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"
29 * Implements Division and Modulo by Consts from "Hackers Delight",
37 #include "irgraph_t.h"
44 #include "dbginfo_t.h"
45 #include "iropt_dbg.h"
54 /** The bit mask, which optimizations to apply. */
55 static arch_dep_opts_t opts;
57 void arch_dep_set_opts(arch_dep_opts_t the_opts)
62 /** check, whether a mode allows a Mulh instruction. */
63 static int allow_Mulh(const ir_settings_arch_dep_t *params, ir_mode *mode)
65 if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
67 return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
73 typedef struct instruction instruction;
75 insn_kind kind; /**< the instruction kind */
76 instruction *in[2]; /**< the ins */
77 unsigned shift_count; /**< shift count for LEA and SHIFT */
78 ir_node *irn; /**< the generated node for this instruction if any. */
79 int costs; /**< the costs for this instruction */
83 * The environment for the strength reduction of multiplications.
85 typedef struct mul_env {
86 struct obstack obst; /**< an obstack for local space. */
87 const ir_settings_arch_dep_t *params;
88 ir_mode *mode; /**< the mode of the multiplication constant */
89 unsigned bits; /**< number of bits in the mode */
90 unsigned max_S; /**< the maximum LEA shift value. */
91 instruction *root; /**< the root of the instruction tree */
92 ir_node *op; /**< the operand that is multiplied */
93 ir_node *blk; /**< the block where the new graph is built */
95 dbg_info *dbg; /**< the debug info for the new graph. */
96 ir_mode *shf_mode; /**< the (unsigned) mode for the shift constants */
97 int fail; /**< set to 1 if the instruction sequence fails the constraints */
98 int n_shift; /**< maximum number of allowed shift instructions */
100 evaluate_costs_func evaluate; /**< the evaluate callback */
104 * Some kind of default evaluator. Return the cost of
107 static int default_evaluate(insn_kind kind, tarval *tv)
117 * emit a LEA (or an Add) instruction
119 static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift)
121 instruction *res = OALLOC(&env->obst, instruction);
122 res->kind = shift > 0 ? LEA : ADD;
125 res->shift_count = shift;
132 * emit a SHIFT (or an Add or a Zero) instruction
134 static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift)
136 instruction *res = OALLOC(&env->obst, instruction);
137 if (shift == env->bits) {
138 /* a 2^bits with bits resolution is a zero */
142 res->shift_count = 0;
143 } else if (shift != 1) {
147 res->shift_count = shift;
152 res->shift_count = 0;
160 * emit a SUB instruction
162 static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b)
164 instruction *res = OALLOC(&env->obst, instruction);
168 res->shift_count = 0;
175 * emit the ROOT instruction
177 static instruction *emit_ROOT(mul_env *env, ir_node *root_op)
179 instruction *res = OALLOC(&env->obst, instruction);
183 res->shift_count = 0;
191 * Returns the condensed representation of the tarval tv
193 static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr)
195 ir_mode *mode = get_tarval_mode(tv);
196 int bits = get_mode_size_bits(mode);
197 char *bitstr = get_tarval_bitpattern(tv);
199 unsigned char *R = obstack_alloc(&env->obst, bits);
202 for (i = 0; bitstr[i] != '\0'; ++i) {
203 if (bitstr[i] == '1') {
216 * Calculate the gain when using the generalized complementary technique
218 static int calculate_gain(unsigned char *R, int r)
224 /* the gain for r == 1 */
226 for (i = 2; i < r; ++i) {
227 /* calculate the gain for r from the gain for r-1 */
228 gain += 2 - R[i - 1];
230 if (gain > max_gain) {
239 * Calculates the condensed complement of a given (R,r) tuple
241 static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs)
243 unsigned char *value = obstack_alloc(&env->obst, env->bits);
247 memset(value, 0, env->bits);
250 for (i = 0; i < gain; ++i) {
255 /* negate and propagate 1 */
257 for (i = 0; i <= j; ++i) {
258 unsigned char v = !value[i];
264 /* condense it again */
267 for (i = 0; i <= j; ++i) {
280 * creates a tarval from a condensed representation.
