#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);
+ }
}
/**************************************************
if (is_ia32_irn(irn)) {
const arch_register_t **slots;
slots = get_ia32_slots(irn);
+ assert(pos < get_ia32_n_res(irn));
reg = slots[pos];
} else {
reg = ia32_get_firm_reg(irn, cur_reg_set);
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, *mem, curr_bp, curr_sp);
+ push = new_rd_ia32_Push(NULL, env->irg, bl, noreg, noreg, *mem, curr_sp, curr_bp);
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);
} else {
const ia32_isa_t *isa = (ia32_isa_t *)env->isa;
ia32_code_gen_t *cg = isa->cg;
- ir_mode *mode_bp = env->isa->bp->reg_class->mode;
+ ir_mode *mode_bp = env->isa->bp->reg_class->mode;
+ ir_graph *irg = current_ir_graph;
- /* gcc always emits a leave at the end of a routine */
- if (1 || ARCH_AMD(isa->opt_arch)) {
+ if (ia32_cg_config.use_leave) {
ir_node *leave;
/* leave */
- leave = new_rd_ia32_Leave(NULL, env->irg, bl, curr_sp, curr_bp);
+ leave = new_rd_ia32_Leave(NULL, irg, bl, curr_sp, curr_bp);
set_ia32_flags(leave, arch_irn_flags_ignore);
- curr_bp = new_r_Proj(current_ir_graph, bl, leave, mode_bp, pn_ia32_Leave_frame);
- curr_sp = new_r_Proj(current_ir_graph, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
+ curr_bp = new_r_Proj(irg, bl, leave, mode_bp, pn_ia32_Leave_frame);
+ curr_sp = new_r_Proj(irg, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
} else {
ir_node *noreg = ia32_new_NoReg_gp(cg);
ir_node *pop;
+ /* the old SP is not needed anymore (kill the proj) */
+ assert(is_Proj(curr_sp));
+ be_kill_node(curr_sp);
+
/* copy ebp to esp */
- curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem);
+ curr_sp = be_new_Copy(&ia32_reg_classes[CLASS_ia32_gp], irg, bl, curr_bp);
+ 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);
/* pop ebp */
pop = new_rd_ia32_Pop(NULL, env->irg, bl, noreg, noreg, *mem, curr_sp);
set_ia32_flags(pop, arch_irn_flags_ignore);
- curr_bp = new_r_Proj(current_ir_graph, bl, pop, mode_bp, pn_ia32_Pop_res);
- curr_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
+ curr_bp = new_r_Proj(irg, bl, pop, mode_bp, pn_ia32_Pop_res);
+ curr_sp = new_r_Proj(irg, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
- *mem = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M);
+ *mem = new_r_Proj(irg, bl, pop, mode_M, pn_ia32_Pop_M);
}
arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
*/
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) {
(void) self;
if (! is_ia32_irn(irn) || /* must be an ia32 irn */
- get_ia32_am_arity(irn) != 2 || /* must be a binary operation TODO is this necessary? */
+ get_ia32_am_arity(irn) != ia32_am_binary || /* must be a binary operation TODO is this necessary? */
get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */
! (get_ia32_am_support(irn) & ia32_am_Source) || /* must be capable of source addressmode */
! ia32_is_spillmode_compatible(mode, spillmode) ||
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, &lower_mode_b_config);
+ if(cg->dump)
+ be_dump(cg->irg, "-lower_modeb", dump_ir_block_graph_sched);
+}
+
/**
* Transforms the standard firm graph into
* an ia32 firm graph
static void ia32_prepare_graph(void *self) {
ia32_code_gen_t *cg = self;
- ir_lower_mode_b(cg->irg, mode_Iu, 0);
/* do local optimisations */
optimize_graph_df(cg->irg);
+
+ /* TODO: we often have dead code reachable through out-edges here. So for
+ * now we rebuild edges (as we need correct user count for code selection)
