X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firarch.c;h=7f15232ba9e6c3d0cf565b9cb97fc173739d50a5;hb=169fd803ea2ed08171113c1fd7ab4e528e1ebc26;hp=b65eaea9d8f3c68fcde59c962092cf2b460ca56d;hpb=ccc92d26458413986cb5ac221d4b83c8fa740628;p=libfirm diff --git a/ir/ir/irarch.c b/ir/ir/irarch.c index b65eaea9d..7f15232ba 100644 --- a/ir/ir/irarch.c +++ b/ir/ir/irarch.c @@ -16,37 +16,98 @@ #include "ircons_t.h" #include "irgmod.h" #include "irvrfy.h" -#include "tv.h" +#include "tv_t.h" #include "dbginfo_t.h" #include "iropt_dbg.h" #include "irflag_t.h" -#include "firmstat.h" +#include "irhooks.h" #include "ircons.h" #include "irarch.h" -#include "firmstat.h" +#include "irreflect.h" #undef DEB #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) { + rflct_sig_t *sig; + int mulh_opc = get_next_ir_opcode(); + + op_Mulh = new_ir_op(mulh_opc, "Mulh", op_pin_state_floats, irop_flag_commutative, oparity_binary, 0, 0); + sig = rflct_signature_allocate(1, 3); + rflct_signature_set_arg(sig, 0, 0, "Res", RFLCT_MC(Int), 0, 0); + rflct_signature_set_arg(sig, 1, 0, "Block", RFLCT_MC(BB), 0, 0); + rflct_signature_set_arg(sig, 1, 1, "Op 0", RFLCT_MC(Int), 0, 0); + rflct_signature_set_arg(sig, 1, 2, "Op 1", RFLCT_MC(Int), 0, 0); + + rflct_new_opcode(mulh_opc, "Mulh", false); + rflct_opcode_add_signature(mulh_opc, sig); + + } + + 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) { opts = the_opts; } +/* check, whether a mode allows a Mulh instruction */ +static int allow_Mulh(ir_mode *mode) +{ + if (get_mode_size_bits(mode) > params->max_bits_for_mulh) + return 0; + return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu); +} + ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) { ir_node *res = irn; @@ -54,18 +115,18 @@ ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) /* If the architecture dependent optimizations were not initialized or this optimization was not enabled. */ - if(params == NULL || (opts & arch_dep_mul_to_shift) == 0) + if (params == NULL || (opts & arch_dep_mul_to_shift) == 0) return irn; - if(get_irn_opcode(irn) == iro_Mul && mode_is_int(mode)) { - ir_node *block = get_nodes_block(irn); + if (get_irn_op(irn) == op_Mul && mode_is_int(mode)) { + ir_node *block = get_irn_n(irn, -1); ir_node *left = get_binop_left(irn); ir_node *right = get_binop_right(irn); tarval *tv = NULL; ir_node *operand = NULL; /* Look, if one operand is a constant. */ - if(get_irn_opcode(left) == iro_Const) { + if (get_irn_opcode(left) == iro_Const) { tv = get_Const_tarval(left); operand = right; } else if(get_irn_opcode(right) == iro_Const) { @@ -73,7 +134,7 @@ ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) operand = left; } - if(tv != NULL) { + if (tv != NULL) { int maximum_shifts = params->maximum_shifts; int also_use_subs = params->also_use_subs; int highest_shift_amount = params->highest_shift_amount; @@ -246,17 +307,15 @@ ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) int amount = abs(curr_shift) - 1; ir_node *aux = operand; - assert(amount >= 0 && "What is a negative shift??"); - if(amount != 0) { - tarval *shift_amount = new_tarval_from_long(amount, mode_Iu); - ir_node *cnst = new_r_Const(current_ir_graph, block, mode_Iu, shift_amount); + if (amount != 0) { + ir_node *cnst = new_r_Const_long(current_ir_graph, block, mode_Iu, amount); aux = new_r_Shl(current_ir_graph, block, operand, cnst, mode); } - if(curr) { - if(sub) + if (curr) { + if (sub) curr = new_r_Sub(current_ir_graph, block, curr, aux, mode); else curr = new_r_Add(current_ir_graph, block, curr, aux, mode); @@ -272,7 +331,7 @@ ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) #ifdef DEB { const char *prefix = ""; - for(i = 0; i < n; i++) { + for (i = 0; i < n; ++i) { fprintf(stderr, "%s%d", prefix, shifts[i]); prefix = ", "; } @@ -289,150 +348,439 @@ ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) } if (res != irn) - stat_arch_dep_replace_mul_with_shifts(irn); + hook_arch_dep_replace_mul_with_shifts(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 */ + pn_Cmp d_cmp, M_cmp; + + tarval *bits_minus_1, *two_bits_1; + + struct ms mag; + + tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode(); + + /* we need overflow mode to work correctly */ + tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP); + + /* 2^(bits-1) */ + bits_minus_1 = new_tarval_from_long(bits - 1, u_mode); + 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) & pn_Cmp_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) & pn_Cmp_Ge) { + q2 = ADD(q2, ONE(u_mode)); + r2 = SUB(r2, ad); + } + + delta = SUB(ad, r2); + } while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq)); + + d_cmp = CMP(d, ZERO(mode)); + + if (d_cmp & pn_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 & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt; + + /* need a sub if d < 0 && M > 0 */ + mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt; + + tarval_set_integer_overflow_mode(rem); + + 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; + tarval *bits_minus_1, *two_bits_1, *seven_ff; + + struct mu magu; + + tarval_int_overflow_mode_t rem = tarval_get_integer_overflow_mode(); + + /* we need overflow mode to work correctly */ + tarval_set_integer_overflow_mode(TV_OVERFLOW_WRAP); + + bits_minus_1 = new_tarval_from_long(bits - 1, mode); + two_bits_1 = SHL(get_mode_one(mode), bits_minus_1); + 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)) & pn_Cmp_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)) & pn_Cmp_Ge) { + if (CMP(q2, seven_ff) & pn_Cmp_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) & pn_Cmp_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) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq))); + + magu.M = ADD(q2, ONE(mode)); /* Magic number */ + magu.s = p - bits; /* and shift amount */ + + tarval_set_integer_overflow_mode(rem); + + return magu; +} + +/** + * build the Mulh replacement code for n / tv + * + * Note that '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_irn_n(div, -1); + ir_mode *mode = get_irn_mode(n); + int bits = get_mode_size_bits(mode); + ir_node *q, *t, *c; + + /* Beware: do not transform bad code */ + if (is_Bad(n) || is_Bad(block)) + return div; + + 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_long(current_ir_graph, block, mode_Iu, mag.s); + q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode); + } + + /* final */ + c = new_r_Const_long(current_ir_graph, block, mode_Iu, bits-1); + 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_long(current_ir_graph, block, mode_Iu, mag.s-1); + 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_long(current_ir_graph, block, mode_Iu, mag.s); + 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; left = get_Div_left(irn); mode = get_irn_mode(left); - block = get_nodes_block(irn); + block = get_irn_n(irn, -1); dbg = get_irn_dbg_info(irn); tv = get_Const_tarval(c); 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; if (k != 1) { - k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(k - 1, mode_Iu)); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1); curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode); } - k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(bits - k, mode_Iu)); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k); curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode); 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)); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k); 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); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k); + res = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode); } } + else { + /* other constant */ + if (allow_Mulh(mode)) + res = replace_div_by_mulh(irn, tv); + } } if (res != irn) - stat_arch_dep_replace_div_with_shifts(irn); + hook_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; left = get_Mod_left(irn); mode = get_irn_mode(left); - block = get_nodes_block(irn); + block = get_irn_n(irn, -1); dbg = get_irn_dbg_info(irn); tv = get_Const_tarval(c); 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; if (k != 1) { - k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(k - 1, mode_Iu)); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1); curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode); } - k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(bits - k, mode_Iu)); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k); curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode); curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode); - k_node = new_r_Const(current_ir_graph, block, mode, new_tarval_from_long((-1) << k, mode)); + k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k); curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode); res = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode); @@ -440,83 +788,101 @@ ir_node *arch_dep_replace_mod_with_shifts(ir_node *irn) else { /* unsigned case */ ir_node *k_node; - k_node = new_r_Const(current_ir_graph, block, mode, new_tarval_from_long((1 << k) - 1, mode)); + k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1); res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode); } } + else { + /* other constant */ + if (allow_Mulh(mode)) { + 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); + hook_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; left = get_DivMod_left(irn); mode = get_irn_mode(left); - block = get_nodes_block(irn); + block = get_irn_n(irn, -1); dbg = get_irn_dbg_info(irn); tv = get_Const_tarval(c); 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) { - k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(k - 1, mode_Iu)); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1); curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode); } - k_node = new_r_Const(current_ir_graph, block, mode_Iu, new_tarval_from_long(bits - k, mode_Iu)); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k); curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode); 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_long(current_ir_graph, block, mode_Iu, k); + + *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)); + k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k); curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode); *mod = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode); @@ -524,24 +890,42 @@ void arch_dep_replace_divmod_with_shifts(ir_node **div, ir_node **mod, ir_node * 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)); - *div = new_rd_Shl(dbg, current_ir_graph, block, left, k_node, mode); + k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k); + *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)); + k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1); *mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode); } } + else { + /* other constant */ + if (allow_Mulh(mode)) { + 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); + hook_arch_dep_replace_DivMod_by_const(irn); } static const arch_dep_params_t default_params = { - 1, /* also use subs */ - 4, /* maximum shifts */ - 31 /* maximum shift amount */ + 1, /* also use subs */ + 4, /* maximum shifts */ + 31, /* maximum shift amount */ + + 0, /* allow Mulhs */ + 0, /* allow Mulus */ + 32 /* Mulh allowed up to 32 bit */ }; const arch_dep_params_t *arch_dep_default_factory(void) {