From: Michael Beck Date: Fri, 22 Oct 2004 13:14:18 +0000 (+0000) Subject: Added Division by constant elimination algorithm from X-Git-Url: http://nsz.repo.hu/git/?a=commitdiff_plain;h=909c2e6d7426096e0e4df2371fbec7a06d4553ef;p=libfirm Added Division by constant elimination algorithm from "Hacker's Delight" [r4189] --- diff --git a/ir/ir/irarch.c b/ir/ir/irarch.c index 4f43bc608..ab66f3615 100644 --- a/ir/ir/irarch.c +++ b/ir/ir/irarch.c @@ -29,18 +29,59 @@ #define MAX_BITSTR 64 +/* when we need verifying */ +#ifdef NDEBUG +# define IRN_VRFY_IRG(res, irg) +#else +# define IRN_VRFY_IRG(res, irg) irn_vrfy_irg(res, irg) +#endif + /** The params got from the factory in arch_dep_init(...). */ static const arch_dep_params_t *params = NULL; /** The bit mask, which optimizations to apply. */ static arch_dep_opts_t opts; +/* we need this new pseudo op */ +static ir_op *op_Mulh = NULL; + +/** + * construct a Mulh: Mulh(a,b) = (a * b) >> w, w is the with in bits of a, b + */ +static ir_node * +new_rd_Mulh (dbg_info *db, ir_graph *irg, ir_node *block, + ir_node *op1, ir_node *op2, ir_mode *mode) +{ + ir_node *in[2]; + ir_node *res; + + if (! op_Mulh) { + op_Mulh = new_ir_op(get_next_ir_opcode(), "Mulh", op_pin_state_floats, irop_flag_commutative, oparity_binary, 0, 0); + } + + in[0] = op1; + in[1] = op2; + res = new_ir_node(db, irg, block, op_Mulh, mode, 2, in); + res = optimize_node(res); + IRN_VRFY_IRG(res, irg); + return res; +} + +ir_op *get_op_Mulh(void) { return op_Mulh; } + void arch_dep_init(arch_dep_params_factory_t factory) { opts = arch_dep_none; - if(factory != NULL) + if (factory != NULL) params = factory(); + + if (params && (opts & (arch_dep_div_by_const|arch_dep_mod_by_const))) { + if (! op_Mulh) { + /* create the Mulh operation */ + op_Mulh = new_ir_op(get_next_ir_opcode(), "Mulh", op_pin_state_floats, irop_flag_commutative, oparity_binary, 0, 0); + } + } } void arch_dep_set_opts(arch_dep_opts_t the_opts) { @@ -294,23 +335,282 @@ ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) return res; } -ir_node *arch_dep_replace_div_with_shifts(ir_node *irn) +/** + * calculated the ld2 of a tarval if tarval is 2^n, else returns -1. + */ +static int tv_ld2(tarval *tv, int bits) +{ + int i, k, num; + + for (num = i = 0; i < bits; ++i) { + unsigned char v = get_tarval_sub_bits(tv, i); + + if (v) { + int j; + + for (j = 0; j < 8; ++j) + if ((1 << j) & v) { + ++num; + k = 8 * i + j; + } + } + } + if (num == 1) + return k; + return -1; +} + + +/* for shorter lines */ +#define ABS(a) tarval_abs(a) +#define NEG(a) tarval_neg(a) +#define NOT(a) tarval_not(a) +#define SHL(a, b) tarval_shl(a, b) +#define SHR(a, b) tarval_shr(a, b) +#define ADD(a, b) tarval_add(a, b) +#define SUB(a, b) tarval_sub(a, b) +#define MUL(a, b) tarval_mul(a, b) +#define DIV(a, b) tarval_div(a, b) +#define MOD(a, b) tarval_mod(a, b) +#define CMP(a, b) tarval_cmp(a, b) +#define CNV(a, m) tarval_convert_to(a, m) +#define ONE(m) get_mode_one(m) +#define ZERO(m) get_mode_null(m) + +struct ms { + tarval *M; /**< magic number */ + int s; /**< shift amount */ + int need_add; /**< an additional add is needed */ + int need_sub; /**< an additional sub is needed */ +}; + +/** + * Signed division by constant d: calculate the Magic multiplier M and the shift amount s + * + * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler + */ +static struct ms magic(tarval *d) +{ + ir_mode *mode = get_tarval_mode(d); + ir_mode *u_mode = find_unsigned_mode(mode); + int bits = get_mode_size_bits(u_mode); + int p; + tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t; /* unsigned */ + pnc_number d_cmp, M_cmp; + + struct ms mag; + + /* 2^(bits-1) */ + tarval *bits_minus_1 = new_tarval_from_long(bits - 1, u_mode); + tarval *two_bits_1 = SHL(get_mode_one(u_mode), bits_minus_1); + + ad = CNV(ABS(d), u_mode); + t = ADD(two_bits_1, SHR(CNV(d, u_mode), bits_minus_1)); + anc = SUB(SUB(t, ONE(u_mode)), MOD(t, ad)); /* Absolute value of nc */ + p = bits - 1; /* Init: p */ + q1 = DIV(two_bits_1, anc); /* Init: q1 = 2^p/|nc| */ + r1 = SUB(two_bits_1, MUL(q1, anc)); /* Init: r1 = rem(2^p, |nc|) */ + q2 = DIV(two_bits_1, ad); /* Init: q2 = 2^p/|d| */ + r2 = SUB(two_bits_1, MUL(q2, ad)); /* Init: r2 = rem(2^p, |d|) */ + + do { + ++p; + q1 = ADD(q1, q1); /* Update q1 = 2^p/|nc| */ + r1 = ADD(r1, r1); /* Update r1 = rem(2^p, |nc|) */ + + if (CMP(r1, anc) & Ge) { + q1 = ADD(q1, ONE(u_mode)); + r1 = SUB(r1, anc); + } + + q2 = ADD(q2, q2); /* Update q2 = 2^p/|d| */ + r2 = ADD(r2, r2); /* Update r2 = rem(2^p, |d|) */ + + if (CMP(r2, ad) & Ge) { + q2 = ADD(q2, ONE(u_mode)); + r2 = SUB(r2, ad); + } + + delta = SUB(ad, r2); + } while (CMP(q1, delta) & Lt || (CMP(q1, delta) & Eq && CMP(r1, ZERO(u_mode)) & Eq)); + + d_cmp = CMP(d, ZERO(mode)); + + if (d_cmp & Ge) + mag.M = ADD(CNV(q2, mode), ONE(mode)); + else + mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode))); + + M_cmp = CMP(mag.M, ZERO(mode)); + + mag.s = p - bits; + + /* need an add if d > 0 && M < 0 */ + mag.need_add = d_cmp & Gt && M_cmp & Lt; + + /* need a sub if d < 0 && M > 0 */ + mag.need_sub = d_cmp & Lt && M_cmp & Gt; + + return mag; +} + +struct mu { + tarval *M; /**< magic add constant */ + int s; /**< shift amount */ + int need_add; /**< add indicator */ +}; + +/** + * Unsigned division by constant d: calculate the Magic multiplier M and the shift amount s + * + * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned) + */ +static struct mu magicu(tarval *d) +{ + ir_mode *mode = get_tarval_mode(d); + int bits = get_mode_size_bits(mode); + int p; + tarval *nc, *delta, *q1, *r1, *q2, *r2; + + struct mu magu; + + tarval *bits_minus_1 = new_tarval_from_long(bits - 1, mode); + tarval *two_bits_1 = SHL(get_mode_one(mode), bits_minus_1); + tarval *seven_ff = SUB(two_bits_1, ONE(mode)); + + magu.need_add = 0; /* initialize the add indicator */ + nc = SUB(NEG(ONE(mode)), MOD(NEG(d), d)); + p = bits - 1; /* Init: p */ + q1 = DIV(two_bits_1, nc); /* Init: q1 = 2^p/nc */ + r1 = SUB(two_bits_1, MUL(q1, nc)); /* Init: r1 = rem(2^p, nc) */ + q2 = DIV(seven_ff, d); /* Init: q2 = (2^p - 1)/d */ + r2 = SUB(seven_ff, MUL(q2, d)); /* Init: r2 = rem(2^p - 1, d) */ + + do { + ++p; + if (CMP(r1, SUB(nc, r1)) & Ge) { + q1 = ADD(ADD(q1, q1), ONE(mode)); + r1 = SUB(ADD(r1, r1), nc); + } + else { + q1 = ADD(q1, q1); + r1 = ADD(r1, r1); + } + + if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & Ge) { + if (CMP(q2, seven_ff) & Ge) + magu.need_add = 1; + + q2 = ADD(ADD(q2, q2), ONE(mode)); + r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d); + } + else { + if (CMP(q2, two_bits_1) & Ge) + magu.need_add = 1; + + q2 = ADD(q2, q2); + r2 = ADD(ADD(r2, r2), ONE(mode)); + } + delta = SUB(SUB(d, ONE(mode)), r2); + } while (p < 2*bits && + (CMP(q1, delta) & Lt || (CMP(q1, delta) & Eq && CMP(r1, ZERO(mode)) & Eq))); + + magu.M = ADD(q2, ONE(mode)); /* Magic number */ + magu.s = p - bits; /* and shift amount */ + + return magu; +} + +/** + * build the Mulh replacement code for n / tv + * + * Note thet 'div' might be a mod or DivMod operation as well + */ +static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) +{ + dbg_info *dbg = get_irn_dbg_info(div); + ir_node *n = get_binop_left(div); + ir_node *block = get_nodes_block(div); + ir_mode *mode = get_irn_mode(n); + int bits = get_mode_size_bits(mode); + ir_node *q, *t, *c; + + if (mode_is_signed(mode)) { + struct ms mag = magic(tv); + + /* generate the Mulh instruction */ + c = new_r_Const(current_ir_graph, block, mode, mag.M); + q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode); + + /* do we need an Add or Sub */ + if (mag.need_add) + q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode); + else if (mag.need_sub) + q = new_rd_Sub(dbg, current_ir_graph, block, q, n, mode); + + /* Do we need the shift */ + if (mag.s > 0) { + c = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(mag.s, mode_Iu)); + q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode); + } + + /* final */ + c = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(bits-1, mode_Iu)); + t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode); + + q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode); + } + else { + struct mu mag = magicu(tv); + ir_node *c; + + /* generate the Mulh instruction */ + c = new_r_Const(current_ir_graph, block, mode, mag.M); + q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode); + + if (mag.need_add) { + if (mag.s > 0) { + /* use the GM scheme */ + t = new_rd_Sub(dbg, current_ir_graph, block, n, q, mode); + + c = new_r_Const(current_ir_graph, block, mode_Iu, get_mode_one(mode_Iu)); + t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode); + + t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode); + + c = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(mag.s-1, mode_Iu)); + q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode); + } + else { + /* use the default scheme */ + q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode); + } + } + else if (mag.s > 0) { /* default scheme, shift needed */ + c = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(mag.s, mode_Iu)); + q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode); + } + } + return q; +} + +ir_node *arch_dep_replace_div_by_const(ir_node *irn) { ir_node *res = irn; /* If the architecture dependent optimizations were not initialized or this optimization was not enabled. */ - if (params == NULL || (opts & arch_dep_div_to_shift) == 0) + if (params == NULL || (opts & arch_dep_div_by_const) == 0) return irn; if (get_irn_opcode(irn) == iro_Div) { ir_node *c = get_Div_right(irn); ir_node *block, *left; ir_mode *mode; - tarval *tv; + tarval *tv, *ntv; dbg_info *dbg; int n, bits; - int i, k, num; + int k, n_flag; if (get_irn_op(c) != op_Const) return irn; @@ -324,22 +624,20 @@ ir_node *arch_dep_replace_div_with_shifts(ir_node *irn) bits = get_mode_size_bits(mode); n = (bits + 7) / 8; - for (num = i = 0; i < n; ++i) { - unsigned char v = get_tarval_sub_bits(tv, i); - - if (v) { - int j; - - for (j = 0; j < 8; ++j) - if ((1 << j) & v) { - ++num; - k = 8 * i + j; - } - } + k = -1; + if (mode_is_signed(mode)) { + /* for signed divisions, the algorithm works for a / -2^k by negating the result */ + ntv = tarval_neg(tv); + n_flag = 1; + k = tv_ld2(ntv, n); } - if (num == 1) { /* division by 2^k */ + if (k < 0) { + n_flag = 0; + k = tv_ld2(tv, n); + } + if (k >= 0) { /* division by 2^k or -2^k */ if (mode_is_signed(mode)) { ir_node *k_node; ir_node *curr = left; @@ -356,39 +654,52 @@ ir_node *arch_dep_replace_div_with_shifts(ir_node *irn) k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(k, mode_Iu)); res = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode); + + if (n_flag) { /* negate the result */ + ir_node *k_node; + + k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode)); + res = new_rd_Sub(dbg, current_ir_graph, block, k_node, res, mode); + } } else { /* unsigned case */ ir_node *k_node; k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(k, mode_Iu)); - res = new_rd_Shl(dbg, current_ir_graph, block, left, k_node, mode); + res = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode); } } + else { + /* other constant */ + if ((mode_is_signed(mode) && params->allow_mulhs) || + (!mode_is_signed(mode) && params->allow_mulhu)) + res = replace_div_by_mulh(irn, tv); + } } if (res != irn) - stat_arch_dep_replace_div_with_shifts(irn); + stat_arch_dep_replace_div_by_const(irn); return res; } -ir_node *arch_dep_replace_mod_with_shifts(ir_node *irn) +ir_node *arch_dep_replace_mod_by_const(ir_node *irn) { ir_node *res = irn; /* If the architecture dependent optimizations were not initialized or this optimization was not enabled. */ - if (params == NULL || (opts & arch_dep_mod_to_shift) == 0) + if (params == NULL || (opts & arch_dep_mod_by_const) == 0) return irn; if (get_irn_opcode(irn) == iro_Mod) { ir_node *c = get_Mod_right(irn); ir_node *block, *left; ir_mode *mode; - tarval *tv; + tarval *tv, *ntv; dbg_info *dbg; int n, bits; - int i, k, num; + int k; if (get_irn_op(c) != op_Const) return irn; @@ -402,22 +713,21 @@ ir_node *arch_dep_replace_mod_with_shifts(ir_node *irn) bits = get_mode_size_bits(mode); n = (bits + 7) / 8; - for (num = i = 0; i < n; ++i) { - unsigned char v = get_tarval_sub_bits(tv, i); - - if (v) { - int j; - - for (j = 0; j < 8; ++j) - if ((1 << j) & v) { - ++num; - k = 8 * i + j; - } - } + k = -1; + if (mode_is_signed(mode)) { + /* for signed divisions, the algorithm works for a / -2^k by negating the result */ + ntv = tarval_neg(tv); + k = tv_ld2(ntv, n); } - if (num == 1) { /* remainder by 2^k */ + if (k < 0) { + k = tv_ld2(tv, n); + } + if (k >= 0) { + /* division by 2^k or -2^k: + * we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k + */ if (mode_is_signed(mode)) { ir_node *k_node; ir_node *curr = left; @@ -444,32 +754,45 @@ ir_node *arch_dep_replace_mod_with_shifts(ir_node *irn) res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode); } } + else { + /* other constant */ + if ((mode_is_signed(mode) && params->allow_mulhs) || + (!mode_is_signed(mode) && params->allow_mulhu)) { + res = replace_div_by_mulh(irn, tv); + + res = new_rd_Mul(dbg, current_ir_graph, block, res, c, mode); + + /* res = arch_dep_mul_to_shift(res); */ + + res = new_rd_Sub(dbg, current_ir_graph, block, left, res, mode); + } + } } if (res != irn) - stat_arch_dep_replace_mod_with_shifts(irn); + stat_arch_dep_replace_mod_by_const(irn); return res; } -void arch_dep_replace_divmod_with_shifts(ir_node **div, ir_node **mod, ir_node *irn) +void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn) { *div = *mod = NULL; /* If the architecture dependent optimizations were not initialized or this optimization was not enabled. */ if (params == NULL || - ((opts & (arch_dep_div_to_shift|arch_dep_mod_to_shift)) != (arch_dep_div_to_shift|arch_dep_mod_to_shift))) + ((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const))) return; if (get_irn_opcode(irn) == iro_DivMod) { ir_node *c = get_DivMod_right(irn); ir_node *block, *left; ir_mode *mode; - tarval *tv; + tarval *tv, *ntv; dbg_info *dbg; int n, bits; - int i, k, num; + int k, n_flag; if (get_irn_op(c) != op_Const) return; @@ -483,24 +806,22 @@ void arch_dep_replace_divmod_with_shifts(ir_node **div, ir_node **mod, ir_node * bits = get_mode_size_bits(mode); n = (bits + 7) / 8; - for (num = i = 0; i < n; ++i) { - unsigned char v = get_tarval_sub_bits(tv, i); - - if (v) { - int j; - - for (j = 0; j < 8; ++j) - if ((1 << j) & v) { - ++num; - k = 8 * i + j; - } - } + k = -1; + if (mode_is_signed(mode)) { + /* for signed divisions, the algorithm works for a / -2^k by negating the result */ + ntv = tarval_neg(tv); + n_flag = 1; + k = tv_ld2(ntv, n); } - if (num == 1) { /* division & remainder by 2^k */ + if (k < 0) { + n_flag = 0; + k = tv_ld2(tv, n); + } + if (k >= 0) { /* division by 2^k or -2^k */ if (mode_is_signed(mode)) { - ir_node *k_node; + ir_node *k_node, *c_k; ir_node *curr = left; if (k != 1) { @@ -513,8 +834,16 @@ void arch_dep_replace_divmod_with_shifts(ir_node **div, ir_node **mod, ir_node * curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode); - k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(k, mode_Iu)); - *div = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode); + c_k = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(k, mode_Iu)); + + *div = new_rd_Shrs(dbg, current_ir_graph, block, curr, c_k, mode); + + if (n_flag) { /* negate the div result */ + ir_node *k_node; + + k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode)); + *div = new_rd_Sub(dbg, current_ir_graph, block, k_node, *div, mode); + } k_node = new_r_Const(current_ir_graph, block, mode, new_tarval_from_long((-1) << k, mode)); curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode); @@ -525,21 +854,38 @@ void arch_dep_replace_divmod_with_shifts(ir_node **div, ir_node **mod, ir_node * ir_node *k_node; k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(k, mode_Iu)); - *div = new_rd_Shl(dbg, current_ir_graph, block, left, k_node, mode); + *div = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode); k_node = new_r_Const(current_ir_graph, block, mode, new_tarval_from_long((1 << k) - 1, mode)); *mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode); } } + else { + /* other constant */ + if ((mode_is_signed(mode) && params->allow_mulhs) || + (!mode_is_signed(mode) && params->allow_mulhu)) { + ir_node *t; + + *div = replace_div_by_mulh(irn, tv); + + t = new_rd_Mul(dbg, current_ir_graph, block, *div, c, mode); + + /* t = arch_dep_mul_to_shift(t); */ + + *mod = new_rd_Sub(dbg, current_ir_graph, block, left, t, mode); + } + } } if (*div) - stat_arch_dep_replace_DivMod_with_shifts(irn); + stat_arch_dep_replace_DivMod_by_const(irn); } static const arch_dep_params_t default_params = { 1, /* also use subs */ + 0, /* allow Mulhs */ + 0, /* allow Mulus */ 4, /* maximum shifts */ 31 /* maximum shift amount */ }; diff --git a/ir/ir/irarch.h b/ir/ir/irarch.h index 6b733ed4c..4aee97b82 100644 --- a/ir/ir/irarch.h +++ b/ir/ir/irarch.h @@ -18,6 +18,8 @@ */ typedef struct { int also_use_subs : 1; /**< Use also Subs when resolving muls to shifts */ + int allow_mulhs : 1; /**< Use the Mulhs operation for division by constant */ + int allow_mulhu : 1; /**< Use the Mulhu operation for division by constant */ int maximum_shifts; /**< The maximum number of shifts that shall be inserted for a mul. */ @@ -25,7 +27,6 @@ typedef struct { int highest_shift_amount; /**< The