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",
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) {
84 /** check, whether a mode allows a Mulh instruction. */
85 static int allow_Mulh(ir_mode *mode) {
86 if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
88 return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
94 typedef struct instruction instruction;
96 insn_kind kind; /**< the instruction kind */
97 instruction *in[2]; /**< the ins */
98 unsigned shift_count; /**< shift count for LEA and SHIFT */
99 ir_node *irn; /**< the generated node for this instruction if any. */
100 int costs; /**< the costs for this instruction */
104 * The environment for the strength reduction of multiplications.
106 typedef struct _mul_env {
107 struct obstack obst; /**< an obstack for local space. */
108 ir_mode *mode; /**< the mode of the multiplication constant */
109 unsigned bits; /**< number of bits in the mode */
110 unsigned max_S; /**< the maximum LEA shift value. */
111 instruction *root; /**< the root of the instruction tree */
112 ir_node *op; /**< the operand that is multiplied */
113 ir_node *blk; /**< the block where the new graph is built */
114 dbg_info *dbg; /**< the debug info for the new graph. */
115 ir_mode *shf_mode; /**< the (unsigned) mode for the shift constants */
116 int fail; /**< set to 1 if the instruction sequence fails the constraints */
117 int n_shift; /**< maximum number of allowed shift instructions */
119 evaluate_costs_func evaluate; /**< the evaluate callback */
123 * Some kind of default evaluator. Return the cost of
126 static int default_evaluate(insn_kind kind, tarval *tv) {
135 * emit a LEA (or an Add) instruction
137 static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) {
138 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
139 res->kind = shift > 0 ? LEA : ADD;
142 res->shift_count = shift;
149 * emit a SHIFT (or an Add or a Zero) instruction
151 static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) {
152 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
153 if (shift == env->bits) {
154 /* a 2^bits with bits resolution is a zero */
158 res->shift_count = 0;
159 } else if (shift != 1) {
163 res->shift_count = shift;
168 res->shift_count = 0;
176 * emit a SUB instruction
178 static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) {
179 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
183 res->shift_count = 0;
190 * emit the ROOT instruction
192 static instruction *emit_ROOT(mul_env *env, ir_node *root_op) {
193 instruction *res = obstack_alloc(&env->obst, sizeof(*res));
197 res->shift_count = 0;
205 * Returns the condensed representation of the tarval tv
207 static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) {
208 ir_mode *mode = get_tarval_mode(tv);
209 int bits = get_mode_size_bits(mode);
210 char *bitstr = get_tarval_bitpattern(tv);
212 unsigned char *R = obstack_alloc(&env->obst, bits);
215 for (i = 0; bitstr[i] != '\0'; ++i) {
216 if (bitstr[i] == '1') {
229 * Calculate the gain when using the generalized complementary technique
231 static int calculate_gain(unsigned char *R, int r) {
236 /* the gain for r == 1 */
238 for (i = 2; i < r; ++i) {
239 /* calculate the gain for r from the gain for r-1 */
240 gain += 2 - R[i - 1];
242 if (gain > max_gain) {
251 * Calculates the condensed complement of a given (R,r) tuple
253 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) {
298 tv = get_mode_one(env->mode);
300 for (i = 0; i < r; ++i) {
303 tarval *t = new_tarval_from_long(j, mode_Iu);
304 tv = tarval_shl(tv, t);
306 res = res ? tarval_add(res, tv) : tv;
312 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
315 * handle simple cases with up-to 2 bits set
317 static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) {
318 instruction *ins, *ins2;
322 return emit_SHIFT(env, env->root, R[0]);
327 if (R[1] <= env->max_S) {
328 ins = emit_LEA(env, ins, ins, R[1]);
330 ins = emit_SHIFT(env, ins, R[0]);
335 ins = emit_SHIFT(env, ins, R[0]);
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 + 5) / 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 set_arch_dep_running(1);
575 if (is_Mul(irn) && mode_is_int(mode)) {
576 ir_node *left = get_binop_left(irn);
577 ir_node *right = get_binop_right(irn);
579 ir_node *operand = NULL;
581 /* Look, if one operand is a constant. */
582 if (is_Const(left)) {
583 tv = get_Const_tarval(left);
585 } else if (is_Const(right)) {
586 tv = get_Const_tarval(right);
591 res = do_decomposition(irn, operand, tv);
