X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_intrinsics.c;h=2dd97f396bc01e455a3910180527614d04fa55a5;hb=e1397b01aceb38b6bb62c319007146af3b922f39;hp=026cd384bcc202592f5477d0f4729a7730977082;hpb=0fbcef83aa6060534172bb13e71cdadb04428806;p=libfirm diff --git a/ir/be/ia32/ia32_intrinsics.c b/ir/be/ia32/ia32_intrinsics.c index 026cd384b..2dd97f396 100644 --- a/ir/be/ia32/ia32_intrinsics.c +++ b/ir/be/ia32/ia32_intrinsics.c @@ -26,6 +26,7 @@ */ #include "config.h" +#include "iredges.h" #include "irgmod.h" #include "irop.h" #include "irnode_t.h" @@ -61,8 +62,39 @@ 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 *jmp, *res, *in[2]; @@ -70,36 +102,83 @@ static void resolve_call(ir_node *call, ir_node *l_res, ir_node *h_res, ir_graph ir_node *nomem = get_irg_no_mem(irg); int old_cse; - in[0] = l_res; - in[1] = h_res; - res = new_r_Tuple(irg, block, h_res == NULL ? 1 : 2, in); + 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_except: + /* 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); - set_Tuple_pred(call, pn_Call_M_regular, nomem); - /* - * 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(irg, block); - set_opt_cse(old_cse); - - 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_M_except, nomem); - set_Tuple_pred(call, pn_Call_P_value_res_base, bad); + turn_into_tuple(call, pn_Call_max); + set_Tuple_pred(call, pn_Call_M_regular, nomem); + /* + * 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_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_M_except, nomem); + 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; dbg_info *dbg = get_irn_dbg_info(call); ir_node *block = get_nodes_block(call); ir_node **params = get_Call_param_arr(call); @@ -118,14 +197,14 @@ 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; } @@ -134,7 +213,6 @@ 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); @@ -153,14 +231,14 @@ 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; } @@ -189,16 +267,16 @@ 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); @@ -209,37 +287,37 @@ 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); + 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); + h_res = new_r_Phi(block, 2, in, h_mode); set_Phi_next(l_res, h_res); set_Phi_next(h_res, NULL); @@ -277,15 +355,15 @@ 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; @@ -295,37 +373,37 @@ 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, h_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); + 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); + h_res = new_r_Phi(block, 2, in, h_mode); set_Phi_next(l_res, h_res); set_Phi_next(h_res, NULL); @@ -363,17 +441,17 @@ 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; @@ -383,37 +461,37 @@ 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); + 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); + h_res = new_r_Phi(block, 2, in, h_mode); set_Phi_next(l_res, h_res); set_Phi_next(h_res, NULL); @@ -426,6 +504,9 @@ 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)) { @@ -463,7 +544,6 @@ static int is_sign_extend(ir_node *low, ir_node *high) * 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; dbg_info *dbg = get_irn_dbg_info(call); ir_node *block = get_nodes_block(call); ir_node **params = get_Call_param_arr(call); @@ -489,26 +569,23 @@ static int map_Mul(ir_node *call, void *ctx) { /* handle the often used case of 32x32=64 mul */ if (is_sign_extend(a_l, a_h) && is_sign_extend(b_l, b_h)) { - 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; + 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_Mul_EDX); + l_res = new_rd_Proj(dbg, block, mul, l_mode, pn_ia32_l_Mul_EAX); + } 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_EDX); + l_res = new_rd_Proj(dbg, block, mul, l_mode, pn_ia32_l_Mul_EAX); + + 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; } @@ -517,7 +594,6 @@ end: * Map a Minus (a_l, a_h) */ static int map_Minus(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); @@ -529,11 +605,11 @@ 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; } @@ -542,7 +618,6 @@ 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; dbg_info *dbg = get_irn_dbg_info(call); ir_node *block = get_nodes_block(call); ir_node **params = get_Call_param_arr(call); @@ -573,17 +648,17 @@ 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; } @@ -591,7 +666,7 @@ 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) { ia32_intrinsic_env_t *env = ctx; @@ -685,13 +760,69 @@ static int map_Conv(ir_node *call, void *ctx) { assert(mode_is_float(get_irn_mode(a_f)) && "unexpected Conv call"); - float_to_ll = new_rd_ia32_l_FloattoLL(dbg, irg, block, a_f); + 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); + + 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]; + + part_block(call); + upper_blk = get_nodes_block(call); - l_res = new_r_Proj(irg, block, float_to_ll, l_res_mode, - pn_ia32_l_FloattoLL_res_low); - h_res = new_r_Proj(irg, block, float_to_ll, h_res_mode, - pn_ia32_l_FloattoLL_res_high); + 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) { @@ -705,8 +836,7 @@ static int map_Conv(ir_node *call, void *ctx) { assert(! mode_is_float(get_irn_mode(a_l)) && ! mode_is_float(get_irn_mode(a_h))); - ll_to_float = new_rd_ia32_l_LLtoFloat(dbg, irg, block, a_h, a_l, - fres_mode); + ll_to_float = new_bd_ia32_l_LLtoFloat(dbg, block, a_h, a_l, fres_mode); /* lower the call */ resolve_call(call, ll_to_float, NULL, irg, block); @@ -782,7 +912,7 @@ 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 = 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); }