+ */
+#if 1
+ edges_deactivate(cg->irg);
+ edges_activate(cg->irg);
+#endif
+
if(cg->dump)
- be_dump(cg->irg, "-lower_modeb", dump_ir_block_graph_sched);
+ be_dump(cg->irg, "-pre_transform", dump_ir_block_graph_sched);
/* transform nodes into assembler instructions */
ia32_transform_graph(cg);
ir_node *base = get_irn_n(node, n_ia32_base);
ir_node *index = get_irn_n(node, n_ia32_index);
ir_node *mem = get_irn_n(node, n_ia32_mem);
+ ir_node *noreg = ia32_new_NoReg_gp(ia32_current_cg);
ir_node *load;
ir_node *load_res;
ir_node *mem_proj;
const ir_edge_t *edge;
- ir_fprintf(stderr, "truning back AM in %+F\n", node);
-
load = new_rd_ia32_Load(dbgi, irg, block, base, index, mem);
load_res = new_rd_Proj(dbgi, irg, block, load, mode_Iu, pn_ia32_Load_res);
ia32_copy_am_attrs(load, node);
set_irn_n(node, n_ia32_mem, new_NoMem());
- if(get_ia32_am_arity(node) == ia32_am_unary) {
- set_irn_n(node, n_ia32_unary_op, load_res);
- } else if(get_ia32_am_arity(node) == ia32_am_binary) {
- set_irn_n(node, n_ia32_binary_right, load_res);
- } else if(get_ia32_am_arity(node) == ia32_am_ternary) {
- set_irn_n(node, n_ia32_binary_right, load_res);
+ switch (get_ia32_am_arity(node)) {
+ case ia32_am_unary:
+ set_irn_n(node, n_ia32_unary_op, load_res);
+ break;
+
+ case ia32_am_binary:
+ if (is_ia32_Immediate(get_irn_n(node, n_ia32_Cmp_right))) {
+ assert(is_ia32_Cmp(node) || is_ia32_Cmp8Bit(node) ||
+ is_ia32_Test(node) || is_ia32_Test8Bit(node));
+ set_irn_n(node, n_ia32_binary_left, load_res);
+ } else {
+ set_irn_n(node, n_ia32_binary_right, load_res);
+ }
+ break;
+
+ case ia32_am_ternary:
+ set_irn_n(node, n_ia32_binary_right, load_res);
+ break;
+
+ default: break;
}
+ set_irn_n(node, n_ia32_base, noreg);
+ set_irn_n(node, n_ia32_index, noreg);
+ set_ia32_am_offs_int(node, 0);
+ set_ia32_am_sc(node, NULL);
+ set_ia32_am_scale(node, 0);
+ clear_ia32_am_sc_sign(node);
/* rewire mem-proj */
if(get_irn_mode(node) == mode_T) {
static ir_node *flags_remat(ir_node *node, ir_node *after)
{
/* we should turn back source address mode when rematerializing nodes */
- ia32_op_type_t type = get_ia32_op_type(node);
+ ia32_op_type_t type = get_ia32_op_type(node);
+ ir_node *block;
ir_node *copy;
- if (type == ia32_AddrModeS) {
- turn_back_am(node);
- } else if (type == ia32_AddrModeD) {
- /* TODO implement this later... */
- panic("found DestAM with flag user %+F this should not happen", node);
+ if(is_Block(after)) {
+ block = after;
} else {
- assert(type == ia32_Normal);
+ block = get_nodes_block(after);
+ }
+
+ switch (type) {
+ case ia32_AddrModeS: turn_back_am(node); break;
+
+ case ia32_AddrModeD:
+ /* TODO implement this later... */
+ panic("found DestAM with flag user %+F this should not happen", node);
+ break;
+
+ default: assert(type == ia32_Normal); break;
}
copy = exact_copy(node);
+ set_nodes_block(copy, block);
sched_add_after(after, copy);
return copy;
}
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);
ir_node *noreg = ia32_new_NoReg_gp(cg);
ir_node *frame = get_irg_frame(irg);
- ir_node *push = new_rd_ia32_Push(dbg, irg, block, frame, noreg, mem, noreg, sp);
+ ir_node *push = new_rd_ia32_Push(dbg, irg, block, frame, noreg, mem, sp, noreg);
set_ia32_frame_ent(push, ent);
set_ia32_use_frame(push);
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);
}
/**
static const arch_code_generator_if_t ia32_code_gen_if = {
ia32_cg_init,
- NULL, /* before abi introduce hook */
