2^bits with bits resolution is a zero */
154 res->shift_count = 0;
155 } else if (shift != 1) {
159 res->shift_count = shift;
164 res->shift_count = 0;
172 * emit a SUB instruction
174 static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) {
175 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
179 res->shift_count = 0;
186 * emit the ROOT instruction
188 static instruction *emit_ROOT(mul_env *env, ir_node *root_op) {
189 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
193 res->shift_count = 0;
201 * Returns the condensed representation of the tarval tv
203 static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) {
204 ir_mode *mode = get_tarval_mode(tv);
205 int bits = get_mode_size_bits(mode);
206 char *bitstr = get_tarval_bitpattern(tv);
208 unsigned char *R = obstack_alloc(&env->obst, bits);
211 for (i = 0; bitstr[i] != '\0'; ++i) {
212 if (bitstr[i] == '1') {
225 * Calculate the gain when using the generalized complementary technique
227 static int calculate_gain(unsigned char *R, int r) {
232 /* the gain for r == 1 */
234 for (i = 2; i < r; ++i) {
235 /* calculate the gain for r from the gain for r-1 */
236 gain += 2 - R[i - 1];
238 if (gain > max_gain) {
247 * Calculates the condensed complement of a given (R,r) tuple
249 static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs) {
250 unsigned char *value = obstack_alloc(&env->obst, env->bits);
254 memset(value, 0, env->bits);
257 for (i = 0; i < gain; ++i) {
262 /* negate and propagate 1 */
264 for (i = 0; i <= j; ++i) {
265 unsigned char v = !value[i];
271 /* condense it again */
274 for (i = 0; i <= j; ++i) {
287 * creates a tarval from a condensed representation.
289 static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) {
294 tv = get_mode_one(env->mode);
296 for (i = 0; i < r; ++i) {
299 tarval *t = new_tarval_from_long(j, mode_Iu);
300 tv = tarval_shl(tv, t);
302 res = res ? tarval_add(res, tv) : tv;
308 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
311 * handle simple cases with up-to 2 bits set
313 static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) {
314 instruction *ins, *ins2;
318 return emit_SHIFT(env, env->root, R[0]);
323 if (R[1] <= env->max_S) {
324 ins = emit_LEA(env, ins, ins, R[1]);
326 ins = emit_SHIFT(env, ins, R[0]);
331 ins = emit_SHIFT(env, ins, R[0]);
334 ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
335 return emit_LEA(env, ins, ins2, 0);
340 * Main decompose driver.
342 static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
347 return decompose_simple_cases(env, R, r, N);
349 if (params->also_use_subs) {
350 gain = calculate_gain(R, r);
352 instruction *instr1, *instr2;
353 unsigned char *R1, *R2;
356 R1 = complement_condensed(env, R, r, gain, &r1);
358 R2 = obstack_alloc(&env->obst, r2);
361 for (i = 0; i < gain; ++i) {
368 /* Two identical bits: normalize */
373 for (i = gain + 1; i < r; ++i) {
377 instr1 = decompose_mul(env, R1, r1, NULL);
378 instr2 = decompose_mul(env, R2, r2, NULL);
379 return emit_SUB(env, instr2, instr1);
384 N = condensed_to_value(env, R, r);
386 for (i = env->max_S; i > 0; --i) {
387 tarval *div_res, *mod_res;
388 tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
390 div_res = tarval_divmod(N, tv, &mod_res);
391 if (mod_res == get_mode_null(env->mode)) {
395 Rs = value_to_condensed(env, div_res, &rs);
397 instruction *N1 = decompose_mul(env, Rs, rs, div_res);
398 return emit_LEA(env, N1, N1, i);
402 return basic_decompose_mul(env, R, r, N);
405 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
408 * basic decomposition routine
410 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
414 if (R[0] == 0) { /* Case 1 */
415 t = R[1] > IMAX(env->max_S, R[1]);
417 Ns = decompose_mul(env, &R[1], r - 1, N);
418 return emit_LEA(env, env->root, Ns, t);
419 } else if (R[0] <= env->max_S) { /* Case 2 */
422 Ns = decompose_mul(env, &R[1], r - 1, N);
423 return emit_LEA(env, Ns, env->root, t);
427 Ns = decompose_mul(env, R, r, N);
428 return emit_SHIFT(env, Ns, t);
433 * recursive build the graph form the instructions.
