#include "error.h"
#include "xmalloc.h"
#include "irtools.h"
+#include "iroptimize.h"
#include "../beabi.h"
#include "../beirg_t.h"
#include "ia32_finish.h"
#include "ia32_util.h"
#include "ia32_fpu.h"
+#include "ia32_architecture.h"
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
&ia32_xmm_regs[REG_XMM_NOREG]);
}
-/* Creates the unique per irg FP NoReg node. */
-ir_node *ia32_new_NoReg_fp(ia32_code_gen_t *cg) {
- return USE_SSE2(cg) ? ia32_new_NoReg_xmm(cg) : ia32_new_NoReg_vfp(cg);
-}
-
ir_node *ia32_new_Unknown_gp(ia32_code_gen_t *cg) {
return create_const(cg, &cg->unknown_gp, new_rd_ia32_Unknown_GP,
&ia32_gp_regs[REG_GP_UKNWN]);
if (req->cls == &ia32_reg_classes[CLASS_ia32_gp])
return ia32_new_NoReg_gp(cg);
- return ia32_new_NoReg_fp(cg);
+ if (ia32_cg_config.use_sse2) {
+ return ia32_new_NoReg_xmm(cg);
+ } else {
+ return ia32_new_NoReg_vfp(cg);
+ }
}
/**************************************************
ir_mode *mode_bp = env->isa->bp->reg_class->mode;
ir_graph *irg = current_ir_graph;
- if (ARCH_AMD(isa->opt_arch)) {
+ if (ia32_cg_config.use_leave) {
ir_node *leave;
/* leave */
*/
static int ia32_get_op_estimated_cost(const void *self, const ir_node *irn)
{
- int cost;
+ int cost;
ia32_op_type_t op_tp;
- const ia32_irn_ops_t *ops = self;
+ (void) self;
if (is_Proj(irn))
return 0;
if (is_ia32_CopyB(irn)) {
cost = 250;
- if (ARCH_INTEL(ops->cg->arch))
- cost += 150;
}
else if (is_ia32_CopyB_i(irn)) {
- int size = get_ia32_pncode(irn);
+ int size = get_ia32_copyb_size(irn);
cost = 20 + (int)ceil((4/3) * size);
- if (ARCH_INTEL(ops->cg->arch))
- cost += 150;
}
/* in case of address mode operations add additional cycles */
else if (op_tp == ia32_AddrModeD || op_tp == ia32_AddrModeS) {
ia32_perform_memory_operand,
};
-ia32_irn_ops_t ia32_irn_ops = {
+static ia32_irn_ops_t ia32_irn_ops = {
&ia32_irn_ops_if,
NULL
};
**************************************************/
static void ia32_before_abi(void *self) {
+ lower_mode_b_config_t lower_mode_b_config = {
+ mode_Iu, /* lowered mode */
+ mode_Bu, /* prefered mode for set */
+ 0, /* don't lower direct compares */
+ };
ia32_code_gen_t *cg = self;
- ir_lower_mode_b(cg->irg, mode_Iu, 0);
+ ir_lower_mode_b(cg->irg, &lower_mode_b_config);
if(cg->dump)
be_dump(cg->irg, "-lower_modeb", dump_ir_block_graph_sched);
}
}
if (mode_is_float(spillmode)) {
- if (USE_SSE2(cg))
+ if (ia32_cg_config.use_sse2)
new_op = new_rd_ia32_xLoad(dbg, irg, block, ptr, noreg, mem, spillmode);
else
new_op = new_rd_ia32_vfld(dbg, irg, block, ptr, noreg, mem, spillmode);
}
if (mode_is_float(mode)) {
- if (USE_SSE2(cg))
+ if (ia32_cg_config.use_sse2)
store = new_rd_ia32_xStore(dbg, irg, block, ptr, noreg, nomem, val);
else
store = new_rd_ia32_vfst(dbg, irg, block, ptr, noreg, nomem, val, mode);
arity = be_get_MemPerm_entity_arity(node);
pops = alloca(arity * sizeof(pops[0]));
- // create pushs
+ /* create Pushs */
for(i = 0; i < arity; ++i) {
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));
+ unsigned entsize = get_type_size_bytes(enttype);
+ unsigned entsize2 = get_type_size_bytes(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");
+ if(entsize2 < entsize)
+ entsize = entsize2;
+ assert( (entsize == 4 || entsize == 8) && "spillslot on x86 should be 32 or 64 bit");
push = create_push(cg, node, node, sp, mem, inent);
sp = create_spproj(cg, node, push, pn_ia32_Push_stack);
- if(entbits == 64) {
- // add another push after the first one
+ if(entsize == 8) {
+ /* add another push after the first one */
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);
