+/**
+ * This is the main ia32 firm backend driver.
+ *
+ * $Id$
+ */
+
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
-#ifdef _WIN32
+#ifdef HAVE_MALLOC_H
#include <malloc.h>
-#else
+#endif
+
+#ifdef HAVE_ALLOCA_H
#include <alloca.h>
#endif
+#ifdef WITH_LIBCORE
+#include <libcore/lc_opts.h>
+#include <libcore/lc_opts_enum.h>
+#endif /* WITH_LIBCORE */
+
#include "pseudo_irg.h"
#include "irgwalk.h"
#include "irprog.h"
#include "iredges_t.h"
#include "ircons.h"
#include "irgmod.h"
+#include "irgopt.h"
#include "bitset.h"
#include "debug.h"
#include "../benode_t.h"
#include "../belower.h"
#include "../besched_t.h"
+#include "../be.h"
#include "bearch_ia32_t.h"
#include "ia32_new_nodes.h" /* ia32 nodes interface */
#include "ia32_emitter.h"
#include "ia32_map_regs.h"
#include "ia32_optimize.h"
+#include "ia32_x87.h"
+#include "ia32_dbg_stat.h"
#define DEBUG_MODULE "firm.be.ia32.isa"
#undef is_Start
#define is_Start(irn) (get_irn_opcode(irn) == iro_Start)
+/* Creates the unique per irg GP NoReg node. */
ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg) {
- if (! cg->noreg_gp) {
- cg->noreg_gp = be_new_NoReg(&ia32_gp_regs[REG_XXX], cg->irg, get_irg_start_block(cg->irg));
- }
-
- return cg->noreg_gp;
+ return be_abi_get_callee_save_irn(cg->birg->abi, &ia32_gp_regs[REG_GP_NOREG]);
}
+/* Creates the unique per irg FP NoReg node. */
ir_node *ia32_new_NoReg_fp(ia32_code_gen_t *cg) {
- if (! cg->noreg_fp) {
- cg->noreg_fp = be_new_NoReg(&ia32_fp_regs[REG_XXXX], cg->irg, get_irg_start_block(cg->irg));
- }
-
- return cg->noreg_fp;
+ return be_abi_get_callee_save_irn(cg->birg->abi,
+ USE_SSE2(cg) ? &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG]);
}
/**************************************************
return (ir_node *)n;
}
-static int is_Call_Proj(const ir_node *n) {
- if (is_Proj(n) &&
- is_Proj(get_Proj_pred(n)) &&
- get_irn_mode(get_Proj_pred(n)) == mode_T &&
- is_ia32_Call(get_Proj_pred(get_Proj_pred(n))))
- {
- return 1;
- }
-
- return 0;
-}
-
-static int is_Start_Proj(const ir_node *n) {
- if (is_Proj(n) &&
- is_Proj(get_Proj_pred(n)) &&
- get_irn_mode(get_Proj_pred(n)) == mode_T &&
- is_Start(get_Proj_pred(get_Proj_pred(n))))
- {
- return 1;
- }
-
- return 0;
-}
-
-static int is_P_frame_base_Proj(const ir_node *n) {
- if (is_Proj(n) &&
- is_Start(get_Proj_pred(n)) &&
- get_Proj_proj(n) == pn_Start_P_frame_base)
- {
- return 1;
- }
-
- return 0;
-}
-
-static int is_used_by_Keep(const ir_node *n) {
- return be_is_Keep(get_edge_src_irn(get_irn_out_edge_first(n)));
-}
/**
* Return register requirements for an ia32 node.
* will be asked for this information.
*/
static const arch_register_req_t *ia32_get_irn_reg_req(const void *self, arch_register_req_t *req, const ir_node *irn, int pos) {
+ const ia32_irn_ops_t *ops = self;
const ia32_register_req_t *irn_req;
long node_pos = pos == -1 ? 0 : pos;
- ir_mode *mode = get_irn_mode(irn);
- firm_dbg_module_t *mod = firm_dbg_register(DEBUG_MODULE);
- const ia32_irn_ops_t *ops = self;
+ ir_mode *mode = is_Block(irn) ? NULL : get_irn_mode(irn);
+ FIRM_DBG_REGISTER(firm_dbg_module_t *mod, DEBUG_MODULE);
- if (mode == mode_T || mode == mode_M) {
- DBG((mod, LEVEL_1, "ignoring mode_T, mode_M node %+F\n", irn));
+ if (is_Block(irn) || mode == mode_M || mode == mode_X) {
+ DBG((mod, LEVEL_1, "ignoring Block, mode_M, mode_X node %+F\n", irn));
return NULL;
}
- DBG((mod, LEVEL_1, "get requirements at pos %d for %+F ... ", pos, irn));
-
+ if (mode == mode_T && pos < 0) {
+ DBG((mod, LEVEL_1, "ignoring request OUT requirements for node %+F\n", irn));
+ return NULL;
+ }
- if (is_Call_Proj(irn) && is_used_by_Keep(irn)) {
- if (pos >= 0) {
- req = NULL;
- }
- else {
- irn_req = ia32_projnum_reg_req_map[get_Proj_proj(irn)];
- memcpy(req, &(irn_req->req), sizeof(*req));
- }
+ DBG((mod, LEVEL_1, "get requirements at pos %d for %+F ... ", pos, irn));
- return req;
- }
- else if (is_Start_Proj(irn)) {
-/* irn_req = ops->cg->reg_param_req[get_Proj_proj(irn)];
- assert(irn_req && "missing requirement for regparam");
- memcpy(req, &(irn_req->req), sizeof(*req)); */
- memcpy(req, &(ia32_default_req_ia32_gp.req), sizeof(*req));
- return req;
- //return NULL;
- }
- else if (is_Proj(irn)) {
+ if (is_Proj(irn)) {
if (pos == -1) {
node_pos = ia32_translate_proj_pos(irn);
}
}
}
else {
- /* treat Phi like Const with default requirements */
- if (is_Phi(irn)) {
- DB((mod, LEVEL_1, "returning standard reqs for %+F\n", irn));
- if (mode_is_float(mode))
- memcpy(req, &(ia32_default_req_ia32_fp.req), sizeof(*req));
+ /* treat Unknowns like Const with default requirements */
+ if (is_Unknown(irn)) {
+ DB((mod, LEVEL_1, "returning UKNWN reqs for %+F\n", irn));
+ if (mode_is_float(mode)) {
+ if (USE_SSE2(ops->cg))
+ memcpy(req, &(ia32_default_req_ia32_xmm_xmm_UKNWN), sizeof(*req));
+ else
+ memcpy(req, &(ia32_default_req_ia32_vfp_vfp_UKNWN), sizeof(*req));
+ }
else if (mode_is_int(mode) || mode_is_reference(mode))
- memcpy(req, &(ia32_default_req_ia32_gp.req), sizeof(*req));
+ memcpy(req, &(ia32_default_req_ia32_gp_gp_UKNWN), sizeof(*req));
else if (mode == mode_T || mode == mode_M) {
- DBG((mod, LEVEL_1, "ignoring Phi node %+F\n", irn));
+ DBG((mod, LEVEL_1, "ignoring Unknown node %+F\n", irn));
return NULL;
}
else
- assert(0 && "unsupported Phi-Mode");
- }
- else if (is_Start(irn)) {
- DB((mod, LEVEL_1, "returning reqs none for ProjX -> Start (%+F )\n", irn));
- switch (node_pos) {
- case pn_Start_X_initial_exec:
- case pn_Start_P_value_arg_base:
- case pn_Start_P_globals:
- case pn_Start_P_frame_base:
- memcpy(req, &(ia32_default_req_none.req), sizeof(*req));
- break;
- case pn_Start_T_args:
- assert(0 && "ProjT(pn_Start_T_args) should not be asked");
- }
- }