282 static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r)
288 tv = get_mode_one(env->mode);
290 for (i = 0; i < r; ++i) {
293 tarval *t = new_tarval_from_long(j, mode_Iu);
294 tv = tarval_shl(tv, t);
296 res = res ? tarval_add(res, tv) : tv;
302 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
305 * handle simple cases with up-to 2 bits set
307 static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N)
309 instruction *ins, *ins2;
313 return emit_SHIFT(env, env->root, R[0]);
318 if (R[1] <= env->max_S) {
319 ins = emit_LEA(env, ins, ins, R[1]);
321 ins = emit_SHIFT(env, ins, R[0]);
326 ins = emit_SHIFT(env, ins, R[0]);
329 ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
330 return emit_LEA(env, ins, ins2, 0);
335 * Main decompose driver.
337 static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N)
343 return decompose_simple_cases(env, R, r, N);
345 if (env->params->also_use_subs) {
346 gain = calculate_gain(R, r);
348 instruction *instr1, *instr2;
349 unsigned char *R1, *R2;
352 R1 = complement_condensed(env, R, r, gain, &r1);
354 R2 = obstack_alloc(&env->obst, r2);
357 for (i = 0; i < gain; ++i) {
364 /* Two identical bits: normalize */
369 for (i = gain + 1; i < r; ++i) {
373 instr1 = decompose_mul(env, R1, r1, NULL);
374 instr2 = decompose_mul(env, R2, r2, NULL);
375 return emit_SUB(env, instr2, instr1);
380 N = condensed_to_value(env, R, r);
382 for (i = env->max_S; i > 0; --i) {
383 tarval *div_res, *mod_res;
384 tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
386 div_res = tarval_divmod(N, tv, &mod_res);
387 if (mod_res == get_mode_null(env->mode)) {
391 Rs = value_to_condensed(env, div_res, &rs);
393 instruction *N1 = decompose_mul(env, Rs, rs, div_res);
394 return emit_LEA(env, N1, N1, i);
398 return basic_decompose_mul(env, R, r, N);
401 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
404 * basic decomposition routine
406 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N)
411 if (R[0] == 0) { /* Case 1 */
412 t = R[1] > IMAX(env->max_S, R[1]);
414 Ns = decompose_mul(env, &R[1], r - 1, N);
415 return emit_LEA(env, env->root, Ns, t);
416 } else if (R[0] <= env->max_S) { /* Case 2 */
419 Ns = decompose_mul(env, &R[1], r - 1, N);
420 return emit_LEA(env, Ns, env->root, t);
424 Ns = decompose_mul(env, R, r, N);
425 return emit_SHIFT(env, Ns, t);
430 * recursive build the graph form the instructions.
432 * @param env the environment
433 * @param inst the instruction
435 static ir_node *build_graph(mul_env *env, instruction *inst)
438 ir_graph *irg = env->irg;
443 switch (inst->kind) {
445 l = build_graph(env, inst->in[0]);
446 r = build_graph(env, inst->in[1]);
447 c = new_r_Const_long(irg, env->shf_mode, inst->shift_count);
448 r = new_rd_Shl(env->dbg, env->blk, r, c, env->mode);
449 return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
451 l = build_graph(env, inst->in[0]);
452 c = new_r_Const_long(irg, env->shf_mode, inst->shift_count);
453 return inst->irn = new_rd_Shl(env->dbg, env->blk, l, c, env->mode);
455 l = build_graph(env, inst->in[0]);
456 r = build_graph(env, inst->in[1]);
457 return inst->irn = new_rd_Sub(env->dbg, env->blk, l, r, env->mode);
459 l = build_graph(env, inst->in[0]);
460 r = build_graph(env, inst->in[1]);
461 return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
463 return inst->irn = new_r_Const(irg, get_mode_null(env->mode));
465 panic("Unsupported instruction kind");
470 * Calculate the costs for the given instruction sequence.
471 * Note that additional costs due to higher register pressure are NOT evaluated yet
473 static int evaluate_insn(mul_env *env, instruction *inst)
477 if (inst->costs >= 0) {
478 /* was already evaluated */
482 switch (inst->kind) {
486 costs = evaluate_insn(env, inst->in[0]);
487 costs += evaluate_insn(env, inst->in[1]);
488 costs += env->evaluate(inst->kind, NULL);
492 if (inst->shift_count > env->params->highest_shift_amount)
494 if (env->n_shift <= 0)
498 costs = evaluate_insn(env, inst->in[0]);
499 costs += env->evaluate(inst->kind, NULL);
503 inst->costs = costs = env->evaluate(inst->kind, NULL);
509 panic("Unsupported instruction kind");
513 * Evaluate the replacement instructions and build a new graph
514 * if faster than the Mul.
