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, const ir_mode *mode, ir_tarval *tv)
118 * emit a LEA (or an Add) instruction
120 static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift)
122 instruction *res = OALLOC(&env->obst, instruction);
123 res->kind = shift > 0 ? LEA : ADD;
126 res->shift_count = shift;
133 * emit a SHIFT (or an Add or a Zero) instruction
135 static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift)
137 instruction *res = OALLOC(&env->obst, instruction);
138 if (shift == env->bits) {
139 /* a 2^bits with bits resolution is a zero */
143 res->shift_count = 0;
144 } else if (shift != 1) {
148 res->shift_count = shift;
153 res->shift_count = 0;
161 * emit a SUB instruction
163 static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b)
165 instruction *res = OALLOC(&env->obst, instruction);
169 res->shift_count = 0;
176 * emit the ROOT instruction
178 static instruction *emit_ROOT(mul_env *env, ir_node *root_op)
180 instruction *res = OALLOC(&env->obst, instruction);
184 res->shift_count = 0;
192 * Returns the condensed representation of the tarval tv
194 static unsigned char *value_to_condensed(mul_env *env, ir_tarval *tv, int *pr)
196 ir_mode *mode = get_tarval_mode(tv);
197 int bits = get_mode_size_bits(mode);
198 char *bitstr = get_tarval_bitpattern(tv);
200 unsigned char *R = (unsigned char*)obstack_alloc(&env->obst, bits);
203 for (i = 0; bitstr[i] != '\0'; ++i) {
204 if (bitstr[i] == '1') {
217 * Calculate the gain when using the generalized complementary technique
219 static int calculate_gain(unsigned char *R, int r)
225 /* the gain for r == 1 */
227 for (i = 2; i < r; ++i) {
228 /* calculate the gain for r from the gain for r-1 */
229 gain += 2 - R[i - 1];
231 if (gain > max_gain) {
240 * Calculates the condensed complement of a given (R,r) tuple
242 static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs)
244 unsigned char *value = (unsigned char*)obstack_alloc(&env->obst, env->bits);
248 memset(value, 0, env->bits);
251 for (i = 0; i < gain; ++i) {
256 /* negate and propagate 1 */
258 for (i = 0; i <= j; ++i) {
259 unsigned char v = !value[i];
265 /* condense it again */
268 for (i = 0; i <= j; ++i) {
281 * creates a tarval from a condensed representation.
283 static ir_tarval *condensed_to_value(mul_env *env, unsigned char *R, int r)
289 tv = get_mode_one(env->mode);
291 for (i = 0; i < r; ++i) {
294 ir_tarval *t = new_tarval_from_long(j, mode_Iu);
295 tv = tarval_shl(tv, t);
297 res = res ? tarval_add(res, tv) : tv;
303 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N);
306 * handle simple cases with up-to 2 bits set
308 static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, ir_tarval *N)
310 instruction *ins, *ins2;
314 return emit_SHIFT(env, env->root, R[0]);
319 if (R[1] <= env->max_S) {
320 ins = emit_LEA(env, ins, ins, R[1]);
322 ins = emit_SHIFT(env, ins, R[0]);
327 ins = emit_SHIFT(env, ins, R[0]);
330 ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
331 return emit_LEA(env, ins, ins2, 0);
336 * Main decompose driver.
338 static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N)
344 return decompose_simple_cases(env, R, r, N);
346 if (env->params->also_use_subs) {
347 gain = calculate_gain(R, r);
349 instruction *instr1, *instr2;
350 unsigned char *R1, *R2;
353 R1 = complement_condensed(env, R, r, gain, &r1);
355 R2 = (unsigned char*)obstack_alloc(&env->obst, r2);
358 for (i = 0; i < gain; ++i) {
365 /* Two identical bits: normalize */
370 for (i = gain + 1; i < r; ++i) {
374 instr1 = decompose_mul(env, R1, r1, NULL);
375 instr2 = decompose_mul(env, R2, r2, NULL);
376 return emit_SUB(env, instr2, instr1);
381 N = condensed_to_value(env, R, r);
383 for (i = env->max_S; i > 0; --i) {
384 ir_tarval *div_res, *mod_res;
385 ir_tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
387 div_res = tarval_divmod(N, tv, &mod_res);
388 if (mod_res == get_mode_null(env->mode)) {
392 Rs = value_to_condensed(env, div_res, &rs);
394 instruction *N1 = decompose_mul(env, Rs, rs, div_res);
395 return emit_LEA(env, N1, N1, i);
399 return basic_decompose_mul(env, R, r, N);
402 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
405 * basic decomposition routine
407 static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N)
412 if (R[0] == 0) { /* Case 1 */
413 t = R[1] > IMAX(env->max_S, R[1]);
415 Ns = decompose_mul(env, &R[1], r - 1, N);
416 return emit_LEA(env, env->root, Ns, t);
417 } else if (R[0] <= env->max_S) { /* Case 2 */
420 Ns = decompose_mul(env, &R[1], r - 1, N);
421 return emit_LEA(env, Ns, env->root, t);
425 Ns = decompose_mul(env, R, r, N);
426 return emit_SHIFT(env, Ns, t);
431 * recursive build the graph form the instructions.