594 hook_arch_dep_replace_mul_with_shifts(irn);
600 //set_arch_dep_running(0);
606 * calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
608 static int tv_ld2(tarval *tv, int bits) {
611 for (num = i = 0; i < bits; ++i) {
612 unsigned char v = get_tarval_sub_bits(tv, i);
617 for (j = 0; j < 8; ++j)
630 /* for shorter lines */
631 #define ABS(a) tarval_abs(a)
632 #define NEG(a) tarval_neg(a)
633 #define NOT(a) tarval_not(a)
634 #define SHL(a, b) tarval_shl(a, b)
635 #define SHR(a, b) tarval_shr(a, b)
636 #define ADD(a, b) tarval_add(a, b)
637 #define SUB(a, b) tarval_sub(a, b)
638 #define MUL(a, b) tarval_mul(a, b)
639 #define DIV(a, b) tarval_div(a, b)
640 #define MOD(a, b) tarval_mod(a, b)
641 #define CMP(a, b) tarval_cmp(a, b)
642 #define CNV(a, m) tarval_convert_to(a, m)
643 #define ONE(m) get_mode_one(m)
644 #define ZERO(m) get_mode_null(m)
646 /** The result of a the magic() function. */
648 tarval *M; /**< magic number */
649 int s; /**< shift amount */
650 int need_add; /**< an additional add is needed */
651 int need_sub; /**< an additional sub is needed */
655 * Signed division by constant d: calculate the Magic multiplier M and the shift amount s
657 * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
659 static struct ms magic(tarval *d) {
660 ir_mode *mode = get_tarval_mode(d);
661 ir_mode *u_mode = find_unsigned_mode(mode);
662 int bits = get_mode_size_bits(u_mode);
664 tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
667 tarval *bits_minus_1, *two_bits_1;
671 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
673 /* we need overflow mode to work correctly */
674 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
677 bits_minus_1 = new_tarval_from_long(bits - 1, u_mode);
678 two_bits_1 = SHL(get_mode_one(u_mode), bits_minus_1);
680 ad = CNV(ABS(d), u_mode);
681 t = ADD(two_bits_1, SHR(CNV(d, u_mode), bits_minus_1));
682 anc = SUB(SUB(t, ONE(u_mode)), MOD(t, ad)); /* Absolute value of nc */
683 p = bits - 1; /* Init: p */
684 q1 = DIV(two_bits_1, anc); /* Init: q1 = 2^p/|nc| */
685 r1 = SUB(two_bits_1, MUL(q1, anc)); /* Init: r1 = rem(2^p, |nc|) */
686 q2 = DIV(two_bits_1, ad); /* Init: q2 = 2^p/|d| */
687 r2 = SUB(two_bits_1, MUL(q2, ad)); /* Init: r2 = rem(2^p, |d|) */
691 q1 = ADD(q1, q1); /* Update q1 = 2^p/|nc| */
692 r1 = ADD(r1, r1); /* Update r1 = rem(2^p, |nc|) */
694 if (CMP(r1, anc) & pn_Cmp_Ge) {
695 q1 = ADD(q1, ONE(u_mode));
699 q2 = ADD(q2, q2); /* Update q2 = 2^p/|d| */
700 r2 = ADD(r2, r2); /* Update r2 = rem(2^p, |d|) */
702 if (CMP(r2, ad) & pn_Cmp_Ge) {
703 q2 = ADD(q2, ONE(u_mode));
708 } while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq));
710 d_cmp = CMP(d, ZERO(mode));
712 if (d_cmp & pn_Cmp_Ge)
713 mag.M = ADD(CNV(q2, mode), ONE(mode));
715 mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode)));
717 M_cmp = CMP(mag.M, ZERO(mode));
721 /* need an add if d > 0 && M < 0 */
722 mag.need_add = d_cmp & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt;
724 /* need a sub if d < 0 && M > 0 */
725 mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt;
727 tarval_set_integer_overflow_mode(rem);
732 /** The result of the magicu() function. */
734 tarval *M; /**< magic add constant */
735 int s; /**< shift amount */
736 int need_add; /**< add indicator */
740 * Unsigned division by constant d: calculate the Magic multiplier M and the shift amount s
742 * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned)
744 static struct mu magicu(tarval *d) {
745 ir_mode *mode = get_tarval_mode(d);
746 int bits = get_mode_size_bits(mode);
748 tarval *nc, *delta, *q1, *r1, *q2, *r2;
749 tarval *bits_minus_1, *two_bits_1, *seven_ff;
753 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
755 /* we need overflow mode to work correctly */
756 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
758 bits_minus_1 = new_tarval_from_long(bits - 1, mode);
759 two_bits_1 = SHL(get_mode_one(mode), bits_minus_1);
760 seven_ff = SUB(two_bits_1, ONE(mode));
762 magu.need_add = 0; /* initialize the add indicator */
763 nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d));
764 p = bits - 1; /* Init: p */
765 q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */
766 r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */
767 q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */
768 r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */
772 if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
773 q1 = ADD(ADD(q1, q1), ONE(mode));
774 r1 = SUB(ADD(r1, r1), nc);
781 if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
