X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_fpu.c;h=2a56a7462bd3da7a70a949d14b978948a8a92a82;hb=4e809225145abac9fd89758b6226012f3e8f7b1e;hp=68d31612f6378456c9ae44dbf35eb1c3b458425d;hpb=5057c4b3b6c48dc75dee684cd172ce4cfed00da7;p=libfirm

diff --git a/ir/be/ia32/ia32_fpu.c b/ir/be/ia32/ia32_fpu.c
index 68d31612f..2a56a7462 100644
--- a/ir/be/ia32/ia32_fpu.c
+++ b/ir/be/ia32/ia32_fpu.c
@@ -34,6 +34,7 @@
 
 #include "ia32_fpu.h"
 #include "ia32_new_nodes.h"
+#include "ia32_architecture.h"
 #include "gen_ia32_regalloc_if.h"
 
 #include "ircons.h"
@@ -51,12 +52,59 @@
 #include "../bessaconstr.h"
 #include "../beirg_t.h"
 
+static ir_entity *fpcw_round    = NULL;
+static ir_entity *fpcw_truncate = NULL;
+
+static ir_entity *create_ent(int value, const char *name)
+{
+	ir_mode   *mode = mode_Hu;
+	ir_type   *type = new_type_primitive(new_id_from_str("_fpcw_type"), mode);
+	ir_type   *glob = get_glob_type();
+	ir_graph  *cnst_irg;
+	ir_entity *ent;
+	ir_node   *cnst;
+	tarval    *tv;
+
+	set_type_alignment_bytes(type, 4);
+
+	tv  = new_tarval_from_long(value, mode);
+	ent = new_entity(glob, new_id_from_str(name), type);
+	set_entity_ld_ident(ent, get_entity_ident(ent));
+	set_entity_visibility(ent, visibility_local);
+	set_entity_variability(ent, variability_constant);
+	set_entity_allocation(ent, allocation_static);
+
+	cnst_irg = get_const_code_irg();
+	cnst     = new_r_Const(cnst_irg, get_irg_start_block(cnst_irg), mode, tv);
+	set_atomic_ent_value(ent, cnst);
+
+	return ent;
+}
+
+static void create_fpcw_entities(void)
+{
+	fpcw_round    = create_ent(0xc7f, "_fpcw_round");
+	fpcw_truncate = create_ent(0x37f, "_fpcw_truncate");
+}
+
 static ir_node *create_fpu_mode_spill(void *env, ir_node *state, int force,
                                       ir_node *after)
 {
 	ia32_code_gen_t *cg = env;
 	ir_node *spill = NULL;
 
+	/* we don't spill the fpcw in unsafe mode */
+	if(ia32_cg_config.use_unsafe_floatconv) {
+		ir_graph *irg = get_irn_irg(state);
+		ir_node *block = get_nodes_block(state);
+		if(force == 1 || !is_ia32_ChangeCW(state)) {
+			ir_node *spill = new_rd_ia32_FnstCWNOP(NULL, irg, block, state);
+			sched_add_after(after, spill);
+			return spill;
+		}
+		return NULL;
+	}
+
 	if(force == 1 || !is_ia32_ChangeCW(state)) {
 		ir_graph *irg = get_irn_irg(state);
 		ir_node *block = get_nodes_block(state);
@@ -64,11 +112,10 @@ static ir_node *create_fpu_mode_spill(void *env, ir_node *state, int force,
 		ir_node *nomem = new_NoMem();
 		ir_node *frame = get_irg_frame(irg);
 
-		spill = new_rd_ia32_FnstCW(NULL, irg, block, frame, noreg, state,
-		                           nomem);
+		spill = new_rd_ia32_FnstCW(NULL, irg, block, frame, noreg, nomem, state);
 		set_ia32_op_type(spill, ia32_AddrModeD);
-		set_ia32_am_flavour(spill, ia32_B);
-		set_ia32_ls_mode(spill, ia32_reg_classes[CLASS_ia32_fp_cw].mode);
+		/* use mode_Iu, as movl has a shorter opcode than movw */
+		set_ia32_ls_mode(spill, mode_Iu);
 		set_ia32_use_frame(spill);
 
 		sched_add_after(after, spill);
@@ -77,21 +124,51 @@ static ir_node *create_fpu_mode_spill(void *env, ir_node *state, int force,
 	return spill;
 }
 
+static ir_node *create_fldcw_ent(ia32_code_gen_t *cg, ir_node *block,
+                                 ir_entity *entity)
+{
+	ir_graph *irg   = get_irn_irg(block);
+	ir_node  *nomem = new_NoMem();
+	ir_node  *noreg = ia32_new_NoReg_gp(cg);
+	ir_node  *reload;
+
+	reload = new_rd_ia32_FldCW(NULL, irg, block, noreg, noreg, nomem);
+	set_ia32_op_type(reload, ia32_AddrModeS);
+	set_ia32_ls_mode(reload, ia32_reg_classes[CLASS_ia32_fp_cw].mode);
+	set_ia32_am_sc(reload, entity);
+	set_ia32_use_frame(reload);
+	arch_set_irn_register(cg->arch_env, reload, &ia32_fp_cw_regs[REG_FPCW]);
+
+	return reload;
+}
+
 static ir_node *create_fpu_mode_reload(void *env, ir_node *state,
                                        ir_node *spill, ir_node *before,
                                        ir_node *last_state)
 {
-	ia32_code_gen_t *cg = env;
-	ir_graph *irg = get_irn_irg(state);
-	ir_node *block = get_nodes_block(before);
-	ir_node *frame = get_irg_frame(irg);
-	ir_node *noreg = ia32_new_NoReg_gp(cg);
-	ir_node *reload = NULL;
+	ia32_code_gen_t *cg    = env;
+	ir_graph        *irg   = get_irn_irg(state);
+	ir_node         *block = get_nodes_block(before);
+	ir_node         *frame = get_irg_frame(irg);
+	ir_node         *noreg = ia32_new_NoReg_gp(cg);
+	ir_node         *reload = NULL;
+
+	if(ia32_cg_config.use_unsafe_floatconv) {
+		if(fpcw_round == NULL) {
+			create_fpcw_entities();
+		}
+		if(spill != NULL) {
+			reload = create_fldcw_ent(cg, block, fpcw_round);
+		} else {
+			reload = create_fldcw_ent(cg, block, fpcw_truncate);
+		}
+		sched_add_before(before, reload);
+		return reload;
+	}
 
