#include "ircons.h"
#include "irprintf.h"
#include "typerep.h"
+#include "bitset.h"
#include "../betranshlp.h"
-#include "../beirg_t.h"
+#include "../beirg.h"
+#include "../beabi.h"
#include "ia32_architecture.h"
#include "ia32_common_transform.h"
heights_t *heights = NULL;
-static const arch_register_req_t no_register_req = {
- arch_register_req_type_none,
- NULL, /* regclass */
- NULL, /* limit bitset */
- 0, /* same pos */
- 0 /* different pos */
-};
-
static int check_immediate_constraint(long val, char immediate_constraint_type)
{
switch (immediate_constraint_type) {
}
}
-/* creates a unique ident by adding a number to a tag */
-ident *ia32_unique_id(const char *tag)
-{
- static unsigned id = 0;
- char str[256];
-
- snprintf(str, sizeof(str), tag, ++id);
- return new_id_from_str(str);
-}
-
/**
* Get a primitive type for a mode with alignment 16.
*/
static ir_type *ia32_get_prim_type(pmap *types, ir_mode *mode)
{
- pmap_entry *e = pmap_find(types, mode);
- ir_type *res;
-
- if (! e) {
- char buf[64];
- snprintf(buf, sizeof(buf), "prim_type_%s", get_mode_name(mode));
- res = new_type_primitive(new_id_from_str(buf), mode);
- if (get_mode_size_bits(mode) >= 80) {
- set_type_alignment_bytes(res, 16);
- }
- pmap_insert(types, mode, res);
+ ir_type *res = pmap_get(types, mode);
+ if (res != NULL)
+ return res;
+
+ res = new_type_primitive(mode);
+ if (get_mode_size_bits(mode) >= 80) {
+ set_type_alignment_bytes(res, 16);
}
- else
- res = e->value;
+ pmap_insert(types, mode, res);
return res;
}
ir_entity *create_float_const_entity(ir_node *cnst)
{
- ia32_isa_t *isa = env_cg->isa;
- tarval *key = get_Const_tarval(cnst);
- pmap_entry *e = pmap_find(isa->tv_ent, key);
- ir_entity *res;
- ir_graph *rem;
-
- if (e == NULL) {
- tarval *tv = key;
- ir_mode *mode = get_tarval_mode(tv);
- ir_type *tp;
-
- if (! ia32_cg_config.use_sse2) {
- /* try to reduce the mode to produce smaller sized entities */
- if (mode != mode_F) {
- if (tarval_ieee754_can_conv_lossless(tv, mode_F)) {
- mode = mode_F;
+ ia32_isa_t *isa = env_cg->isa;
+ tarval *tv = get_Const_tarval(cnst);
+ ir_entity *res = pmap_get(isa->tv_ent, tv);
+ ir_initializer_t *initializer;
+ ir_mode *mode;
+ ir_type *tp;
+
+ if (res != NULL)
+ return res;
+
+ mode = get_tarval_mode(tv);
+
+ if (! ia32_cg_config.use_sse2) {
+ /* try to reduce the mode to produce smaller sized entities */
+ if (mode != mode_F) {
+ if (tarval_ieee754_can_conv_lossless(tv, mode_F)) {
+ mode = mode_F;
+ tv = tarval_convert_to(tv, mode);
+ } else if (mode != mode_D) {
+ if (tarval_ieee754_can_conv_lossless(tv, mode_D)) {
+ mode = mode_D;
tv = tarval_convert_to(tv, mode);
- } else if (mode != mode_D) {
- if (tarval_ieee754_can_conv_lossless(tv, mode_D)) {
- mode = mode_D;
- tv = tarval_convert_to(tv, mode);
- }
}
}
}
-
- if (mode == get_irn_mode(cnst)) {
- /* mode was not changed */
- tp = get_Const_type(cnst);
- if (tp == firm_unknown_type)
- tp = ia32_get_prim_type(isa->types, mode);
- } else
- tp = ia32_get_prim_type(isa->types, mode);
-
- res = new_entity(get_glob_type(), ia32_unique_id(".LC%u"), tp);
-
- set_entity_ld_ident(res, get_entity_ident(res));
- set_entity_visibility(res, visibility_local);
- set_entity_variability(res, variability_constant);
- set_entity_allocation(res, allocation_static);
-
- /* we create a new entity here: It's initialization must resist on the