set_irn_n(node, i, new_Bad());
}
- // create pops
+ /* create pops */
for(i = arity - 1; i >= 0; --i) {
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));
+ unsigned entsize = get_type_size_bytes(enttype);
+ unsigned entsize2 = get_type_size_bytes(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");
+ if(entsize2 < entsize)
+ entsize = entsize2;
+ assert( (entsize == 4 || entsize == 8) && "spillslot on x86 should be 32 or 64 bit");
pop = create_pop(cg, node, node, sp, outent);
sp = create_spproj(cg, node, pop, pn_ia32_Pop_stack);
- if(entbits == 64) {
+ if(entsize == 8) {
add_ia32_am_offs_int(pop, 4);
- // add another pop after the first one
+ /* add another pop after the first one */
pop = create_pop(cg, node, node, sp, outent);
sp = create_spproj(cg, node, pop, pn_ia32_Pop_stack);
}
}
in[0] = sp;
- keep = be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1, in);
+ keep = be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1, in);
sched_add_before(node, keep);
- // exchange memprojs
+ /* exchange memprojs */
foreach_out_edge_safe(node, edge, next) {
ir_node *proj = get_edge_src_irn(edge);
int p = get_Proj_proj(proj);
set_Proj_proj(proj, pn_ia32_Pop_M);
}
- // remove memperm
+ /* remove memperm */
arity = get_irn_arity(node);
for(i = 0; i < arity; ++i) {
set_irn_n(node, i, new_Bad());
be_node_needs_frame_entity(env, node, mode, align);
} else if (is_ia32_vfild(node) || is_ia32_xLoad(node)
|| is_ia32_vfld(node)) {
- const ir_mode *mode = get_ia32_ls_mode(node);
- int align = 4;
+ 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;
+ /* although 2 byte would be enough 4 byte performs best */
+ const ir_mode *mode = mode_Iu;
+ int align = 4;
be_node_needs_frame_entity(env, node, mode, align);
} else {
#ifndef NDEBUG
x87_simulate_graph(cg->arch_env, cg->birg);
}
+ /* do peephole optimisations */
+ ia32_peephole_optimization(cg);
+
/* create block schedule, this also removes empty blocks which might
* produce critical edges */
cg->blk_sched = be_create_block_schedule(irg, cg->birg->exec_freq);
-
- /* do peephole optimisations */
- ia32_peephole_optimization(irg, cg);
}
/**
cg->isa = isa;
cg->birg = birg;
cg->blk_sched = NULL;
- cg->fp_kind = isa->fp_kind;
cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
- /* copy optimizations from isa for easier access */
- cg->opt = isa->opt;
- cg->arch = isa->arch;
- cg->opt_arch = isa->opt_arch;
-
/* enter it */
isa->cg = cg;
NULL, /* 8bit register names high */
NULL, /* types */
NULL, /* tv_ents */
- (0 |
- IA32_OPT_INCDEC | /* optimize add 1, sub 1 into inc/dec default: on */
- IA32_OPT_CC),
- arch_pentium_4, /* instruction architecture */
- arch_pentium_4, /* optimize for architecture */
- fp_x87, /* floating point mode */
NULL, /* current code generator */
+ NULL, /* abstract machine */
#ifndef NDEBUG
NULL, /* name obstack */
- 0 /* name obst size */
#endif
};
-static void set_arch_costs(enum cpu_support arch);
-
/**
* Initializes the backend ISA.
*/
ia32_register_init();
ia32_create_opcodes();
- set_arch_costs(isa->opt_arch);
-
- if ((ARCH_INTEL(isa->arch) && isa->arch < arch_pentium_4) ||
- (ARCH_AMD(isa->arch) && isa->arch < arch_athlon))
- /* no SSE2 for these cpu's */
- isa->fp_kind = fp_x87;
-
- if (ARCH_INTEL(isa->opt_arch) && isa->opt_arch >= arch_pentium_4) {
- /* Pentium 4 don't like inc and dec instructions */
- isa->opt &= ~IA32_OPT_INCDEC;
- }
-
be_emit_init(file_handle);
isa->regs_16bit = pmap_create();
isa->regs_8bit = pmap_create();
* @param mode The mode in question.