 	if(spill != NULL) {
 		reload = new_rd_ia32_FldCW(NULL, irg, block, frame, noreg, spill);
 		set_ia32_op_type(reload, ia32_AddrModeS);
-		set_ia32_am_flavour(reload, ia32_B);
 		set_ia32_ls_mode(reload, ia32_reg_classes[CLASS_ia32_fp_cw].mode);
 		set_ia32_use_frame(reload);
 		arch_set_irn_register(cg->arch_env, reload, &ia32_fp_cw_regs[REG_FPCW]);
@@ -101,19 +178,18 @@ static ir_node *create_fpu_mode_reload(void *env, ir_node *state,
 		ir_mode *lsmode = ia32_reg_classes[CLASS_ia32_fp_cw].mode;
 		ir_node *nomem = new_NoMem();
 		ir_node *cwstore, *load, *load_res, *or, *store, *fldcw;
+		ir_node *or_const;
 
 		assert(last_state != NULL);
-		cwstore = new_rd_ia32_FnstCW(NULL, irg, block, frame, noreg, last_state,
-		                             nomem);
+		cwstore = new_rd_ia32_FnstCW(NULL, irg, block, frame, noreg, nomem,
+		                             last_state);
 		set_ia32_op_type(cwstore, ia32_AddrModeD);
-		set_ia32_am_flavour(cwstore, ia32_B);
 		set_ia32_ls_mode(cwstore, lsmode);
 		set_ia32_use_frame(cwstore);
 		sched_add_before(before, cwstore);
 
 		load = new_rd_ia32_Load(NULL, irg, block, frame, noreg, cwstore);
 		set_ia32_op_type(load, ia32_AddrModeS);
-		set_ia32_am_flavour(load, ia32_B);
 		set_ia32_ls_mode(load, lsmode);
 		set_ia32_use_frame(load);
 		sched_add_before(before, load);
@@ -121,21 +197,23 @@ static ir_node *create_fpu_mode_reload(void *env, ir_node *state,
 		load_res = new_r_Proj(irg, block, load, mode_Iu, pn_ia32_Load_res);
 
 		/* TODO: make the actual mode configurable in ChangeCW... */
-		or = new_rd_ia32_Or(NULL, irg, block, noreg, noreg, load_res, noreg,
-		                    nomem);
-		set_ia32_Immop_tarval(or, new_tarval_from_long(3072, mode_Iu));
+		or_const = new_rd_ia32_Immediate(NULL, irg, get_irg_start_block(irg),
+		                                 NULL, 0, 3072);
+		arch_set_irn_register(cg->arch_env, or_const,
+		                      &ia32_gp_regs[REG_GP_NOREG]);
+		or = new_rd_ia32_Or(NULL, irg, block, noreg, noreg, nomem, load_res,
+		                    or_const);
 		sched_add_before(before, or);
 
-		store = new_rd_ia32_Store(NULL, irg, block, frame, noreg, or, nomem);
+		store = new_rd_ia32_Store(NULL, irg, block, frame, noreg, nomem, or);
 		set_ia32_op_type(store, ia32_AddrModeD);
-		set_ia32_am_flavour(store, ia32_B);
-		set_ia32_ls_mode(store, lsmode);
+		/* use mode_Iu, as movl has a shorter opcode than movw */
+		set_ia32_ls_mode(store, mode_Iu);
 		set_ia32_use_frame(store);
 		sched_add_before(before, store);
 
 		fldcw = new_rd_ia32_FldCW(NULL, irg, block, frame, noreg, store);
 		set_ia32_op_type(fldcw, ia32_AddrModeS);
-		set_ia32_am_flavour(fldcw, ia32_B);
 		set_ia32_ls_mode(fldcw, lsmode);
 		set_ia32_use_frame(fldcw);
 		arch_set_irn_register(cg->arch_env, fldcw, &ia32_fp_cw_regs[REG_FPCW]);
@@ -152,12 +230,15 @@ typedef struct collect_fpu_mode_nodes_env_t {
 	ir_node         **state_nodes;
 } collect_fpu_mode_nodes_env_t;
 
-static
-void collect_fpu_mode_nodes_walker(ir_node *node, void *data)
+static void collect_fpu_mode_nodes_walker(ir_node *node, void *data)
 {
 	collect_fpu_mode_nodes_env_t *env = data;
+	const arch_register_t *reg;
+
+	if(!mode_is_data(get_irn_mode(node)))
+		return;
 
-	const arch_register_t *reg = arch_get_irn_register(env->arch_env, node);
+	reg = arch_get_irn_register(env->arch_env, node);
 	if(reg == &ia32_fp_cw_regs[REG_FPCW] && !is_ia32_ChangeCW(node)) {
 		ARR_APP1(ir_node*, env->state_nodes, node);
 	}