+/*
+ * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
+ *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
/**
- * This is the main ia32 firm backend driver.
- * @author Christian Wuerdig
- * $Id$
+ * @file
+ * @brief This is the main ia32 firm backend driver.
+ * @author Christian Wuerdig
+ * @version $Id$
*/
#ifdef HAVE_CONFIG_H
-#include <config.h>
-#endif
-
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
-#ifdef HAVE_ALLOCA_H
-#include <alloca.h>
+#include "config.h"
#endif
#include <libcore/lc_opts.h>
#include "pdeq.h"
#include "pset.h"
#include "debug.h"
+#include "error.h"
+#include "xmalloc.h"
-#include "../beabi.h" /* the general register allocator interface */
+#include "../beabi.h"
+#include "../beirg_t.h"
#include "../benode_t.h"
#include "../belower.h"
#include "../besched_t.h"
#include "../bespillslots.h"
#include "../bemodule.h"
#include "../begnuas.h"
+#include "../bestate.h"
#include "bearch_ia32_t.h"
-#include "ia32_new_nodes.h" /* ia32 nodes interface */
-#include "gen_ia32_regalloc_if.h" /* the generated interface (register type and class defenitions) */
+#include "ia32_new_nodes.h"
+#include "gen_ia32_regalloc_if.h"
#include "gen_ia32_machine.h"
#include "ia32_transform.h"
#include "ia32_emitter.h"
#include "ia32_dbg_stat.h"
#include "ia32_finish.h"
#include "ia32_util.h"
+#include "ia32_fpu.h"
+
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
/* TODO: ugly */
static set *cur_reg_set = NULL;
+ir_mode *mode_fpcw = NULL;
+
typedef ir_node *(*create_const_node_func) (dbg_info *dbg, ir_graph *irg, ir_node *block);
static INLINE ir_node *create_const(ia32_code_gen_t *cg, ir_node **place,
- create_const_node_func func, arch_register_t* reg)
+ create_const_node_func func,
+ const arch_register_t* reg)
{
ir_node *block, *res;
- ir_node *in[1];
- ir_node *startnode;
- ir_node *keep;
if(*place != NULL)
return *place;
arch_set_irn_register(cg->arch_env, res, reg);
*place = res;
- /* keep the node so it isn't accidently removed when unused ... */
- in[0] = res;
- keep = be_new_Keep(arch_register_get_class(reg), cg->irg, block, 1, in);
-
- /* schedule the node if we already have a scheduled program */
- startnode = get_irg_start(cg->irg);
- if(sched_is_scheduled(startnode)) {
- sched_add_after(startnode, res);
- sched_add_after(res, keep);
- }
+ add_irn_dep(get_irg_end(cg->irg), res);
+ /* add_irn_dep(get_irg_start(cg->irg), res); */
return res;
}
&ia32_xmm_regs[REG_XMM_UKNWN]);
}
+ir_node *ia32_new_Fpu_truncate(ia32_code_gen_t *cg) {
+ return create_const(cg, &cg->fpu_trunc_mode, new_rd_ia32_ChangeCW,
+ &ia32_fp_cw_regs[REG_FPCW]);
+}
+
/**
* Returns gp_noreg or fp_noreg, depending in input requirements.
if (is_ia32_Ld(irn))
classification |= arch_irn_class_load;
- if (is_ia32_St(irn) || is_ia32_Store8Bit(irn))
+ if (is_ia32_St(irn))
classification |= arch_irn_class_store;
if (is_ia32_need_stackent(irn))
pset_insert_ptr(s, env->isa->bp);
}
-#if 0
-static unsigned count_callee_saves(ia32_code_gen_t *cg)
-{
- unsigned callee_saves = 0;
- int c, num_reg_classes;
- arch_isa_if_t *isa;
-
- num_reg_classes = arch_isa_get_n_reg_class(isa);
- for(c = 0; c < num_reg_classes; ++c) {
- int r, num_registers;
- arch_register_class_t *regclass = arch_isa_get_reg_class(isa, c);
-
- num_registers = arch_register_class_n_regs(regclass);
- for(r = 0; r < num_registers; ++r) {
- arch_register_t *reg = arch_register_for_index(regclass, r);
- if(arch_register_type_is(reg, callee_save))
- callee_saves++;
- }
- }
-
- return callee_saves;
-}
-
-static void create_callee_save_regprojs(ia32_code_gen_t *cg, ir_node *regparams)
-{
- int c, num_reg_classes;
- arch_isa_if_t *isa;
- long n = 0;
-
- num_reg_classes = arch_isa_get_n_reg_class(isa);
- cg->initial_regs = obstack_alloc(cg->obst,
- num_reg_classes * sizeof(cg->initial_regs[0]));
-
- for(c = 0; c < num_reg_classes; ++c) {
- int r, num_registers;
- ir_node **initial_regclass;
- arch_register_class_t *regclass = arch_isa_get_reg_class(isa, c);
-
- num_registers = arch_register_class_n_regs(regclass);
- initial_regclass = obstack_alloc(num_registers * sizeof(initial_regclass[0]));
- for(r = 0; r < num_registers; ++r) {
- ir_node *proj;
- arch_register_t *reg = arch_register_for_index(regclass, r);
- if(!arch_register_type_is(reg, callee_save))
- continue;
-
- proj = new_r_Proj(irg, start_block, regparams, n);