- else if (get_irn_op(irn) == op_Return && pos > 0) {
- DB((mod, LEVEL_1, "returning reqs EAX for %+F\n", irn));
- memcpy(req, &(ia32_default_req_ia32_gp_eax.req), sizeof(*req));
+ assert(0 && "unsupported Unknown-Mode");
}
else {
DB((mod, LEVEL_1, "returning NULL for %+F (not ia32)\n", irn));
}
static void ia32_set_irn_reg(const void *self, ir_node *irn, const arch_register_t *reg) {
- int pos = 0;
+ int pos = 0;
+ const ia32_irn_ops_t *ops = self;
+
+ if (get_irn_mode(irn) == mode_X) {
+ return;
+ }
+
+ DBG((ops->cg->mod, LEVEL_1, "ia32 assigned register %s to node %+F\n", reg->name, irn));
if (is_Proj(irn)) {
pos = ia32_translate_proj_pos(irn);
const arch_register_t *reg = NULL;
if (is_Proj(irn)) {
+
+ if (get_irn_mode(irn) == mode_X) {
+ return NULL;
+ }
+
pos = ia32_translate_proj_pos(irn);
irn = my_skip_proj(irn);
}
irn = my_skip_proj(irn);
if (is_cfop(irn))
return arch_irn_class_branch;
- else if (is_ia32_Call(irn))
- return arch_irn_class_call;
else if (is_ia32_irn(irn))
return arch_irn_class_normal;
else
if (is_ia32_irn(irn))
return get_ia32_flags(irn);
else {
+ if (is_Unknown(irn))
+ return arch_irn_flags_ignore;
return 0;
}
}
+static entity *ia32_get_frame_entity(const void *self, const ir_node *irn) {
+ return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL;
+}
+
static void ia32_set_stack_bias(const void *self, ir_node *irn, int bias) {
- if (get_ia32_use_frame(irn)) {
- /* TODO: correct offset */
+ char buf[64];
+ const ia32_irn_ops_t *ops = self;
+
+ if (get_ia32_frame_ent(irn)) {
+ ia32_am_flavour_t am_flav = get_ia32_am_flavour(irn);
+
+ DBG((ops->cg->mod, LEVEL_1, "stack biased %+F with %d\n", irn, bias));
+ snprintf(buf, sizeof(buf), "%d", bias);
+
+ if (get_ia32_op_type(irn) == ia32_Normal) {
+ set_ia32_cnst(irn, buf);
+ }
+ else {
+ add_ia32_am_offs(irn, buf);
+ am_flav |= ia32_O;
+ set_ia32_am_flavour(irn, am_flav);
+ }
}
}
+typedef struct {
+ be_abi_call_flags_bits_t flags;
+ const arch_isa_t *isa;
+ const arch_env_t *aenv;
+ ir_graph *irg;
+} ia32_abi_env_t;
+
+static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
+{
+ ia32_abi_env_t *env = xmalloc(sizeof(env[0]));
+ be_abi_call_flags_t fl = be_abi_call_get_flags(call);
+ env->flags = fl.bits;
+ env->irg = irg;
+ env->aenv = aenv;
+ env->isa = aenv->isa;
+ return env;
+}
+
+static void ia32_abi_dont_save_regs(void *self, pset *s)
+{
+ ia32_abi_env_t *env = self;
+ if(env->flags.try_omit_fp)
+ pset_insert_ptr(s, env->isa->bp);
+}
+
+/**
+ * Generate the prologue.
+ * @param self The callback object.
+ * @param mem A pointer to the mem node. Update this if you define new memory.
+ * @param reg_map A mapping mapping all callee_save/ignore/parameter registers to their defining nodes.
+ * @return The register which shall be used as a stack frame base.
+ *
+ * All nodes which define registers in @p reg_map must keep @p reg_map current.
+ */
+static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map)
+{
+ ia32_abi_env_t *env = self;
+
+ if (!env->flags.try_omit_fp) {
+ int reg_size = get_mode_size_bytes(env->isa->bp->reg_class->mode);
+ ir_node *bl = get_irg_start_block(env->irg);
+ ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
+ ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
+ ir_node *push;
+
+ /* push ebp */
+ push = new_rd_ia32_Push(NULL, env->irg, bl, curr_sp, curr_bp, *mem, mode_T);
+ curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), 0);
+ *mem = new_r_Proj(env->irg, bl, push, mode_M, 1);
+
+ /* the push must have SP out register */
+ arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
+
+ /* move esp to ebp */
+ curr_bp = be_new_Copy(env->isa->bp->reg_class, env->irg, bl, curr_sp);
+ be_set_constr_single_reg(curr_bp, BE_OUT_POS(0), env->isa->bp);
+ arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
+ be_node_set_flags(curr_bp, BE_OUT_POS(0), arch_irn_flags_ignore);
+
+ /* beware: the copy must be done before any other sp use */
+ curr_sp = be_new_CopyKeep_single(env->isa->sp->reg_class, env->irg, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
+ be_set_constr_single_reg(curr_sp, BE_OUT_POS(0), env->isa->sp);
+ arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
+ be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore);
+
+ be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
+ be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
+
+ return env->isa->bp;
+ }
+
+ return env->isa->sp;
+}
+
+static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
+{
+ ia32_abi_env_t *env = self;
+ ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
+ ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
+
+ if (env->flags.try_omit_fp) {
+ /* simply remove the stack frame here */
+ curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, *mem, BE_STACK_FRAME_SIZE, be_stack_dir_shrink);
+ }
+ else {
+ ir_node *pop;
+ ir_mode *mode_bp = env->isa->bp->reg_class->mode;
+ int reg_size = get_mode_size_bytes(env->isa->bp->reg_class->mode);
+
+ /* copy ebp to esp */
+ curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem);
+
+ /* pop ebp */
+ pop = new_rd_ia32_Pop(NULL, env->irg, bl, curr_sp, *mem, mode_T);
+ set_ia32_flags(pop, arch_irn_flags_ignore);
+ curr_bp = new_r_Proj(current_ir_graph, bl, pop, mode_bp, 0);
+ curr_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(curr_sp), 1);
+ *mem = new_r_Proj(current_ir_graph, bl, pop, mode_M, 2);
+ arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
+ arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
+ }
+
+ be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
+ be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
+}
+
+/**
+ * Produces the type which sits between the stack args and the locals on the stack.