515 * Returns the root of the new graph then or irn otherwise.
517 * @param irn the Mul operation
518 * @param operand the multiplication operand
519 * @param tv the multiplication constant
521 * @return the new graph
523 static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv)
532 obstack_init(&env.obst);
533 env.params = be_get_backend_param()->dep_param;
534 env.mode = get_tarval_mode(tv);
535 env.bits = (unsigned)get_mode_size_bits(env.mode);
537 env.root = emit_ROOT(&env, operand);
539 env.n_shift = env.params->maximum_shifts;
540 env.evaluate = env.params->evaluate != NULL ? env.params->evaluate : default_evaluate;
541 env.irg = get_irn_irg(irn);
543 R = value_to_condensed(&env, tv, &r);
544 inst = decompose_mul(&env, R, r, tv);
546 /* the paper suggests 70% here */
547 mul_costs = (env.evaluate(MUL, tv) * 7 + 5) / 10;
548 if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) {
550 env.blk = get_nodes_block(irn);
551 env.dbg = get_irn_dbg_info(irn);
552 env.shf_mode = find_unsigned_mode(env.mode);
553 if (env.shf_mode == NULL)
554 env.shf_mode = mode_Iu;
556 res = build_graph(&env, inst);
558 obstack_free(&env.obst, NULL);
562 /* Replace Muls with Shifts and Add/Subs. */
563 ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn)
567 ir_mode *mode = get_irn_mode(irn);
572 const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
575 /* If the architecture dependent optimizations were not initialized
576 or this optimization was not enabled. */
577 if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
580 if (!is_Mul(irn) || !mode_is_int(mode))
583 /* we should never do the reverse transformations again
585 irg = get_irn_irg(irn);
586 set_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP);
588 left = get_binop_left(irn);
589 right = get_binop_right(irn);
593 /* Look, if one operand is a constant. */
594 if (is_Const(left)) {
595 tv = get_Const_tarval(left);
597 } else if (is_Const(right)) {
598 tv = get_Const_tarval(right);
603 res = do_decomposition(irn, operand, tv);
606 hook_arch_dep_replace_mul_with_shifts(irn);
615 * calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
617 static int tv_ld2(tarval *tv, int bits)
621 for (num = i = 0; i < bits; ++i) {
622 unsigned char v = get_tarval_sub_bits(tv, i);
627 for (j = 0; j < 8; ++j)
640 /* for shorter lines */
641 #define ABS(a) tarval_abs(a)
642 #define NEG(a) tarval_neg(a)
643 #define NOT(a) tarval_not(a)
644 #define SHL(a, b) tarval_shl(a, b)
645 #define SHR(a, b) tarval_shr(a, b)
646 #define ADD(a, b) tarval_add(a, b)
647 #define SUB(a, b) tarval_sub(a, b, NULL)
648 #define MUL(a, b) tarval_mul(a, b)
649 #define DIV(a, b) tarval_div(a, b)
650 #define MOD(a, b) tarval_mod(a, b)
651 #define CMP(a, b) tarval_cmp(a, b)
652 #define CNV(a, m) tarval_convert_to(a, m)
653 #define ONE(m) get_mode_one(m)
654 #define ZERO(m) get_mode_null(m)
656 /** The result of a the magic() function. */
658 tarval *M; /**< magic number */
659 int s; /**< shift amount */
660 int need_add; /**< an additional add is needed */
661 int need_sub; /**< an additional sub is needed */
665 * Signed division by constant d: calculate the Magic multiplier M and the shift amount s
667 * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
669 static struct ms magic(tarval *d)
671 ir_mode *mode = get_tarval_mode(d);
672 ir_mode *u_mode = find_unsigned_mode(mode);
673 int bits = get_mode_size_bits(u_mode);
675 tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
678 tarval *bits_minus_1, *two_bits_1;
682 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
