X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_intrinsics.c;h=6778f966fbf129f1ca70fd7d3b0ba3580635de3a;hb=e9788b1182b08e303f11c0df45a8a2f72955e520;hp=c55f21a3e2b6b41330f11ac32eef6f690ffe4f1f;hpb=80b16ed5806282f5030754ce55509bceea1f953b;p=libfirm diff --git a/ir/be/ia32/ia32_intrinsics.c b/ir/be/ia32/ia32_intrinsics.c index c55f21a3e..6778f966f 100644 --- a/ir/be/ia32/ia32_intrinsics.c +++ b/ir/be/ia32/ia32_intrinsics.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. * @@ -24,10 +24,9 @@ * @author Michael Beck * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif +#include "iredges.h" #include "irgmod.h" #include "irop.h" #include "irnode_t.h" @@ -35,6 +34,7 @@ #include "irprog_t.h" #include "lowering.h" #include "array.h" +#include "error.h" #include "ia32_new_nodes.h" #include "bearch_ia32_t.h" @@ -43,15 +43,16 @@ /** The array of all intrinsics that must be mapped. */ static i_record *intrinsics; -/** An array to cache all entities */ -static ir_entity *i_ents[iro_MaxOpcode]; +/** An array to cache all entities. */ +static ir_entity *i_ents[iro_Last + 1]; /* * Maps all intrinsic calls that the backend support * and map all instructions the backend did not support * to runtime calls. */ -void ia32_handle_intrinsics(void) { +void ia32_handle_intrinsics(void) +{ if (intrinsics && ARR_LEN(intrinsics) > 0) { lower_intrinsics(intrinsics, ARR_LEN(intrinsics), /*part_block_used=*/1); } @@ -62,34 +63,125 @@ void ia32_handle_intrinsics(void) { #define BINOP_Right_Low 2 #define BINOP_Right_High 3 +/** + * 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 irg the graph to replace on + */ +static void reroute_result(ir_node *proj, ir_node *l_res, ir_node *h_res, ir_graph *irg) +{ + const ir_edge_t *edge, *next; + + foreach_out_edge_safe(proj, edge, next) { + ir_node *proj = get_edge_src_irn(edge); + long pn = get_Proj_proj(proj); + + if (pn == 0) { + edges_reroute(proj, l_res, irg); + } else if (pn == 1 && h_res != NULL) { + edges_reroute(proj, h_res, irg); + } else { + panic("Unsupported Result-Proj from Call found"); + } + } +} + /** * Replace a call be a tuple of l_res, h_res. + * + * @param call the call node to replace + * @param l_res the lower 32 bit result + * @param h_res the upper 32 bit result or NULL + * @param irg the graph to replace on + * @param block the block to replace on (always the call block) */ -static void resolve_call(ir_node *call, ir_node *l_res, ir_node *h_res, ir_graph *irg, ir_node *block) { - ir_node *res, *in[2]; - - in[0] = l_res; - in[1] = h_res; - res = new_r_Tuple(irg, block, h_res == NULL ? 1 : 2, in); - - turn_into_tuple(call, pn_Call_max); - set_Tuple_pred(call, pn_Call_M_regular, get_irg_no_mem(irg)); - /* Matze: the new_r_Jmp here sometimes CSEs and then bad things happen - * (in movgen.c from 186.crafty for example) I don't know why it is here - * and if this fix is correct... */ - /*set_Tuple_pred(call, pn_Call_X_regular, new_r_Jmp(irg, block));*/ - set_Tuple_pred(call, pn_Call_X_regular, get_irg_bad(irg)); - set_Tuple_pred(call, pn_Call_X_except, get_irg_bad(irg)); - set_Tuple_pred(call, pn_Call_T_result, res); - set_Tuple_pred(call, pn_Call_M_except, get_irg_no_mem(irg)); - set_Tuple_pred(call, pn_Call_P_value_res_base, get_irg_bad(irg)); +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) { + ir_node *proj = get_edge_src_irn(edge); + pn_Call pn = get_Proj_proj(proj); + + switch (pn) { + case pn_Call_X_regular: + /* Beware: + * We do not check here if this call really has exception and regular Proj's. + * new_r_Jmp might than be CSEd with the real exit jmp and then bad things happen + * (in movgen.c from 186.crafty for example). + * So be sure the newly created Jmp cannot CSE. + */ + old_cse = get_opt_cse(); + set_opt_cse(0); + jmp = new_r_Jmp(block); + set_opt_cse(old_cse); + edges_reroute(proj, jmp, irg); + break; + + case pn_Call_X_except: + case pn_Call_P_value_res_base: + /* should not happen here */ + edges_reroute(proj, bad, irg); + break; + case pn_Call_M: + /* should not happen here */ + edges_reroute(proj, nomem, irg); + break; + case pn_Call_T_result: + reroute_result(proj, l_res, h_res, irg); + break; + default: + panic("Wrong Proj from Call"); + } + kill_node(proj); + } + kill_node(call); + } else { + /* no edges, build Tuple */ + if (h_res == NULL) + res = l_res; + else { + in[0] = l_res; + in[1] = h_res; + 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. + * new_r_Jmp might than be CSEd with the real exit jmp and then bad things happen + * (in movgen.c from 186.crafty for example). + * So be sure the newly created Jmp cannot CSE. + */ + old_cse = get_opt_cse(); + set_opt_cse(0); + 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); + set_Tuple_pred(call, pn_Call_P_value_res_base, bad); + } } /** * Map an Add (a_l, a_h, b_l, b_h) */ -static int map_Add(ir_node *call, void *ctx) { - ir_graph *irg = current_ir_graph; +static int map_Add(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); @@ -108,14 +200,14 @@ static int map_Add(ir_node *call, void *ctx) { /* l_res = a_l + b_l */ /* h_res = a_h + b_h + carry */ - add_low = new_rd_ia32_l_Add(dbg, irg, block, a_l, b_l, mode_T); - flags = new_r_Proj(irg, block, add_low, mode_flags, pn_ia32_flags); - add_high = new_rd_ia32_l_Adc(dbg, irg, block, a_h, b_h, flags, h_mode); + add_low = new_bd_ia32_l_Add(dbg, block, a_l, b_l, mode_T); + flags = new_r_Proj(block, add_low, mode_flags, pn_ia32_flags); + add_high = new_bd_ia32_l_Adc(dbg, block, a_h, b_h, flags, h_mode); - l_res = new_r_Proj(irg, block, add_low, l_mode, pn_ia32_res); + l_res = new_r_Proj(block, add_low, l_mode, pn_ia32_res); h_res = add_high; - resolve_call(call, l_res, h_res, irg, block); + resolve_call(call, l_res, h_res, current_ir_graph, block); return 1; } @@ -124,7 +216,6 @@ static int map_Add(ir_node *call, void *ctx) { */ static int map_Sub(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); @@ -143,21 +234,22 @@ static int map_Sub(ir_node *call, void *ctx) /* l_res = a_l - b_l */ /* h_res = a_h - b_h - carry */ - sub_low = new_rd_ia32_l_Sub(dbg, irg, block, a_l, b_l, mode_T); - flags = new_r_Proj(irg, block, sub_low, mode_flags, pn_ia32_flags); - sub_high = new_rd_ia32_l_Sbb(dbg, irg, block, a_h, b_h, flags, h_mode); + sub_low = new_bd_ia32_l_Sub(dbg, block, a_l, b_l, mode_T); + flags = new_r_Proj(block, sub_low, mode_flags, pn_ia32_flags); + sub_high = new_bd_ia32_l_Sbb(dbg, block, a_h, b_h, flags, h_mode); - l_res = new_r_Proj(irg, block, sub_low, l_mode, pn_ia32_res); + l_res = new_r_Proj( block, sub_low, l_mode, pn_ia32_res); h_res = sub_high; - resolve_call(call, l_res, h_res, irg, block); + resolve_call(call, l_res, h_res, current_ir_graph, block); return 1; } /** * Map a Shl (a_l, a_h, count) */ -static int map_Shl(ir_node *call, void *ctx) { +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); @@ -179,16 +271,16 @@ static int map_Shl(ir_node *call, void *ctx) { if (tarval_cmp(tv, new_tarval_from_long(32, l_mode)) & (pn_Cmp_Gt|pn_Cmp_Eq)) { /* 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, irg, block, a_l, h_mode); - h_res = new_rd_Shl(dbg, irg, block, conv, cnt, h_mode); - l_res = new_rd_Const(dbg, irg, block, l_mode, get_mode_null(l_mode)); + 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_rd_ia32_l_ShlD(dbg, irg, block, a_h, a_l, cnt, h_mode); + 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_rd_ia32_l_ShlDep(dbg, irg, block, a_l, cnt, h_res, l_mode); + 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); @@ -199,39 +291,39 @@ static int map_Shl(ir_node *call, void *ctx) { upper = get_nodes_block(call); /* h_res = SHLD a_h, a_l, cnt */ - h1 = new_rd_ia32_l_ShlD(dbg, irg, upper, a_h, a_l, cnt, h_mode); + h1 = new_bd_ia32_l_ShlD(dbg, upper, a_h, a_l, cnt, h_mode); /* l_res = SHL a_l, cnt */ - l1 = new_rd_ia32_l_ShlDep(dbg, irg, upper, a_l, cnt, h1, l_mode); + 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, upper, c_mode, 32); - irn = new_rd_And(dbg, irg, upper, cnt, irn, c_mode); - irn = new_rd_Cmp(dbg, irg, upper, irn, new_r_Const(irg, upper, c_mode, get_mode_null(c_mode))); - irn = new_r_Proj(irg, upper, irn, mode_b, pn_Cmp_Eq); - cond = new_rd_Cond(dbg, irg, upper, irn); + 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))); + irn = new_r_Proj(upper, irn, mode_b, pn_Cmp_Eq); + cond = new_rd_Cond(dbg, upper, irn); - in[0] = new_r_Proj(irg, upper, cond, mode_X, pn_Cond_true); - in[1] = new_r_Proj(irg, upper, cond, mode_X, pn_Cond_false); + in[0] = new_r_Proj(upper, cond, mode_X, pn_Cond_true); + in[1] = new_r_Proj(upper, 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, irg, n_block, l1, h_mode); - l2 = new_r_Const(irg, n_block, l_mode, get_mode_null(l_mode)); - in[1] = new_r_Jmp(irg, n_block); + 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(irg, block, 2, in, l_mode); - set_irn_link(block, l_res); + 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(irg, block, 2, in, h_mode); - set_irn_link(l_res, h_res); - set_irn_link(h_res, NULL); + 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); @@ -245,7 +337,8 @@ static int map_Shl(ir_node *call, void *ctx) { /** * Map a Shr (a_l, a_h, count) */ -static int map_Shr(ir_node *call, void *ctx) { +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); @@ -267,15 +360,15 @@ static int map_Shr(ir_node *call, void *ctx) { if (tarval_cmp(tv, new_tarval_from_long(32, l_mode)) & (pn_Cmp_Gt|pn_Cmp_Eq)) { /* 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, irg, block, a_h, l_mode); - h_res = new_rd_Const(dbg, irg, block, h_mode, get_mode_null(h_mode)); - l_res = new_rd_Shr(dbg, irg, block, conv, cnt, l_mode); + 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_rd_ia32_l_ShrD(dbg, irg, block, a_l, a_h, cnt, l_mode); + 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_rd_ia32_l_ShrDep(dbg, irg, block, a_h, cnt, l_res, h_mode); + 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; @@ -285,39 +378,39 @@ static int map_Shr(ir_node *call, void *ctx) { upper = get_nodes_block(call); /* l_res = SHRD a_h:a_l, cnt */ - l1 = new_rd_ia32_l_ShrD(dbg, irg, upper, a_l, a_h, cnt, l_mode); + l1 = new_bd_ia32_l_ShrD(dbg, upper, a_l, a_h, cnt, l_mode); /* h_res = SHR a_h, cnt */ - h1 = new_rd_ia32_l_ShrDep(dbg, irg, upper, a_h, cnt, l1, h_mode); + 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, upper, c_mode, 32); - irn = new_rd_And(dbg, irg, upper, cnt, irn, c_mode); - irn = new_rd_Cmp(dbg, irg, upper, irn, new_r_Const(irg, upper, c_mode, get_mode_null(c_mode))); - irn = new_r_Proj(irg, upper, irn, mode_b, pn_Cmp_Eq); - cond = new_rd_Cond(dbg, irg, upper, irn); + 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))); + irn = new_r_Proj(upper, irn, mode_b, pn_Cmp_Eq); + cond = new_rd_Cond(dbg, upper, irn); - in[0] = new_r_Proj(irg, upper, cond, mode_X, pn_Cond_true); - in[1] = new_r_Proj(irg, upper, cond, mode_X, pn_Cond_false); + in[0] = new_r_Proj(upper, cond, mode_X, pn_Cond_true); + in[1] = new_r_Proj(upper, 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, irg, n_block, h1, l_mode); - h2 = new_r_Const(irg, n_block, l_mode, get_mode_null(h_mode)); - in[1] = new_r_Jmp(irg, n_block); + 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(irg, block, 2, in, l_mode); - set_irn_link(block, l_res); + 