+ * it will contain the return address and space to store the old base pointer.
+ * @return The Firm type modeling the ABI between type.
+ */
+static ir_type *ia32_abi_get_between_type(void *self)
+{
+ static ir_type *omit_fp_between_type = NULL;
+ static ir_type *between_type = NULL;
+
+ ia32_abi_env_t *env = self;
+
+ if(!between_type) {
+ entity *old_bp_ent;
+ entity *ret_addr_ent;
+ entity *omit_fp_ret_addr_ent;
+
+ ir_type *old_bp_type = new_type_primitive(new_id_from_str("bp"), mode_P);
+ ir_type *ret_addr_type = new_type_primitive(new_id_from_str("return_addr"), mode_P);
+
+ between_type = new_type_class(new_id_from_str("ia32_between_type"));
+ old_bp_ent = new_entity(between_type, new_id_from_str("old_bp"), old_bp_type);
+ ret_addr_ent = new_entity(between_type, new_id_from_str("ret_addr"), ret_addr_type);
+
+ set_entity_offset_bytes(old_bp_ent, 0);
+ set_entity_offset_bytes(ret_addr_ent, get_type_size_bytes(old_bp_type));
+ set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
+
+ omit_fp_between_type = new_type_class(new_id_from_str("ia32_between_type_omit_fp"));
+ omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, new_id_from_str("ret_addr"), ret_addr_type);
+
+ set_entity_offset_bytes(omit_fp_ret_addr_ent, 0);
+ set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
+ }
+
+ return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
+}
+
+static const be_abi_callbacks_t ia32_abi_callbacks = {
+ ia32_abi_init,
+ free,
+ ia32_abi_get_between_type,
+ ia32_abi_dont_save_regs,
+ ia32_abi_prologue,
+ ia32_abi_epilogue,
+};
+
/* fill register allocator interface */
static const arch_irn_ops_if_t ia32_irn_ops_if = {
ia32_get_irn_reg,
ia32_classify,
ia32_get_flags,
+ ia32_get_frame_entity,
ia32_set_stack_bias
};
*/
static void ia32_prepare_graph(void *self) {
ia32_code_gen_t *cg = self;
+ DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;)
- irg_walk_blkwise_graph(cg->irg, ia32_place_consts, ia32_transform_node, cg);
- dump_ir_block_graph_sched(cg->irg, "-transformed");
- edges_deactivate(cg->irg);
- edges_activate(cg->irg);
- irg_walk_blkwise_graph(cg->irg, NULL, ia32_optimize_am, cg);
- dump_ir_block_graph_sched(cg->irg, "-am");
-}
+ FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
+ ia32_register_transformers();
+ irg_walk_blkwise_graph(cg->irg, ia32_place_consts_set_modes, ia32_transform_node, cg);
+ be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
+ if (cg->opt & IA32_OPT_DOAM) {
+ edges_deactivate(cg->irg);
+ //dead_node_elimination(cg->irg);
+ edges_activate(cg->irg);
+ FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
-/**
- * Stack reservation and StackParam lowering.
- */
-static void ia32_finish_irg(ir_graph *irg, ia32_code_gen_t *cg) {
-#if 0
- firm_dbg_module_t *mod = cg->mod;
- ir_node *frame = get_irg_frame(irg);
- ir_node *end_block = get_irg_end_block(irg);
- ir_node **returns, **in, **new_in;
- ir_node *stack_reserve, *sched_point;
- ir_node *stack_free, *new_ret, *return_block;
- int stack_size = 0, i, n_arg;
- arch_register_t *stack_reg;
- tarval *stack_size_tv;
- dbg_info *frame_dbg;
-
- /* Determine stack register */
- if (cg->has_alloca) {
- stack_reg = &ia32_gp_regs[REG_EBP];
- }
- else {
- stack_reg = &ia32_gp_regs[REG_ESP];
+ irg_walk_blkwise_graph(cg->irg, NULL, ia32_optimize_am, cg);
+ be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
}
- /* If frame is used, then we need to reserve some stackspace. */
- if (get_irn_n_edges(frame) > 0) {
- /* The initial stack reservation. */
- stack_size = get_type_size_bytes(get_irg_frame_type(irg));
- frame_dbg = get_irn_dbg_info(frame);
- stack_reserve = new_rd_ia32_Sub_i(frame_dbg, irg, get_nodes_block(frame), new_NoMem(), mode_Is);
- stack_size_tv = new_tarval_from_long(stack_size, mode_Is);
- set_ia32_Immop_tarval(stack_reserve, stack_size_tv);
+ DEBUG_ONLY(cg->mod = old_mod;)
+}
- assert(stack_size && "bOrken stack layout");
- /* reroute all edges from frame pointer to corrected frame pointer */
- edges_reroute(frame, stack_reserve, irg);
- set_irn_n(stack_reserve, 0, frame);
+/**
+ * Insert copies for all ia32 nodes where the should_be_same requirement
+ * is not fulfilled.