782 if (CMP(q2, seven_ff) & pn_Cmp_Ge)
785 q2 = ADD(ADD(q2, q2), ONE(mode));
786 r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
789 if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
793 r2 = ADD(ADD(r2, r2), ONE(mode));
795 delta = SUB(SUB(d, ONE(mode)), r2);
796 } while (p < 2*bits &&
797 (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
799 magu.M = ADD(q2, ONE(mode)); /* Magic number */
800 magu.s = p - bits; /* and shift amount */
802 tarval_set_integer_overflow_mode(rem);
808 * Build the Mulh replacement code for n / tv.
810 * Note that 'div' might be a mod or DivMod operation as well
812 static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
813 dbg_info *dbg = get_irn_dbg_info(div);
814 ir_node *n = get_binop_left(div);
815 ir_node *block = get_irn_n(div, -1);
816 ir_mode *mode = get_irn_mode(n);
817 int bits = get_mode_size_bits(mode);
820 /* Beware: do not transform bad code */
821 if (is_Bad(n) || is_Bad(block))
824 if (mode_is_signed(mode)) {
825 struct ms mag = magic(tv);
827 /* generate the Mulh instruction */
828 c = new_r_Const(current_ir_graph, block, mode, mag.M);
829 q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
831 /* do we need an Add or Sub */
833 q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
834 else if (mag.need_sub)
835 q = new_rd_Sub(dbg, current_ir_graph, block, q, n, mode);
837 /* Do we need the shift */
839 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
840 q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
844 c = new_r_Const_long(current_ir_graph, block, mode_Iu, bits-1);
845 t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
847 q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode);
849 struct mu mag = magicu(tv);
852 /* generate the Mulh instruction */
853 c = new_r_Const(current_ir_graph, block, mode, mag.M);
854 q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
858 /* use the GM scheme */
859 t = new_rd_Sub(dbg, current_ir_graph, block, n, q, mode);
861 c = new_r_Const(current_ir_graph, block, mode_Iu, get_mode_one(mode_Iu));
862 t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
864 t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode);
866 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s-1);
867 q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
869 /* use the default scheme */
870 q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
872 } else if (mag.s > 0) { /* default scheme, shift needed */
873 c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
874 q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
880 /* Replace Divs with Shifts and Add/Subs and Mulh. */
881 ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
884 /* If the architecture dependent optimizations were not initialized
885 or this optimization was not enabled. */
886 if (params == NULL || (opts & arch_dep_div_by_const) == 0)
890 ir_node *c = get_Div_right(irn);
891 ir_node *block, *left;
901 tv = get_Const_tarval(c);
903 /* check for division by zero */
904 if (tarval_is_null(tv))
907 left = get_Div_left(irn);
908 mode = get_irn_mode(left);
909 block = get_irn_n(irn, -1);
910 dbg = get_irn_dbg_info(irn);
912 bits = get_mode_size_bits(mode);
916 if (mode_is_signed(mode)) {
917 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
918 ntv = tarval_neg(tv);
928 if (k >= 0) { /* division by 2^k or -2^k */
929 if (mode_is_signed(mode)) {
931 ir_node *curr = left;
933 /* create the correction code for signed values only if there might be a remainder */
934 if (! is_Div_remainderless(irn)) {
936 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
937 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
940 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
941 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
943 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
948 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
949 res = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode);
951 if (n_flag) { /* negate the result */
954 k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode));
955 res = new_rd_Sub(dbg, current_ir_graph, block, k_node, res, mode);
957 } else { /* unsigned case */
960 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
961 res = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
965 if (allow_Mulh(mode))
966 res = replace_div_by_mulh(irn, tv);
971 hook_arch_dep_replace_division_by_const(irn);
976 /* Replace Mods with Shifts and Add/Subs and Mulh. */
977 ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