* @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) {
- const ia32_isa_t *isa = self;
+const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self,
+ const ir_mode *mode)
+{
+ (void) self;
+
if (mode_is_float(mode)) {
- return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
+ return ia32_cg_config.use_sse2 ? &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
*/
-static void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
- const ia32_isa_t *isa = self;
+static void ia32_get_call_abi(const void *self, ir_type *method_type,
+ be_abi_call_t *abi)
+{
ir_type *tp;
ir_mode *mode;
unsigned cc;
int n, i, regnum;
be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
+ (void) self;
/* set abi flags for calls */
call_flags.bits.left_to_right = 0; /* always last arg first on stack */
} else {
cc = get_method_calling_convention(method_type);
if (get_method_additional_properties(method_type) & mtp_property_private
- && (ia32_isa_template.opt & IA32_OPT_CC)) {
+ && (ia32_cg_config.optimize_cc)) {
/* set the calling conventions to register parameter */
cc = (cc & ~cc_bits) | cc_reg_param;
}
}
+
+ /* we have to pop the shadow parameter ourself for compound calls */
+ if( (get_method_calling_convention(method_type) & cc_compound_ret)
+ && !(cc & cc_reg_param)) {
+ be_abi_call_set_pop(abi, get_mode_size_bytes(mode_P_data));
+ }
+
n = get_method_n_params(method_type);
for (i = regnum = 0; i < n; i++) {
- const ir_mode *mode;
+ ir_mode *mode;
const arch_register_t *reg = NULL;
tp = get_method_param_type(method_type, i);
mode = get_type_mode(tp);
if (mode != NULL) {
- reg = ia32_get_RegParam_reg(isa->cg, cc, regnum, mode);
+ reg = ia32_get_RegParam_reg(cc, regnum, mode);
}
if (reg != NULL) {
be_abi_call_param_reg(abi, i, reg);
++regnum;
} else {
- be_abi_call_param_stack(abi, i, 4, 0, 0);
+ /* Micro optimisation: if the mode is shorter than 4 bytes, load 4 bytes.
+ * movl has a shorter opcode than mov[sz][bw]l */
+ ir_mode *load_mode = mode;
+ if (mode != NULL && get_mode_size_bytes(mode) < 4) load_mode = mode_Iu;
+ be_abi_call_param_stack(abi, i, load_mode, 4, 0, 0);
}
}
(void)i;
(void)j;
+ if(!ia32_cg_config.use_cmov) {
+ /* TODO: we could still handle abs(x)... */
+ return 0;
+ }
+
/* we can't handle psis with 64bit compares yet */
if(is_Proj(sel)) {
ir_node *pred = get_Proj_pred(sel);
return 1;
}
-typedef struct insn_const {
- int add_cost; /**< cost of an add instruction */
- int lea_cost; /**< cost of a lea instruction */
- int const_shf_cost; /**< cost of a constant shift instruction */
- int cost_mul_start; /**< starting cost of a multiply instruction */
- int cost_mul_bit; /**< cost of multiply for every set bit */
-} insn_const;
-
-/* costs for the i386 */
-static const insn_const i386_cost = {
- 1, /* cost of an add instruction */
- 1, /* cost of a lea instruction */
- 2, /* cost of a constant shift instruction */
- 6, /* starting cost of a multiply instruction */
- 1 /* cost of multiply for every set bit */
-};
-
-/* costs for the i486 */
-static const insn_const i486_cost = {
- 1, /* cost of an add instruction */
- 1, /* cost of a lea instruction */
- 2, /* cost of a constant shift instruction */
- 12, /* starting cost of a multiply instruction */
- 1 /* cost of multiply for every set bit */
-};
-
-/* costs for the Pentium */
-static const insn_const pentium_cost = {
- 1, /* cost of an add instruction */
- 1, /* cost of a lea instruction */
- 1, /* cost of a constant shift instruction */
- 11, /* starting cost of a multiply instruction */
- 0 /* cost of multiply for every set bit */
-};
-
-/* costs for the Pentium Pro */
-static const insn_const pentiumpro_cost = {
- 1, /* cost of an add instruction */
- 1, /* cost of a lea instruction */
- 1, /* cost of a constant shift instruction */
- 4, /* starting cost of a multiply instruction */
- 0 /* cost of multiply for every set bit */
-};
-
-/* costs for the K6 */
-static const insn_const k6_cost = {
- 1, /* cost of an add instruction */
- 2, /* cost of a lea instruction */
- 1, /* cost of a constant shift instruction */