435 * @param env the environment
436 * @param inst the instruction
438 static ir_node *build_graph(mul_env *env, instruction *inst) {
444 switch (inst->kind) {
446 l = build_graph(env, inst->in[0]);
447 r = build_graph(env, inst->in[1]);
448 c = new_Const_long(env->shf_mode, inst->shift_count);
449 r = new_rd_Shl(env->dbg, current_ir_graph, env->blk, r, c, env->mode);
450 return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
452 l = build_graph(env, inst->in[0]);
453 c = new_Const_long(env->shf_mode, inst->shift_count);
454 return inst->irn = new_rd_Shl(env->dbg, current_ir_graph, env->blk, l, c, env->mode);
456 l = build_graph(env, inst->in[0]);
457 r = build_graph(env, inst->in[1]);
458 return inst->irn = new_rd_Sub(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
460 l = build_graph(env, inst->in[0]);
461 r = build_graph(env, inst->in[1]);
462 return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
464 return inst->irn = new_Const(get_mode_null(env->mode));
466 panic("Unsupported instruction kind");
472 * Calculate the costs for the given instruction sequence.
473 * Note that additional costs due to higher register pressure are NOT evaluated yet
475 static int evaluate_insn(mul_env *env, instruction *inst) {
478 if (inst->costs >= 0) {
479 /* was already evaluated */
483 switch (inst->kind) {
487 costs = evaluate_insn(env, inst->in[0]);
488 costs += evaluate_insn(env, inst->in[1]);
489 costs += env->evaluate(inst->kind, NULL);
493 if (inst->shift_count > params->highest_shift_amount)
495 if (env->n_shift <= 0)
499 costs = evaluate_insn(env, inst->in[0]);
500 costs += env->evaluate(inst->kind, NULL);
504 inst->costs = costs = env->evaluate(inst->kind, NULL);
510 panic("Unsupported instruction kind");
514 * Evaluate the replacement instructions and build a new graph
515 * if faster than the Mul.
516 * Returns the root of the new graph then or irn otherwise.
518 * @param irn the Mul operation
519 * @param operand the multiplication operand
520 * @param tv the multiplication constant
522 * @return the new graph
524 static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) {
532 obstack_init(&env.obst);
533 env.mode = get_tarval_mode(tv);
534 env.bits = (unsigned)get_mode_size_bits(env.mode);
536 env.root = emit_ROOT(&env, operand);
538 env.n_shift = params->maximum_shifts;
539 env.evaluate = params->evaluate != NULL ? params->evaluate : default_evaluate;
541 R = value_to_condensed(&env, tv, &r);
542 inst = decompose_mul(&env, R, r, tv);
544 /* the paper suggests 70% here */
545 mul_costs = (env.evaluate(MUL, tv) * 7 + 5) / 10;
546 if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) {
548 env.blk = get_nodes_block(irn);
549 env.dbg = get_irn_dbg_info(irn);
550 env.shf_mode = find_unsigned_mode(env.mode);
551 if (env.shf_mode == NULL)
552 env.shf_mode = mode_Iu;
554 res = build_graph(&env, inst);
556 obstack_free(&env.obst, NULL);
560 /* Replace Muls with Shifts and Add/Subs. */
561 ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) {
564 ir_mode *mode = get_irn_mode(irn);
571 /* If the architecture dependent optimizations were not initialized
572 or this optimization was not enabled. */
573 if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
576 if (!is_Mul(irn) || !mode_is_int(mode))
579 /* we should never do the reverse transformations again
581 irg = get_irn_irg(irn);
582 set_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP);
584 left = get_binop_left(irn);
585 right = get_binop_right(irn);
589 /* Look, if one operand is a constant. */
590 if (is_Const(left)) {
591 tv = get_Const_tarval(left);
593 } else if (is_Const(right)) {
594 tv = get_Const_tarval(right);
599 res = do_decomposition(irn, operand, tv);
602 hook_arch_dep_replace_mul_with_shifts(irn);
611 * calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
613 static int tv_ld2(tarval *tv, int bits) {
616 for (num = i = 0; i < bits; ++i) {
617 unsigned char v = get_tarval_sub_bits(tv, i);
622 for (j = 0; j < 8; ++j)
635 /* for shorter lines */
636 #define ABS(a) tarval_abs(a)
637 #define NEG(a) tarval_neg(a)
638 #define NOT(a) tarval_not(a)
639 #define SHL(a, b) tarval_shl(a, b)
640 #define SHR(a, b) tarval_shr(a, b)
641 #define ADD(a, b) tarval_add(a, b)
642 #define SUB(a, b) tarval_sub(a, b, NULL)
643 #define MUL(a, b) tarval_mul(a, b)
644 #define DIV(a, b) tarval_div(a, b)
645 #define MOD(a, b) tarval_mod(a, b)