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(irg, block, 2, in, h_mode); - set_irn_link(l_res, h_res); - set_irn_link(h_res, NULL); + 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); @@ -331,7 +424,8 @@ static int map_Shr(ir_node *call, void *ctx) { /** * Map a Shrs (a_l, a_h, count) */ -static int map_Shrs(ir_node *call, void *ctx) { +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); @@ -353,17 +447,17 @@ static int map_Shrs(ir_node *call, void *ctx) { if (tarval_cmp(tv, new_tarval_from_long(32, l_mode)) & (pn_Cmp_Gt|pn_Cmp_Eq)) { /* 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, irg, block, a_h, l_mode); + 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, irg, block, a_h, new_r_Const_long(irg, block, c_mode, 31), h_mode); - l_res = new_rd_Shrs(dbg, irg, block, conv, cnt, l_mode); + 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_rd_ia32_l_ShrD(dbg, irg, block, a_l, a_h, cnt, l_mode); + 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_rd_ia32_l_SarDep(dbg, irg, block, a_h, cnt, l_res, h_mode); + 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; @@ -373,39 +467,39 @@ static int map_Shrs(ir_node *call, void *ctx) { upper = get_nodes_block(call); /* l_res = SHRD a_h:a_l, cnt */ - l1 = new_rd_ia32_l_ShrD(dbg, irg, upper, a_l, a_h, cnt, l_mode); + l1 = new_bd_ia32_l_ShrD(dbg, upper, a_l, a_h, cnt, l_mode); /* h_res = SAR a_h, cnt */ - h1 = new_rd_ia32_l_SarDep(dbg, irg, upper, a_h, cnt, l1, h_mode); + 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, upper, c_mode, 32); - irn = new_rd_And(dbg, irg, upper, cnt, irn, c_mode); - irn = new_rd_Cmp(dbg, irg, upper, irn, new_r_Const(irg, upper, c_mode, get_mode_null(c_mode))); - irn = new_r_Proj(irg, upper, irn, mode_b, pn_Cmp_Eq); - cond = new_rd_Cond(dbg, irg, upper, irn); + 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))); + irn = new_r_Proj(upper, irn, mode_b, pn_Cmp_Eq); + cond = new_rd_Cond(dbg, upper, irn); - in[0] = new_r_Proj(irg, upper, cond, mode_X, pn_Cond_true); - in[1] = new_r_Proj(irg, upper, cond, mode_X, pn_Cond_false); + in[0] = new_r_Proj(upper, cond, mode_X, pn_Cond_true); + in[1] = new_r_Proj(upper, 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, irg, n_block, h1, l_mode); - h2 = new_rd_Shrs(dbg, irg, n_block, a_h, new_r_Const_long(irg, block, c_mode, 31), h_mode); - in[1] = new_r_Jmp(irg, n_block); + 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(irg, block, 2, in, l_mode); - set_irn_link(block, l_res); + 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(irg, block, 2, in, h_mode); - set_irn_link(l_res, h_res); - set_irn_link(h_res, NULL); + 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); @@ -416,11 +510,47 @@ static int map_Shrs(ir_node *call, void *ctx) { return 1; } +/** + * Checks where node high is a sign extension of low. + */ +static int is_sign_extend(ir_node *low, ir_node *high) +{ + if (is_Shrs(high)) { + ir_node *high_l; + ir_node *high_r; + tarval *shift_count; + + high_r = get_Shrs_right(high); + if (!is_Const(high_r)) return 0; + + shift_count = get_Const_tarval(high_r); + if (!tarval_is_long(shift_count)) return 0; + if (get_tarval_long(shift_count) != 31) return 0; + + high_l = get_Shrs_left(high); + + if (is_Conv(low) && get_Conv_op(low) == high_l) return 1; + if (is_Conv(high_l) && get_Conv_op(high_l) == low) return 1; + } else if (is_Const(low) && is_Const(high)) { + tarval *tl = get_Const_tarval(low); + tarval *th = get_Const_tarval(high); + + if (tarval_is_long(th) && tarval_is_long(tl)) { + long l = get_tarval_long(tl); + long h = get_tarval_long(th); + + return (h == 0 && l >= 0) || (h == -1 && l < 0); + } + } + + return 0; +} + /** * Map a Mul (a_l, a_h, b_l, b_h) */ -static int map_Mul(ir_node *call, void *ctx) { - ir_graph *irg = current_ir_graph; +static int map_Mul(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); @@ -445,59 +575,24 @@ static int map_Mul(ir_node *call, void *ctx) { */ /* handle the often used case of 32x32=64 mul */ - if (is_Shrs(a_h) && get_Shrs_left(a_h) == a_l) { - ir_node *c1 = get_Shrs_right(a_h); - - if (is_Const(c1)) { - tarval *tv = get_Const_tarval(c1); - - if (tarval_is_long(tv) && get_tarval_long(tv) == 31) { - /* a is a sign extend */ - - if (is_Shrs(b_h) && get_Shrs_left(b_h) == b_l && c1 == get_Shrs_right(b_h)) { - /* b is a sign extend: it's a 32 * 32 = 64 signed multiplication */ - mul = new_rd_ia32_l_IMul(dbg, irg, block, a_l, b_l); - h_res = new_rd_Proj(dbg, irg, block, mul, h_mode, pn_ia32_l_Mul_EDX); - l_res = new_rd_Proj(dbg, irg, block, mul, l_mode, pn_ia32_l_Mul_EAX); - - goto end; - } - /* we rely here on Consts being on the right side */ - if (is_Const(b_h) && is_Const(b_l)) { - tarval *th = get_Const_tarval(b_h); - tarval *tl = get_Const_tarval(b_l); - - if (tarval_is_long(th) && tarval_is_long(tl)) { - long h = get_tarval_long(th); - long l = get_tarval_long(tl); - - if ((h == 0 && l >= 0) || (h == -1 && l < 0)) { - /* b is a sign extended const */ - mul = new_rd_ia32_l_IMul(dbg, irg, block, a_l, b_l); - h_res = new_rd_Proj(dbg, irg, block, mul, h_mode, pn_ia32_l_Mul_EDX); - l_res = new_rd_Proj(dbg, irg, block, mul, l_mode, pn_ia32_l_Mul_EAX); - - goto end; - } - } - } - } - } + if (is_sign_extend(a_l, a_h) && is_sign_extend(b_l, b_h)) { + mul = new_bd_ia32_l_IMul(dbg, block, a_l, b_l); + h_res = new_rd_Proj(dbg, block, mul, h_mode, pn_ia32_l_IMul_res_high); + l_res = new_rd_Proj(dbg, block, mul, l_mode, pn_ia32_l_IMul_res_low); + } else { + /* note that zero extension is handled hare efficiently */ + mul = new_bd_ia32_l_Mul(dbg, block, a_l, b_l); + pEDX = new_rd_Proj(dbg, block, mul, h_mode, pn_ia32_l_Mul_res_high); + l_res = new_rd_Proj(dbg, block, mul, l_mode, pn_ia32_l_Mul_res_low); + + b_l = new_rd_Conv(dbg, block, b_l, h_mode); + mul = new_rd_Mul( dbg, block, a_h, b_l, h_mode); + add = new_rd_Add( dbg, block, mul, pEDX, h_mode); + a_l = new_rd_Conv(dbg, block, a_l, h_mode); + mul = new_rd_Mul( dbg, block, a_l, b_h, h_mode); + h_res = new_rd_Add( dbg, block, add, mul, h_mode); } - - mul = new_rd_ia32_l_Mul(dbg, irg, block, a_l, b_l); - pEDX = new_rd_Proj(dbg, irg, block, mul, h_mode, pn_ia32_l_Mul_EDX); - l_res = new_rd_Proj(dbg, irg, block, mul, l_mode, pn_ia32_l_Mul_EAX); - - b_l = new_rd_Conv(dbg, irg, block, b_l, h_mode); - mul = new_rd_Mul( dbg, irg, block, a_h, b_l, h_mode); - add = new_rd_Add( dbg, irg, block, mul, pEDX, h_mode); - a_l = new_rd_Conv(dbg, irg, block, a_l, h_mode); - mul = new_rd_Mul( dbg, irg, block, a_l, b_h, h_mode); - h_res = new_rd_Add( dbg, irg, block, add, mul, h_mode); - -end: - resolve_call(call, l_res, h_res, irg, block); + resolve_call(call, l_res, h_res, current_ir_graph, block); return 1; } @@ -505,8 +600,8 @@ end: /** * Map a Minus (a_l, a_h) */ -static int map_Minus(ir_node *call, void *ctx) { - ir_graph *irg = current_ir_graph; +static int map_Minus(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); @@ -518,11 +613,11 @@ static int map_Minus(ir_node *call, void *ctx) { ir_node *l_res, *h_res, *res; (void) ctx; - res = new_rd_ia32_Minus64Bit(dbg, irg, block, a_l, a_h); - l_res = new_r_Proj(irg, block, res, l_mode, pn_ia32_Minus64Bit_low_res); - h_res = new_r_Proj(irg, block, res, h_mode, pn_ia32_Minus64Bit_high_res); + res = new_bd_ia32_Minus64Bit(dbg, block, a_l, a_h); + l_res = new_r_Proj(block, res, l_mode, pn_ia32_Minus64Bit_low_res); + h_res = new_r_Proj(block, res, h_mode, pn_ia32_Minus64Bit_high_res); - resolve_call(call, l_res, h_res, irg, block); + resolve_call(call, l_res, h_res, current_ir_graph, block); return 