- const code irg */
- rem = current_ir_graph;
- current_ir_graph = get_const_code_irg();
- set_atomic_ent_value(res, new_Const_type(tv, tp));
- current_ir_graph = rem;
-
- pmap_insert(isa->tv_ent, key, res);
- } else {
- res = e->value;
}
+ tp = ia32_get_prim_type(isa->types, mode);
+ res = new_entity(get_glob_type(), id_unique("C%u"), tp);
+ set_entity_ld_ident(res, get_entity_ident(res));
+ set_entity_visibility(res, ir_visibility_private);
+ add_entity_linkage(res, IR_LINKAGE_CONSTANT);
+
+ initializer = create_initializer_tarval(tv);
+ set_entity_initializer(res, initializer);
+
+ pmap_insert(isa->tv_ent, tv, res);
return res;
}
ir_graph *irg = current_ir_graph;
ir_node *start_block = get_irg_start_block(irg);
ir_node *immediate = new_bd_ia32_Immediate(NULL, start_block, symconst,
- symconst_sign, val);
+ symconst_sign, no_pic_adjust, val);
arch_set_irn_register(immediate, &ia32_gp_regs[REG_GP_NOREG]);
return immediate;
/* TODO: construct a hashmap instead of doing linear search for clobber
* register */
- for(c = 0; c < N_CLASSES; ++c) {
+ for (c = 0; c < N_CLASSES; ++c) {
cls = & ia32_reg_classes[c];
- for(r = 0; r < cls->n_regs; ++r) {
+ for (r = 0; r < cls->n_regs; ++r) {
const arch_register_t *temp_reg = arch_register_for_index(cls, r);
if (strcmp(temp_reg->name, clobber) == 0
|| (c == CLASS_ia32_gp && strcmp(temp_reg->name+1, clobber) == 0)) {
return reg;
}
-int ia32_mode_needs_gp_reg(ir_mode *mode) {
+int ia32_mode_needs_gp_reg(ir_mode *mode)
+{
if (mode == mode_fpcw)
return 0;
if (get_mode_size_bits(mode) > 32)
/* TODO: improve error messages with node and source info. (As users can
* easily hit these) */
- while(*c != 0) {
- switch(*c) {
+ while (*c != 0) {
+ switch (*c) {
case ' ':
case '\t':
case '\n':
++c;
break;
case '#':
- while(*c != 0 && *c != ',')
+ while (*c != 0 && *c != ',')
++c;
break;
constraint->immediate_type = immediate_type;
}
+static bool can_match(const arch_register_req_t *in,
+ const arch_register_req_t *out)
+{
+ if (in->cls != out->cls)
+ return false;
+ if ( (in->type & arch_register_req_type_limited) == 0
+ || (out->type & arch_register_req_type_limited) == 0 )
+ return true;
+
+ return (*in->limited & *out->limited) != 0;
+}
+
+static inline ir_node *get_new_node(ir_node *node)
+{
+#ifdef FIRM_GRGEN_BE
+ if (be_transformer == TRANSFORMER_DEFAULT) {
+ return be_transform_node(node);
+ } else {
+ return node;
+ }
+#else
+ return be_transform_node(node);
+#endif
+}
+
ir_node *gen_ASM(ir_node *node)
{
- ir_node *block = NULL;
- ir_node *new_block = NULL;
+ ir_node *block = get_nodes_block(node);
+ ir_node *new_block = get_new_node(block);
dbg_info *dbgi = get_irn_dbg_info(node);
int i, arity;
+ int value_arity;
int out_idx;
ir_node **in;
ir_node *new_node;
ident **clobbers;
int clobbers_flags = 0;
unsigned clobber_bits[N_CLASSES];
+ int out_size;
+ backend_info_t *info;
memset(&clobber_bits, 0, sizeof(clobber_bits));
- switch (be_transformer) {
- case TRANSFORMER_DEFAULT:
- block = get_nodes_block(node);
- new_block = be_transform_node(block);
- break;
-
-#ifdef FIRM_GRGEN_BE
- case TRANSFORMER_PBQP:
- case TRANSFORMER_RAND:
- new_block = get_nodes_block(node);
- break;
-#endif
-
- default:
- panic("invalid transformer");
- }
-
/* workaround for lots of buggy code out there as most people think volatile
* asm is enough for everything and forget the flags (linux kernel, etc.)
