#include "iroptimize.h"
#include "instrument.h"
#include "iropt_t.h"
+#include "lower_dw.h"
#include "../beabi.h"
#include "../beirg.h"
#include "ia32_x87.h"
#include "ia32_dbg_stat.h"
#include "ia32_finish.h"
-#include "ia32_util.h"
#include "ia32_fpu.h"
#include "ia32_architecture.h"
ir_mode *ia32_mode_fpcw = NULL;
/** The current omit-fp state */
-static unsigned ia32_curr_fp_ommitted = 0;
static ir_type *omit_fp_between_type = NULL;
static ir_type *between_type = NULL;
static ir_entity *old_bp_ent = NULL;
if (is_ia32_Pop(node) || is_ia32_PopMem(node))
return -4;
- if (is_ia32_Leave(node) || (be_is_Copy(node)
- && arch_get_irn_register(node) == &ia32_registers[REG_ESP])) {
+ if (is_ia32_Leave(node) || is_ia32_CopyEbpEsp(node)) {
return SP_BIAS_RESET;
}
return 0;
}
-/**
- * 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.
- * @param stack_bias Points to the current stack bias, can be modified if needed.
- *
- * @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, int *stack_bias)
-{
- ia32_abi_env_t *env = (ia32_abi_env_t*)self;
- ir_graph *irg = env->irg;
- const arch_env_t *arch_env = be_get_irg_arch_env(irg);
-
- ia32_curr_fp_ommitted = env->flags.try_omit_fp;
- if (! env->flags.try_omit_fp) {
- ir_node *bl = get_irg_start_block(env->irg);
- ir_node *curr_sp = be_abi_reg_map_get(reg_map, arch_env->sp);
- ir_node *curr_bp = be_abi_reg_map_get(reg_map, arch_env->bp);
- ir_node *noreg = ia32_new_NoReg_gp(irg);
- ir_node *push;
-
- /* mark bp register as ignore */
- be_set_constr_single_reg_out(get_Proj_pred(curr_bp),
- get_Proj_proj(curr_bp), arch_env->bp, arch_register_req_type_ignore);
-
- /* push ebp */
- push = new_bd_ia32_Push(NULL, bl, noreg, noreg, *mem, curr_bp, curr_sp);
- arch_irn_add_flags(push, arch_irn_flags_prolog);
- curr_sp = new_r_Proj(push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
- *mem = new_r_Proj(push, mode_M, pn_ia32_Push_M);
-
- /* the push must have SP out register */
- arch_set_irn_register(curr_sp, arch_env->sp);
-
- /* this modifies the stack bias, because we pushed 32bit */
- *stack_bias -= 4;
-
- /* move esp to ebp */
- curr_bp = be_new_Copy(arch_env->bp->reg_class, bl, curr_sp);
- arch_irn_add_flags(curr_bp, arch_irn_flags_prolog);
- be_set_constr_single_reg_out(curr_bp, 0, arch_env->bp,
- arch_register_req_type_ignore);
-
- /* beware: the copy must be done before any other sp use */
- curr_sp = be_new_CopyKeep_single(arch_env->sp->reg_class, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
- arch_irn_add_flags(curr_sp, arch_irn_flags_prolog);
- be_set_constr_single_reg_out(curr_sp, 0, arch_env->sp,
- arch_register_req_type_produces_sp);
-
- be_abi_reg_map_set(reg_map, arch_env->sp, curr_sp);
- be_abi_reg_map_set(reg_map, arch_env->bp, curr_bp);
-
- return arch_env->bp;
- }
-
- return arch_env->sp;
-}
-
-/**
- * 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.
- *
- * All nodes which define registers in @p reg_map must keep @p reg_map current.