1; } @@ -530,8 +625,8 @@ static int map_Minus(ir_node *call, void *ctx) { /** * Map a Abs (a_l, a_h) */ -static int map_Abs(ir_node *call, void *ctx) { - ir_graph *irg = current_ir_graph; +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); @@ -562,17 +657,17 @@ static int map_Abs(ir_node *call, void *ctx) { */ /* TODO: give a hint to the backend somehow to not create a cltd here... */ - sign = new_rd_Shrs(dbg, irg, block, a_h, new_Const_long(l_mode, 31), h_mode); - sign_l = new_rd_Conv(dbg, irg, block, sign, l_mode); - sub_l = new_rd_Eor(dbg, irg, block, a_l, sign_l, l_mode); - sub_h = new_rd_Eor(dbg, irg, block, a_h, sign, h_mode); + sign = new_rd_Shrs(dbg, block, a_h, new_Const_long(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_rd_ia32_l_Sub(dbg, irg, block, sub_l, sign_l, mode_T); - l_res = new_r_Proj(irg, block, l_sub, l_mode, pn_ia32_res); - flags = new_r_Proj(irg, block, l_sub, mode_flags, pn_ia32_flags); - h_res = new_rd_ia32_l_Sbb(dbg, irg, block, sub_h, sign, flags, h_mode); + l_sub = new_bd_ia32_l_Sub(dbg, block, sub_l, sign_l, mode_T); + l_res = new_r_Proj(block, l_sub, l_mode, pn_ia32_res); + flags = new_r_Proj(block, 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, irg, block); + resolve_call(call, l_res, h_res, current_ir_graph, block); return 1; } @@ -580,14 +675,16 @@ static int map_Abs(ir_node *call, void *ctx) { #define ID(x) new_id_from_chars(x, sizeof(x)-1) /** - * Maps a Div. Change into a library call + * Maps a Div. Change into a library call. */ -static int map_Div(ir_node *call, void *ctx) { +static int map_Div(ir_node *call, void *ctx) +{ ia32_intrinsic_env_t *env = ctx; ir_type *method = get_Call_type(call); ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1)); ir_node *ptr; ir_entity *ent; + ir_graph *irg = get_irn_irg(call); symconst_symbol sym; if (mode_is_signed(h_mode)) { @@ -596,7 +693,7 @@ static int map_Div(ir_node *call, void *ctx) { if (ent == NULL) { /* create library entity */ ent = env->divdi3 = new_entity(get_glob_type(), ID("__divdi3"), method); - set_entity_visibility(ent, visibility_external_allocated); + set_entity_visibility(ent, ir_visibility_external); set_entity_ld_ident(ent, ID("__divdi3")); } } else { @@ -605,25 +702,30 @@ static int map_Div(ir_node *call, void *ctx) { if (ent == NULL) { /* create library entity */ ent = env->udivdi3 = new_entity(get_glob_type(), ID("__udivdi3"), method); - set_entity_visibility(ent, visibility_external_allocated); + set_entity_visibility(ent, ir_visibility_external); set_entity_ld_ident(ent, ID("__udivdi3")); } } - sym.entity_p = ent; + ptr = get_Call_ptr(call); - set_SymConst_symbol(ptr, sym); + sym.entity_p = ent; + ptr = new_r_SymConst(irg, get_irn_mode(ptr), sym, symconst_addr_ent); + set_Call_ptr(call, ptr); + return 1; } /** * Maps a Mod. Change into a library call */ -static int map_Mod(ir_node *call, void *ctx) { +static int map_Mod(ir_node *call, void *ctx) +{ ia32_intrinsic_env_t *env = ctx; ir_type *method = get_Call_type(call); - ir_mode *h_mode = get_type_mode(get_method_res_type(method, 0)); + ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1)); ir_node *ptr; ir_entity *ent; + ir_graph *irg = get_irn_irg(call); symconst_symbol sym; if (mode_is_signed(h_mode)) { @@ -632,7 +734,7 @@ static int map_Mod(ir_node *call, void *ctx) { if (ent == NULL) { /* create library entity */ ent = env->moddi3 = new_entity(get_glob_type(), ID("__moddi3"), method); - set_entity_visibility(ent, visibility_external_allocated); + set_entity_visibility(ent, ir_visibility_external); set_entity_ld_ident(ent, ID("__moddi3")); } } else { @@ -641,34 +743,36 @@ static int map_Mod(ir_node *call, void *ctx) { if (ent == NULL) { /* create library