646 #define CMP(a, b) tarval_cmp(a, b)
647 #define CNV(a, m) tarval_convert_to(a, m)
648 #define ONE(m) get_mode_one(m)
649 #define ZERO(m) get_mode_null(m)
651 /** The result of a the magic() function. */
653 tarval *M; /**< magic number */
654 int s; /**< shift amount */
655 int need_add; /**< an additional add is needed */
656 int need_sub; /**< an additional sub is needed */
660 * Signed division by constant d: calculate the Magic multiplier M and the shift amount s
662 * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
664 static struct ms magic(tarval *d) {
665 ir_mode *mode = get_tarval_mode(d);
666 ir_mode *u_mode = find_unsigned_mode(mode);
667 int bits = get_mode_size_bits(u_mode);
669 tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
672 tarval *bits_minus_1, *two_bits_1;
676 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
678 /* we need overflow mode to work correctly */
679 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
682 bits_minus_1 = new_tarval_from_long(bits - 1, u_mode);
683 two_bits_1 = SHL(get_mode_one(u_mode), bits_minus_1);
685 ad = CNV(ABS(d), u_mode);
686 t = ADD(two_bits_1, SHR(CNV(d, u_mode), bits_minus_1));
687 anc = SUB(SUB(t, ONE(u_mode)), MOD(t, ad)); /* Absolute value of nc */
688 p = bits - 1; /* Init: p */
689 q1 = DIV(two_bits_1, anc); /* Init: q1 = 2^p/|nc| */
690 r1 = SUB(two_bits_1, MUL(q1, anc)); /* Init: r1 = rem(2^p, |nc|) */
691 q2 = DIV(two_bits_1, ad); /* Init: q2 = 2^p/|d| */
692 r2 = SUB(two_bits_1, MUL(q2, ad)); /* Init: r2 = rem(2^p, |d|) */
696 q1 = ADD(q1, q1); /* Update q1 = 2^p/|nc| */
697 r1 = ADD(r1, r1); /* Update r1 = rem(2^p, |nc|) */
699 if (CMP(r1, anc) & pn_Cmp_Ge) {
700 q1 = ADD(q1, ONE(u_mode));
704 q2 = ADD(q2, q2); /* Update q2 = 2^p/|d| */
705 r2 = ADD(r2, r2); /* Update r2 = rem(2^p, |d|) */
707 if (CMP(r2, ad) & pn_Cmp_Ge) {
708 q2 = ADD(q2, ONE(u_mode));
713 } while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq));
715 d_cmp = CMP(d, ZERO(mode));
717 if (d_cmp & pn_Cmp_Ge)
718 mag.M = ADD(CNV(q2, mode), ONE(mode));
720 mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode)));
722 M_cmp = CMP(mag.M, ZERO(mode));
726 /* need an add if d > 0 && M < 0 */
727 mag.need_add = d_cmp & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt;
729 /* need a sub if d < 0 && M > 0 */
730 mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt;
732 tarval_set_integer_overflow_mode(rem);
737 /** The result of the magicu() function. */
739 tarval *M; /**< magic add constant */
740 int s; /**< shift amount */
741 int need_add; /**< add indicator */
745 * Unsigned division by constant d: calculate the Magic multiplier M and the shift amount s
747 * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned)
749 static struct mu magicu(tarval *d) {
750 ir_mode *mode = get_tarval_mode(d);
751 int bits = get_mode_size_bits(mode);
753 tarval *nc, *delta, *q1, *r1, *q2, *r2;
754 tarval *bits_minus_1, *two_bits_1, *seven_ff;
758 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
760 /* we need overflow mode to work correctly */
761 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
763 bits_minus_1 = new_tarval_from_long(bits - 1, mode);
764 two_bits_1 = SHL(get_mode_one(mode), bits_minus_1);
765 seven_ff = SUB(two_bits_1, ONE(mode));
767 magu.need_add = 0; /* initialize the add indicator */
768 nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d));
769 p = bits - 1; /* Init: p */
770 q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */
771 r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */
772 q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */
773 r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */
777 if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
778 q1 = ADD(ADD(q1, q1), ONE(mode));
779 r1 = SUB(ADD(r1, r1), nc);
786 if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
787 if (CMP(q2, seven_ff) & pn_Cmp_Ge)
790 q2 = ADD(ADD(q2, q2), ONE(mode));
791 r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
794 if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
798 r2 = ADD(ADD(r2, r2), ONE(mode));
800 delta = SUB(SUB(d, ONE(mode)), r2);
801 } while (p < 2*bits &&
802 (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
804 magu.M = ADD(q2, ONE(mode)); /* Magic number */
805 magu.s = p - bits; /* and shift amount */
807 tarval_set_integer_overflow_mode(rem);
813 * Build the Mulh replacement code for n / tv.
