X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_intrinsics.c;h=8df4aeb6a44f37702f1636390fcad6d3c89ed39d;hb=2d4cb2720430d39651ddb33d9b342f69bd757a75;hp=f2c1a17159170998840ae986a08a83bc9228a762;hpb=8c1f62ae9c96c62a4192ca022c07889c27ca0c0e;p=libfirm diff --git a/ir/be/ia32/ia32_intrinsics.c b/ir/be/ia32/ia32_intrinsics.c index f2c1a1715..8df4aeb6a 100644 --- a/ir/be/ia32/ia32_intrinsics.c +++ b/ir/be/ia32/ia32_intrinsics.c @@ -1,20 +1,6 @@ /* - * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. - * * This file is part of libFirm. - * - * This file may be distributed and/or modified under the terms of the - * GNU General Public License version 2 as published by the Free Software - * Foundation and appearing in the file LICENSE.GPL included in the - * packaging of this file. - * - * Licensees holding valid libFirm Professional Edition licenses may use - * this file in accordance with the libFirm Commercial License. - * Agreement provided with the Software. - * - * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE - * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR - * PURPOSE. + * Copyright (C) 2012 University of Karlsruhe. */ /** @@ -22,7 +8,6 @@ * @brief This file implements the mapping of 64Bit intrinsic * functions to code or library calls. * @author Michael Beck - * @version $Id$ */ #include "config.h" @@ -32,9 +17,11 @@ #include "irnode_t.h" #include "ircons.h" #include "irprog_t.h" -#include "lowering.h" +#include "iroptimize.h" +#include "lower_dw.h" #include "array.h" #include "error.h" +#include "util.h" #include "ia32_new_nodes.h" #include "bearch_ia32_t.h" @@ -67,15 +54,13 @@ void ia32_handle_intrinsics(void) /** * Reroute edges from the pn_Call_T_result proj of a call. * - * @param proj the pn_Call_T_result Proj - * @param l_res the lower 32 bit result - * @param h_res the upper 32 bit result or NULL + * @param resproj the pn_Call_T_result Proj + * @param l_res the lower 32 bit result + * @param h_res the upper 32 bit result or NULL */ -static void reroute_result(ir_node *proj, ir_node *l_res, ir_node *h_res) +static void reroute_result(ir_node *resproj, ir_node *l_res, ir_node *h_res) { - const ir_edge_t *edge, *next; - - foreach_out_edge_safe(proj, edge, next) { + foreach_out_edge_safe(resproj, edge) { ir_node *proj = get_edge_src_irn(edge); long pn = get_Proj_proj(proj); @@ -101,15 +86,12 @@ static void reroute_result(ir_node *proj, ir_node *l_res, ir_node *h_res) static void resolve_call(ir_node *call, ir_node *l_res, ir_node *h_res, ir_graph *irg, ir_node *block) { ir_node *jmp, *res, *in[2]; - ir_node *bad = get_irg_bad(irg); ir_node *nomem = get_irg_no_mem(irg); int old_cse; if (edges_activated(irg)) { /* use rerouting to prevent some warning in the backend */ - const ir_edge_t *edge, *next; - - foreach_out_edge_safe(call, edge, next) { + foreach_out_edge_safe(call, edge) { ir_node *proj = get_edge_src_irn(edge); pn_Call pn = (pn_Call)get_Proj_proj(proj); @@ -130,7 +112,7 @@ static void resolve_call(ir_node *call, ir_node *l_res, ir_node *h_res, ir_graph case pn_Call_X_except: /* should not happen here */ - edges_reroute(proj, bad); + edges_reroute(proj, new_r_Bad(irg, mode_X)); break; case pn_Call_M: /* should not happen here */ @@ -155,7 +137,6 @@ static void resolve_call(ir_node *call, ir_node *l_res, ir_node *h_res, ir_graph res = new_r_Tuple(block, 2, in); } - turn_into_tuple(call, pn_Call_max); /* * Beware: * We do not check here if this call really has exception and regular Proj's. @@ -168,10 +149,13 @@ static void resolve_call(ir_node *call, ir_node *l_res, ir_node *h_res, ir_graph jmp = new_r_Jmp(block); set_opt_cse(old_cse); - set_Tuple_pred(call, pn_Call_M, nomem); - set_Tuple_pred(call, pn_Call_X_regular, jmp); - set_Tuple_pred(call, pn_Call_X_except, bad); - set_Tuple_pred(call, pn_Call_T_result, res); + ir_node *const in[] = { + [pn_Call_M] = nomem, + [pn_Call_T_result] = res, + [pn_Call_X_regular] = jmp, + [pn_Call_X_except] = new_r_Bad(irg, mode_X), + }; + turn_into_tuple(call, ARRAY_SIZE(in), in); } } @@ -243,275 +227,6 @@ static int map_Sub(ir_node *call, void *ctx) return 1; } -/** - * Map a Shl (a_l, a_h, count) - */ -static int map_Shl(ir_node *call, void *ctx) -{ - ir_graph *irg = current_ir_graph; - dbg_info *dbg = get_irn_dbg_info(call); - ir_node *block = get_nodes_block(call); - ir_node **params = get_Call_param_arr(call); - ir_type *method = get_Call_type(call); - ir_node *a_l = params[BINOP_Left_Low]; - ir_node *a_h = params[BINOP_Left_High]; - ir_node *cnt = params[BINOP_Right_Low]; - ir_mode *l_mode = get_type_mode(get_method_res_type(method, 0)); - ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1)); - ir_mode *c_mode; - ir_node *l_res, *h_res, *irn, *cond, *upper, *n_block, *l1, *l2, *h1, *h2, *in[2]; - (void) ctx; - - if (is_Const(cnt)) { - /* the shift count is a const, create better code */ - ir_tarval *tv = get_Const_tarval(cnt); - - if (tarval_cmp(tv, new_tarval_from_long(32, l_mode)) - & (ir_relation_greater_equal)) { - /* simplest case: shift only the lower bits. Note that there is no - need to reduce the constant here, this is done by the hardware. */ - ir_node *conv = new_rd_Conv(dbg, block, a_l, h_mode); - h_res = new_rd_Shl(dbg, block, conv, cnt, h_mode); - l_res = new_rd_Const(dbg, irg, get_mode_null(l_mode)); - - } else { - /* h_res = SHLD a_h, a_l, cnt */ - h_res = new_bd_ia32_l_ShlD(dbg, block, a_h, a_l, cnt, h_mode); - - /* l_res = SHL a_l, cnt */ - l_res = new_bd_ia32_l_ShlDep(dbg, block, a_l, cnt, h_res, l_mode); - } - - resolve_call(call, l_res, h_res, irg, block); - return 1; - } - - part_block(call); - upper = get_nodes_block(call); - - /* h_res = SHLD a_h, a_l, cnt */ - h1 = new_bd_ia32_l_ShlD(dbg, upper, a_h, a_l, cnt, h_mode); - - /* l_res = SHL a_l, cnt */ - l1 = new_bd_ia32_l_ShlDep(dbg, upper, a_l, cnt, h1, l_mode); - - c_mode = get_irn_mode(cnt); - irn = new_r_Const_long(irg, c_mode, 32); - irn = new_rd_And(dbg, upper, cnt, irn, c_mode); - irn = new_rd_Cmp(dbg, upper, irn, new_r_Const(irg, get_mode_null(c_mode)), ir_relation_equal); - cond = new_rd_Cond(dbg, upper, irn); - - in[0] = new_r_Proj(cond, mode_X, pn_Cond_true); - in[1] = new_r_Proj(cond, mode_X, pn_Cond_false); - - /* the block for cnt >= 32 */ - n_block = new_rd_Block(dbg, irg, 1, &in[1]); - h2 = new_rd_Conv(dbg, n_block, l1, h_mode); - l2 = new_r_Const(irg, get_mode_null(l_mode)); - in[1] = new_r_Jmp(n_block); - - set_irn_in(block, 2, in); - - in[0] = l1; - in[1] = l2; - l_res = new_r_Phi(block, 2, in, l_mode); - set_Block_phis(block, l_res); - - in[0] = h1; - in[1] = h2; - h_res = new_r_Phi(block, 2, in, h_mode); - set_Phi_next(l_res, h_res); - set_Phi_next(h_res, NULL); - - /* move it down */ - set_nodes_block(call, block); - for (irn = (ir_node*)get_irn_link(call); irn != NULL; - irn = (ir_node*)get_irn_link(irn)) { - set_nodes_block(irn, block); - } - - resolve_call(call, l_res, h_res, irg, block); - return 1; -} - -/** - * Map a Shr (a_l, a_h, count) - */ -static int map_Shr(ir_node *call, void *ctx) -{ - ir_graph *irg = current_ir_graph; - dbg_info *dbg = get_irn_dbg_info(call); - ir_node *block = get_nodes_block(call); - ir_node **params = get_Call_param_arr(call); - ir_type *method = get_Call_type(call); - ir_node *a_l = params[BINOP_Left_Low]; - ir_node *a_h = params[BINOP_Left_High]; - ir_node *cnt = params[BINOP_Right_Low]; - ir_mode *l_mode = get_type_mode(get_method_res_type(method, 0)); - ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1)); - ir_mode *c_mode; - ir_node *l_res, *h_res, *irn, *cond, *upper, *n_block, *l1, *l2, *h1, *h2, *in[2]; - (void) ctx; - - if (is_Const(cnt)) { - /* the shift count is a const, create better code */ - ir_tarval *tv = get_Const_tarval(cnt); - - if (tarval_cmp(tv, new_tarval_from_long(32, l_mode)) & (ir_relation_greater_equal)) { - /* simplest case: shift only the higher bits. Note that there is no - need to reduce the constant here, this is done by the hardware. */ - ir_node *conv = new_rd_Conv(dbg, block, a_h, l_mode); - h_res = new_rd_Const(dbg, irg, get_mode_null(h_mode)); - l_res = new_rd_Shr(dbg, block, conv, cnt, l_mode); - } else { - /* l_res = SHRD a_h:a_l, cnt */ - l_res = new_bd_ia32_l_ShrD(dbg, block, a_l, a_h, cnt, l_mode); - - /* h_res = SHR a_h, cnt */ - h_res = new_bd_ia32_l_ShrDep(dbg, block, a_h, cnt, l_res, h_mode); - } - resolve_call(call, l_res, h_res, irg, block); - return 1; - } - - part_block(call); - upper = get_nodes_block(call); - - /* l_res = SHRD a_h:a_l, cnt */ - l1 = new_bd_ia32_l_ShrD(dbg, upper, a_l, a_h, cnt, l_mode); - - /* h_res = SHR a_h, cnt */ - h1 = new_bd_ia32_l_ShrDep(dbg, upper, a_h, cnt, l1, h_mode); - - c_mode = get_irn_mode(cnt); - irn = new_r_Const_long(irg, c_mode, 32); - irn = new_rd_And(dbg, upper, cnt, irn, c_mode); - irn = new_rd_Cmp(dbg, upper, irn, new_r_Const(irg, get_mode_null(c_mode)), ir_relation_equal); - cond = new_rd_Cond(dbg, upper, irn); - - in[0] = new_r_Proj(cond, mode_X, pn_Cond_true); - in[1] = new_r_Proj(cond, mode_X, pn_Cond_false); - - /* the block for cnt >= 32 */ - n_block = new_rd_Block(dbg, irg, 1, &in[1]); - l2 = new_rd_Conv(dbg, n_block, h1, l_mode); - h2 = new_r_Const(irg, get_mode_null(h_mode)); - in[1] = new_r_Jmp(n_block); - - set_irn_in(block, 2, in); - - in[0] = l1; - in[1] = l2; - l_res = new_r_Phi(block, 2, in, l_mode); - set_Block_phis(block, l_res); - - in[0] = h1; - in[1] = h2; - h_res = new_r_Phi(block, 2, in, h_mode); - set_Phi_next(l_res, h_res); - set_Phi_next(h_res, NULL); - - /* move it down */ - set_nodes_block(call, block); - for (irn = (ir_node*)get_irn_link(call); irn != NULL; - irn = (ir_node*)get_irn_link(irn)) { - set_nodes_block(irn, block); - } - - resolve_call(call, l_res, h_res, irg, block); - return 1; -} - -/** - * Map a Shrs (a_l, a_h, count) - */ -static int map_Shrs(ir_node *call, void *ctx) -{ - ir_graph *irg = current_ir_graph; - dbg_info *dbg = get_irn_dbg_info(call); - ir_node *block = get_nodes_block(call); - ir_node **params = get_Call_param_arr(call); - ir_type *method = get_Call_type(call); - ir_node *a_l = params[BINOP_Left_Low]; - ir_node *a_h = params[BINOP_Left_High]; - ir_node *cnt = params[BINOP_Right_Low]; - ir_mode *l_mode = get_type_mode(get_method_res_type(method, 0)); - ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1)); - ir_mode *c_mode; - ir_node *l_res, *h_res, *irn, *cond, *upper, *n_block, *l1, *l2, *h1, *h2, *in[2]; - (void) ctx; - - if (is_Const(cnt)) { - /* the shift count is a const, create better code */ - ir_tarval *tv = get_Const_tarval(cnt); - - if (tarval_cmp(tv, new_tarval_from_long(32, l_mode)) & (ir_relation_greater_equal)) { - /* simplest case: shift only the higher bits. Note that there is no - need to reduce the constant here, this is done by the hardware. */ - ir_node *conv = new_rd_Conv(dbg, block, a_h, l_mode); - ir_mode *c_mode = get_irn_mode(cnt); - - h_res = new_rd_Shrs(dbg, block, a_h, new_r_Const_long(irg, c_mode, 31), h_mode); - l_res = new_rd_Shrs(dbg, block, conv, cnt, l_mode); - } else { - /* l_res = SHRD a_h:a_l, cnt */ - l_res = new_bd_ia32_l_ShrD(dbg, block, a_l, a_h, cnt, l_mode); - - /* h_res = SAR a_h, cnt */ - h_res = new_bd_ia32_l_SarDep(dbg, block, a_h, cnt, l_res, h_mode); - } - resolve_call(call, l_res, h_res, irg, block); - return 1; - } - - part_block(call); - upper = get_nodes_block(call); - - /* l_res = SHRD a_h:a_l, cnt */ - l1 = new_bd_ia32_l_ShrD(dbg, upper, a_l, a_h, cnt, l_mode); - - /* h_res = SAR a_h, cnt */ - h1 = new_bd_ia32_l_SarDep(dbg, upper, a_h, cnt, l1, h_mode); - - c_mode = get_irn_mode(cnt); - irn = new_r_Const_long(irg, c_mode, 32); - irn = new_rd_And(dbg, upper, cnt, irn, c_mode); - irn = new_rd_Cmp(dbg, upper, irn, new_r_Const(irg, get_mode_null(c_mode)), ir_relation_equal); - cond = new_rd_Cond(dbg, upper, irn); - - in[0] = new_r_Proj(cond, mode_X, pn_Cond_true); - in[1] = new_r_Proj(cond, mode_X, pn_Cond_false); - - /* the block for cnt >= 32 */ - n_block = new_rd_Block(dbg, irg, 1, &in[1]); - l2 = new_rd_Conv(dbg, n_block, h1, l_mode); - h2 = new_rd_Shrs(dbg, n_block, a_h, new_r_Const_long(irg, c_mode, 31), h_mode); - in[1] = new_r_Jmp(n_block); - - set_irn_in(block, 2, in); - - in[0] = l1; - in[1] = l2; - l_res = new_r_Phi(block, 2, in, l_mode); - set_Block_phis(block, l_res); - - in[0] = h1; - in[1] = h2; - h_res = new_r_Phi(block, 2, in, h_mode); - set_Phi_next(l_res, h_res); - set_Phi_next(h_res, NULL); - - /* move it down */ - set_nodes_block(call, block); - for (irn = (ir_node*)get_irn_link(call); irn != NULL; - irn = (ir_node*)get_irn_link(irn)) { - set_nodes_block(irn, block); - } - - resolve_call(call, l_res, h_res, irg, block); - return 1; -} - /** * Checks where node high is a sign extension of low. */ @@ -624,81 +339,8 @@ static int map_Minus(ir_node *call, void *ctx) return 1; } -#if 0 -/** - * Map a Abs (a_l, a_h) - */ -static int map_Abs(ir_node *call, void *ctx) -{ - dbg_info *dbg = get_irn_dbg_info(call); - ir_node *block = get_nodes_block(call); - ir_node **params = get_Call_param_arr(call); - ir_type *method = get_Call_type(call); - ir_node *a_l = params[BINOP_Left_Low]; - ir_node *a_h = params[BINOP_Left_High]; - ir_mode *l_mode = get_type_mode(get_method_res_type(method, 0)); - ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1)); - ir_mode *mode_flags = ia32_reg_classes[CLASS_ia32_flags].mode; - ir_node *l_res, *h_res, *sign, *sub_l, *sub_h; - ir_node *sign_l; - ir_node *l_sub; - ir_node *flags; - (void) ctx; - - /* - Code inspired by gcc output :) (although gcc doubles the - operation for t1 as t2 and uses t1 for operations with low part - and t2 for operations with high part which is actually unnecessary - because t1 and t2 represent the same value) - - t1 = SHRS a_h, 31 - t2 = a_l ^ t1 - t3 = a_h ^ t1 - l_res = t2 - t1 - h_res = t3 - t1 - carry - - */ - - /* TODO: give a hint to the backend somehow to not create a cltd here... */ - sign = new_rd_Shrs(dbg, block, a_h, new_r_Const_long(irg, l_mode, 31), h_mode); - sign_l = new_rd_Conv(dbg, block, sign, l_mode); - sub_l = new_rd_Eor(dbg, block, a_l, sign_l, l_mode); - sub_h = new_rd_Eor(dbg, block, a_h, sign, h_mode); - - l_sub = new_bd_ia32_l_Sub(dbg, block, sub_l, sign_l, mode_T); - l_res = new_r_Proj(l_sub, l_mode, pn_ia32_res); - flags = new_r_Proj(l_sub, mode_flags, pn_ia32_flags); - h_res = new_bd_ia32_l_Sbb(dbg, block, sub_h, sign, flags, h_mode); - - resolve_call(call, l_res, h_res, current_ir_graph, block); - - return 1; -} -#endif - #define ID(x) new_id_from_chars(x, sizeof(x)-1) -static ir_entity *create_compiler_lib_entity(const char *name, ir_type *type) -{ - ir_type *glob = get_glob_type(); - ident *id = new_id_from_str(name); - ir_entity *entity; - - /* Hack: we need to know the type of runtime library we use. Strictly - speaking it's not the same as the object-file-format. But in practice - the following should be enough */ - if (be_gas_object_file_format == OBJECT_FILE_FORMAT_MACH_O - || be_gas_object_file_format == OBJECT_FILE_FORMAT_COFF) { - id = id_mangle3("___", id, ""); - } else { - id = id_mangle3("__", id, ""); - } - entity = new_entity(glob, id, type); - set_entity_visibility(entity, ir_visibility_local); - set_entity_ld_ident(entity, id); - return entity; -} - /** * Maps a Div. Change into a library call. */ @@ -716,14 +358,17 @@ static int map_Div(ir_node *call, void *ctx) /* 64bit signed Division */ ent = env->divdi3; if (ent == NULL) { - ent = env->divdi3 = create_compiler_lib_entity("divdi3", method); + /* create library entity */ + ident *id = ID("__divdi3"); + ent = env->divdi3 = create_compilerlib_entity(id, method); } } else { /* 64bit unsigned Division */ ent = env->udivdi3; if (ent == NULL) { /* create library entity */ - ent = env->udivdi3 = create_compiler_lib_entity("udivdi3", method); + ident *id = ID("__udivdi3"); + ent = env->udivdi3 = create_compilerlib_entity(id, method); } } @@ -753,14 +398,16 @@ static int map_Mod(ir_node *call, void *ctx) ent = env->moddi3; if (ent == NULL) { /* create library entity */ - ent = env->moddi3 = create_compiler_lib_entity("moddi3", method); + ident *id = ID("__moddi3"); + ent = env->moddi3 = create_compilerlib_entity(id, method); } } else { /* 64bit signed Modulo */ ent = env->umoddi3; if (ent == NULL) { /* create library entity */ - ent = env->umoddi3 = create_compiler_lib_entity("umoddi3", method); + ident *id = ID("__umoddi3"); + ent = env->umoddi3 = create_compilerlib_entity(id, method); } } @@ -805,9 +452,8 @@ static int map_Conv(ir_node *call, void *ctx) h_res = new_r_Proj(float_to_ll, h_res_mode, pn_ia32_l_FloattoLL_res_high); } else { - /* convert from float to signed 64bit */ - ir_mode *flt_mode = get_irn_mode(a_f); - ir_tarval *flt_tv = new_tarval_from_str("9223372036854775808", 19, flt_mode); + /* Convert from float to unsigned 64bit. */ + ir_tarval *flt_tv = new_tarval_from_str("9223372036854775808", 19, ia32_mode_E); ir_node *flt_corr = new_r_Const(irg, flt_tv); ir_node *lower_blk = block; ir_node *upper_blk; @@ -817,6 +463,7 @@ static int map_Conv(ir_node *call, void *ctx) part_block(call); upper_blk = get_nodes_block(call); + a_f = new_rd_Conv(dbg, upper_blk, a_f, ia32_mode_E); cmp = new_rd_Cmp(dbg, upper_blk, a_f, flt_corr, ir_relation_less); cond = new_rd_Cond(dbg, upper_blk, cmp); in[0] = new_r_Proj(cond, mode_X, pn_Cond_true); @@ -833,14 +480,15 @@ static int map_Conv(ir_node *call, void *ctx) int_phi = new_r_Phi(lower_blk, 2, in, h_res_mode); in[0] = a_f; - in[1] = new_rd_Sub(dbg, upper_blk, a_f, flt_corr, flt_mode); + in[1] = new_rd_Sub(dbg, upper_blk, a_f, flt_corr, ia32_mode_E); - flt_phi = new_r_Phi(lower_blk, 2, in, flt_mode); + flt_phi = new_r_Phi(lower_blk, 2, in, ia32_mode_E); /* fix Phi links for next part_block() */ - set_Block_phis(lower_blk, int_phi); - set_Phi_next(int_phi, flt_phi); - set_Phi_next(flt_phi, NULL); + if (is_Phi(int_phi)) + add_Block_phi(lower_blk, int_phi); + if (is_Phi(flt_phi)) + add_Block_phi(lower_blk, flt_phi); float_to_ll = new_bd_ia32_l_FloattoLL(dbg, lower_blk, flt_phi); @@ -905,18 +553,6 @@ ir_entity *ia32_create_intrinsic_fkt(ir_type *method, const ir_op *op, ent = &i_ents[iro_Sub]; mapper = map_Sub; break; - case iro_Shl: - ent = &i_ents[iro_Shl]; - mapper = map_Shl; - break; - case iro_Shr: - ent = &i_ents[iro_Shr]; - mapper = map_Shr; - break; - case iro_Shrs: - ent = &i_ents[iro_Shrs]; - mapper = map_Shrs; - break; case iro_Mul: ent = &i_ents[iro_Mul]; mapper = map_Mul; @@ -943,9 +579,7 @@ ir_entity *ia32_create_intrinsic_fkt(ir_type *method, const ir_op *op, } if (ent && ! *ent) { -#define IDENT(s) new_id_from_chars(s, sizeof(s)-1) - - ident *id = id_mangle(IDENT("L"), get_op_ident(op)); + ident *id = id_mangle(ID("L"), get_op_ident(op)); *ent = new_entity(get_glob_type(), id, method); set_entity_visibility(*ent, ir_visibility_private); }