*/
reg_map_size = constraint->pos;
}
for (i = 0; i < arity; ++i) {
- const ir_asm_constraint *constraint = &in_constraints[i];
+ const ir_asm_constraint *constraint = &in_constraints[i];
if (constraint->pos > reg_map_size)
reg_map_size = constraint->pos;
}
memset(register_map, 0, reg_map_size * sizeof(register_map[0]));
/* construct output constraints */
- out_reg_reqs = obstack_alloc(obst, out_arity * sizeof(out_reg_reqs[0]));
+ out_size = out_arity + 1;
+ out_reg_reqs = obstack_alloc(obst, out_size * sizeof(out_reg_reqs[0]));
for (out_idx = 0; out_idx < n_out_constraints; ++out_idx) {
const ir_asm_constraint *constraint = &out_constraints[out_idx];
if (r_clobber_bits != 0) {
if (parsed_constraint.all_registers_allowed) {
parsed_constraint.all_registers_allowed = 0;
- be_abi_set_non_ignore_regs(env_cg->birg->abi,
+ be_abi_set_non_ignore_regs(be_get_irg_abi(env_cg->irg),
parsed_constraint.cls,
&parsed_constraint.allowed_registers);
}
}
if (input == NULL) {
- ir_node *pred = NULL;
- switch (be_transformer) {
- case TRANSFORMER_DEFAULT:
- pred = get_irn_n(node, i);
- input = be_transform_node(pred);
- break;
-
-#ifdef FIRM_GRGEN_BE
- case TRANSFORMER_PBQP:
- case TRANSFORMER_RAND:
- input = get_irn_n(node, i);
- break;
-#endif
-
- default: panic("invalid transformer");
- }
+ ir_node *pred = get_irn_n(node, i);
+ input = get_new_node(pred);
if (parsed_constraint.cls == NULL
&& parsed_constraint.same_as < 0) {
++out_idx;
}
+ /* count inputs which are real values (and not memory) */
+ value_arity = 0;
+ for (i = 0; i < arity; ++i) {
+ ir_node *in = get_irn_n(node, i);
+ if (get_irn_mode(in) == mode_M)
+ continue;
+ ++value_arity;
+ }
+
+ /* Attempt to make ASM node register pressure faithful.
+ * (This does not work for complicated cases yet!)
+ *
+ * Algorithm: Check if there are fewer inputs or outputs (I will call this
+ * the smaller list). Then try to match each constraint of the smaller list
+ * to 1 of the other list. If we can't match it, then we have to add a dummy
+ * input/output to the other list
+ *
+ * FIXME: This is still broken in lots of cases. But at least better than
+ * before...
+ * FIXME: need to do this per register class...
+ */
+ if (out_arity <= value_arity) {
+ int orig_arity = arity;
+ int in_size = arity;
+ int o;
+ bitset_t *used_ins = bitset_alloca(arity);
+ for (o = 0; o < out_arity; ++o) {
+ int i;
+ const arch_register_req_t *outreq = out_reg_reqs[o];
+
+ if (outreq->cls == NULL) {
+ continue;
+ }
+
+ for (i = 0; i < orig_arity; ++i) {
+ const arch_register_req_t *inreq;
+ if (bitset_is_set(used_ins, i))
+ continue;
+ inreq = in_reg_reqs[i];
+ if (!can_match(outreq, inreq))
+ continue;
+ bitset_set(used_ins, i);
+ break;
+ }
+ /* did we find any match? */
+ if (i < orig_arity)
+ continue;
+
+ /* we might need more space in the input arrays */
+ if (arity >= in_size) {
+ const arch_register_req_t **new_in_reg_reqs;
+ ir_node **new_in;
+
+ in_size *= 2;
+ new_in_reg_reqs
+ = obstack_alloc(obst, in_size*sizeof(in_reg_reqs[0]));
+ memcpy(new_in_reg_reqs, in_reg_reqs, arity * sizeof(new_in_reg_reqs[0]));
+ new_in = ALLOCANZ(ir_node*, in_size);
+ memcpy(new_in, in, arity*sizeof(new_in[0]));
+
+ in_reg_reqs = new_in_reg_reqs;
+ in = new_in;
+ }
+
+ /* add a new (dummy) input which occupies the register */
+ assert(outreq->type & arch_register_req_type_limited);
+ in_reg_reqs[arity] = outreq;
+ in[arity] = new_bd_ia32_ProduceVal(NULL, block);
+ be_dep_on_frame(in[arity]);
+ ++arity;
+ }