+ * Transform Sub into Neg -- Add if IN2 == OUT
+ */
+static void ia32_finish_node(ir_node *irn, void *env) {
+ ia32_code_gen_t *cg = env;
+ const ia32_register_req_t **reqs;
+ const arch_register_t *out_reg, *in_reg, *in2_reg;
+ int n_res, i;
+ ir_node *copy, *in_node, *block, *in2_node;
+ ia32_op_type_t op_tp;
- /* schedule frame pointer */
- if (! sched_is_scheduled(frame)) {
- sched_add_after(get_irg_start(irg), frame);
+ if (is_ia32_irn(irn)) {
+ /* AM Dest nodes don't produce any values */
+ op_tp = get_ia32_op_type(irn);
+ if (op_tp == ia32_AddrModeD)
+ goto end;
+
+ reqs = get_ia32_out_req_all(irn);
+ n_res = get_ia32_n_res(irn);
+ block = get_nodes_block(irn);
+
+ /* check all OUT requirements, if there is a should_be_same */
+ if (op_tp == ia32_Normal && ! is_ia32_Lea(irn) && ! is_ia32_Conv_I2I(irn) && ! is_ia32_Conv_I2I8Bit(irn)) {
+ for (i = 0; i < n_res; i++) {
+ if (arch_register_req_is(&(reqs[i]->req), should_be_same)) {
+ /* get in and out register */
+ out_reg = get_ia32_out_reg(irn, i);
+ in_node = get_irn_n(irn, reqs[i]->same_pos);
+ in_reg = arch_get_irn_register(cg->arch_env, in_node);
+
+ /* don't copy ignore nodes */
+ if (arch_irn_is(cg->arch_env, in_node, ignore) && is_Proj(in_node))
+ continue;
+
+ /* check if in and out register are equal */
+ if (! REGS_ARE_EQUAL(out_reg, in_reg)) {
+ /* in case of a commutative op: just exchange the in's */
+ /* beware: the current op could be everything, so test for ia32 */
+ /* commutativity first before getting the second in */
+ if (is_ia32_commutative(irn)) {
+ in2_node = get_irn_n(irn, reqs[i]->same_pos ^ 1);
+ in2_reg = arch_get_irn_register(cg->arch_env, in2_node);
+
+ if (REGS_ARE_EQUAL(out_reg, in2_reg)) {
+ set_irn_n(irn, reqs[i]->same_pos, in2_node);
+ set_irn_n(irn, reqs[i]->same_pos ^ 1, in_node);
+ }
+ else
+ goto insert_copy;
+ }
+ else {
+insert_copy:
+ DBG((cg->mod, LEVEL_1, "inserting copy for %+F in_pos %d\n", irn, reqs[i]->same_pos));
+ /* create copy from in register */
+ copy = be_new_Copy(arch_register_get_class(in_reg), cg->irg, block, in_node);
+
+ DBG_OPT_2ADDRCPY(copy);
+
+ /* destination is the out register */
+ arch_set_irn_register(cg->arch_env, copy, out_reg);
+
+ /* insert copy before the node into the schedule */
+ sched_add_before(irn, copy);
+
+ /* set copy as in */
+ set_irn_n(irn, reqs[i]->same_pos, copy);
+ }
+ }
+ }
+ }
}
- /* set register */
- arch_set_irn_register(cg->arch_env, frame, stack_reg);
- arch_set_irn_register(cg->arch_env, stack_reserve, stack_reg);
-
- /* insert into schedule */
- sched_add_after(frame, stack_reserve);
-
- /* Free stack for each Return node */
- returns = get_Block_cfgpred_arr(end_block);
- for (i = 0; i < get_Block_n_cfgpreds(end_block); i++) {
- assert(get_irn_opcode(returns[i]) == iro_Return && "cfgpred of endblock is not a return");
-
- return_block = get_nodes_block(returns[i]);
-
- /* free the stack */
- stack_free = new_rd_ia32_Add_i(frame_dbg, irg, return_block, stack_reserve, mode_Is);
- set_ia32_Immop_tarval(stack_free, stack_size_tv);
- arch_set_irn_register(cg->arch_env, stack_free, stack_reg);
-
- DBG((mod, LEVEL_1, "examining %+F, %+F created, block %+F", returns[i], stack_free, return_block));
-
- /* get the old Return arguments */
- n_arg = get_Return_n_ress(returns[i]);
- in = get_Return_res_arr(returns[i]);
- new_in = alloca((n_arg + 2) * sizeof(new_in[0]));
-
- /* copy the old to the new in's */
- memcpy(new_in, in, n_arg * sizeof(in[0]));
- new_in[n_arg++] = stack_free;
- new_in[n_arg++] = get_Return_mem(returns[i]);
+ /* If we have a CondJmp with immediate, we need to */
+ /* check if it's the right operand, otherwise we have */
+ /* to change it, as CMP doesn't support immediate as */
+ /* left operands. */
+ if (is_ia32_CondJmp(irn) && (is_ia32_ImmConst(irn) || is_ia32_ImmSymConst(irn)) && op_tp == ia32_AddrModeS) {
+ long pnc = get_negated_pnc(get_ia32_pncode(irn), get_ia32_res_mode(irn));
+ set_ia32_op_type(irn, ia32_AddrModeD);
+ set_ia32_pncode(irn, pnc);
+ }
- /* create the new return node */
- new_ret = new_rd_ia32_Return(get_irn_dbg_info(returns[i]), irg, return_block, n_arg, new_in);
+ /* check if there is a sub which need to be transformed */
+ ia32_transform_sub_to_neg_add(irn, cg);
- /* In case the return node is the only node in the block, */
- /* it is not scheduled, so we need this work-around. */
- if (! sched_is_scheduled(returns[i])) {
- sched_point = return_block;
- }
- else {
- sched_point = sched_prev(returns[i]);
- sched_remove(returns[i]);
- }
+ /* transform a LEA into an Add if possible */
+ ia32_transform_lea_to_add(irn, cg);
+ }
+end:
- /* exchange the old return with the new one */
- exchange(returns[i], new_ret);
+ /* check for peephole optimization */
+ ia32_peephole_optimization(irn, cg);
+}
- DB((mod, LEVEL_1, " ... replaced with %+F\n", new_ret));
+static void ia32_finish_irg_walker(ir_node *block, void *env) {
+ ir_node *irn, *next;
- /* remove the old one from schedule and add the new nodes properly */
- sched_add_after(sched_point, new_ret);
- sched_add_after(sched_point, stack_free);
- }
+ for (irn = sched_first(block); !sched_is_end(irn); irn = next) {
+ next = sched_next(irn);
+ ia32_finish_node(irn, env);
}
-#endif
+}
+
+/**
+ * Add Copy nodes for not fulfilled should_be_equal constraints
+ */
+static void ia32_finish_irg(ir_graph *irg, ia32_code_gen_t *cg) {
+ irg_block_walk_graph(irg, NULL, ia32_finish_irg_walker, cg);
}
* Dummy functions for hooks we don't need but which must be filled.
*/
static void ia32_before_sched(void *self) {
- ia32_code_gen_t *cg = self;
-
- lower_nodes_before_sched(cg->irg, cg->arch_env);
}
+/**
+ * Called before the register allocator.
+ * Calculate a block schedule here. We need it for the x87
+ * simulator and the emitter.