- 3, /* starting cost of a multiply instruction */
- 0 /* cost of multiply for every set bit */
-};
-
-/* costs for the Athlon */
-static const insn_const athlon_cost = {
- 1, /* cost of an add instruction */
- 2, /* cost of a lea instruction */
- 1, /* cost of a constant shift instruction */
- 5, /* starting cost of a multiply instruction */
- 0 /* cost of multiply for every set bit */
-};
-
-/* costs for the Pentium 4 */
-static const insn_const pentium4_cost = {
- 1, /* cost of an add instruction */
- 3, /* cost of a lea instruction */
- 4, /* cost of a constant shift instruction */
- 15, /* starting cost of a multiply instruction */
- 0 /* cost of multiply for every set bit */
-};
-
-/* costs for the Core */
-static const insn_const core_cost = {
- 1, /* cost of an add instruction */
- 1, /* cost of a lea instruction */
- 1, /* cost of a constant shift instruction */
- 10, /* starting cost of a multiply instruction */
- 0 /* cost of multiply for every set bit */
-};
-
-/* costs for the generic */
-static const insn_const generic_cost = {
- 1, /* cost of an add instruction */
- 2, /* cost of a lea instruction */
- 1, /* cost of a constant shift instruction */
- 4, /* starting cost of a multiply instruction */
- 0 /* cost of multiply for every set bit */
-};
-
-static const insn_const *arch_costs = &generic_cost;
-
-static void set_arch_costs(enum cpu_support arch) {
- switch (arch) {
- case arch_i386:
- arch_costs = &i386_cost;
- break;
- case arch_i486:
- arch_costs = &i486_cost;
- break;
- case arch_pentium:
- case arch_pentium_mmx:
- arch_costs = &pentium_cost;
- break;
- case arch_pentium_pro:
- case arch_pentium_2:
- case arch_pentium_3:
- arch_costs = &pentiumpro_cost;
- break;
- case arch_pentium_4:
- arch_costs = &pentium4_cost;
- break;
- case arch_pentium_m:
- arch_costs = &pentiumpro_cost;
- break;
- case arch_core:
- arch_costs = &core_cost;
- break;
- case arch_k6:
- arch_costs = &k6_cost;
- break;
- case arch_athlon:
- case arch_athlon_xp:
- case arch_athlon_64:
- case arch_opteron:
- arch_costs = &athlon_cost;
- break;
- case arch_generic:
- default:
- arch_costs = &generic_cost;
- }
-}
-
-/**
- * Evaluate a given simple instruction.
- */
-static int ia32_evaluate_insn(insn_kind kind, tarval *tv) {
- int cost;
-
- switch (kind) {
- case MUL:
- cost = arch_costs->cost_mul_start;
- if (arch_costs->cost_mul_bit > 0) {
- char *bitstr = get_tarval_bitpattern(tv);
- int i;
-
- for (i = 0; bitstr[i] != '\0'; ++i) {
- if (bitstr[i] == '1') {
- cost += arch_costs->cost_mul_bit;
- }
- }
- free(bitstr);
- }
- return cost;
- case LEA:
- return arch_costs->lea_cost;
- case ADD:
- case SUB:
- return arch_costs->add_cost;
- case SHIFT:
- return arch_costs->const_shf_cost;
- case ZERO:
- return arch_costs->add_cost;
- default:
- return 1;
- }
-}
-
/**
* Returns the libFirm configuration parameter for this backend.
*/
NULL, /* will be set below */
};
+ ia32_setup_cg_config();
+
p.dep_param = &ad;
p.if_conv_info = &ifconv;
return &p;
}
-/* 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, },
- { "athlon-xp", arch_athlon_xp, },
- { "athlon64", arch_athlon_64, },
- { "opteron", arch_opteron, },
- { "generic", arch_generic, },
- { 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_enum_int_items_t gas_items[] = {
{ "normal", GAS_FLAVOUR_NORMAL },
{ "mingw", GAS_FLAVOUR_MINGW },
};
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("nooptcc", "do not optimize calling convention", &ia32_isa_template.opt, IA32_OPT_CC),
- LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
+ LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
LC_OPT_LAST
};
void be_init_arch_ia32(void)
{
- lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
+ lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
lc_opt_entry_t *ia32_grp = lc_opt_get_grp(be_grp, "ia32");
lc_opt_add_table(ia32_grp, ia32_options);
ia32_init_optimize();
ia32_init_transform();
ia32_init_x87();
+ ia32_init_architecture();
}
BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_ia32);