- be_set_constr_single_reg(regparams, n, reg);
- arch_set_irn_register(cg->arch_env, proj, reg);
-
- initial_regclass[r] = proj;
- n++;
- }
- cg->initial_regs[c] = initial_regclass;
- }
-}
-
-static void callee_saves_obstack_grow(ia32_code_gen_t *cg)
-{
- int c, num_reg_classes;
- arch_isa_if_t *isa;
-
- for(c = 0; c < num_reg_classes; ++c) {
- int r, num_registers;
-
- num_registers = arch_register_class_n_regs(regclass);
- for(r = 0; r < num_registers; ++r) {
- ir_node *proj;
- arch_register_t *reg = arch_register_for_index(regclass, r);
- if(!arch_register_type_is(reg, callee_save))
- continue;
-
- proj = cg->initial_regs[c][r];
- obstack_ptr_grow(cg->obst, proj);
- }
- }
-}
-
-static unsigned count_parameters_in_regs(ia32_code_gen_t *cg)
-{
- return 0;
-}
-
-static void ia32_gen_prologue(ia32_code_gen_t *cg)
-{
- ir_graph *irg = cg->irg;
- ir_node *start_block = get_irg_start_block(irg);
- ir_node *sp;
- ir_node *regparams;
- int n_regparams_out;
-
- /* Create the regparams node */
- n_regparams_out = count_callee_saves(cg) + count_parameters_in_regs(cg);
- regparams = be_new_RegParams(irg, start_block, n_regparams_out);
-
- create_callee_save_regprojs(cg, regparams);
-
- /* Setup the stack */
- if(!omit_fp) {
- 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 *noreg = ia32_new_NoReg_gp(cg);
- ir_node *push;
-
- /* push ebp */
- push = new_rd_ia32_Push(NULL, env->irg, bl, noreg, noreg, curr_bp, curr_sp, *mem);
- curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
- *mem = new_r_Proj(env->irg, bl, push, mode_M, pn_ia32_Push_M);
-
- /* the push must have SP out register */
- arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
- set_ia32_flags(push, arch_irn_flags_ignore);
-
- /* 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);
- }
-
- sp = be_new_IncSP(sp, irg, start_block, initialsp, BE_STACK_FRAME_SIZE_EXPAND);
- set_irg_frame(irg, sp);
-}
-
-static void ia32_gen_epilogue(ia32_code_gen_t *cg)
-{
- int n_callee_saves = count_callee_saves(cg);
- int n_results_regs = 0;
- int barrier_size;
- ir_node *barrier;
- ir_node *end_block = get_irg_end_block(irg);
- ir_node **in;
-
- /* We have to make sure that all reloads occur before the stack frame
- gets destroyed, so we create a barrier for all callee-save and return
- values here */
- barrier_size = n_callee_saves + n_results_regs;
- barrier = be_new_Barrier(irg, end_block, barrier_size,
-
- /* simply remove the stack frame here */
- curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK);
- add_irn_dep(curr_sp, *mem);
-}
-#endif
-
/**
* Generate the routine prologue.
*
ir_node *noreg = ia32_new_NoReg_gp(cg);
ir_node *push;
+ /* ALL nodes representing bp must be set to ignore. */
+ be_node_set_flags(get_Proj_pred(curr_bp), BE_OUT_POS(get_Proj_proj(curr_bp)), arch_irn_flags_ignore);
+
/* push ebp */
push = new_rd_ia32_Push(NULL, env->irg, bl, noreg, noreg, curr_bp, curr_sp, *mem);
curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
ia32_abi_get_between_type,
ia32_abi_dont_save_regs,
ia32_abi_prologue,
- ia32_abi_epilogue,
+ ia32_abi_epilogue
};
/* fill register allocator interface */
*/
static void ia32_prepare_graph(void *self) {
ia32_code_gen_t *cg = self;
- DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;)
-
- FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
/* transform psi condition trees */
ia32_pre_transform_phase(cg);
/* transform all remaining nodes */
ia32_transform_graph(cg);
+ //add_fpu_edges(cg->birg);
// Matze: disabled for now. Because after transformation start block has no
// self-loop anymore so it might be merged with its successor block. This
// will bring several nodes to the startblock which sometimes get scheduled
// before the initial IncSP/Barrier
- //local_optimize_graph(cg->irg);
+ local_optimize_graph(cg->irg);
if (cg->dump)
be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
/* optimize address mode */
- FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
- ia32_optimize_addressmode(cg);
+ ia32_optimize_graph(cg);
+
+ if (cg->dump)
+ be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
/* do code placement, to optimize the position of constants */
place_code(cg->irg);
if (cg->dump)
- be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
-
- DEBUG_ONLY(cg->mod = old_mod;)
+ be_dump(cg->irg, "-place", dump_ir_block_graph_sched);
}
/**
after removing the Load from schedule.
