/**
* This is the main ia32 firm backend driver.
- *
+ * @author Christian Wuerdig
* $Id$
*/
#include "ircons.h"
#include "irgmod.h"
#include "irgopt.h"
-
-#include "bitset.h"
+#include "irbitset.h"
+#include "pdeq.h"
#include "debug.h"
#include "../beabi.h" /* the general register allocator interface */
USE_SSE2(cg) ? &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG]);
}
+/* returns the first Proj with given mode from mode_T node */
+static ir_node *get_proj_for_mode(ir_node *node, ir_mode *mode) {
+ const ir_edge_t *edge;
+
+ assert(get_irn_mode(node) == mode_T && "Need mode_T node.");
+
+ foreach_out_edge(node, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ if (get_irn_mode(proj) == mode)
+ return proj;
+ }
+
+ return NULL;
+}
+
/**************************************************
* _ _ _ __
* | | | (_)/ _|
return arch_irn_class_branch;
else if (is_ia32_Cnst(irn))
return arch_irn_class_const;
+ else if (is_ia32_Ld(irn))
+ return arch_irn_class_load;
+ else if (is_ia32_St(irn) || is_ia32_Store8Bit(irn))
+ return arch_irn_class_store;
else if (is_ia32_irn(irn))
return arch_irn_class_normal;
else
return env;
}
+/**
+ * Put all registers which are saved by the prologue/epilogue in a set.
+ *
+ * @param self The callback object.
+ * @param s The result set.
+ */
static void ia32_abi_dont_save_regs(void *self, pset *s)
{
ia32_abi_env_t *env = self;
/**
* Generate the routine 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 map 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.
/**
* Generate the routine epilogue.
* @param self The callback object.
+ * @param bl The block for the epilog
* @param mem A pointer to the mem node. Update this if you define new memory.
* @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
* @return The register which shall be used as a stack frame base.
ir_mode *mode_bp = env->isa->bp->reg_class->mode;
int reg_size = get_mode_size_bytes(env->isa->bp->reg_class->mode);
- /* AMD processors prefer leave at the end of a routine */
- if (ARCH_AMD(isa->opt_arch)) {
+ /* gcc always emits a leave at the end of a routine */
+ if (1 || ARCH_AMD(isa->opt_arch)) {
ir_node *leave;
/* leave */
*/
static ir_type *ia32_abi_get_between_type(void *self)
{
+#define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
static ir_type *omit_fp_between_type = NULL;
static ir_type *between_type = NULL;
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);
+ ir_type *old_bp_type = new_type_primitive(IDENT("bp"), mode_P);
+ ir_type *ret_addr_type = new_type_primitive(IDENT("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);
+ between_type = new_type_struct(IDENT("ia32_between_type"));
+ old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type);
+ ret_addr_ent = new_entity(between_type, IDENT("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));
+ set_type_state(between_type, layout_fixed);
- 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);
+ omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp"));
+ omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("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));
+ set_type_state(omit_fp_between_type, layout_fixed);
}
return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
+#undef IDENT
+}
+
+/**
+ * Get the estimated cycle count for @p irn.
+ *
+ * @param self The this pointer.
+ * @param irn The node.
+ *
+ * @return The estimated cycle count for this operation
+ */
+static int ia32_get_op_estimated_cost(const void *self, const ir_node *irn)
+{
+ int cost;
+ switch (get_ia32_irn_opcode(irn)) {
+ case iro_ia32_xDiv:
+ case iro_ia32_DivMod:
+ cost = 8;
+ break;
+
+ case iro_ia32_xLoad:
+ case iro_ia32_l_Load:
+ case iro_ia32_Load:
+ case iro_ia32_Push:
+ case iro_ia32_Pop:
+ cost = 10;
+ break;
+
+ case iro_ia32_xStore:
+ case iro_ia32_l_Store:
+ case iro_ia32_Store:
+ case iro_ia32_Store8Bit:
+ cost = 50;
+ break;
+
+ case iro_ia32_MulS:
+ case iro_ia32_Mul:
+ case iro_ia32_Mulh:
+ case iro_ia32_xMul:
+ case iro_ia32_l_MulS:
+ case iro_ia32_l_Mul:
+ cost = 2;
+ break;
+
+ default:
+ cost = 1;
+ }
+
+ return cost;
+}
+
+/**
+ * Returns the inverse operation if @p irn, recalculating the argument at position @p i.