815 * Note that 'div' might be a mod or DivMod operation as well
817 static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
818 dbg_info *dbg = get_irn_dbg_info(div);
819 ir_node *n = get_binop_left(div);
820 ir_node *block = get_irn_n(div, -1);
821 ir_mode *mode = get_irn_mode(n);
822 int bits = get_mode_size_bits(mode);
825 /* Beware: do not transform bad code */
826 if (is_Bad(n) || is_Bad(block))
829 if (mode_is_signed(mode)) {
830 struct ms mag = magic(tv);
832 /* generate the Mulh instruction */
833 c = new_Const(mag.M);
834 q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
836 /* do we need an Add or Sub */
838 q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
839 else if (mag.need_sub)
840 q = new_rd_Sub(dbg, current_ir_graph, block, q, n, mode);
842 /* Do we need the shift */
844 c = new_Const_long(mode_Iu, mag.s);
845 q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
849 c = new_Const_long(mode_Iu, bits - 1);
850 t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
852 q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode);
854 struct mu mag = magicu(tv);
857 /* generate the Mulh instruction */
858 c = new_Const(mag.M);
859 q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
863 /* use the GM scheme */
864 t = new_rd_Sub(dbg, current_ir_graph, block, n, q, mode);
866 c = new_Const(get_mode_one(mode_Iu));
867 t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
869 t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode);
871 c = new_Const_long(mode_Iu, mag.s - 1);
872 q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
874 /* use the default scheme */
875 q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
877 } else if (mag.s > 0) { /* default scheme, shift needed */
878 c = new_Const_long(mode_Iu, mag.s);
879 q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
885 /* Replace Divs with Shifts and Add/Subs and Mulh. */
886 ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
889 /* If the architecture dependent optimizations were not initialized
890 or this optimization was not enabled. */
891 if (params == NULL || (opts & arch_dep_div_by_const) == 0)
895 ir_node *c = get_Div_right(irn);
896 ir_node *block, *left;
906 tv = get_Const_tarval(c);
908 /* check for division by zero */
909 if (tarval_is_null(tv))
912 left = get_Div_left(irn);
913 mode = get_irn_mode(left);
914 block = get_irn_n(irn, -1);
915 dbg = get_irn_dbg_info(irn);
917 bits = get_mode_size_bits(mode);
921 if (mode_is_signed(mode)) {
922 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
923 ntv = tarval_neg(tv);
933 if (k >= 0) { /* division by 2^k or -2^k */
934 if (mode_is_signed(mode)) {
936 ir_node *curr = left;
938 /* create the correction code for signed values only if there might be a remainder */
939 if (! get_Div_no_remainder(irn)) {
941 k_node = new_Const_long(mode_Iu, k - 1);
942 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
945 k_node = new_Const_long(mode_Iu, bits - k);
946 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
948 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
953 k_node = new_Const_long(mode_Iu, k);
954 res = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode);
956 if (n_flag) { /* negate the result */
959 k_node = new_Const(get_mode_null(mode));
960 res = new_rd_Sub(dbg, current_ir_graph, block, k_node, res, mode);
962 } else { /* unsigned case */
965 k_node = new_Const_long(mode_Iu, k);
966 res = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
970 if (allow_Mulh(mode))
971 res = replace_div_by_mulh(irn, tv);
976 hook_arch_dep_replace_division_by_const(irn);
981 /* Replace Mods with Shifts and Add/Subs and Mulh. */
982 ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
985 /* If the architecture dependent optimizations were not initialized
986 or this optimization was not enabled. */
987 if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
991 ir_node *c = get_Mod_right(irn);
992 ir_node *block, *left;
1002 tv = get_Const_tarval(c);
1004 /* check for division by zero */
1005 if (tarval_is_null(tv))