684 /* we need overflow mode to work correctly */
685 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
688 bits_minus_1 = new_tarval_from_long(bits - 1, u_mode);
689 two_bits_1 = SHL(get_mode_one(u_mode), bits_minus_1);
691 ad = CNV(ABS(d), u_mode);
692 t = ADD(two_bits_1, SHR(CNV(d, u_mode), bits_minus_1));
693 anc = SUB(SUB(t, ONE(u_mode)), MOD(t, ad)); /* Absolute value of nc */
694 p = bits - 1; /* Init: p */
695 q1 = DIV(two_bits_1, anc); /* Init: q1 = 2^p/|nc| */
696 r1 = SUB(two_bits_1, MUL(q1, anc)); /* Init: r1 = rem(2^p, |nc|) */
697 q2 = DIV(two_bits_1, ad); /* Init: q2 = 2^p/|d| */
698 r2 = SUB(two_bits_1, MUL(q2, ad)); /* Init: r2 = rem(2^p, |d|) */
702 q1 = ADD(q1, q1); /* Update q1 = 2^p/|nc| */
703 r1 = ADD(r1, r1); /* Update r1 = rem(2^p, |nc|) */
705 if (CMP(r1, anc) & pn_Cmp_Ge) {
706 q1 = ADD(q1, ONE(u_mode));
710 q2 = ADD(q2, q2); /* Update q2 = 2^p/|d| */
711 r2 = ADD(r2, r2); /* Update r2 = rem(2^p, |d|) */
713 if (CMP(r2, ad) & pn_Cmp_Ge) {
714 q2 = ADD(q2, ONE(u_mode));
719 } while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq));
721 d_cmp = CMP(d, ZERO(mode));
723 if (d_cmp & pn_Cmp_Ge)
724 mag.M = ADD(CNV(q2, mode), ONE(mode));
726 mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode)));
728 M_cmp = CMP(mag.M, ZERO(mode));
732 /* need an add if d > 0 && M < 0 */
733 mag.need_add = d_cmp & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt;
735 /* need a sub if d < 0 && M > 0 */
736 mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt;
738 tarval_set_integer_overflow_mode(rem);
743 /** The result of the magicu() function. */
745 tarval *M; /**< magic add constant */
746 int s; /**< shift amount */
747 int need_add; /**< add indicator */
751 * Unsigned division by constant d: calculate the Magic multiplier M and the shift amount s
753 * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned)
755 static struct mu magicu(tarval *d)
757 ir_mode *mode = get_tarval_mode(d);
758 int bits = get_mode_size_bits(mode);
760 tarval *nc, *delta, *q1, *r1, *q2, *r2;
761 tarval *bits_minus_1, *two_bits_1, *seven_ff;
765 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
767 /* we need overflow mode to work correctly */
768 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
770 bits_minus_1 = new_tarval_from_long(bits - 1, mode);
771 two_bits_1 = SHL(get_mode_one(mode), bits_minus_1);
772 seven_ff = SUB(two_bits_1, ONE(mode));
774 magu.need_add = 0; /* initialize the add indicator */
775 nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d));
776 p = bits - 1; /* Init: p */
777 q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */
778 r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */
779 q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */
780 r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */
784 if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
785 q1 = ADD(ADD(q1, q1), ONE(mode));
786 r1 = SUB(ADD(r1, r1), nc);
793 if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
794 if (CMP(q2, seven_ff) & pn_Cmp_Ge)
797 q2 = ADD(ADD(q2, q2), ONE(mode));
798 r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
801 if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
805 r2 = ADD(ADD(r2, r2), ONE(mode));
807 delta = SUB(SUB(d, ONE(mode)), r2);
808 } while (p < 2*bits &&
809 (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
811 magu.M = ADD(q2, ONE(mode)); /* Magic number */
812 magu.s = p - bits; /* and shift amount */
814 tarval_set_integer_overflow_mode(rem);
820 * Build the Mulh replacement code for n / tv.