433 * @param env the environment
434 * @param inst the instruction
436 static ir_node *build_graph(mul_env *env, instruction *inst)
439 ir_graph *irg = env->irg;
444 switch (inst->kind) {
446 l = build_graph(env, inst->in[0]);
447 r = build_graph(env, inst->in[1]);
448 c = new_r_Const_long(irg, env->shf_mode, inst->shift_count);
449 r = new_rd_Shl(env->dbg, env->blk, r, c, env->mode);
450 return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
452 l = build_graph(env, inst->in[0]);
453 c = new_r_Const_long(irg, env->shf_mode, inst->shift_count);
454 return inst->irn = new_rd_Shl(env->dbg, 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, 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, env->blk, l, r, env->mode);
464 return inst->irn = new_r_Const(irg, get_mode_null(env->mode));
466 panic("Unsupported instruction kind");
471 * Calculate the costs for the given instruction sequence.
472 * Note that additional costs due to higher register pressure are NOT evaluated yet
474 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, env->mode, NULL);
493 if (inst->shift_count > env->params->highest_shift_amount)
495 if (env->n_shift <= 0)
499 costs = evaluate_insn(env, inst->in[0]);
500 costs += env->evaluate(inst->kind, env->mode, NULL);
504 inst->costs = costs = env->evaluate(inst->kind, env->mode, 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, ir_tarval *tv)
533 obstack_init(&env.obst);
534 env.params = be_get_backend_param()->dep_param;
535 env.mode = get_tarval_mode(tv);
536 env.bits = (unsigned)get_mode_size_bits(env.mode);
538 env.root = emit_ROOT(&env, operand);
540 env.n_shift = env.params->maximum_shifts;
541 env.evaluate = env.params->evaluate != NULL ? env.params->evaluate : default_evaluate;
542 env.irg = get_irn_irg(irn);
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, env.mode, 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)
567 ir_mode *mode = get_irn_mode(irn);
573 const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
576 /* If the architecture dependent optimizations were not initialized
577 or this optimization was not enabled. */
578 if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
581 if (!is_Mul(irn) || !mode_is_int(mode))
584 /* we should never do the reverse transformations again
586 irg = get_irn_irg(irn);
587 set_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP);
589 left = get_binop_left(irn);
590 right = get_binop_right(irn);
594 /* Look, if one operand is a constant. */
595 if (is_Const(left)) {
596 tv = get_Const_tarval(left);
598 } else if (is_Const(right)) {
599 tv = get_Const_tarval(right);
604 res = do_decomposition(irn, operand, tv);
607 hook_arch_dep_replace_mul_with_shifts(irn);
616 * calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
618 static int tv_ld2(ir_tarval *tv, int bits)
622 for (num = i = 0; i < bits; ++i) {
623 unsigned char v = get_tarval_sub_bits(tv, i);
628 for (j = 0; j < 8; ++j)
641 /* for shorter lines */
642 #define ABS(a) tarval_abs(a)
643 #define NEG(a) tarval_neg(a)
644 #define NOT(a) tarval_not(a)
645 #define SHL(a, b) tarval_shl(a, b)
646 #define SHR(a, b) tarval_shr(a, b)
647 #define ADD(a, b) tarval_add(a, b)
648 #define SUB(a, b) tarval_sub(a, b, NULL)
649 #define MUL(a, b) tarval_mul(a, b)
650 #define DIV(a, b) tarval_div(a, b)
651 #define MOD(a, b) tarval_mod(a, b)
652 #define CMP(a, b) tarval_cmp(a, b)
653 #define CNV(a, m) tarval_convert_to(a, m)
654 #define ONE(m) get_mode_one(m)
655 #define ZERO(m) get_mode_null(m)
657 /** The result of a the magic() function. */
659 ir_tarval *M; /**< magic number */
660 int s; /**< shift amount */
661 int need_add; /**< an additional add is needed */
662 int need_sub; /**< an additional sub is needed */
666 * Signed division by constant d: calculate the Magic multiplier M and the shift amount s
668 * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