- */
-static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
-{
- ia32_abi_env_t *env = (ia32_abi_env_t*)self;
- const arch_env_t *arch_env = be_get_irg_arch_env(env->irg);
- ir_node *curr_sp = be_abi_reg_map_get(reg_map, arch_env->sp);
- ir_node *curr_bp = be_abi_reg_map_get(reg_map, arch_env->bp);
-
- if (env->flags.try_omit_fp) {
- /* simply remove the stack frame here */
- curr_sp = be_new_IncSP(arch_env->sp, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK, 0);
- arch_irn_add_flags(curr_sp, arch_irn_flags_epilog);
- } else {
- ir_mode *mode_bp = arch_env->bp->reg_class->mode;
-
- if (ia32_cg_config.use_leave) {
- ir_node *leave;
-
- /* leave */
- leave = new_bd_ia32_Leave(NULL, bl, curr_bp);
- curr_bp = new_r_Proj(leave, mode_bp, pn_ia32_Leave_frame);
- curr_sp = new_r_Proj(leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
- arch_irn_add_flags(leave, arch_irn_flags_epilog);
- } else {
- ir_node *pop;
-
- /* copy ebp to esp */
- curr_sp = be_new_Copy(&ia32_reg_classes[CLASS_ia32_gp], bl, curr_bp);
- arch_set_irn_register(curr_sp, arch_env->sp);
- be_set_constr_single_reg_out(curr_sp, 0, arch_env->sp,
- arch_register_req_type_ignore);
- arch_irn_add_flags(curr_sp, arch_irn_flags_epilog);
-
- /* pop ebp */
- pop = new_bd_ia32_PopEbp(NULL, bl, *mem, curr_sp);
- curr_bp = new_r_Proj(pop, mode_bp, pn_ia32_Pop_res);
- curr_sp = new_r_Proj(pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
- arch_irn_add_flags(pop, arch_irn_flags_epilog);
-
- *mem = new_r_Proj(pop, mode_M, pn_ia32_Pop_M);
- }
- arch_set_irn_register(curr_sp, arch_env->sp);
- arch_set_irn_register(curr_bp, arch_env->bp);
- }
-
- be_abi_reg_map_set(reg_map, arch_env->sp, curr_sp);
- be_abi_reg_map_set(reg_map, arch_env->bp, curr_bp);
-}
-
-/**
- * Initialize the callback object.
- * @param call The call object.
- * @param irg The graph with the method.
- * @return Some pointer. This pointer is passed to all other callback functions as self object.
- */
-static void *ia32_abi_init(const be_abi_call_t *call, ir_graph *irg)
-{
- ia32_abi_env_t *env = XMALLOC(ia32_abi_env_t);
- be_abi_call_flags_t fl = be_abi_call_get_flags(call);
- env->flags = fl.bits;
- env->irg = irg;
- return env;
-}
-
-/**
- * Destroy the callback object.
- * @param self The callback object.
- */
-static void ia32_abi_done(void *self)
-{
- free(self);
-}
-
/**
* Build the between type and entities if not already build.
*/
* 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 *ia32_abi_get_between_type(ir_graph *irg)
{
- ia32_abi_env_t *env = (ia32_abi_env_t*)self;
-
+ const be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
ia32_build_between_type();
- return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
+ return layout->sp_relative ? omit_fp_between_type : between_type;
}
/**
* Return the stack entity that contains the return address.
*/
-ir_entity *ia32_get_return_address_entity(void)
+ir_entity *ia32_get_return_address_entity(ir_graph *irg)
{
+ const be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
ia32_build_between_type();
- return ia32_curr_fp_ommitted ? omit_fp_ret_addr_ent : ret_addr_ent;
+ return layout->sp_relative ? omit_fp_ret_addr_ent : ret_addr_ent;
}
/**
* Return the stack entity that contains the frame address.
*/
-ir_entity *ia32_get_frame_address_entity(void)
+ir_entity *ia32_get_frame_address_entity(ir_graph *irg)
{
+ const be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
ia32_build_between_type();
- return ia32_curr_fp_ommitted ? NULL : old_bp_ent;
+ return layout->sp_relative ? NULL : old_bp_ent;