980 /* If the architecture dependent optimizations were not initialized
981 or this optimization was not enabled. */
982 if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
986 ir_node *c = get_Mod_right(irn);
987 ir_node *block, *left;
997 tv = get_Const_tarval(c);
999 /* check for division by zero */
1000 if (tarval_is_null(tv))
1003 left = get_Mod_left(irn);
1004 mode = get_irn_mode(left);
1005 block = get_irn_n(irn, -1);
1006 dbg = get_irn_dbg_info(irn);
1007 bits = get_mode_size_bits(mode);
1011 if (mode_is_signed(mode)) {
1012 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1013 ntv = tarval_neg(tv);
1022 /* division by 2^k or -2^k:
1023 * we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k
1025 if (mode_is_signed(mode)) {
1027 ir_node *curr = left;
1030 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
1031 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
1034 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
1035 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
1037 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
1039 k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
1040 curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
1042 res = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
1043 } else { /* unsigned case */
1046 k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1);
1047 res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
1050 /* other constant */
1051 if (allow_Mulh(mode)) {
1052 res = replace_div_by_mulh(irn, tv);
1054 res = new_rd_Mul(dbg, current_ir_graph, block, res, c, mode);
1056 /* res = arch_dep_mul_to_shift(res); */
1058 res = new_rd_Sub(dbg, current_ir_graph, block, left, res, mode);
1064 hook_arch_dep_replace_division_by_const(irn);
1069 /* Replace DivMods with Shifts and Add/Subs and Mulh. */
1070 void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn) {
1073 /* If the architecture dependent optimizations were not initialized
1074 or this optimization was not enabled. */
1075 if (params == NULL ||
1076 ((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
1079 if (is_DivMod(irn)) {
1080 ir_node *c = get_DivMod_right(irn);
1081 ir_node *block, *left;
1091 tv = get_Const_tarval(c);
1093 /* check for division by zero */
1094 if (tarval_is_null(tv))
1097 left = get_DivMod_left(irn);
1098 mode = get_irn_mode(left);
1099 block = get_irn_n(irn, -1);
1100 dbg = get_irn_dbg_info(irn);
1102 bits = get_mode_size_bits(mode);
1106 if (mode_is_signed(mode)) {
1107 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1108 ntv = tarval_neg(tv);
1118 if (k >= 0) { /* division by 2^k or -2^k */
1119 if (mode_is_signed(mode)) {
1120 ir_node *k_node, *c_k;
1121 ir_node *curr = left;
1124 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
1125 curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
1128 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
1129 curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
1131 curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
1133 c_k = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
1135 *div = new_rd_Shrs(dbg, current_ir_graph, block, curr, c_k, mode);
1137 if (n_flag) { /* negate the div result */
1140 k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode));
1141 *div = new_rd_Sub(dbg, current_ir_graph, block, k_node, *div, mode);
1144 k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
1145 curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
1147 *mod = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
1148 } else { /* unsigned case */
1151 k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
1152 *div = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
1154 k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1);
1155 *mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
1158 /* other constant */
1159 if (allow_Mulh(mode)) {
1162 *div = replace_div_by_mulh(irn, tv);
1164 t = new_rd_Mul(dbg, current_ir_graph, block, *div, c, mode);
1166 /* t = arch_dep_mul_to_shift(t); */
1168 *mod = new_rd_Sub(dbg, current_ir_graph, block, left, t, mode);
1174 hook_arch_dep_replace_division_by_const(irn);
1178 static const ir_settings_arch_dep_t default_params = {
1179 1, /* also use subs */
1180 4, /* maximum shifts */
1181 31, /* maximum shift amount */
1182 default_evaluate, /* default evaluator */
1184 0, /* allow Mulhs */
1185 0, /* allow Mulus */
1186 32 /* Mulh allowed up to 32 bit */
1189 /* A default parameter factory for testing purposes. */
1190 const ir_settings_arch_dep_t *arch_dep_default_factory(void) {
1191 return &default_params;