+ } else {
+ int i;
+ bitset_t *used_outs = bitset_alloca(out_arity);
+ int orig_out_arity = out_arity;
+ for (i = 0; i < arity; ++i) {
+ int o;
+ const arch_register_req_t *inreq = in_reg_reqs[i];
+
+ if (inreq->cls == NULL) {
+ continue;
+ }
+
+ for (o = 0; o < orig_out_arity; ++o) {
+ const arch_register_req_t *outreq;
+ if (bitset_is_set(used_outs, o))
+ continue;
+ outreq = out_reg_reqs[o];
+ if (!can_match(outreq, inreq))
+ continue;
+ bitset_set(used_outs, i);
+ break;
+ }
+ /* did we find any match? */
+ if (o < orig_out_arity)
+ continue;
+
+ /* we might need more space in the output arrays */
+ if (out_arity >= out_size) {
+ const arch_register_req_t **new_out_reg_reqs;
+
+ out_size *= 2;
+ new_out_reg_reqs
+ = obstack_alloc(obst, out_size*sizeof(out_reg_reqs[0]));
+ memcpy(new_out_reg_reqs, out_reg_reqs,
+ out_arity * sizeof(new_out_reg_reqs[0]));
+ out_reg_reqs = new_out_reg_reqs;
+ }
+
+ /* add a new (dummy) output which occupies the register */
+ assert(inreq->type & arch_register_req_type_limited);
+ out_reg_reqs[out_arity] = inreq;
+ ++out_arity;
+ }
+ }
+
+ /* append none register requirement for the memory output */
+ if (out_arity + 1 >= out_size) {
+ const arch_register_req_t **new_out_reg_reqs;
+
+ out_size = out_arity + 1;
+ new_out_reg_reqs
+ = obstack_alloc(obst, out_size*sizeof(out_reg_reqs[0]));
+ memcpy(new_out_reg_reqs, out_reg_reqs,
+ out_arity * sizeof(new_out_reg_reqs[0]));
+ out_reg_reqs = new_out_reg_reqs;
+ }
+
+ /* add a new (dummy) output which occupies the register */
+ out_reg_reqs[out_arity] = arch_no_register_req;
+ ++out_arity;
+
new_node = new_bd_ia32_Asm(dbgi, new_block, arity, in, out_arity,
get_ASM_text(node), register_map);
if (arity == 0)
be_dep_on_frame(new_node);
- set_ia32_out_req_all(new_node, out_reg_reqs);
+ info = be_get_info(new_node);
+ for (i = 0; i < out_arity; ++i) {
+ info->out_infos[i].req = out_reg_reqs[i];
+ }
set_ia32_in_req_all(new_node, in_reg_reqs);
SET_IA32_ORIG_NODE(new_node, node);
return new_node;
}
-ir_node *gen_CopyB(ir_node *node) {
- ir_node *block = NULL;
- ir_node *src = NULL;
- ir_node *new_src = NULL;
- ir_node *dst = NULL;
- ir_node *new_dst = NULL;
- ir_node *mem = NULL;
- ir_node *new_mem = NULL;
+ir_node *gen_CopyB(ir_node *node)
+{
+ ir_node *block = get_new_node(get_nodes_block(node));
+ ir_node *src = get_CopyB_src(node);
+ ir_node *new_src = get_new_node(src);
+ ir_node *dst = get_CopyB_dst(node);
+ ir_node *new_dst = get_new_node(dst);
+ ir_node *mem = get_CopyB_mem(node);
+ ir_node *new_mem = get_new_node(mem);
ir_node *res = NULL;
dbg_info *dbgi = get_irn_dbg_info(node);
int size = get_type_size_bytes(get_CopyB_type(node));
int rem;
- switch (be_transformer) {
- case TRANSFORMER_DEFAULT:
- block = be_transform_node(get_nodes_block(node));
- src = get_CopyB_src(node);
- new_src = be_transform_node(src);
- dst = get_CopyB_dst(node);
- new_dst = be_transform_node(dst);
- mem = get_CopyB_mem(node);
- new_mem = be_transform_node(mem);
- break;
-
-#ifdef FIRM_GRGEN_BE
- case TRANSFORMER_PBQP:
- case TRANSFORMER_RAND:
- block = get_nodes_block(node);
- new_src = get_CopyB_src(node);
- new_dst = get_CopyB_dst(node);
- new_mem = get_CopyB_mem(node);
- break;
-#endif
-
- default: panic("invalid transformer");
- }
-
/* If we have to copy more than 32 bytes, we use REP MOVSx and */
/* then we need the size explicitly in ECX. */
if (size >= 32 * 4) {
rem = size & 0x3; /* size % 4 */
size >>= 2;
- res = new_bd_ia32_Const(dbgi, block, NULL, 0, size);