+ */
static void ia32_before_ra(void *self) {
+ ia32_code_gen_t *cg = self;
+
+ cg->blk_sched = sched_create_block_schedule(cg->irg);
}
/**
- * Creates a Store for a Spill
+ * Transforms a be node into a Load.
*/
-static ir_node *ia32_lower_spill(void *self, ir_node *spill) {
- ia32_code_gen_t *cg = self;
- ir_graph *irg = cg->irg;
- dbg_info *dbg = get_irn_dbg_info(spill);
- ir_node *block = get_nodes_block(spill);
- ir_node *ptr = get_irg_frame(irg);
- ir_node *val = be_get_Spill_context(spill);
- ir_node *mem = new_rd_NoMem(irg);
- ir_mode *mode = get_irn_mode(spill);
- entity *ent = be_get_spill_entity(spill);
- unsigned offs = get_entity_offset_bytes(ent);
- ir_node *noreg, *res;
- char buf[64];
-
- DB((cg->mod, LEVEL_1, "lower_spill: got offset %d for %+F\n", offs, ent));
+static void transform_to_Load(ia32_transform_env_t *env) {
+ ir_node *irn = env->irn;
+ entity *ent = be_get_frame_entity(irn);
+ ir_mode *mode = env->mode;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *nomem = new_rd_NoMem(env->irg);
+ ir_node *sched_point = NULL;
+ ir_node *ptr = get_irn_n(irn, 0);
+ ir_node *mem = be_is_Reload(irn) ? get_irn_n(irn, 1) : nomem;
+ ir_node *new_op, *proj;
+ const arch_register_t *reg;
+
+ if (sched_is_scheduled(irn)) {
+ sched_point = sched_prev(irn);
+ }
if (mode_is_float(mode)) {
- noreg = ia32_new_NoReg_fp(cg);
- res = new_rd_ia32_fStore(dbg, irg, block, ptr, noreg, val, mem, mode);
+ if (USE_SSE2(env->cg))
+ new_op = new_rd_ia32_fLoad(env->dbg, env->irg, env->block, ptr, noreg, mem, mode_T);
+ else
+ new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem, mode_T);
}
else {
- noreg = ia32_new_NoReg_gp(cg);
- res = new_rd_ia32_Store(dbg, irg, block, ptr, noreg, val, mem, mode);
+ new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem, mode_T);
}
- snprintf(buf, sizeof(buf), "%d", offs);
- add_ia32_am_offs(res, buf);
+ set_ia32_am_support(new_op, ia32_am_Source);
+ set_ia32_op_type(new_op, ia32_AddrModeS);
+ set_ia32_am_flavour(new_op, ia32_B);
+ set_ia32_ls_mode(new_op, mode);
+ set_ia32_frame_ent(new_op, ent);
+ set_ia32_use_frame(new_op);
+
+ DBG_OPT_RELOAD2LD(irn, new_op);
+
+ proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, pn_Load_res);
+
+ if (sched_point) {
+ sched_add_after(sched_point, new_op);
+ sched_add_after(new_op, proj);
- return res;
+ sched_remove(irn);
+ }
+
+ /* copy the register from the old node to the new Load */
+ reg = arch_get_irn_register(env->cg->arch_env, irn);
+ arch_set_irn_register(env->cg->arch_env, new_op, reg);
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, new_op));
+
+ exchange(irn, proj);
}
/**
- * Create a Load for a Spill
+ * Transforms a be node into a Store.
*/
-static ir_node *ia32_lower_reload(void *self, ir_node *reload) {
- ia32_code_gen_t *cg = self;
- ir_graph *irg = cg->irg;
- dbg_info *dbg = get_irn_dbg_info(reload);
- ir_node *block = get_nodes_block(reload);
- ir_node *ptr = get_irg_frame(irg);
- ir_mode *mode = get_irn_mode(reload);
- ir_node *pred = get_irn_n(reload, 0);
- char buf[64];
- char *ofs;
- ir_node *noreg, *res;
-
- /* Get the offset to Load from. It can either be a Spill or a Store. */
- if (be_is_Spill(pred)) {
- entity *ent = be_get_spill_entity(pred);
- unsigned offs = get_entity_offset_bytes(ent);
- DB((cg->mod, LEVEL_1, "lower_reload: got offset %d for %+F\n", offs, ent));
-
- snprintf(buf, sizeof(buf), "%d", offs);
- }
- else if (is_ia32_Store(pred) || is_ia32_fStore(pred)) {
- ofs = get_ia32_am_offs(pred);
- strncpy(buf, ofs, sizeof(buf));
- free(ofs);
+static void transform_to_Store(ia32_transform_env_t *env) {
+ ir_node *irn = env->irn;
+ entity *ent = be_get_frame_entity(irn);
+ ir_mode *mode = env->mode;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *nomem = new_rd_NoMem(env->irg);
+ ir_node *ptr = get_irn_n(irn, 0);
+ ir_node *val = get_irn_n(irn, 1);
+ ir_node *new_op, *proj;
+ ir_node *sched_point = NULL;
+
+ if (sched_is_scheduled(irn)) {
+ sched_point = sched_prev(irn);
+ }
+
+ if (mode_is_float(mode)) {
+ if (USE_SSE2(env->cg))
+ new_op = new_rd_ia32_fStore(env->dbg, env->irg, env->block, ptr, noreg, val, nomem, mode_T);
+ else
+ new_op = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, nomem, mode_T);
+ }
+ else if (get_mode_size_bits(mode) == 8) {
+ new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, nomem, mode_T);
}
else {
- assert(0 && "unsupported Reload predecessor");
+ new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, nomem, mode_T);
}
- /* Create the Load */
+ set_ia32_am_support(new_op, ia32_am_Dest);
+ set_ia32_op_type(new_op, ia32_AddrModeD);
+ set_ia32_am_flavour(new_op, ia32_B);
+ set_ia32_ls_mode(new_op, mode);
+ set_ia32_frame_ent(new_op, ent);
+ set_ia32_use_frame(new_op);
+
+ DBG_OPT_SPILL2ST(irn, new_op);
+
+ proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode_M, 0);
+
+ if (sched_point) {
+ sched_add_after(sched_point, new_op);
+ sched_add_after(new_op, proj);
+
+ sched_remove(irn);
+ }
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, new_op));
+
+ exchange(irn, proj);
+}
+
+/**
+ * Fix the mode of Spill/Reload
+ */
+static ir_mode *fix_spill_mode(ia32_code_gen_t *cg, ir_mode *mode)
+{
if (mode_is_float(mode)) {
- noreg = ia32_new_NoReg_fp(cg);
- res = new_rd_ia32_fLoad(dbg, irg, block, ptr, noreg, pred, mode_T);
+ if (USE_SSE2(cg))
+ mode = mode_D;
+ else
+ mode = mode_E;
}
- else {
- noreg = ia32_new_NoReg_gp(cg);
- res = new_rd_ia32_Load(dbg, irg, block, ptr, noreg, pred, mode_T);
+ else
+ mode = mode_Is;
+ return mode;
+}
+
+/**
+ * Block-Walker: Calls the transform functions Spill and Reload.
