X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firarch.c;h=b650a3e674fbf0548d8f2a688bab314b1650f8c6;hb=a6e674f74ea93c3661f100088db520acb701571f;hp=a3b7413aa8d4390668557b35df7b9323368009b6;hpb=06a5aaa9fa2bdff237781b9725e58b7298fabe64;p=libfirm diff --git a/ir/ir/irarch.c b/ir/ir/irarch.c index a3b7413aa..b650a3e67 100644 --- a/ir/ir/irarch.c +++ b/ir/ir/irarch.c @@ -24,15 +24,12 @@ * @author Sebastian Hack, Michael Beck * @version $Id$ * - * Implements "Strenght Reduction of Multiplications by Integer Constants" by Youfeng Wu. + * Implements "Strength Reduction of Multiplications by Integer Constants" by Youfeng Wu. * Implements Division and Modulo by Consts from "Hackers Delight", */ #include "config.h" -#ifdef HAVE_STDLIB_H -# include -#endif - +#include #include #include "irnode_t.h" @@ -69,19 +66,22 @@ static const ir_settings_arch_dep_t *params = NULL; /** The bit mask, which optimizations to apply. */ static arch_dep_opts_t opts; -void arch_dep_init(arch_dep_params_factory_t factory) { +void arch_dep_init(arch_dep_params_factory_t factory) +{ opts = arch_dep_none; if (factory != NULL) params = factory(); } -void arch_dep_set_opts(arch_dep_opts_t the_opts) { +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) { +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); @@ -122,7 +122,8 @@ typedef struct _mul_env { * Some kind of default evaluator. Return the cost of * instructions. */ -static int default_evaluate(insn_kind kind, tarval *tv) { +static int default_evaluate(insn_kind kind, tarval *tv) +{ (void) tv; if (kind == MUL) @@ -133,8 +134,9 @@ static int default_evaluate(insn_kind kind, tarval *tv) { /** * emit a LEA (or an Add) instruction */ -static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) { - instruction *res = obstack_alloc(&env->obst, sizeof(*res)); +static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) +{ + instruction *res = OALLOC(&env->obst, instruction); res->kind = shift > 0 ? LEA : ADD; res->in[0] = a; res->in[1] = b; @@ -147,8 +149,9 @@ static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsig /** * emit a SHIFT (or an Add or a Zero) instruction */ -static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) { - instruction *res = obstack_alloc(&env->obst, sizeof(*res)); +static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) +{ + instruction *res = OALLOC(&env->obst, instruction); if (shift == env->bits) { /* a 2^bits with bits resolution is a zero */ res->kind = ZERO; @@ -174,8 +177,9 @@ static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) { /** * emit a SUB instruction */ -static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) { - instruction *res = obstack_alloc(&env->obst, sizeof(*res)); +static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) +{ + instruction *res = OALLOC(&env->obst, instruction); res->kind = SUB; res->in[0] = a; res->in[1] = b; @@ -188,8 +192,9 @@ static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) { /** * emit the ROOT instruction */ -static instruction *emit_ROOT(mul_env *env, ir_node *root_op) { - instruction *res = obstack_alloc(&env->obst, sizeof(*res)); +static instruction *emit_ROOT(mul_env *env, ir_node *root_op) +{ + instruction *res = OALLOC(&env->obst, instruction); res->kind = ROOT; res->in[0] = NULL; res->in[1] = NULL; @@ -203,7 +208,8 @@ static instruction *emit_ROOT(mul_env *env, ir_node *root_op) { /** * Returns the condensed representation of the tarval tv */ -static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) { +static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) +{ ir_mode *mode = get_tarval_mode(tv); int bits = get_mode_size_bits(mode); char *bitstr = get_tarval_bitpattern(tv); @@ -227,7 +233,8 @@ static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) { /** * Calculate the gain when using the generalized complementary technique */ -static int calculate_gain(unsigned char *R, int r) { +static int calculate_gain(unsigned char *R, int r) +{ int max_gain = 0; int idx = -1, i; int gain; @@ -249,7 +256,8 @@ static int calculate_gain(unsigned char *R, int r) { /** * Calculates the condensed complement of a given (R,r) tuple */ -static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs) { +static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs) +{ unsigned char *value = obstack_alloc(&env->obst, env->bits); int i, l, j; unsigned char c; @@ -289,7 +297,8 @@ static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r /** * creates a tarval from a condensed representation. */ -static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) { +static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) +{ tarval *res, *tv; int i, j; @@ -313,7 +322,8 @@ static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, t /* * handle simple cases with up-to 2 bits set */ -static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) { +static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) +{ instruction *ins, *ins2; (void) N; @@ -342,7 +352,8 @@ static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r /** * Main decompose driver. */ -static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) { +static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) +{ unsigned i; int gain; @@ -410,7 +421,8 @@ static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval /** * basic decomposition routine */ -static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) { +static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) +{ instruction *Ns; unsigned t; @@ -438,7 +450,8 @@ static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, t * @param env the environment * @param inst the instruction */ -static ir_node *build_graph(mul_env *env, instruction *inst) { +static ir_node *build_graph(mul_env *env, instruction *inst) +{ ir_node *l, *r, *c; if (inst->irn) @@ -449,25 +462,24 @@ static ir_node *build_graph(mul_env *env, instruction *inst) { l = build_graph(env, inst->in[0]); r = build_graph(env, inst->in[1]); c = new_Const_long(env->shf_mode, inst->shift_count); - r = new_rd_Shl(env->dbg, current_ir_graph, env->blk, r, c, env->mode); - return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode); + r = new_rd_Shl(env->dbg, env->blk, r, c, env->mode); + return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode); case SHIFT: l = build_graph(env, inst->in[0]); c = new_Const_long(env->shf_mode, inst->shift_count); - return inst->irn = new_rd_Shl(env->dbg, current_ir_graph, env->blk, l, c, env->mode); + return inst->irn = new_rd_Shl(env->dbg, env->blk, l, c, env->mode); case SUB: l = build_graph(env, inst->in[0]); r = build_graph(env, inst->in[1]); - return inst->irn = new_rd_Sub(env->dbg, current_ir_graph, env->blk, l, r, env->mode); + return inst->irn = new_rd_Sub(env->dbg, env->blk, l, r, env->mode); case ADD: l = build_graph(env, inst->in[0]); r = build_graph(env, inst->in[1]); - return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode); + return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode); case ZERO: return inst->irn = new_Const(get_mode_null(env->mode)); default: panic("Unsupported instruction kind"); - return NULL; } } @@ -475,7 +487,8 @@ static ir_node *build_graph(mul_env *env, instruction *inst) { * Calculate the costs for the given instruction sequence. * Note that additional costs due to higher register pressure are NOT evaluated yet */ -static int evaluate_insn(mul_env *env, instruction *inst) { +static int evaluate_insn(mul_env *env, instruction *inst) +{ int costs; if (inst->costs >= 0) { @@ -524,7 +537,8 @@ static int evaluate_insn(mul_env *env, instruction *inst) { * * @return the new graph */ -static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) { +static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) +{ mul_env env; instruction *inst; unsigned char *R; @@ -561,43 +575,52 @@ static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) { } /* Replace Muls with Shifts and Add/Subs. */ -ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) { - ir_node *res = irn; +ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) +{ + ir_graph *irg; + ir_node *res = irn; ir_mode *mode = get_irn_mode(irn); + ir_node *left; + ir_node *right; + ir_node *operand; + tarval *tv; + /* If the architecture dependent optimizations were not initialized or this optimization was not enabled. */ if (params == NULL || (opts & arch_dep_mul_to_shift) == 0) return irn; - set_arch_dep_running(1); - { - if (is_Mul(irn) && mode_is_int(mode)) { - 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 (is_Const(left)) { - tv = get_Const_tarval(left); - operand = right; - } else if (is_Const(right)) { - tv = get_Const_tarval(right); - operand = left; - } + if (!is_Mul(irn) || !mode_is_int(mode)) + return res; + + /* we should never do the reverse transformations again + (like x+x -> 2*x) */ + irg = get_irn_irg(irn); + set_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP); + + left = get_binop_left(irn); + right = get_binop_right(irn); + tv = NULL; + operand = NULL; + + /* Look, if one operand is a constant. */ + if (is_Const(left)) { + tv = get_Const_tarval(left); + operand = right; + } else if (is_Const(right)) { + tv = get_Const_tarval(right); + operand = left; + } - if (tv != NULL) { - res = do_decomposition(irn, operand, tv); + if (tv != NULL) { + res = do_decomposition(irn, operand, tv); - if (res != irn) { - hook_arch_dep_replace_mul_with_shifts(irn); - exchange(irn, res); - } - } + if (res != irn) { + hook_arch_dep_replace_mul_with_shifts(irn); + exchange(irn, res); } } - //set_arch_dep_running(0); return res; } @@ -605,7 +628,8 @@ ir_node *arch_dep_replace_mul_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) { +static int tv_ld2(tarval *tv, int bits) +{ int i, k = 0, num; for (num = i = 0; i < bits; ++i) { @@ -656,7 +680,8 @@ struct ms { * * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler */ -static struct ms magic(tarval *d) { +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); @@ -741,7 +766,8 @@ struct mu { * * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned) */ -static struct mu magicu(tarval *d) { +static struct mu magicu(tarval *d) +{ ir_mode *mode = get_tarval_mode(d); int bits = get_mode_size_bits(mode); int p; @@ -809,7 +835,8 @@ static struct mu magicu(tarval *d) { * * Note that 'div' might be a mod or DivMod operation as well */ -static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) { +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); @@ -826,59 +853,60 @@ static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) { /* generate the Mulh instruction */ c = new_Const(mag.M); - q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode); + q = new_rd_Mulh(dbg, 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); + q = new_rd_Add(dbg, block, q, n, mode); else if (mag.need_sub) - q = new_rd_Sub(dbg, current_ir_graph, block, q, n, mode); + q = new_rd_Sub(dbg, block, q, n, mode); /* Do we need the shift */ if (mag.s > 0) { c = new_Const_long(mode_Iu, mag.s); - q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode); + q = new_rd_Shrs(dbg, block, q, c, mode); } /* final */ c = new_Const_long(mode_Iu, bits - 1); - t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode); + t = new_rd_Shr(dbg, block, q, c, mode); - q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode); + q = new_rd_Add(dbg, block, q, t, mode); } else { struct mu mag = magicu(tv); ir_node *c; /* generate the Mulh instruction */ c = new_Const(mag.M); - q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode); + q = new_rd_Mulh(dbg, 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); + t = new_rd_Sub(dbg, block, n, q, mode); c = new_Const(get_mode_one(mode_Iu)); - t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode); + t = new_rd_Shr(dbg, block, t, c, mode); - t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode); + t = new_rd_Add(dbg, block, t, q, mode); c = new_Const_long(mode_Iu, mag.s - 1); - q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode); + q = new_rd_Shr(dbg, block, t, c, mode); } else { /* use the default scheme */ - q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode); + q = new_rd_Add(dbg, block, q, n, mode); } } else if (mag.s > 0) { /* default scheme, shift needed */ c = new_Const_long(mode_Iu, mag.s); - q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode); + q = new_rd_Shr(dbg, block, q, c, mode); } } return q; } /* Replace Divs with Shifts and Add/Subs and Mulh. */ -ir_node *arch_dep_replace_div_by_const(ir_node *irn) { +ir_node *arch_dep_replace_div_by_const(ir_node *irn) +{ ir_node *res = irn; /* If the architecture dependent optimizations were not initialized @@ -893,7 +921,8 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) { tarval *tv, *ntv; dbg_info *dbg; int n, bits; - int k, n_flag; + int k; + int n_flag = 0; if (! is_Const(c)) return irn; @@ -931,34 +960,34 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) { ir_node *curr = left; /* create the correction code for signed values only if there might be a remainder */ - if (! is_Div_remainderless(irn)) { + if (! get_Div_no_remainder(irn)) { if (k != 1) { k_node = new_Const_long(mode_Iu, k - 1); - curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode); + curr = new_rd_Shrs(dbg, block, left, k_node, mode); } k_node = new_Const_long(mode_Iu, bits - k); - curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode); + curr = new_rd_Shr(dbg, block, curr, k_node, mode); - curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode); + curr = new_rd_Add(dbg, block, left, curr, mode); } else { k_node = left; } k_node = new_Const_long(mode_Iu, k); - res = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode); + res = new_rd_Shrs(dbg, block, curr, k_node, mode); if (n_flag) { /* negate the result */ ir_node *k_node; k_node = new_Const(get_mode_null(mode)); - res = new_rd_Sub(dbg, current_ir_graph, block, k_node, res, mode); + res = new_rd_Sub(dbg, block, k_node, res, mode); } } else { /* unsigned case */ ir_node *k_node; k_node = new_Const_long(mode_Iu, k); - res = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode); + res = new_rd_Shr(dbg, block, left, k_node, mode); } } else { /* other constant */ @@ -974,7 +1003,8 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) { } /* Replace Mods with Shifts and Add/Subs and Mulh. */ -ir_node *arch_dep_replace_mod_by_const(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 @@ -1028,34 +1058,34 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) { if (k != 1) { k_node = new_Const_long(mode_Iu, k - 1); - curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode); + curr = new_rd_Shrs(dbg, block, left, k_node, mode); } k_node = new_Const_long(mode_Iu, bits - k); - curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode); + curr = new_rd_Shr(dbg, block, curr, k_node, mode); - curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode); + curr = new_rd_Add(dbg, block, left, curr, mode); k_node = new_Const_long(mode, (-1) << k); - curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode); + curr = new_rd_And(dbg, block, curr, k_node, mode); - res = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode); + res = new_rd_Sub(dbg, block, left, curr, mode); } else { /* unsigned case */ ir_node *k_node; k_node = new_Const_long(mode, (1 << k) - 1); - res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode); + res = new_rd_And(dbg, 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 = new_rd_Mul(dbg, block, res, c, mode); /* res = arch_dep_mul_to_shift(res); */ - res = new_rd_Sub(dbg, current_ir_graph, block, left, res, mode); + res = new_rd_Sub(dbg, block, left, res, mode); } } } @@ -1067,7 +1097,8 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) { } /* Replace DivMods with Shifts and Add/Subs and Mulh. */ -void arch_dep_replace_divmod_by_const(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 @@ -1083,7 +1114,8 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn tarval *tv, *ntv; dbg_info *dbg; int n, bits; - int k, n_flag; + int k; + int n_flag = 0; if (! is_Const(c)) return; @@ -1122,37 +1154,37 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn if (k != 1) { k_node = new_Const_long(mode_Iu, k - 1); - curr = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode); + curr = new_rd_Shrs(dbg, block, left, k_node, mode); } k_node = new_Const_long(mode_Iu, bits - k); - curr = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode); + curr = new_rd_Shr(dbg, block, curr, k_node, mode); - curr = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode); + curr = new_rd_Add(dbg, block, left, curr, mode); c_k = new_Const_long(mode_Iu, k); - *div = new_rd_Shrs(dbg, current_ir_graph, block, curr, c_k, mode); + *div = new_rd_Shrs(dbg, block, curr, c_k, mode); if (n_flag) { /* negate the div result */ ir_node *k_node; k_node = new_Const(get_mode_null(mode)); - *div = new_rd_Sub(dbg, current_ir_graph, block, k_node, *div, mode); + *div = new_rd_Sub(dbg, block, k_node, *div, mode); } k_node = new_Const_long(mode, (-1) << k); - curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode); + curr = new_rd_And(dbg, block, curr, k_node, mode); - *mod = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode); + *mod = new_rd_Sub(dbg, block, left, curr, mode); } else { /* unsigned case */ ir_node *k_node; k_node = new_Const_long(mode_Iu, k); - *div = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode); + *div = new_rd_Shr(dbg, block, left, k_node, mode); k_node = new_Const_long(mode, (1 << k) - 1); - *mod = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode); + *mod = new_rd_And(dbg, block, left, k_node, mode); } } else { /* other constant */ @@ -1161,11 +1193,11 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn *div = replace_div_by_mulh(irn, tv); - t = new_rd_Mul(dbg, current_ir_graph, block, *div, c, mode); + t = new_rd_Mul(dbg, block, *div, c, mode); /* t = arch_dep_mul_to_shift(t); */ - *mod = new_rd_Sub(dbg, current_ir_graph, block, left, t, mode); + *mod = new_rd_Sub(dbg, block, left, t, mode); } } } @@ -1187,6 +1219,7 @@ static const ir_settings_arch_dep_t default_params = { }; /* A default parameter factory for testing purposes. */ -const ir_settings_arch_dep_t *arch_dep_default_factory(void) { +const ir_settings_arch_dep_t *arch_dep_default_factory(void) +{ return &default_params; }