highest shift amount you want to tolerate. Muls which would require a higher shift constant are left. */ - } arch_dep_params_t; /** @@ -45,8 +46,8 @@ const arch_dep_params_t *arch_dep_default_factory(void); typedef enum { arch_dep_none = 0, arch_dep_mul_to_shift = 1, /**< optimize Mul into Shift/Add/Sub */ - arch_dep_div_to_shift = 2, /**< optimize Div into Shift/Add/Mul */ - arch_dep_mod_to_shift = 4 /**< optimize Mod into Shift/Add/Mul */ + arch_dep_div_by_const = 2, /**< optimize Div into Shift/Add/Mulh */ + arch_dep_mod_by_const = 4 /**< optimize Mod into Shift/Add/Mulh */ } arch_dep_opts_t; /** @@ -82,56 +83,52 @@ void arch_dep_set_opts(arch_dep_opts_t opts); ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn); /** - * Replace Divs with Shifts and Add/Subs. + * Replace Divs with Shifts and Add/Subs and Mulh. * This function is driven by the 3 parameters: - * - also_use_subs - * - maximum_shifts - * - highest_shift_amount + * - allow_mulhu + * - allow_mulhs * * If irn is a Div with a Const, The constant is inspected, if it meets the - * requirements of the three variables stated above. If a Shl/Add/Sub + * requirements of the variables stated above. If a Shl/Add/Sub/Mulh * sequence can be generated, that meets these requirements, this expression * is returned. In each other case, irn is returned unmodified. * * @param irn The Firm node to inspect. * @return A replacement expression for irn. */ -ir_node *arch_dep_replace_div_with_shifts(ir_node *irn); +ir_node *arch_dep_replace_div_by_const(ir_node *irn); /** - * Replace Mods with Shifts and Add/Subs. + * Replace Mods with Shifts and Add/Subs and Mulh. * This function is driven by the 3 parameters: - * - also_use_subs - * - maximum_shifts - * - highest_shift_amount + * - allow_mulhu + * - allow_mulhs * - * If irn is a Div with a Const, The constant is inspected, if it meets the - * requirements of the three variables stated above. If a Shl/Add/Sub + * If irn is a Mod with a Const, The constant is inspected, if it meets the + * requirements of the variables stated above. If a Shl/Add/Sub/Mulh * sequence can be generated, that meets these requirements, this expression * is returned. In each other case, irn is returned unmodified. * * @param irn The Firm node to inspect. * @return A replacement expression for irn. */ -ir_node *arch_dep_replace_mod_with_shifts(ir_node *irn); +ir_node *arch_dep_replace_mod_by_const(ir_node *irn); /** - * Replace Mods with Shifts and Add/Subs. + * Replace DivMods with Shifts and Add/Subs and Mulh. * This function is driven by the 3 parameters: - * - also_use_subs - * - maximum_shifts - * - highest_shift_amount + * - allow_mulhu + * - allow_mulhs * - * If irn is a Div with a Const, The constant is inspected, if it meets the - * requirements of the three variables stated above. If a Shl/Add/Sub + * If irn is a DivMod with a Const, The constant is inspected, if it meets the + * requirements of the variables stated above. If a Shl/Add/Sub/Mulh * sequence can be generated, that meets these requirements, this expression * is returned. In each other case, irn is returned unmodified. * * @param div After call contains the Firm node div result or NULL. * @param mod After call contains the Firm node mod result or NULL. * @param irn The Firm node to inspect. - * @return A replacement expression for irn. */ -void arch_dep_replace_divmod_with_shifts(ir_node **div, ir_node **mod, ir_node *irn); +void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn); #endif