X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firarch.c;h=8f753f089041db943cb77028403f242a4ed4518f;hb=8bae7e4a8cf8f1a323ccf7b60c4091da49324f94;hp=de70b03c4b8c086ed754652ef3757efb7754e597;hpb=958a0df58adc11a4301d72142e30b6ec76f9b708;p=libfirm diff --git a/ir/ir/irarch.c b/ir/ir/irarch.c index de70b03c4..8f753f089 100644 --- a/ir/ir/irarch.c +++ b/ir/ir/irarch.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -24,17 +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", */ -#ifdef HAVE_CONFIG_H -# include "config.h" -#endif - -#ifdef HAVE_STDLIB_H -# include -#endif +#include "config.h" +#include #include #include "irnode_t.h" @@ -52,6 +47,7 @@ #include "ircons.h" #include "irarch.h" #include "irflag.h" +#include "error.h" #undef DEB @@ -135,7 +131,7 @@ 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)); + instruction *res = OALLOC(&env->obst, instruction); res->kind = shift > 0 ? LEA : ADD; res->in[0] = a; res->in[1] = b; @@ -149,7 +145,7 @@ 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)); + instruction *res = OALLOC(&env->obst, instruction); if (shift == env->bits) { /* a 2^bits with bits resolution is a zero */ res->kind = ZERO; @@ -176,7 +172,7 @@ 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)); + instruction *res = OALLOC(&env->obst, instruction); res->kind = SUB; res->in[0] = a; res->in[1] = b; @@ -190,7 +186,7 @@ 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)); + instruction *res = OALLOC(&env->obst, instruction); res->kind = ROOT; res->in[0] = NULL; res->in[1] = NULL; @@ -449,25 +445,25 @@ static ir_node *build_graph(mul_env *env, instruction *inst) { case LEA: l = build_graph(env, inst->in[0]); r = build_graph(env, inst->in[1]); - c = new_r_Const(current_ir_graph, env->blk, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode)); - 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); + c = new_Const_long(env->shf_mode, inst->shift_count); + 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_r_Const(current_ir_graph, env->blk, env->shf_mode, new_tarval_from_long(inst->shift_count, env->shf_mode)); - return inst->irn = new_rd_Shl(env->dbg, current_ir_graph, env->blk, l, c, env->mode); + c = new_Const_long(env->shf_mode, inst->shift_count); + 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_r_Const(current_ir_graph, env->blk, env->mode, get_mode_null(env->mode)); + return inst->irn = new_Const(get_mode_null(env->mode)); default: - assert(0); + panic("Unsupported instruction kind"); return NULL; } } @@ -507,16 +503,17 @@ static int evaluate_insn(mul_env *env, instruction *inst) { case ZERO: inst->costs = costs = env->evaluate(inst->kind, NULL); return costs; - default: - assert(0); - return 0; + case MUL: + case ROOT: + break; } + panic("Unsupported instruction kind"); } /** * Evaluate the replacement instructions and build a new graph * if faster than the Mul. - * returns the root of the new graph then or irn otherwise. + * Returns the root of the new graph then or irn otherwise. * * @param irn the Mul operation * @param operand the multiplication operand @@ -545,7 +542,7 @@ static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) { inst = decompose_mul(&env, R, r, tv); /* the paper suggests 70% here */ - mul_costs = (env.evaluate(MUL, tv) * 7) / 10; + mul_costs = (env.evaluate(MUL, tv) * 7 + 5) / 10; if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) { env.op = operand; env.blk = get_nodes_block(irn); @@ -562,42 +559,50 @@ 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_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; } @@ -634,7 +639,7 @@ static int tv_ld2(tarval *tv, int bits) { #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 SUB(a, b) tarval_sub(a, b, NULL) #define MUL(a, b) tarval_mul(a, b) #define DIV(a, b) tarval_div(a, b) #define MOD(a, b) tarval_mod(a, b) @@ -825,53 +830,53 @@ static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) { 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); + c = new_Const(mag.M); + 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_r_Const_long(current_ir_graph, block, mode_Iu, mag.s); - q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode); + c = new_Const_long(mode_Iu, mag.s); + q = new_rd_Shrs(dbg, 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); + c = new_Const_long(mode_Iu, bits - 1); + 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_r_Const(current_ir_graph, block, mode, mag.M); - q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode); + c = new_Const(mag.M); + 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_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); + c = new_Const(get_mode_one(mode_Iu)); + 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_r_Const_long(current_ir_graph, block, mode_Iu, mag.s-1); - q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode); + c = new_Const_long(mode_Iu, mag.s - 1); + 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_r_Const_long(current_ir_graph, block, mode_Iu, mag.s); - q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode); + c = new_Const_long(mode_Iu, mag.s); + q = new_rd_Shr(dbg, block, q, c, mode); } } return q; @@ -886,7 +891,7 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) { if (params == NULL || (opts & arch_dep_div_by_const) == 0) return irn; - if (get_irn_opcode(irn) == iro_Div) { + if (is_Div(irn)) { ir_node *c = get_Div_right(irn); ir_node *block, *left; ir_mode *mode; @@ -895,7 +900,7 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) { int n, bits; int k, n_flag; - if (get_irn_op(c) != op_Const) + if (! is_Const(c)) return irn; tv = get_Const_tarval(c); @@ -930,30 +935,35 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) { ir_node *k_node; ir_node *curr = left; - if (k != 1) { - 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); - } + /* create the correction code for signed values only if there might be a remainder */ + if (! get_Div_no_remainder(irn)) { + if (k != 1) { + k_node = new_Const_long(mode_Iu, k - 1); + curr = new_rd_Shrs(dbg, block, left, k_node, mode); + } - 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); + k_node = new_Const_long(mode_Iu, bits - k); + 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_r_Const_long(current_ir_graph, block, mode_Iu, k); - res = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode); + k_node = new_Const_long(mode_Iu, k); + res = new_rd_Shrs(dbg, 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); + k_node = new_Const(get_mode_null(mode)); + res = new_rd_Sub(dbg, block, k_node, res, mode); } } else { /* unsigned case */ ir_node *k_node; - 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); + k_node = new_Const_long(mode_Iu, k); + res = new_rd_Shr(dbg, block, left, k_node, mode); } } else { /* other constant */ @@ -977,7 +987,7 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) { if (params == NULL || (opts & arch_dep_mod_by_const) == 0) return irn; - if (get_irn_opcode(irn) == iro_Mod) { + if (is_Mod(irn)) { ir_node *c = get_Mod_right(irn); ir_node *block, *left; ir_mode *mode; @@ -986,7 +996,7 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) { int n, bits; int k; - if (get_irn_op(c) != op_Const) + if (! is_Const(c)) return irn; tv = get_Const_tarval(c); @@ -1022,35 +1032,35 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) { ir_node *curr = left; if (k != 1) { - 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_Const_long(mode_Iu, k - 1); + curr = new_rd_Shrs(dbg, block, left, k_node, mode); } - 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); + k_node = new_Const_long(mode_Iu, bits - k); + 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_r_Const_long(current_ir_graph, block, mode, (-1) << k); - curr = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode); + k_node = new_Const_long(mode, (-1) << k); + 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_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1); - res = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode); + k_node = new_Const_long(mode, (1 << k) - 1); + 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); } } } @@ -1071,7 +1081,7 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn ((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) { + if (is_DivMod(irn)) { ir_node *c = get_DivMod_right(irn); ir_node *block, *left; ir_mode *mode; @@ -1080,7 +1090,7 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn int n, bits; int k, n_flag; - if (get_irn_op(c) != op_Const) + if (! is_Const(c)) return; tv = get_Const_tarval(c); @@ -1116,38 +1126,38 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn ir_node *curr = left; if (k != 1) { - 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_Const_long(mode_Iu, k - 1); + curr = new_rd_Shrs(dbg, block, left, k_node, mode); } - 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); + k_node = new_Const_long(mode_Iu, bits - k); + 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_r_Const_long(current_ir_graph, block, mode_Iu, k); + 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_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_Const(get_mode_null(mode)); + *div = new_rd_Sub(dbg, block, k_node, *div, 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); + k_node = new_Const_long(mode, (-1) << k); + 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_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_Const_long(mode_Iu, k); + *div = new_rd_Shr(dbg, block, left, k_node, 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); + k_node = new_Const_long(mode, (1 << k) - 1); + *mod = new_rd_And(dbg, block, left, k_node, mode); } } else { /* other constant */ @@ -1156,11 +1166,11 @@ 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); } } }