670 static struct ms magic(ir_tarval *d)
672 ir_mode *mode = get_tarval_mode(d);
673 ir_mode *u_mode = find_unsigned_mode(mode);
674 int bits = get_mode_size_bits(u_mode);
676 ir_tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */
679 ir_tarval *bits_minus_1, *two_bits_1;
683 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
685 /* we need overflow mode to work correctly */
686 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
689 bits_minus_1 = new_tarval_from_long(bits - 1, u_mode);
690 two_bits_1 = SHL(get_mode_one(u_mode), bits_minus_1);
692 ad = CNV(ABS(d), u_mode);
693 t = ADD(two_bits_1, SHR(CNV(d, u_mode), bits_minus_1));
694 anc = SUB(SUB(t, ONE(u_mode)), MOD(t, ad)); /* Absolute value of nc */
695 p = bits - 1; /* Init: p */
696 q1 = DIV(two_bits_1, anc); /* Init: q1 = 2^p/|nc| */
697 r1 = SUB(two_bits_1, MUL(q1, anc)); /* Init: r1 = rem(2^p, |nc|) */
698 q2 = DIV(two_bits_1, ad); /* Init: q2 = 2^p/|d| */
699 r2 = SUB(two_bits_1, MUL(q2, ad)); /* Init: r2 = rem(2^p, |d|) */
703 q1 = ADD(q1, q1); /* Update q1 = 2^p/|nc| */
704 r1 = ADD(r1, r1); /* Update r1 = rem(2^p, |nc|) */
706 if (CMP(r1, anc) & pn_Cmp_Ge) {
707 q1 = ADD(q1, ONE(u_mode));
711 q2 = ADD(q2, q2); /* Update q2 = 2^p/|d| */
712 r2 = ADD(r2, r2); /* Update r2 = rem(2^p, |d|) */
714 if (CMP(r2, ad) & pn_Cmp_Ge) {
715 q2 = ADD(q2, ONE(u_mode));
720 } while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq));
722 d_cmp = CMP(d, ZERO(mode));
724 if (d_cmp & pn_Cmp_Ge)
725 mag.M = ADD(CNV(q2, mode), ONE(mode));
727 mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode)));
729 M_cmp = CMP(mag.M, ZERO(mode));
733 /* need an add if d > 0 && M < 0 */
734 mag.need_add = d_cmp & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt;
736 /* need a sub if d < 0 && M > 0 */
737 mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt;
739 tarval_set_integer_overflow_mode(rem);
744 /** The result of the magicu() function. */
746 ir_tarval *M; /**< magic add constant */
747 int s; /**< shift amount */
748 int need_add; /**< add indicator */
752 * Unsigned division by constant d: calculate the Magic multiplier M and the shift amount s
754 * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned)
756 static struct mu magicu(ir_tarval *d)
758 ir_mode *mode = get_tarval_mode(d);
759 int bits = get_mode_size_bits(mode);
761 ir_tarval *nc, *delta, *q1, *r1, *q2, *r2;
762 ir_tarval *bits_minus_1, *two_bits_1, *seven_ff;
766 tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode();
768 /* we need overflow mode to work correctly */
769 tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP);
771 bits_minus_1 = new_tarval_from_long(bits - 1, mode);
772 two_bits_1 = SHL(get_mode_one(mode), bits_minus_1);
773 seven_ff = SUB(two_bits_1, ONE(mode));
775 magu.need_add = 0; /* initialize the add indicator */
776 nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d));
777 p = bits - 1; /* Init: p */
778 q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */
779 r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */
780 q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */
781 r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */
785 if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
786 q1 = ADD(ADD(q1, q1), ONE(mode));
787 r1 = SUB(ADD(r1, r1), nc);
794 if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
795 if (CMP(q2, seven_ff) & pn_Cmp_Ge)
798 q2 = ADD(ADD(q2, q2), ONE(mode));
799 r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
802 if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
806 r2 = ADD(ADD(r2, r2), ONE(mode));
808 delta = SUB(SUB(d, ONE(mode)), r2);
809 } while (p < 2*bits &&
810 (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
812 magu.M = ADD(q2, ONE(mode)); /* Magic number */
813 magu.s = p - bits; /* and shift amount */
815 tarval_set_integer_overflow_mode(rem);
821 * Build the Mulh replacement code for n / tv.