+ ia32_before_abi, /* before abi introduce hook */
ia32_prepare_graph,
NULL, /* spill */
ia32_before_sched, /* before scheduling hook */
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;
7, /* costs for a spill instruction */
5, /* costs for a reload instruction */
},
- NULL_EMITTER, /* emitter environment */
NULL, /* 16bit register names */
NULL, /* 8bit register names */
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_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_PUSHARGS), /* create pushs for function argument passing, default: on */
- 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_env(&isa->emit, file_handle);
+ be_emit_init(file_handle);
isa->regs_16bit = pmap_create();
isa->regs_8bit = pmap_create();
isa->regs_8bit_high = pmap_create();
ia32_handle_intrinsics();
/* needed for the debug support */
- be_gas_emit_switch_section(&isa->emit, GAS_SECTION_TEXT);
- be_emit_cstring(&isa->emit, ".Ltext0:\n");
- be_emit_write_line(&isa->emit);
+ be_gas_emit_switch_section(GAS_SECTION_TEXT);
+ be_emit_cstring(".Ltext0:\n");
+ be_emit_write_line();
/* we mark referenced global entities, so we can only emit those which
* are actually referenced. (Note: you mustn't use the type visited flag
ia32_isa_t *isa = self;
/* emit now all global declarations */
- be_gas_emit_decls(&isa->emit, isa->arch_isa.main_env, 1);
+ be_gas_emit_decls(isa->arch_isa.main_env, 1);
pmap_destroy(isa->regs_16bit);
pmap_destroy(isa->regs_8bit);
obstack_free(isa->name_obst, NULL);
#endif /* NDEBUG */
- be_emit_destroy_env(&isa->emit);
+ be_emit_exit();
free(self);
}
* - the virtual floating point registers
* - the SSE vector register set
*/
-static int ia32_get_n_reg_class(const void *self) {
+static unsigned ia32_get_n_reg_class(const void *self) {
(void) self;
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)
+static const arch_register_class_t *ia32_get_reg_class(const void *self,
+ unsigned i)
{
(void) self;
- assert(i >= 0 && i < N_CLASSES);
+ assert(i < N_CLASSES);
return &ia32_reg_classes[i];
}
* @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);
-
- unsigned use_push = !IS_P6_ARCH(isa->opt_arch);
+ (void) self;
/* set abi flags for calls */
call_flags.bits.left_to_right = 0; /* always last arg first on stack */
- call_flags.bits.store_args_sequential = use_push;
+ call_flags.bits.store_args_sequential = 0;
/* 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 */
cc = cc_cdecl_set;
} else {
cc = get_method_calling_convention(method_type);
- if (get_method_additional_properties(method_type) & mtp_property_private) {
+ if (get_method_additional_properties(method_type) & mtp_property_private
+ && (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 1
+ 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);
if(is_Cmp(pred)) {
ir_node *left = get_Cmp_left(pred);
ir_mode *cmp_mode = get_irn_mode(left);
- if(mode_is_float(cmp_mode))
+ if(!mode_is_float(cmp_mode) && get_mode_size_bits(cmp_mode) > 32)
return 0;
}
}
-#endif
/* check the Phi nodes */
for (phi = phi_list; phi; phi = get_irn_link(phi)) {
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_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, },
- { "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("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("nopushargs", "do not create pushs for function arguments", &ia32_isa_template.opt, IA32_OPT_PUSHARGS),
- 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);