entity */ ent = env->umoddi3 = new_entity(get_glob_type(), ID("__umoddi3"), method); - set_entity_visibility(ent, visibility_external_allocated); + set_entity_visibility(ent, ir_visibility_external); set_entity_ld_ident(ent, ID("__umoddi3")); } } - sym.entity_p = ent; + ptr = get_Call_ptr(call); - set_SymConst_symbol(ptr, sym); + sym.entity_p = ent; + ptr = new_r_SymConst(irg, get_irn_mode(ptr), sym, symconst_addr_ent); + set_Call_ptr(call, ptr); + return 1; } /** * Maps a Conv. */ -static int map_Conv(ir_node *call, void *ctx) { - ia32_intrinsic_env_t *env = 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); - int n = get_Call_n_params(call); - int gp_bytes = get_mode_size_bytes(ia32_reg_classes[CLASS_ia32_gp].mode); - ir_entity *ent; - ir_node *l_res, *h_res, *frame, *fres; - ir_node *store_l, *store_h; - ir_node *op_mem[2], *mem; +static int map_Conv(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); + int n = get_Call_n_params(call); + ir_node *l_res, *h_res; + (void) ctx; if (n == 1) { + ir_node *float_to_ll; + /* We have a Conv float -> long long here */ ir_node *a_f = params[0]; ir_mode *l_res_mode = get_type_mode(get_method_res_type(method, 0)); @@ -676,113 +780,88 @@ static int map_Conv(ir_node *call, void *ctx) { assert(mode_is_float(get_irn_mode(a_f)) && "unexpected Conv call"); - /* allocate memory on frame to store args */ - ent = env->irg == irg ? env->d_ll_conv : NULL; - if (! ent) { - ent = env->d_ll_conv = frame_alloc_area(get_irg_frame_type(irg), 2 * gp_bytes, 16, 0); - env->irg = irg; - } + if (mode_is_signed(h_res_mode)) { + /* convert from float to signed 64bit */ + float_to_ll = new_bd_ia32_l_FloattoLL(dbg, block, a_f); - /* Store arg */ - frame = get_irg_frame(irg); + l_res = new_r_Proj(block, float_to_ll, l_res_mode, + pn_ia32_l_FloattoLL_res_low); + h_res = new_r_Proj(block, 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); + tarval *flt_tv = new_tarval_from_str("9223372036854775808", 19, flt_mode); + ir_node *flt_corr = new_Const(flt_tv); + ir_node *lower_blk = block; + ir_node *upper_blk; + ir_node *cmp, *proj, *cond, *blk, *int_phi, *flt_phi; + ir_node *in[2]; - /* - Now we create a node to move the value from a XMM register into - x87 FPU because it is unknown here, which FPU is used. - This node is killed in transformation phase when not needed. - Otherwise it is split up into a movsd + fld - */ - a_f = new_rd_ia32_l_SSEtoX87(dbg, irg, block, frame, a_f, get_irg_no_mem(irg), mode_D); - set_ia32_frame_ent(a_f, ent); - set_ia32_use_frame(a_f); - set_ia32_ls_mode(a_f, mode_D); - - /* store from FPU as Int */ - a_f = new_rd_ia32_l_vfist(dbg, irg, block, frame, a_f, get_irg_no_mem(irg)); - set_ia32_frame_ent(a_f, ent); - set_ia32_use_frame(a_f); - set_ia32_ls_mode(a_f, mode_Ls); - mem = a_f; - - /* load low part of the result */ - l_res = new_rd_ia32_l_Load(dbg, irg, block, frame, mem); - set_ia32_frame_ent(l_res, ent); - set_ia32_use_frame(l_res); - set_ia32_ls_mode(l_res, l_res_mode); - l_res = new_r_Proj(irg, block, l_res, l_res_mode, pn_ia32_l_Load_res); - - /* load hight part of the result */ - h_res = new_rd_ia32_l_Load(dbg, irg, block, frame, mem); - set_ia32_frame_ent(h_res, ent); - add_ia32_am_offs_int(h_res, gp_bytes); - set_ia32_use_frame(h_res); - set_ia32_ls_mode(h_res, h_res_mode); - h_res = new_r_Proj(irg, block, h_res, h_res_mode, pn_ia32_l_Load_res); + part_block(call); + upper_blk = get_nodes_block(call); + + cmp = new_rd_Cmp(dbg, upper_blk, a_f, flt_corr); + proj = new_r_Proj(upper_blk, cmp, mode_b, pn_Cmp_Lt); + cond = new_rd_Cond(dbg, upper_blk, proj); + in[0] = new_r_Proj(upper_blk, cond, mode_X, pn_Cond_true); + in[1] = new_r_Proj(upper_blk, cond, mode_X, pn_Cond_false); + blk = new_r_Block(irg, 