822 * Note that 'div' might be a mod or DivMod operation as well
824 static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv)
826 dbg_info *dbg = get_irn_dbg_info(div);
827 ir_node *n = get_binop_left(div);
828 ir_node *block = get_irn_n(div, -1);
829 ir_mode *mode = get_irn_mode(n);
830 int bits = get_mode_size_bits(mode);
833 /* Beware: do not transform bad code */
834 if (is_Bad(n) || is_Bad(block))
837 if (mode_is_signed(mode)) {
838 ir_graph *irg = get_irn_irg(div);
839 struct ms mag = magic(tv);
841 /* generate the Mulh instruction */
842 c = new_r_Const(irg, mag.M);
843 q = new_rd_Mulh(dbg, block, n, c, mode);
845 /* do we need an Add or Sub */
847 q = new_rd_Add(dbg, block, q, n, mode);
848 else if (mag.need_sub)
849 q = new_rd_Sub(dbg, block, q, n, mode);
851 /* Do we need the shift */
853 c = new_r_Const_long(irg, mode_Iu, mag.s);
854 q = new_rd_Shrs(dbg, block, q, c, mode);
858 c = new_r_Const_long(irg, mode_Iu, bits - 1);
859 t = new_rd_Shr(dbg, block, q, c, mode);
861 q = new_rd_Add(dbg, block, q, t, mode);
863 struct mu mag = magicu(tv);
865 ir_graph *irg = get_irn_irg(div);
867 /* generate the Mulh instruction */
868 c = new_r_Const(irg, mag.M);
869 q = new_rd_Mulh(dbg, block, n, c, mode);
873 /* use the GM scheme */
874 t = new_rd_Sub(dbg, block, n, q, mode);
876 c = new_r_Const(irg, get_mode_one(mode_Iu));
877 t = new_rd_Shr(dbg, block, t, c, mode);
879 t = new_rd_Add(dbg, block, t, q, mode);
881 c = new_r_Const_long(irg, mode_Iu, mag.s - 1);
882 q = new_rd_Shr(dbg, block, t, c, mode);
884 /* use the default scheme */
885 q = new_rd_Add(dbg, block, q, n, mode);
887 } else if (mag.s > 0) { /* default scheme, shift needed */
888 c = new_r_Const_long(irg, mode_Iu, mag.s);
889 q = new_rd_Shr(dbg, block, q, c, mode);
895 /* Replace Divs with Shifts and Add/Subs and Mulh. */
896 ir_node *arch_dep_replace_div_by_const(ir_node *irn)
898 const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
901 /* If the architecture dependent optimizations were not initialized
902 or this optimization was not enabled. */
903 if (params == NULL || (opts & arch_dep_div_by_const) == 0)
907 ir_node *c = get_Div_right(irn);
908 ir_node *block, *left;
919 tv = get_Const_tarval(c);
921 /* check for division by zero */
922 if (tarval_is_null(tv))
925 left = get_Div_left(irn);
926 mode = get_irn_mode(left);
927 block = get_irn_n(irn, -1);
928 dbg = get_irn_dbg_info(irn);
930 bits = get_mode_size_bits(mode);
934 if (mode_is_signed(mode)) {
935 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
936 ntv = tarval_neg(tv);
946 if (k >= 0) { /* division by 2^k or -2^k */
947 ir_graph *irg = get_irn_irg(irn);
948 if (mode_is_signed(mode)) {
950 ir_node *curr = left;
952 /* create the correction code for signed values only if there might be a remainder */
953 if (! get_Div_no_remainder(irn)) {
955 k_node = new_r_Const_long(irg, mode_Iu, k - 1);
956 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
959 k_node = new_r_Const_long(irg, mode_Iu, bits - k);
960 curr = new_rd_Shr(dbg, block, curr, k_node, mode);
962 curr = new_rd_Add(dbg, block, left, curr, mode);
967 k_node = new_r_Const_long(irg, mode_Iu, k);
968 res = new_rd_Shrs(dbg, block, curr, k_node, mode);
970 if (n_flag) { /* negate the result */
973 k_node = new_r_Const(irg, get_mode_null(mode));
974 res = new_rd_Sub(dbg, block, k_node, res, mode);
976 } else { /* unsigned case */
979 k_node = new_r_Const_long(irg, mode_Iu, k);
980 res = new_rd_Shr(dbg, block, left, k_node, mode);
984 if (allow_Mulh(params, mode))
985 res = replace_div_by_mulh(irn, tv);