}
/**
mode = get_irn_mode(irn);
irn_mode = get_irn_mode(irn);
noreg = get_irn_n(irn, 0);
- nomem = new_r_NoMem(irg);
+ nomem = get_irg_no_mem(irg);
dbg = get_irn_dbg_info(irn);
/* initialize structure */
switch (get_ia32_irn_opcode(irn)) {
case iro_ia32_Add:
-#if 0
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_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, i ^ 1));
inverse->costs += 2;
}
-#endif
break;
case iro_ia32_Sub:
-#if 0
if (get_ia32_immop_type(irn) != ia32_ImmNone) {
/* we have a sub with a const/symconst here */
/* invers == add with this const */
}
inverse->costs += 1;
}
-#endif
break;
case iro_ia32_Xor:
-#if 0
if (get_ia32_immop_type(irn) != ia32_ImmNone) {
/* xor with const: inverse = xor */
inverse->nodes[0] = new_bd_ia32_Xor(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
inverse->nodes[0] = new_bd_ia32_Xor(dbg, block, noreg, noreg, nomem, (ir_node *) irn, get_irn_n(irn, i));
inverse->costs += 1;
}
-#endif
break;
case iro_ia32_Not: {
inverse->nodes[0] = new_bd_ia32_Not(dbg, block, (ir_node*) irn);
}
static const be_abi_callbacks_t ia32_abi_callbacks = {
- ia32_abi_init,
- ia32_abi_done,
ia32_abi_get_between_type,
- ia32_abi_prologue,
- ia32_abi_epilogue
};
/* register allocator interface */
ia32_copy_am_attrs(load, node);
if (is_ia32_is_reload(node))
set_ia32_is_reload(load);
- set_irn_n(node, n_ia32_mem, new_r_NoMem(irg));
+ set_irn_n(node, n_ia32_mem, get_irg_no_mem(irg));
switch (get_ia32_am_support(node)) {
case ia32_am_unary:
ir_node *noreg = ia32_new_NoReg_gp(irg);
ir_node *sched_point = NULL;
ir_node *ptr = get_irg_frame(irg);
- ir_node *mem = get_irn_n(node, be_pos_Reload_mem);
+ ir_node *mem = get_irn_n(node, n_be_Reload_mem);
ir_node *new_op, *proj;
const arch_register_t *reg;
dbg_info *dbg = get_irn_dbg_info(node);
ir_node *block = get_nodes_block(node);
ir_entity *ent = be_get_frame_entity(node);
- const ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
+ const ir_node *spillval = get_irn_n(node, n_be_Spill_val);
ir_mode *mode = get_spill_mode(spillval);
ir_node *noreg = ia32_new_NoReg_gp(irg);
- ir_node *nomem = new_r_NoMem(irg);
+ ir_node *nomem = get_irg_no_mem(irg);
ir_node *ptr = get_irg_frame(irg);
- ir_node *val = get_irn_n(node, be_pos_Spill_val);
+ ir_node *val = get_irn_n(node, n_be_Spill_val);
+ ir_node *res;
ir_node *store;
ir_node *sched_point = NULL;
}
if (mode_is_float(mode)) {
- if (ia32_cg_config.use_sse2)
+ if (ia32_cg_config.use_sse2) {
store = new_bd_ia32_xStore(dbg, block, ptr, noreg, nomem, val);
- else
+ res = new_r_Proj(store, mode_M, pn_ia32_xStore_M);
+ } else {
store = new_bd_ia32_vfst(dbg, block, ptr, noreg, nomem, val, mode);
+ res = new_r_Proj(store, mode_M, pn_ia32_vfst_M);
+ }
} else if (get_mode_size_bits(mode) == 128) {
/* Spill 128 bit SSE registers */
store = new_bd_ia32_xxStore(dbg, block, ptr, noreg, nomem, val);
+ res = new_r_Proj(store, mode_M, pn_ia32_xxStore_M);
} else if (get_mode_size_bits(mode) == 8) {
store = new_bd_ia32_Store8Bit(dbg, block, ptr, noreg, nomem, val);
+ res = new_r_Proj(store, mode_M, pn_ia32_Store8Bit_M);
} else {
store = new_bd_ia32_Store(dbg, block, ptr, noreg, nomem, val);
+ res = new_r_Proj(store, mode_M, pn_ia32_Store_M);
}
set_ia32_op_type(store, ia32_AddrModeD);
sched_remove(node);
}
- exchange(node, store);
+ exchange(node, res);
}
static ir_node *create_push(ir_node *node, ir_node *schedpoint, ir_node *sp, ir_node *mem, ir_entity *ent)
ir_node *noreg = ia32_new_NoReg_gp(irg);
ir_node *frame = get_irg_frame(irg);
- ir_node *pop = new_bd_ia32_PopMem(dbg, block, frame, noreg, new_r_NoMem(irg), sp);