+ */
+static void ia32_after_ra_walker(ir_node *block, void *env) {
+ ir_node *node, *prev;
+ ia32_code_gen_t *cg = env;
+ ia32_transform_env_t tenv;
+
+ tenv.block = block;
+ tenv.irg = current_ir_graph;
+ tenv.cg = cg;
+ DEBUG_ONLY(tenv.mod = cg->mod;)
+
+ /* beware: the schedule is changed here */
+ for (node = sched_last(block); !sched_is_begin(node); node = prev) {
+ prev = sched_prev(node);
+ if (be_is_Reload(node)) {
+ /* we always reload the whole register */
+ tenv.dbg = get_irn_dbg_info(node);
+ tenv.irn = node;
+ tenv.mode = fix_spill_mode(cg, get_irn_mode(node));
+ transform_to_Load(&tenv);
+ }
+ else if (be_is_Spill(node)) {
+ /* we always spill the whole register */
+ tenv.dbg = get_irn_dbg_info(node);
+ tenv.irn = node;
+ tenv.mode = fix_spill_mode(cg, get_irn_mode(be_get_Spill_context(node)));
+ transform_to_Store(&tenv);
+ }
}
+}
- /* Set offset */
- add_ia32_am_offs(res, buf);
+/**
+ * We transform Spill and Reload here. This needs to be done before
+ * stack biasing otherwise we would miss the corrected offset for these nodes.
+ *
+ * If x87 instruction should be emitted, run the x87 simulator and patch
+ * the virtual instructions. This must obviously be done after register allocation.
+ */
+static void ia32_after_ra(void *self) {
+ ia32_code_gen_t *cg = self;
+ irg_block_walk_graph(cg->irg, NULL, ia32_after_ra_walker, self);
- /* Return the result Proj */
- return new_rd_Proj(dbg, irg, block, res, mode, 0);
+ /* if we do x87 code generation, rewrite all the virtual instructions and registers */
+ if (cg->used_fp == fp_x87) {
+ x87_simulate_graph(cg->arch_env, cg->irg, cg->blk_sched);
+ }
}
+
/**
* Emits the code, closes the output file and frees
* the code generator interface.
*/
static void ia32_codegen(void *self) {
ia32_code_gen_t *cg = self;
- ir_graph *irg = cg->irg;
- FILE *out = cg->out;
-
- if (cg->emit_decls) {
- ia32_gen_decls(cg->out);
- cg->emit_decls = 0;
- }
+ ir_graph *irg = cg->irg;
ia32_finish_irg(irg, cg);
- //dump_ir_block_graph_sched(irg, "-finished");
- ia32_gen_routine(out, irg, cg);
+ be_dump(irg, "-finished", dump_ir_block_graph_sched);
+ ia32_gen_routine(cg->isa->out, irg, cg);
cur_reg_set = NULL;
- pmap_destroy(cg->tv_ent);
- pmap_destroy(cg->types);
+ /* remove it from the isa */
+ cg->isa->cg = NULL;
/* de-allocate code generator */
del_set(cg->reg_set);
free(self);
+
}
-static void *ia32_cg_init(FILE *F, ir_graph *irg, const arch_env_t *arch_env);
+static void *ia32_cg_init(const be_irg_t *birg);
static const arch_code_generator_if_t ia32_code_gen_if = {
ia32_cg_init,
+ NULL, /* before abi introduce hook */
ia32_prepare_graph,
ia32_before_sched, /* before scheduling hook */
ia32_before_ra, /* before register allocation hook */
- ia32_lower_spill,
- ia32_lower_reload,
+ ia32_after_ra, /* after register allocation hook */
ia32_codegen /* emit && done */
};
/**
- * Initializes the code generator.
+ * Initializes a IA32 code generator.
*/
-static void *ia32_cg_init(FILE *F, ir_graph *irg, const arch_env_t *arch_env) {
- ia32_isa_t *isa = (ia32_isa_t *)arch_env->isa;
+static void *ia32_cg_init(const be_irg_t *birg) {
+ ia32_isa_t *isa = (ia32_isa_t *)birg->main_env->arch_env->isa;
ia32_code_gen_t *cg = xcalloc(1, sizeof(*cg));
- cg->impl = &ia32_code_gen_if;
- cg->irg = irg;
- cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024);
- cg->mod = firm_dbg_register("firm.be.ia32.cg");
- cg->out = F;
- cg->arch_env = arch_env;
- cg->types = pmap_create();
- cg->tv_ent = pmap_create();
- cg->noreg_gp = NULL;
- cg->noreg_fp = NULL;
-
- isa->num_codegens++;
-
- if (isa->num_codegens > 1)
- cg->emit_decls = 0;
- else
- cg->emit_decls = 1;
+ cg->impl = &ia32_code_gen_if;
+ cg->irg = birg->irg;
+ cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024);
+ cg->arch_env = birg->main_env->arch_env;
+ cg->isa = isa;
+ cg->birg = birg;
+ cg->blk_sched = NULL;
+ cg->fp_to_gp = NULL;
+ cg->gp_to_fp = NULL;
+ cg->fp_kind = isa->fp_kind;
+ cg->used_fp = fp_none;
+
+ FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg");
+
+ /* copy optimizations from isa for easier access */
+ cg->opt = isa->opt;
+
+ /* enter it */
+ isa->cg = cg;
+
+#ifndef NDEBUG
+ if (isa->name_obst_size) {
+ //printf("freed %d bytes from name obst\n", isa->name_obst_size);
+ isa->name_obst_size = 0;
+ obstack_free(isa->name_obst, NULL);
+ obstack_init(isa->name_obst);
+ }
+#endif /* NDEBUG */
cur_reg_set = cg->reg_set;
*
*****************************************************************/
+/**
+ * The template that generates a new ISA object.
+ * Note that this template can be changed by command line
+ * arguments.