+ *
+ * @param irn The original operation
+ * @param i Index of the argument we want the inverse operation to yield
+ * @param inverse struct to be filled with the resulting inverse op
+ * @param obstack The obstack to use for allocation of the returned nodes array
+ * @return The inverse operation or NULL if operation invertible
+ */
+static arch_inverse_t *ia32_get_inverse(const void *self, const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst) {
+ ir_graph *irg;
+ ir_mode *mode;
+ ir_node *block, *noreg, *nomem;
+ int pnc;
+
+ /* we cannot invert non-ia32 irns */
+ if (! is_ia32_irn(irn))
+ return NULL;
+
+ /* operand must always be a real operand (not base, index or mem) */
+ if (i != 2 && i != 3)
+ return NULL;
+
+ /* we don't invert address mode operations */
+ if (get_ia32_op_type(irn) != ia32_Normal)
+ return NULL;
+
+ irg = get_irn_irg(irn);
+ block = get_nodes_block(irn);
+ mode = get_ia32_res_mode(irn);
+ noreg = get_irn_n(irn, 0);
+ nomem = new_r_NoMem(irg);
+
+ /* initialize structure */
+ inverse->nodes = obstack_alloc(obst, 2 * sizeof(inverse->nodes[0]));
+ inverse->costs = 0;
+ inverse->n = 2;
+
+ switch (get_ia32_irn_opcode(irn)) {
+ case iro_ia32_Add:
+ if (get_ia32_immop_type(irn) == ia32_ImmConst) {
+ /* we have an add with a const here */
+ /* invers == add with negated const */
+ inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
+ pnc = pn_ia32_Add_res;
+ inverse->costs += 1;
+ copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
+ set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn)));
+ set_ia32_commutative(inverse->nodes[0]);
+ }
+ else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) {
+ /* we have an add with a symconst here */
+ /* invers == sub with const */
+ inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
+ pnc = pn_ia32_Sub_res;
+ inverse->costs += 2;
+ copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
+ }
+ else {
+ /* normal add: inverse == sub */
+ ir_node *proj = get_irn_out_edge_first(irn)->src;
+ assert(proj && is_Proj(proj));
+
+ inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, i ^ 1), nomem);
+ pnc = pn_ia32_Sub_res;
+ inverse->costs += 2;
+ }
+ break;
+ case iro_ia32_Sub:
+ if (get_ia32_immop_type(irn) != ia32_ImmNone) {
+ /* we have a sub with a const/symconst here */
+ /* invers == add with this const */
+ inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
+ pnc = pn_ia32_Add_res;
+ inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
+ copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
+ }
+ else {
+ /* normal sub */
+ ir_node *proj = get_irn_out_edge_first(irn)->src;
+ assert(proj && is_Proj(proj));
+
+ if (i == 2) {
+ inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, 3), nomem);
+ }
+ else {
+ inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, 2), proj, nomem);
+ }
+ pnc = pn_ia32_Sub_res;
+ inverse->costs += 1;
+ }
+ break;
+ case iro_ia32_Eor:
+ if (get_ia32_immop_type(irn) != ia32_ImmNone) {
+ /* xor with const: inverse = xor */
+ inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
+ pnc = pn_ia32_Eor_res;
+ inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
+ copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
+ }
+ else {
+ /* normal xor */
+ inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, (ir_node *)irn, get_irn_n(irn, i), nomem);
+ pnc = pn_ia32_Eor_res;
+ inverse->costs += 1;
+ }
+ break;
+ case iro_ia32_Not:
+ inverse->nodes[0] = new_rd_ia32_Not(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), nomem);
+ pnc = pn_ia32_Not_res;
+ inverse->costs += 1;
+ break;
+ case iro_ia32_Minus:
+ inverse->nodes[0] = new_rd_ia32_Minus(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), nomem);
+ pnc = pn_ia32_Minus_res;
+ inverse->costs += 1;
+ break;
+ default:
+ /* inverse operation not supported */
+ return NULL;
+ }
+
+ set_ia32_res_mode(inverse->nodes[0], mode);
+ inverse->nodes[1] = new_r_Proj(irg, block, inverse->nodes[0], mode, pnc);
+
+ return inverse;
}
static const be_abi_callbacks_t ia32_abi_callbacks = {
ia32_classify,
ia32_get_flags,
ia32_get_frame_entity,
- ia32_set_stack_bias
+ ia32_set_stack_bias,
+ ia32_get_inverse,
+ ia32_get_op_estimated_cost
};
ia32_irn_ops_t ia32_irn_ops = {
*/
static void ia32_prepare_graph(void *self) {
ia32_code_gen_t *cg = self;
+ dom_front_info_t *dom;
DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;)
FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
+
+ /* 1st: transform constants and psi condition trees */
+ irg_walk_blkwise_graph(cg->irg, ia32_place_consts_set_modes, ia32_transform_psi_cond_tree, cg);
+
+ /* 2nd: transform all remaining nodes */
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);
+ dom = be_compute_dominance_frontiers(cg->irg);
+ irg_walk_blkwise_graph(cg->irg, NULL, ia32_transform_node, cg);
+ be_free_dominance_frontiers(dom);
- if (cg->opt & IA32_OPT_DOAM) {
- edges_deactivate(cg->irg);
- //dead_node_elimination(cg->irg);
- edges_activate(cg->irg);
+ if (cg->dump)
+ be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
- FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
+ /* 3rd: optimize address mode */
+ FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
+ ia32_optimize_addressmode(cg);
- irg_walk_blkwise_graph(cg->irg, NULL, ia32_optimize_am, cg);
+ if (cg->dump)
be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
- }
DEBUG_ONLY(cg->mod = old_mod;)
}
+static INLINE int need_constraint_copy(ir_node *irn) {
+ return \
+ ! is_ia32_Lea(irn) && \
+ ! is_ia32_Conv_I2I(irn) && \
+ ! is_ia32_Conv_I2I8Bit(irn) && \
+ ! is_ia32_CmpCMov(irn) && \
+ ! is_ia32_CmpSet(irn);
+}
/**
* Insert copies for all ia32 nodes where the should_be_same requirement
block = get_nodes_block(irn);
/* check all OUT requirements, if there is a should_be_same */
- if ((op_tp == ia32_Normal || op_tp == ia32_AddrModeS) &&
- ! is_ia32_Lea(irn) && ! is_ia32_Conv_I2I(irn) && ! is_ia32_Conv_I2I8Bit(irn))
+ if ((op_tp == ia32_Normal || op_tp == ia32_AddrModeS) && need_constraint_copy(irn))
{
for (i = 0; i < n_res; i++) {
if (arch_register_req_is(&(reqs[i]->req), should_be_same)) {
}
}
- /* If we have a CondJmp with immediate, we need to */
+ /* If we have a CondJmp/CmpSet/xCmpSet 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) {
+ if ((is_ia32_CondJmp(irn) || is_ia32_CmpSet(irn) || is_ia32_xCmpSet(irn)) &&
+ (is_ia32_ImmConst(irn) || is_ia32_ImmSymConst(irn)) &&
+ op_tp == ia32_AddrModeS)
+ {
set_ia32_op_type(irn, ia32_AddrModeD);
set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
}
static void ia32_finish_irg_walker(ir_node *block, void *env) {
ir_node *irn, *next;
- for (irn = sched_first(block); !sched_is_end(irn); irn = next) {
+ for (irn = sched_first(block); ! sched_is_end(irn); irn = next) {
next = sched_next(irn);
ia32_finish_node(irn, env);
}
}
+static void ia32_push_on_queue_walker(ir_node *block, void *env) {