+ ir_node *pop = new_bd_ia32_PopMem(dbg, block, frame, noreg,
+ get_irg_no_mem(irg), sp);
set_ia32_frame_ent(pop, ent);
set_ia32_use_frame(pop);
{
ir_node *block = get_nodes_block(node);
ir_graph *irg = get_irn_irg(node);
- ir_node *sp = be_abi_get_ignore_irn(be_get_irg_abi(irg), &ia32_registers[REG_ESP]);
+ ir_node *sp = be_get_initial_reg_value(irg, &ia32_registers[REG_ESP]);
int arity = be_get_MemPerm_entity_arity(node);
ir_node **pops = ALLOCAN(ir_node*, arity);
ir_node *in[1];
sp = create_spproj(node, push, pn_ia32_Push_stack);
}
- set_irn_n(node, i, new_r_Bad(irg));
+ set_irn_n(node, i, new_r_Bad(irg, mode_X));
}
/* create pops */
}
/* remove memperm */
- arity = get_irn_arity(node);
- for (i = 0; i < arity; ++i) {
- set_irn_n(node, i, new_r_Bad(irg));
- }
sched_remove(node);
+ kill_node(node);
}
/**
be_node_needs_frame_entity(env, node, mode, align);
}
+static int determine_ebp_input(ir_node *ret)
+{
+ const arch_register_t *bp = &ia32_registers[REG_EBP];
+ int arity = get_irn_arity(ret);
+ int i;
+
+ for (i = 0; i < arity; ++i) {
+ ir_node *input = get_irn_n(ret, i);
+ if (arch_get_irn_register(input) == bp)
+ return i;
+ }
+ panic("no ebp input found at %+F", ret);
+}
+
+static void introduce_epilog(ir_node *ret)
+{
+ const arch_register_t *sp = &ia32_registers[REG_ESP];
+ const arch_register_t *bp = &ia32_registers[REG_EBP];
+ ir_graph *irg = get_irn_irg(ret);
+ ir_type *frame_type = get_irg_frame_type(irg);
+ unsigned frame_size = get_type_size_bytes(frame_type);
+ be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
+ ir_node *block = get_nodes_block(ret);
+ ir_node *first_sp = get_irn_n(ret, n_be_Return_sp);
+ ir_node *curr_sp = first_sp;
+ ir_mode *mode_gp = mode_Iu;
+
+ if (!layout->sp_relative) {
+ int n_ebp = determine_ebp_input(ret);
+ ir_node *curr_bp = get_irn_n(ret, n_ebp);
+ if (ia32_cg_config.use_leave) {
+ ir_node *leave = new_bd_ia32_Leave(NULL, block, curr_bp);
+ curr_bp = new_r_Proj(leave, mode_gp, pn_ia32_Leave_frame);
+ curr_sp = new_r_Proj(leave, mode_gp, pn_ia32_Leave_stack);
+ arch_set_irn_register(curr_bp, bp);
+ arch_set_irn_register(curr_sp, sp);
+ sched_add_before(ret, leave);
+ } else {
+ ir_node *pop;
+ ir_node *curr_mem = get_irn_n(ret, n_be_Return_mem);
+ /* copy ebp to esp */
+ curr_sp = new_bd_ia32_CopyEbpEsp(NULL, block, curr_bp);
+ arch_set_irn_register(curr_sp, sp);
+ sched_add_before(ret, curr_sp);
+
+ /* pop ebp */
+ pop = new_bd_ia32_PopEbp(NULL, block, curr_mem, curr_sp);
+ curr_bp = new_r_Proj(pop, mode_gp, pn_ia32_PopEbp_res);
+ curr_sp = new_r_Proj(pop, mode_gp, pn_ia32_PopEbp_stack);
+ curr_mem = new_r_Proj(pop, mode_M, pn_ia32_Pop_M);
+ arch_set_irn_register(curr_bp, bp);
+ arch_set_irn_register(curr_sp, sp);
+ sched_add_before(ret, pop);
+
+ set_irn_n(ret, n_be_Return_mem, curr_mem);
+ }
+ set_irn_n(ret, n_ebp, curr_bp);
+ } else {
+ ir_node *incsp = be_new_IncSP(sp, block, curr_sp, -(int)frame_size, 0);
+ sched_add_before(ret, incsp);
+ curr_sp = incsp;
+ }
+ set_irn_n(ret, n_be_Return_sp, curr_sp);
+
+ /* keep verifier happy... */
+ if (get_irn_n_edges(first_sp) == 0 && is_Proj(first_sp)) {
+ kill_node(first_sp);
+ }
+}
+
+/**
+ * put the Prolog code at the beginning, epilog code before each return
+ */
+static void introduce_prolog_epilog(ir_graph *irg)
+{
+ const arch_register_t *sp = &ia32_registers[REG_ESP];
+ const arch_register_t *bp = &ia32_registers[REG_EBP];
+ ir_node *start = get_irg_start(irg);
+ ir_node *block = get_nodes_block(start);
+ ir_type *frame_type = get_irg_frame_type(irg);