+ */
static ia32_isa_t ia32_isa_template = {
- &ia32_isa_if,
- &ia32_gp_regs[REG_ESP],
- &ia32_gp_regs[REG_EBP],
- -1,
- 0,
- NULL
+ {
+ &ia32_isa_if, /* isa interface implementation */
+ &ia32_gp_regs[REG_ESP], /* stack pointer register */
+ &ia32_gp_regs[REG_EBP], /* base pointer register */
+ -1, /* stack direction */
+ },
+ NULL, /* 16bit register names */
+ NULL, /* 8bit register names */
+ NULL, /* types */
+ NULL, /* tv_ents */
+ (0 |
+ IA32_OPT_DOAM | /* optimize address mode default: on */
+ IA32_OPT_PLACECNST | /* place constants immediately before instructions, default: on */
+ IA32_OPT_IMMOPS | /* operations can use immediates, default: on */
+ IA32_OPT_EXTBB), /* use extended basic block scheduling, default: on */
+ arch_pentium_4, /* instruction architecture */
+ arch_pentium_4, /* optimize for architecture */
+ fp_sse2, /* use sse2 unit */
+ NULL, /* current code generator */
+#ifndef NDEBUG
+ NULL, /* name obstack */
+ 0 /* name obst size */
+#endif
};
/**
* Initializes the backend ISA.
*/
-static void *ia32_init(void) {
+static void *ia32_init(FILE *file_handle) {
static int inited = 0;
ia32_isa_t *isa;
- if(inited)
+ if (inited)
return NULL;
- isa = xcalloc(1, sizeof(*isa));
+ isa = xmalloc(sizeof(*isa));
memcpy(isa, &ia32_isa_template, sizeof(*isa));
- isa->reg_projnum_map = new_set(ia32_cmp_reg_projnum_assoc, 1024);
-
ia32_register_init(isa);
ia32_create_opcodes();
+ ia32_register_copy_attr_func();
+
+ isa->regs_16bit = pmap_create();
+ isa->regs_8bit = pmap_create();
+ isa->types = pmap_create();
+ isa->tv_ent = pmap_create();
+ isa->out = file_handle;
+
+ ia32_build_16bit_reg_map(isa->regs_16bit);
+ ia32_build_8bit_reg_map(isa->regs_8bit);
+
+ /* patch register names of x87 registers */
+ if (USE_x87(isa)) {
+ ia32_st_regs[0].name = "st";
+ ia32_st_regs[1].name = "st(1)";
+ ia32_st_regs[2].name = "st(2)";
+ ia32_st_regs[3].name = "st(3)";
+ ia32_st_regs[4].name = "st(4)";
+ ia32_st_regs[5].name = "st(5)";
+ ia32_st_regs[6].name = "st(6)";
+ ia32_st_regs[7].name = "st(7)";
+ }
+
+#ifndef NDEBUG
+ isa->name_obst = xmalloc(sizeof(*isa->name_obst));
+ obstack_init(isa->name_obst);
+ isa->name_obst_size = 0;
+#endif /* NDEBUG */
+
+ fprintf(isa->out, "\t.intel_syntax\n");
inited = 1;
* Closes the output file and frees the ISA structure.
*/
static void ia32_done(void *self) {
+ ia32_isa_t *isa = self;
+
+ /* emit now all global declarations */
+ ia32_gen_decls(isa->out);
+
+ pmap_destroy(isa->regs_16bit);
+ pmap_destroy(isa->regs_8bit);
+ pmap_destroy(isa->tv_ent);
+ pmap_destroy(isa->types);
+
+#ifndef NDEBUG
+ //printf("name obst size = %d bytes\n", isa->name_obst_size);
+ obstack_free(isa->name_obst, NULL);
+#endif /* NDEBUG */
+
free(self);
}
-
+/**
+ * Return the number of register classes for this architecture.
+ * We report always these:
+ * - the general purpose registers
+ * - the floating point register set (depending on the unit used for FP)
+ * - MMX/SSE registers (currently not supported)
+ */
static int ia32_get_n_reg_class(const void *self) {
- return N_CLASSES;
+ return 2;
}
+/**
+ * Return the register class for index i.
+ */
static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
- assert(i >= 0 && i < N_CLASSES && "Invalid ia32 register class requested.");
- return &ia32_reg_classes[i];
+ const ia32_isa_t *isa = self;
+ assert(i >= 0 && i < 2 && "Invalid ia32 register class requested.");
+ if (i == 0)
+ return &ia32_reg_classes[CLASS_ia32_gp];
+ return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
}
/**
* @return A register class which can hold values of the given mode.
*/
const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self, const ir_mode *mode) {
- if (mode_is_float(mode))
- return &ia32_reg_classes[CLASS_ia32_fp];
+ const ia32_isa_t *isa = self;
+ if (mode_is_float(mode)) {
+ return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
+ }
else
return &ia32_reg_classes[CLASS_ia32_gp];
}
* @param method_type The type of the method (procedure) in question.
* @param abi The abi object to be modified
*/
-void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
+static void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
+ const ia32_isa_t *isa = self;
ir_type *tp;
ir_mode *mode;
unsigned cc = get_method_calling_convention(method_type);
int n = get_method_n_params(method_type);
int biggest_n = -1;
int stack_idx = 0;
- int i, ignore;
+ int i, ignore_1, ignore_2;
ir_mode **modes;
const arch_register_t *reg;
+ be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
+
+ /* set abi flags for calls */
+ call_flags.bits.left_to_right = 0; /* always last arg first on stack */
+ call_flags.bits.store_args_sequential = 0; /* use stores instead of push */
+ /* call_flags.bits.try_omit_fp not changed: can handle both settings */
+ call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */
+ call_flags.bits.call_has_imm = 1; /* IA32 calls can have immediate address */
/* set stack parameter passing style */
- be_abi_call_set_flags(abi, BE_ABI_FRAME_POINTER_DEDICATED, 4);
+ be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
/* collect the mode for each type */
modes = alloca(n * sizeof(modes[0]));
/* set register parameters */
if (cc & cc_reg_param) {
/* determine the number of parameters passed via registers */
- biggest_n = ia32_get_n_regparam_class(n, modes, &ignore, &ignore);
+ biggest_n = ia32_get_n_regparam_class(n, modes, &ignore_1, &ignore_2);
/* loop over all parameters and set the register requirements */
for (i = 0; i <= biggest_n; i++) {
/* set stack parameters */
for (i = stack_idx; i < n; i++) {
- be_abi_call_param_stack(abi, i);
+ be_abi_call_param_stack(abi, i, 1, 0, 0);
}
be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX]);
}
else if (n == 1) {
+ const arch_register_t *reg;
+
tp = get_method_res_type(method_type, 0);
+ assert(is_atomic_type(tp));
mode = get_type_mode(tp);
- if (mode_is_float(mode)) {
- be_abi_call_res_reg(abi, 1, &ia32_fp_regs[REG_XMM0]);
- }
- else {
- be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EAX]);
- }
+ reg = mode_is_float(mode) ?