990 hook_arch_dep_replace_division_by_const(irn);
995 /* Replace Mods with Shifts and Add/Subs and Mulh. */
996 ir_node *arch_dep_replace_mod_by_const(ir_node *irn)
998 const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
1001 /* If the architecture dependent optimizations were not initialized
1002 or this optimization was not enabled. */
1003 if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
1007 ir_node *c = get_Mod_right(irn);
1008 ir_node *block, *left;
1018 tv = get_Const_tarval(c);
1020 /* check for division by zero */
1021 if (tarval_is_null(tv))
1024 left = get_Mod_left(irn);
1025 mode = get_irn_mode(left);
1026 block = get_irn_n(irn, -1);
1027 dbg = get_irn_dbg_info(irn);
1028 bits = get_mode_size_bits(mode);
1032 if (mode_is_signed(mode)) {
1033 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1034 ntv = tarval_neg(tv);
1043 ir_graph *irg = get_irn_irg(irn);
1044 /* division by 2^k or -2^k:
1045 * we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k
1047 if (mode_is_signed(mode)) {
1049 ir_node *curr = left;
1052 k_node = new_r_Const_long(irg, mode_Iu, k - 1);
1053 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
1056 k_node = new_r_Const_long(irg, mode_Iu, bits - k);
1057 curr = new_rd_Shr(dbg, block, curr, k_node, mode);
1059 curr = new_rd_Add(dbg, block, left, curr, mode);
1061 k_node = new_r_Const_long(irg, mode, (-1) << k);
1062 curr = new_rd_And(dbg, block, curr, k_node, mode);
1064 res = new_rd_Sub(dbg, block, left, curr, mode);
1065 } else { /* unsigned case */
1068 k_node = new_r_Const_long(irg, mode, (1 << k) - 1);
1069 res = new_rd_And(dbg, block, left, k_node, mode);
1072 /* other constant */
1073 if (allow_Mulh(params, mode)) {
1074 res = replace_div_by_mulh(irn, tv);
1076 res = new_rd_Mul(dbg, block, res, c, mode);
1078 /* res = arch_dep_mul_to_shift(res); */
1080 res = new_rd_Sub(dbg, block, left, res, mode);
1086 hook_arch_dep_replace_division_by_const(irn);
1091 /* Replace DivMods with Shifts and Add/Subs and Mulh. */
1092 void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn)
1094 const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
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 ir_graph *irg = get_irn_irg(irn);
1145 if (mode_is_signed(mode)) {
1146 ir_node *k_node, *c_k;
1147 ir_node *curr = left;
1150 k_node = new_r_Const_long(irg, mode_Iu, k - 1);
1151 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
1154 k_node = new_r_Const_long(irg, mode_Iu, bits - k);
1155 curr = new_rd_Shr(dbg, block, curr, k_node, mode);
1157 curr = new_rd_Add(dbg, block, left, curr, mode);
1159 c_k = new_r_Const_long(irg, mode_Iu, k);
1161 *div = new_rd_Shrs(dbg, block, curr, c_k, mode);
1163 if (n_flag) { /* negate the div result */
1164 ir_node *k_node = new_r_Const(irg, get_mode_null(mode));
1165 *div = new_rd_Sub(dbg, block, k_node, *div, mode);
1168 k_node = new_r_Const_long(irg, mode, (-1) << k);
1169 curr = new_rd_And(dbg, block, curr, k_node, mode);
1171 *mod = new_rd_Sub(dbg, block, left, curr, mode);
1172 } else { /* unsigned case */
1173 ir_node *k_node = new_r_Const_long(irg, mode_Iu, k);
1174 *div = new_rd_Shr(dbg, block, left, k_node, mode);
1176 k_node = new_r_Const_long(irg, mode, (1 << k) - 1);
1177 *mod = new_rd_And(dbg, block, left, k_node, mode);
1180 /* other constant */
1181 if (allow_Mulh(params, mode)) {
1184 *div = replace_div_by_mulh(irn, tv);
1186 t = new_rd_Mul(dbg, block, *div, c, mode);
1188 /* t = arch_dep_mul_to_shift(t); */
1190 *mod = new_rd_Sub(dbg, block, left, t, mode);
1196 hook_arch_dep_replace_division_by_const(irn);