*/
irg_walk_graph(cg->irg, NULL, remove_unused_loads_walker, already_visited);
+
+ /* setup fpu rounding modes */
+ ia32_setup_fpu_mode(cg);
}
sched_point = sched_prev(node);
}
+ /* No need to spill unknown values... */
+ if(is_ia32_Unknown_GP(val) ||
+ is_ia32_Unknown_VFP(val) ||
+ is_ia32_Unknown_XMM(val)) {
+ store = nomem;
+ if(sched_point)
+ sched_remove(node);
+
+ exchange(node, store);
+ return;
+ }
+
if (mode_is_float(mode)) {
if (USE_SSE2(cg))
store = new_rd_ia32_xStore(dbg, irg, block, ptr, noreg, val, nomem);
else
store = new_rd_ia32_vfst(dbg, irg, block, ptr, noreg, val, nomem);
- }
- else if (get_mode_size_bits(mode) == 128) {
+ } else if (get_mode_size_bits(mode) == 128) {
// Spill 128 bit SSE registers
store = new_rd_ia32_xxStore(dbg, irg, block, ptr, noreg, val, nomem);
- }
- else if (get_mode_size_bits(mode) == 8) {
+ } else if (get_mode_size_bits(mode) == 8) {
store = new_rd_ia32_Store8Bit(dbg, irg, block, ptr, noreg, val, nomem);
- }
- else {
+ } else {
store = new_rd_ia32_Store(dbg, irg, block, ptr, noreg, val, nomem);
}
set_ia32_ls_mode(store, mode);
set_ia32_frame_ent(store, ent);
set_ia32_use_frame(store);
-
- DBG_OPT_SPILL2ST(node, store);
SET_IA32_ORIG_NODE(store, ia32_get_old_node_name(cg, node));
+ DBG_OPT_SPILL2ST(node, store);
if (sched_point) {
sched_add_after(sched_point, store);
// create pushs
for(i = 0; i < arity; ++i) {
- ir_entity *ent = be_get_MemPerm_in_entity(node, i);
- ir_type *enttype = get_entity_type(ent);
+ ir_entity *inent = be_get_MemPerm_in_entity(node, i);
+ ir_entity *outent = be_get_MemPerm_out_entity(node, i);
+ ir_type *enttype = get_entity_type(inent);
int entbits = get_type_size_bits(enttype);
+ int entbits2 = get_type_size_bits(get_entity_type(outent));
ir_node *mem = get_irn_n(node, i + 1);
ir_node *push;
+ /* work around cases where entities have different sizes */
+ if(entbits2 < entbits)
+ entbits = entbits2;
assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit");
- push = create_push(cg, node, node, sp, mem, ent);
+ push = create_push(cg, node, node, sp, mem, inent);
sp = create_spproj(cg, node, push, pn_ia32_Push_stack, node);
if(entbits == 64) {
// add another push after the first one
- push = create_push(cg, node, node, sp, mem, ent);
+ push = create_push(cg, node, node, sp, mem, inent);
add_ia32_am_offs_int(push, 4);
sp = create_spproj(cg, node, push, pn_ia32_Push_stack, node);
}
// create pops
for(i = arity - 1; i >= 0; --i) {
- ir_entity *ent = be_get_MemPerm_out_entity(node, i);
- ir_type *enttype = get_entity_type(ent);
+ ir_entity *inent = be_get_MemPerm_in_entity(node, i);
+ ir_entity *outent = be_get_MemPerm_out_entity(node, i);
+ ir_type *enttype = get_entity_type(outent);
int entbits = get_type_size_bits(enttype);
-
+ int entbits2 = get_type_size_bits(get_entity_type(inent));
ir_node *pop;
+ /* work around cases where entities have different sizes */
+ if(entbits2 < entbits)
+ entbits = entbits2;
assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit");
- pop = create_pop(cg, node, node, sp, ent);
+ pop = create_pop(cg, node, node, sp, outent);
sp = create_spproj(cg, node, pop, pn_ia32_Pop_stack, node);
if(entbits == 64) {
add_ia32_am_offs_int(pop, 4);
// add another pop after the first one
- pop = create_pop(cg, node, node, sp, ent);
+ pop = create_pop(cg, node, node, sp, outent);
sp = create_spproj(cg, node, pop, pn_ia32_Pop_stack, node);
}
const ir_mode *mode = get_ia32_ls_mode(node);
int align = 4;
be_node_needs_frame_entity(env, node, mode, align);
+ } else if(is_ia32_FldCW(node)) {
+ const ir_mode *mode = ia32_reg_classes[CLASS_ia32_fp_cw].mode;