+ (USE_SSE2(isa) ? &ia32_xmm_regs[REG_XMM0] : &ia32_vfp_regs[REG_VF0]) :
+ &ia32_gp_regs[REG_EAX];
+
+ be_abi_call_res_reg(abi, 0, reg);
}
}
return &ia32_irn_handler;
}
-long ia32_handle_call_proj(const void *self, ir_node *proj, int is_keep) {
- ia32_isa_t *isa = (ia32_isa_t *)self;
- long pn = get_Proj_proj(proj);
-
- if (!is_keep) {
- /* It's not a Keep proj, which means, that it is a result proj. */
- /* Possible result proj numbers are 0 and 1 */
- /* Set the correct register (depends on the mode) and the */
- /* corresponding proj number */
- if (mode_is_float(get_irn_mode(proj))) {
- assert(pn == 0 && "only one floating point result supported");
-
- /* Get the proj number for the floating point result */
- pn = ia32_get_reg_projnum(&ia32_fp_regs[REG_XMM0], isa->reg_projnum_map);
- }
- else {
- /* In case of 64bit return value, the result is */
- /* in EDX:EAX and we have two result projs. */
- switch (pn) {
- case 0:
- pn = ia32_get_reg_projnum(&ia32_gp_regs[REG_EAX], isa->reg_projnum_map);
- break;
- case 1:
- pn = ia32_get_reg_projnum(&ia32_gp_regs[REG_EDX], isa->reg_projnum_map);
- break;
- default:
- assert(0 && "only two int results supported");
- }
- }
-
- /* Set the correct proj number */
- set_Proj_proj(proj, pn);
- }
- else {
- /* Set mode to floating point if required */
- if (!strcmp(ia32_reg_classes[CLASS_ia32_fp].name,
- ia32_projnum_reg_req_map[pn]->req.cls->name)) {
- set_irn_mode(proj, mode_F);
- }
- }
-
- return pn;
-}
-
int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
return is_ia32_irn(irn);
}
* Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded
*/
static const list_sched_selector_t *ia32_get_list_sched_selector(const void *self) {
+// memcpy(&ia32_sched_selector, reg_pressure_selector, sizeof(list_sched_selector_t));
memcpy(&ia32_sched_selector, trivial_selector, sizeof(list_sched_selector_t));
ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule;
return &ia32_sched_selector;
}
+/**
+ * Returns the necessary byte alignment for storing a register of given class.
+ */
+static int ia32_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) {
+ ir_mode *mode = arch_register_class_mode(cls);
+ int bytes = get_mode_size_bytes(mode);
+
+ if (mode_is_float(mode) && bytes > 8)
+ return 16;
+ return bytes;
+}
+
#ifdef WITH_LIBCORE
+
+/* instruction set architectures. */
+static const lc_opt_enum_int_items_t arch_items[] = {
+ { "386", arch_i386, },
+ { "486", arch_i486, },
+ { "pentium", arch_pentium, },
+ { "586", arch_pentium, },
+ { "pentiumpro", arch_pentium_pro, },
+ { "686", arch_pentium_pro, },
+ { "pentiummmx", arch_pentium_mmx, },
+ { "pentium2", arch_pentium_2, },
+ { "p2", arch_pentium_2, },
+ { "pentium3", arch_pentium_3, },
+ { "p3", arch_pentium_3, },
+ { "pentium4", arch_pentium_4, },
+ { "p4", arch_pentium_4, },
+ { "pentiumm", arch_pentium_m, },
+ { "pm", arch_pentium_m, },
+ { "core", arch_core, },
+ { "k6", arch_k6, },
+ { "athlon", arch_athlon, },
+ { "athlon64", arch_athlon_64, },
+ { "opteron", arch_opteron, },
+ { NULL, 0 }
+};
+
+static lc_opt_enum_int_var_t arch_var = {
+ &ia32_isa_template.arch, arch_items
+};
+
+static lc_opt_enum_int_var_t opt_arch_var = {
+ &ia32_isa_template.opt_arch, arch_items
+};
+
+static const lc_opt_enum_int_items_t fp_unit_items[] = {
+ { "x87" , fp_x87 },
+ { "sse2", fp_sse2 },
+ { NULL, 0 }
+};
+
+static lc_opt_enum_int_var_t fp_unit_var = {
+ &ia32_isa_template.fp_kind, fp_unit_items
+};
+
+static const lc_opt_table_entry_t ia32_options[] = {
+ LC_OPT_ENT_ENUM_INT("arch", "select the instruction architecture", &arch_var),
+ LC_OPT_ENT_ENUM_INT("opt", "optimize for instruction architecture", &opt_arch_var),
+ LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &fp_unit_var),
+ LC_OPT_ENT_BIT("incdec", "optimize for inc/dec", &ia32_isa_template.opt, IA32_OPT_INCDEC),
+ LC_OPT_ENT_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM),
+ LC_OPT_ENT_NEGBIT("noplacecnst", "do not place constants", &ia32_isa_template.opt, IA32_OPT_PLACECNST),
+ LC_OPT_ENT_NEGBIT("noimmop", "no operations with immediates", &ia32_isa_template.opt, IA32_OPT_IMMOPS),
+ LC_OPT_ENT_NEGBIT("noextbb", "do not use extended basic block scheduling", &ia32_isa_template.opt, IA32_OPT_EXTBB),
+ { NULL }
+};
+
+/**
+ * Register command line options for the ia32 backend.
+ *
+ * Options so far:
+ *
+ * ia32-arch=arch create instruction for arch
+ * ia32-opt=arch optimize for run on arch
+ * ia32-fpunit=unit select floating point unit (x87 or SSE2)
+ * ia32-incdec optimize for inc/dec
+ * ia32-noaddrmode do not use address mode
+ * ia32-noplacecnst do not place constants,
+ * ia32-noimmop no operations with immediates
+ * ia32-noextbb do not use extended basic block scheduling
+ */
static void ia32_register_options(lc_opt_entry_t *ent)
{
+ lc_opt_entry_t *be_grp_ia32 = lc_opt_get_grp(ent, "ia32");
+ lc_opt_add_table(be_grp_ia32, ia32_options);
}
#endif /* WITH_LIBCORE */
const arch_isa_if_t ia32_isa_if = {
-#ifdef WITH_LIBCORE
- ia32_register_options,
-#endif
ia32_init,
ia32_done,
ia32_get_n_reg_class,
ia32_get_irn_handler,
ia32_get_code_generator_if,
ia32_get_list_sched_selector,
- ia32_handle_call_proj
+ ia32_get_reg_class_alignment,
+#ifdef WITH_LIBCORE
+ ia32_register_options
+#endif
};