823 * Note that 'div' might be a mod or DivMod operation as well
825 static ir_node *replace_div_by_mulh(ir_node *div, ir_tarval *tv)
827 dbg_info *dbg = get_irn_dbg_info(div);
828 ir_node *n = get_binop_left(div);
829 ir_node *block = get_irn_n(div, -1);
830 ir_mode *mode = get_irn_mode(n);
831 int bits = get_mode_size_bits(mode);
834 /* Beware: do not transform bad code */
835 if (is_Bad(n) || is_Bad(block))
838 if (mode_is_signed(mode)) {
839 ir_graph *irg = get_irn_irg(div);
840 struct ms mag = magic(tv);
842 /* generate the Mulh instruction */
843 c = new_r_Const(irg, mag.M);
844 q = new_rd_Mulh(dbg, block, n, c, mode);
846 /* do we need an Add or Sub */
848 q = new_rd_Add(dbg, block, q, n, mode);
849 else if (mag.need_sub)
850 q = new_rd_Sub(dbg, block, q, n, mode);
852 /* Do we need the shift */
854 c = new_r_Const_long(irg, mode_Iu, mag.s);
855 q = new_rd_Shrs(dbg, block, q, c, mode);
859 c = new_r_Const_long(irg, mode_Iu, bits - 1);
860 t = new_rd_Shr(dbg, block, q, c, mode);
862 q = new_rd_Add(dbg, block, q, t, mode);
864 struct mu mag = magicu(tv);
866 ir_graph *irg = get_irn_irg(div);
868 /* generate the Mulh instruction */
869 c = new_r_Const(irg, mag.M);
870 q = new_rd_Mulh(dbg, block, n, c, mode);
874 /* use the GM scheme */
875 t = new_rd_Sub(dbg, block, n, q, mode);
877 c = new_r_Const(irg, get_mode_one(mode_Iu));
878 t = new_rd_Shr(dbg, block, t, c, mode);
880 t = new_rd_Add(dbg, block, t, q, mode);
882 c = new_r_Const_long(irg, mode_Iu, mag.s - 1);
883 q = new_rd_Shr(dbg, block, t, c, mode);
885 /* use the default scheme */
886 q = new_rd_Add(dbg, block, q, n, mode);
888 } else if (mag.s > 0) { /* default scheme, shift needed */
889 c = new_r_Const_long(irg, mode_Iu, mag.s);
890 q = new_rd_Shr(dbg, block, q, c, mode);
896 /* Replace Divs with Shifts and Add/Subs and Mulh. */
897 ir_node *arch_dep_replace_div_by_const(ir_node *irn)
899 const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
902 /* If the architecture dependent optimizations were not initialized
903 or this optimization was not enabled. */
904 if (params == NULL || (opts & arch_dep_div_by_const) == 0)
908 ir_node *c = get_Div_right(irn);
909 ir_node *block, *left;
920 tv = get_Const_tarval(c);
922 /* check for division by zero */
923 if (tarval_is_null(tv))
926 left = get_Div_left(irn);
927 mode = get_irn_mode(left);
928 block = get_irn_n(irn, -1);
929 dbg = get_irn_dbg_info(irn);
931 bits = get_mode_size_bits(mode);
935 if (mode_is_signed(mode)) {
936 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
937 ntv = tarval_neg(tv);
947 if (k >= 0) { /* division by 2^k or -2^k */
948 ir_graph *irg = get_irn_irg(irn);
949 if (mode_is_signed(mode)) {
951 ir_node *curr = left;
953 /* create the correction code for signed values only if there might be a remainder */
954 if (! get_Div_no_remainder(irn)) {
956 k_node = new_r_Const_long(irg, mode_Iu, k - 1);
957 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
960 k_node = new_r_Const_long(irg, mode_Iu, bits - k);
961 curr = new_rd_Shr(dbg, block, curr, k_node, mode);
963 curr = new_rd_Add(dbg, block, left, curr, mode);
968 k_node = new_r_Const_long(irg, mode_Iu, k);
969 res = new_rd_Shrs(dbg, block, curr, k_node, mode);
971 if (n_flag) { /* negate the result */
974 k_node = new_r_Const(irg, get_mode_null(mode));
975 res = new_rd_Sub(dbg, block, k_node, res, mode);
977 } else { /* unsigned case */
980 k_node = new_r_Const_long(irg, mode_Iu, k);
981 res = new_rd_Shr(dbg, block, left, k_node, mode);
985 if (allow_Mulh(params, mode))
986 res = replace_div_by_mulh(irn, tv);