+ int align = 4;
+ be_node_needs_frame_entity(env, node, mode, align);
} else if (is_ia32_SetST0(node)) {
const ir_mode *mode = get_ia32_ls_mode(node);
int align = 4;
be_node_needs_frame_entity(env, node, mode, align);
} else {
#ifndef NDEBUG
- if(!is_ia32_Store(node)
- && !is_ia32_xStore(node)
- && !is_ia32_xStoreSimple(node)
+ if(!is_ia32_St(node) && !is_ia32_xStoreSimple(node)
&& !is_ia32_vfist(node)
- && !is_ia32_GetST0(node)) {
+ && !is_ia32_GetST0(node)
+ && !is_ia32_FnstCW(node)) {
assert(0);
}
#endif
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");
-
/* copy optimizations from isa for easier access */
cg->opt = isa->opt;
cg->arch = isa->arch;
-1, /* stack direction */
NULL, /* main environment */
},
- {}, /* emitter environment */
+ { NULL, }, /* emitter environment */
NULL, /* 16bit register names */
NULL, /* 8bit register names */
NULL, /* types */
if (inited)
return NULL;
+ inited = 1;
set_tarval_output_modes();
isa = xmalloc(sizeof(*isa));
memcpy(isa, &ia32_isa_template, sizeof(*isa));
+ if(mode_fpcw == NULL) {
+ mode_fpcw = new_ir_mode("Fpcw", irms_int_number, 16, 0, irma_none, 0);
+ }
+
ia32_register_init(isa);
ia32_create_opcodes();
ia32_register_copy_attr_func();
ia32_build_16bit_reg_map(isa->regs_16bit);
ia32_build_8bit_reg_map(isa->regs_8bit);
- /* patch register names of x87 registers */
- 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);
be_emit_cstring(&isa->emit, ".Ltext0:\n");
be_emit_write_line(&isa->emit);
- inited = 1;
+ /* we mark referenced global entities, so we can only emit those which
+ * are actually referenced. (Note: you mustn't use the type visited flag
+ * elsewhere in the backend)
+ */
+ inc_master_type_visited();
return isa;
}
ia32_isa_t *isa = self;
/* emit now all global declarations */
- be_gas_emit_decls(&isa->emit, isa->arch_isa.main_env);
+ be_gas_emit_decls(&isa->emit, isa->arch_isa.main_env, 1);
pmap_destroy(isa->regs_16bit);
pmap_destroy(isa->regs_8bit);
* - the SSE vector register set
*/
static int ia32_get_n_reg_class(const void *self) {
- return 3;
+ return N_CLASSES;
}
/**
* Return the register class for index i.
*/
-static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
- switch (i) {
- case 0:
- return &ia32_reg_classes[CLASS_ia32_gp];
- case 1:
- return &ia32_reg_classes[CLASS_ia32_xmm];
- case 2:
- return &ia32_reg_classes[CLASS_ia32_vfp];
- default:
- assert(0 && "Invalid ia32 register class requested.");
- return NULL;
- }
+static const arch_register_class_t *ia32_get_reg_class(const void *self, int i)
+{
+ assert(i >= 0 && i < N_CLASSES);
+ return &ia32_reg_classes[i];
}
/**
}
int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
- return is_ia32_irn(irn) ? 1 : -1;
+ if(!is_ia32_irn(irn))
+ return -1;
+
+ if(is_ia32_NoReg_GP(irn) || is_ia32_NoReg_VFP(irn) || is_ia32_NoReg_XMM(irn)
+ || is_ia32_Unknown_GP(irn) || is_ia32_Unknown_XMM(irn)
+ || is_ia32_Unknown_VFP(irn) || is_ia32_ChangeCW(irn))
+ return 0;
+
+ return 1;
}
/**
ia32_get_irg_list,
};
+void ia32_init_emitter(void);
+void ia32_init_finish(void);
+void ia32_init_optimize(void);
+void ia32_init_transform(void);
+void ia32_init_x87(void);
+
void be_init_arch_ia32(void)
{
lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
lc_opt_add_table(ia32_grp, ia32_options);
be_register_isa_if("ia32", &ia32_isa_if);
+
+ FIRM_DBG_REGISTER(dbg, "firm.be.ia32.cg");
+
+ ia32_init_emitter();
+ ia32_init_finish();
+ ia32_init_optimize();
+ ia32_init_transform();
+ ia32_init_x87();
}
BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_ia32);
projects / libfirm / blobdiff