1, &in[1]); + in[1] = new_r_Jmp(blk); + + set_irn_in(lower_blk, 2, in); + + /* create to Phis */ + in[0] = new_Const(get_mode_null(h_res_mode)); + in[1] = new_Const_long(h_res_mode, 0x80000000); + + 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); + + flt_phi = new_r_Phi(lower_blk, 2, in, flt_mode); + + /* 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); + + float_to_ll = new_bd_ia32_l_FloattoLL(dbg, lower_blk, flt_phi); + + l_res = new_r_Proj(lower_blk, float_to_ll, l_res_mode, + pn_ia32_l_FloattoLL_res_low); + h_res = new_r_Proj(lower_blk, float_to_ll, h_res_mode, + pn_ia32_l_FloattoLL_res_high); + h_res = new_rd_Add(dbg, lower_blk, h_res, int_phi, h_res_mode); + + /* move the call and its Proj's to the lower block */ + set_nodes_block(call, lower_blk); + + for (proj = get_irn_link(call); proj != NULL; proj = get_irn_link(proj)) + set_nodes_block(proj, lower_blk); + block = lower_blk; + } /* lower the call */ resolve_call(call, l_res, h_res, irg, block); - } - else if (n == 2) { + } else if (n == 2) { + ir_node *ll_to_float; + /* We have a Conv long long -> float here */ ir_node *a_l = params[BINOP_Left_Low]; ir_node *a_h = params[BINOP_Left_High]; - ir_mode *mode_a_l = get_irn_mode(a_l); - ir_mode *mode_a_h = get_irn_mode(a_h); ir_mode *fres_mode = get_type_mode(get_method_res_type(method, 0)); - assert(! mode_is_float(mode_a_l) && ! mode_is_float(mode_a_h) && "unexpected Conv call"); + assert(! mode_is_float(get_irn_mode(a_l)) + && ! mode_is_float(get_irn_mode(a_h))); - /* allocate memory on frame to store args */ - ent = env->irg == irg ? env->ll_d_conv : NULL; - if (! ent) { - ent = env->ll_d_conv = frame_alloc_area(get_irg_frame_type(irg), 2 * gp_bytes, 16, 0); - env->irg = irg; - } - - /* Store arg */ - frame = get_irg_frame(irg); - - /* store first arg (low part) */ - store_l = new_rd_ia32_l_Store(dbg, irg, block, frame, a_l, get_irg_no_mem(irg)); - set_ia32_frame_ent(store_l, ent); - set_ia32_use_frame(store_l); - set_ia32_ls_mode(store_l, get_irn_mode(a_l)); - op_mem[0] = store_l; - - /* store second arg (high part) */ - store_h = new_rd_ia32_l_Store(dbg, irg, block, frame, a_h, get_irg_no_mem(irg)); - set_ia32_frame_ent(store_h, ent); - add_ia32_am_offs_int(store_h, gp_bytes); - set_ia32_use_frame(store_h); - set_ia32_ls_mode(store_h, get_irn_mode(a_h)); - op_mem[1] = store_h; - - mem = new_r_Sync(irg, block, 2, op_mem); - - /* Load arg into x87 FPU (implicit convert) */ - fres = new_rd_ia32_l_vfild(dbg, irg, block, frame, mem); - set_ia32_frame_ent(fres, ent); - set_ia32_use_frame(fres); - set_ia32_ls_mode(fres, mode_D); - mem = new_r_Proj(irg, block, fres, mode_M, pn_ia32_l_vfild_M); - fres = new_r_Proj(irg, block, fres, fres_mode, pn_ia32_l_vfild_res); - - /* - Now we create a node to move the loaded value into a XMM - register because it is unknown here, which FPU is used. - This node is killed in transformation phase when not needed. - Otherwise it is split up into a fst + movsd - */ - fres = new_rd_ia32_l_X87toSSE(dbg, irg, block, frame, fres, mem, fres_mode); - set_ia32_frame_ent(fres, ent); - set_ia32_use_frame(fres); - set_ia32_ls_mode(fres, fres_mode); + ll_to_float = new_bd_ia32_l_LLtoFloat(dbg, block, a_h, a_l, fres_mode); /* lower the call */ - resolve_call(call, fres, NULL, irg, block); - } - else { - assert(0 && "unexpected Conv call"); + resolve_call(call, ll_to_float, NULL, irg, block); + } else { + panic("unexpected Conv call %+F", call); } return 1; @@ -853,8 +932,9 @@ 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 = mangle(IDENT("L"), get_op_ident(op)); + ident *id = id_mangle(IDENT("L"), get_op_ident(op)); *ent = new_entity(get_glob_type(), id, method); + set_entity_visibility(*ent, ir_visibility_private); } elt.i_call.kind = INTRINSIC_CALL;