+ waitq *wq = env;
+ waitq_put(wq, block);
+}
+
+
/**
* 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);
+ waitq *wq = new_waitq();
+
+ /* Push the blocks on the waitq because ia32_finish_irg_walker starts more walks ... */
+ irg_block_walk_graph(irg, NULL, ia32_push_on_queue_walker, wq);
+
+ while (! waitq_empty(wq)) {
+ ir_node *block = waitq_get(wq);
+ ia32_finish_irg_walker(block, cg);
+ }
+ del_waitq(wq);
}
static void ia32_before_sched(void *self) {
}
+static void remove_unused_nodes(ir_node *irn, bitset_t *already_visited) {
+ int i;
+ ir_mode *mode;
+ ir_node *mem_proj;
+
+ if (is_Block(irn))
+ return;
+
+ mode = get_irn_mode(irn);
+
+ /* check if we already saw this node or the node has more than one user */
+ if (bitset_contains_irn(already_visited, irn) || get_irn_n_edges(irn) > 1)
+ return;
+
+ /* mark irn visited */
+ bitset_add_irn(already_visited, irn);
+
+ /* non-Tuple nodes with one user: ok, return */
+ if (get_irn_n_edges(irn) >= 1 && mode != mode_T)
+ return;
+
+ /* tuple node has one user which is not the mem proj-> ok */
+ if (mode == mode_T && get_irn_n_edges(irn) == 1) {
+ mem_proj = get_proj_for_mode(irn, mode_M);
+ if (! mem_proj)
+ return;
+ }
+
+ for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
+ ir_node *pred = get_irn_n(irn, i);
+
+ /* do not follow memory edges or we will accidentally remove stores */
+ if (is_Proj(pred) && get_irn_mode(pred) == mode_M)
+ continue;
+
+ set_irn_n(irn, i, new_Bad());
+
+ /*
+ The current node is about to be removed: if the predecessor
+ has only this node as user, it need to be removed as well.
+ */
+ if (get_irn_n_edges(pred) <= 1)
+ remove_unused_nodes(pred, already_visited);
+ }
+
+ if (sched_is_scheduled(irn))
+ sched_remove(irn);
+}
+
+static void remove_unused_loads_walker(ir_node *irn, void *env) {
+ bitset_t *already_visited = env;
+ if (is_ia32_Ld(irn) && ! bitset_contains_irn(already_visited, irn))
+ remove_unused_nodes(irn, 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;
+ ia32_code_gen_t *cg = self;
+ bitset_t *already_visited = bitset_irg_malloc(cg->irg);
cg->blk_sched = sched_create_block_schedule(cg->irg);
+
+ /*
+ Handle special case:
+ There are sometimes unused loads, only pinned by memory.
+ We need to remove those Loads and all other nodes which won't be used
+ after removing the Load from schedule.
+ */
+ irg_walk_graph(cg->irg, remove_unused_loads_walker, NULL, already_visited);
+ bitset_free(already_visited);
}
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));
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
exchange(irn, proj);
}
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));
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
exchange(irn, proj);
}
irg_block_walk_graph(cg->irg, NULL, ia32_after_ra_walker, self);
/* if we do x87 code generation, rewrite all the virtual instructions and registers */
- if (cg->used_fp == fp_x87) {
+ if (cg->used_fp == fp_x87 || cg->force_sim) {
x87_simulate_graph(cg->arch_env, cg->irg, cg->blk_sched);
}
}
ir_graph *irg = cg->irg;
ia32_finish_irg(irg, cg);
- be_dump(irg, "-finished", dump_ir_block_graph_sched);
+ if (cg->dump)
+ be_dump(irg, "-finished", dump_ir_block_graph_sched);
ia32_gen_routine(cg->isa->out, irg, cg);
cur_reg_set = NULL;
cg->gp_to_fp = NULL;
cg->fp_kind = isa->fp_kind;
cg->used_fp = fp_none;
+ cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg");
* arguments.