+ unsigned frame_size = get_type_size_bytes(frame_type);
+ be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
+ ir_node *initial_sp = be_get_initial_reg_value(irg, sp);
+ ir_node *curr_sp = initial_sp;
+ ir_mode *mode_gp = mode_Iu;
+
+ if (!layout->sp_relative) {
+ /* push ebp */
+ ir_node *mem = get_irg_initial_mem(irg);
+ ir_node *noreg = ia32_new_NoReg_gp(irg);
+ ir_node *initial_bp = be_get_initial_reg_value(irg, bp);
+ ir_node *curr_bp = initial_bp;
+ ir_node *push = new_bd_ia32_Push(NULL, block, noreg, noreg, mem, curr_bp, curr_sp);
+ ir_node *incsp;
+
+ curr_sp = new_r_Proj(push, mode_gp, pn_ia32_Push_stack);
+ mem = new_r_Proj(push, mode_M, pn_ia32_Push_M);
+ arch_set_irn_register(curr_sp, sp);
+ sched_add_after(start, push);
+
+ /* move esp to ebp */
+ curr_bp = be_new_Copy(bp->reg_class, block, curr_sp);
+ sched_add_after(push, curr_bp);
+ be_set_constr_single_reg_out(curr_bp, 0, bp, arch_register_req_type_ignore);
+ curr_sp = be_new_CopyKeep_single(sp->reg_class, block, curr_sp, curr_bp, mode_gp);
+ sched_add_after(curr_bp, curr_sp);
+ be_set_constr_single_reg_out(curr_sp, 0, sp, arch_register_req_type_produces_sp);
+ edges_reroute(initial_bp, curr_bp);
+ set_irn_n(push, n_ia32_Push_val, initial_bp);
+
+ incsp = be_new_IncSP(sp, block, curr_sp, frame_size, 0);
+ edges_reroute(initial_sp, incsp);
+ set_irn_n(push, n_ia32_Push_stack, initial_sp);
+ sched_add_after(curr_sp, incsp);
+
+ layout->initial_bias = -4;
+ } else {
+ ir_node *incsp = be_new_IncSP(sp, block, curr_sp, frame_size, 0);
+ edges_reroute(initial_sp, incsp);
+ be_set_IncSP_pred(incsp, curr_sp);
+ sched_add_after(start, incsp);
+ }
+
+ /* introduce epilog for every return node */
+ {
+ ir_node *end_block = get_irg_end_block(irg);
+ int arity = get_irn_arity(end_block);
+ int i;
+
+ for (i = 0; i < arity; ++i) {
+ ir_node *ret = get_irn_n(end_block, i);
+ assert(be_is_Return(ret));
+ introduce_epilog(ret);
+ }
+ }
+}
+
/**
* We transform Spill and Reload here. This needs to be done before
* stack biasing otherwise we would miss the corrected offset for these nodes.
be_free_frame_entity_coalescer(fec_env);
irg_block_walk_graph(irg, NULL, ia32_after_ra_walker, NULL);
+
+ introduce_prolog_epilog(irg);
}
/**
&ia32_registers[REG_ESP], /* stack pointer register */
&ia32_registers[REG_EBP], /* base pointer register */
&ia32_reg_classes[CLASS_ia32_gp], /* static link pointer register class */
- -1, /* stack direction */
2, /* power of two stack alignment, 2^2 == 4 */
NULL, /* main environment */
7, /* costs for a spill instruction */
set_tarval_output_modes();
- memcpy(isa, &ia32_isa_template, sizeof(*isa));
+ *isa = ia32_isa_template;
if (ia32_mode_fpcw == NULL) {
ia32_mode_fpcw = new_ir_mode("Fpcw", irms_int_number, 16, 0, irma_none, 0);
if (get_mode_size_bits(mode) > 32)
return false;
/* we can handle Abs for all modes and compares (except 64bit) */
- if (be_mux_is_abs(sel, mux_true, mux_false) != 0)
+ if (ir_mux_is_abs(sel, mux_true, mux_false) != 0)
return true;
/* we can't handle MuxF yet */
if (mode_is_float(mode))
/* lower compound param handling */
lower_calls_with_compounds(¶ms);
- lower_dw_ops(&lower_dw_params);
+ ir_prepare_dw_lowering(&lower_dw_params);
+ ir_lower_dw_ops();
for (i = 0; i < n_irgs; ++i) {
ir_graph *irg = get_irp_irg(i);
ia32_emit, /* emit && done */
};
-BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_ia32);
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_ia32)
void be_init_arch_ia32(void)
{
lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
projects / libfirm / blobdiff