1008 left = get_Mod_left(irn);
1009 mode = get_irn_mode(left);
1010 block = get_irn_n(irn, -1);
1011 dbg = get_irn_dbg_info(irn);
1012 bits = get_mode_size_bits(mode);
1016 if (mode_is_signed(mode)) {
1017 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1018 ntv = tarval_neg(tv);
1027 /* division by 2^k or -2^k:
1028 * we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k
1030 if (mode_is_signed(mode)) {
1032 ir_node *curr = left;
1035 k_node = new_Const_long(mode_Iu, k - 1);
1036 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
1039 k_node = new_Const_long(mode_Iu, bits - k);
1040 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
1042 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
1044 k_node = new_Const_long(mode, (-1) << k);
1045 curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
1047 res = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
1048 } else { /* unsigned case */
1051 k_node = new_Const_long(mode, (1 << k) - 1);
1052 res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
1055 /* other constant */
1056 if (allow_Mulh(mode)) {
1057 res = replace_div_by_mulh(irn, tv);
1059 res = new_rd_Mul(dbg, current_ir_graph, block, res, c, mode);
1061 /* res = arch_dep_mul_to_shift(res); */
1063 res = new_rd_Sub(dbg, current_ir_graph, block, left, res, mode);
1069 hook_arch_dep_replace_division_by_const(irn);
1074 /* Replace DivMods with Shifts and Add/Subs and Mulh. */
1075 void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn) {
1078 /* If the architecture dependent optimizations were not initialized
1079 or this optimization was not enabled. */
1080 if (params == NULL ||
1081 ((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
1084 if (is_DivMod(irn)) {
1085 ir_node *c = get_DivMod_right(irn);
1086 ir_node *block, *left;
1096 tv = get_Const_tarval(c);
1098 /* check for division by zero */
1099 if (tarval_is_null(tv))
1102 left = get_DivMod_left(irn);
1103 mode = get_irn_mode(left);
1104 block = get_irn_n(irn, -1);
1105 dbg = get_irn_dbg_info(irn);
1107 bits = get_mode_size_bits(mode);
1111 if (mode_is_signed(mode)) {
1112 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1113 ntv = tarval_neg(tv);
1123 if (k >= 0) { /* division by 2^k or -2^k */
1124 if (mode_is_signed(mode)) {
1125 ir_node *k_node, *c_k;
1126 ir_node *curr = left;
1129 k_node = new_Const_long(mode_Iu, k - 1);
1130 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
1133 k_node = new_Const_long(mode_Iu, bits - k);
1134 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
1136 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
1138 c_k = new_Const_long(mode_Iu, k);
1140 *div = new_rd_Shrs(dbg, current_ir_graph, block, curr, c_k, mode);
1142 if (n_flag) { /* negate the div result */
1145 k_node = new_Const(get_mode_null(mode));
1146 *div = new_rd_Sub(dbg, current_ir_graph, block, k_node, *div, mode);
1149 k_node = new_Const_long(mode, (-1) << k);
1150 curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
1152 *mod = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
1153 } else { /* unsigned case */
1156 k_node = new_Const_long(mode_Iu, k);
1157 *div = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
1159 k_node = new_Const_long(mode, (1 << k) - 1);
1160 *mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
1163 /* other constant */
1164 if (allow_Mulh(mode)) {
1167 *div = replace_div_by_mulh(irn, tv);
1169 t = new_rd_Mul(dbg, current_ir_graph, block, *div, c, mode);
1171 /* t = arch_dep_mul_to_shift(t); */
1173 *mod = new_rd_Sub(dbg, current_ir_graph, block, left, t, mode);
1179 hook_arch_dep_replace_division_by_const(irn);
1183 static const ir_settings_arch_dep_t default_params = {
1184 1, /* also use subs */
1185 4, /* maximum shifts */
1186 31, /* maximum shift amount */
1187 default_evaluate, /* default evaluator */
1189 0, /* allow Mulhs */
1190 0, /* allow Mulus */
1191 32 /* Mulh allowed up to 32 bit */
1194 /* A default parameter factory for testing purposes. */
1195 const ir_settings_arch_dep_t *arch_dep_default_factory(void) {
1196 return &default_params;