991 hook_arch_dep_replace_division_by_const(irn);
996 /* Replace Mods with Shifts and Add/Subs and Mulh. */
997 ir_node *arch_dep_replace_mod_by_const(ir_node *irn)
999 const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
1002 /* If the architecture dependent optimizations were not initialized
1003 or this optimization was not enabled. */
1004 if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
1008 ir_node *c = get_Mod_right(irn);
1009 ir_node *block, *left;
1011 ir_tarval *tv, *ntv;
1019 tv = get_Const_tarval(c);
1021 /* check for division by zero */
1022 if (tarval_is_null(tv))
1025 left = get_Mod_left(irn);
1026 mode = get_irn_mode(left);
1027 block = get_irn_n(irn, -1);
1028 dbg = get_irn_dbg_info(irn);
1029 bits = get_mode_size_bits(mode);
1033 if (mode_is_signed(mode)) {
1034 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1035 ntv = tarval_neg(tv);
1044 ir_graph *irg = get_irn_irg(irn);
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_r_Const_long(irg, mode_Iu, k - 1);
1054 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
1057 k_node = new_r_Const_long(irg, 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_r_Const_long(irg, 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_r_Const_long(irg, mode, (1 << k) - 1);
1070 res = new_rd_And(dbg, block, left, k_node, mode);
1073 /* other constant */
1074 if (allow_Mulh(params, 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)
1095 const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
1098 /* If the architecture dependent optimizations were not initialized
1099 or this optimization was not enabled. */
1100 if (params == NULL ||
1101 ((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
1104 if (is_DivMod(irn)) {
1105 ir_node *c = get_DivMod_right(irn);
1106 ir_node *block, *left;
1108 ir_tarval *tv, *ntv;
1117 tv = get_Const_tarval(c);
1119 /* check for division by zero */
1120 if (tarval_is_null(tv))
1123 left = get_DivMod_left(irn);
1124 mode = get_irn_mode(left);
1125 block = get_irn_n(irn, -1);
1126 dbg = get_irn_dbg_info(irn);
1128 bits = get_mode_size_bits(mode);
1132 if (mode_is_signed(mode)) {
1133 /* for signed divisions, the algorithm works for a / -2^k by negating the result */
1134 ntv = tarval_neg(tv);
1144 if (k >= 0) { /* division by 2^k or -2^k */
1145 ir_graph *irg = get_irn_irg(irn);
1146 if (mode_is_signed(mode)) {
1147 ir_node *k_node, *c_k;
1148 ir_node *curr = left;
1151 k_node = new_r_Const_long(irg, mode_Iu, k - 1);
1152 curr = new_rd_Shrs(dbg, block, left, k_node, mode);
1155 k_node = new_r_Const_long(irg, mode_Iu, bits - k);
1156 curr = new_rd_Shr(dbg, block, curr, k_node, mode);
1158 curr = new_rd_Add(dbg, block, left, curr, mode);
1160 c_k = new_r_Const_long(irg, mode_Iu, k);
1162 *div = new_rd_Shrs(dbg, block, curr, c_k, mode);
1164 if (n_flag) { /* negate the div result */
1165 ir_node *k_node = new_r_Const(irg, get_mode_null(mode));
1166 *div = new_rd_Sub(dbg, block, k_node, *div, mode);
1169 k_node = new_r_Const_long(irg, 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 */
1174 ir_node *k_node = new_r_Const_long(irg, mode_Iu, k);
1175 *div = new_rd_Shr(dbg, block, left, k_node, mode);
1177 k_node = new_r_Const_long(irg, mode, (1 << k) - 1);
1178 *mod = new_rd_And(dbg, block, left, k_node, mode);
1181 /* other constant */
1182 if (allow_Mulh(params, mode)) {
1185 *div = replace_div_by_mulh(irn, tv);
1187 t = new_rd_Mul(dbg, block, *div, c, mode);
1189 /* t = arch_dep_mul_to_shift(t); */
1191 *mod = new_rd_Sub(dbg, block, left, t, mode);
1197 hook_arch_dep_replace_division_by_const(irn);