*/
static ia32_isa_t ia32_isa_template = {
- {
- &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 */
- },
+ {
+ &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 |
+ (0 |
IA32_OPT_INCDEC | /* optimize add 1, sub 1 into inc/dec default: on */
IA32_OPT_DOAM | /* optimize address mode default: on */
+ IA32_OPT_LEA | /* optimize for LEAs 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 */
ia32_register_init(isa);
ia32_create_opcodes();
- ia32_register_copy_attr_func();
if ((ARCH_INTEL(isa->arch) && isa->arch < arch_pentium_4) ||
(ARCH_AMD(isa->arch) && isa->arch < arch_athlon))
/* 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)";
+ 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_size = 0;
#endif /* NDEBUG */
+ ia32_handle_intrinsics();
+ ia32_switch_section(NULL, NO_SECTION);
fprintf(isa->out, "\t.intel_syntax\n");
inited = 1;
* 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)
+ * - the SSE floating point register set
+ * - the virtual floating point registers
*/
static int ia32_get_n_reg_class(const void *self) {
- return 2;
+ return 3;
}
/**
*/
static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
const ia32_isa_t *isa = self;
- assert(i >= 0 && i < 2 && "Invalid ia32 register class requested.");
+ assert(i >= 0 && i < 3 && "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];
+ else if (i == 1)
+ return &ia32_reg_classes[CLASS_ia32_xmm];
+ else
+ return &ia32_reg_classes[CLASS_ia32_vfp];
}
/**
const arch_register_t *reg;
be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
+ unsigned use_push = !IS_P6_ARCH(isa->opt_arch);
+
/* 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.store_args_sequential = use_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 */
}
int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
- return is_ia32_irn(irn);
+ return is_ia32_irn(irn) ? 1 : -1;
}
/**
return bytes;
}
+static ia32_intrinsic_env_t intrinsic_env = { NULL, NULL };
+
+/**
+ * Returns the libFirm configuration parameter for this backend.
+ */
+static const backend_params *ia32_get_libfirm_params(void) {
+ static const arch_dep_params_t ad = {
+ 1, /* also use subs */
+ 4, /* maximum shifts */
+ 31, /* maximum shift amount */
+
+ 1, /* allow Mulhs */
+ 1, /* allow Mulus */
+ 32 /* Mulh allowed up to 32 bit */
+ };
+ static backend_params p = {
+ NULL, /* no additional opcodes */
+ NULL, /* will be set later */
+ 1, /* need dword lowering */
+ ia32_create_intrinsic_fkt,
+ &intrinsic_env, /* context for ia32_create_intrinsic_fkt */
+ };
+
+ p.dep_param = &ad;
+ return &p;
+}
#ifdef WITH_LIBCORE
/* instruction set architectures. */
&ia32_isa_template.fp_kind, fp_unit_items
};
+static const lc_opt_enum_int_items_t gas_items[] = {
+ { "linux", ASM_LINUX_GAS },
+ { "mingw", ASM_MINGW_GAS },
+ { NULL, 0 }
+};
+
+static lc_opt_enum_int_var_t gas_var = {
+ (int *)&asm_flavour, gas_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_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM),
+ 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_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM),
+ LC_OPT_ENT_NEGBIT("nolea", "do not optimize for LEAs", &ia32_isa_template.opt, IA32_OPT_LEA),
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),
+ 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),
+ LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
{ NULL }
};
* ia32-fpunit=unit select floating point unit (x87 or SSE2)
* ia32-incdec optimize for inc/dec
* ia32-noaddrmode do not use address mode
+ * ia32-nolea do not optimize for LEAs
* ia32-noplacecnst do not place constants,
* ia32-noimmop no operations with immediates
* ia32-noextbb do not use extended basic block scheduling
+ * ia32-gasmode set the GAS compatibility mode
*/
static void ia32_register_options(lc_opt_entry_t *ent)
{
ia32_get_code_generator_if,
ia32_get_list_sched_selector,
ia32_get_reg_class_alignment,
+ ia32_get_libfirm_params,
#ifdef WITH_LIBCORE
ia32_register_options
#endif
projects / libfirm / blobdiff