+ res = new_bd_ia32_Const(dbgi, block, NULL, 0, 0, size);
be_dep_on_frame(res);
res = new_bd_ia32_CopyB(dbgi, block, new_dst, new_src, res, new_mem, rem);
return res;
}
-ir_node *gen_Proj_tls(ir_node *node) {
- ir_node *block = NULL;
- dbg_info *dbgi = NULL;
- ir_node *res = NULL;
-
- switch (be_transformer) {
- case TRANSFORMER_DEFAULT:
- block = be_transform_node(get_nodes_block(node));
- break;
-
-#ifdef FIRM_GRGEN_BE
- case TRANSFORMER_PBQP:
- case TRANSFORMER_RAND:
- block = get_nodes_block(node);
- break;
-#endif
-
- default: panic("invalid transformer");
- }
+ir_node *gen_Proj_tls(ir_node *node)
+{
+ ir_node *block = get_new_node(get_nodes_block(node));
+ ir_node *res = NULL;
- res = new_bd_ia32_LdTls(dbgi, block, mode_Iu);
+ res = new_bd_ia32_LdTls(NULL, block, mode_Iu);
return res;
}
ir_node *gen_Unknown(ir_node *node)
{
- ir_mode *mode = get_irn_mode(node);
+ ir_mode *mode = get_irn_mode(node);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = get_irg_start_block(irg);
+ ir_node *res = NULL;
if (mode_is_float(mode)) {
if (ia32_cg_config.use_sse2) {
- return ia32_new_Unknown_xmm(env_cg);
+ res = new_bd_ia32_xUnknown(dbgi, block);
} else {
- /* Unknown nodes are buggy in x87 simulator, use zero for now... */
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *block = get_irg_start_block(irg);
- ir_node *ret = new_bd_ia32_vfldz(dbgi, block);
-
- be_dep_on_frame(ret);
- return ret;
+ res = new_bd_ia32_vfldz(dbgi, block);
}
} else if (ia32_mode_needs_gp_reg(mode)) {
- return ia32_new_Unknown_gp(env_cg);
+ res = new_bd_ia32_Unknown(dbgi, block);
} else {
panic("unsupported Unknown-Mode");
}
- return NULL;
+
+ be_dep_on_frame(res);
+ return res;
}
const arch_register_req_t *make_register_req(const constraint_t *constraint,
/* pure memory ops */
if (constraint->cls == NULL) {
- return &no_register_req;
+ return arch_no_register_req;
}
if (constraint->allowed_registers != 0
ir_node *try_create_Immediate(ir_node *node, char immediate_constraint_type)
{
- int minus = 0;
- tarval *offset = NULL;
- int offset_sign = 0;
long val = 0;
ir_entity *symconst_ent = NULL;
- int symconst_sign = 0;
ir_mode *mode;
ir_node *cnst = NULL;
ir_node *symconst = NULL;
return NULL;
}
- if (is_Minus(node)) {
- minus = 1;
- node = get_Minus_op(node);
- }
-
if (is_Const(node)) {
- cnst = node;
- symconst = NULL;
- offset_sign = minus;
- } else if (is_SymConst(node)) {
- cnst = NULL;
- symconst = node;
- symconst_sign = minus;
+ cnst = node;
+ symconst = NULL;
+ } else if (is_Global(node)) {
+ cnst = NULL;
+ symconst = node;
} else if (is_Add(node)) {
ir_node *left = get_Add_left(node);
ir_node *right = get_Add_right(node);
- if (is_Const(left) && is_SymConst(right)) {
- cnst = left;
- symconst = right;
- symconst_sign = minus;
- offset_sign = minus;
- } else if (is_SymConst(left) && is_Const(right)) {
- cnst = right;
- symconst = left;
- symconst_sign = minus;
- offset_sign = minus;
- }
- } else if (is_Sub(node)) {
- ir_node *left = get_Sub_left(node);
- ir_node *right = get_Sub_right(node);
- if (is_Const(left) && is_SymConst(right)) {
- cnst = left;
- symconst = right;
- symconst_sign = !minus;
- offset_sign = minus;
- } else if (is_SymConst(left) && is_Const(right)) {
- cnst = right;
- symconst = left;
- symconst_sign = minus;
- offset_sign = !minus;
+ if (is_Const(left) && is_Global(right)) {
+ cnst = left;
+ symconst = right;
+ } else if (is_Global(left) && is_Const(right)) {
+ cnst = right;
+ symconst = left;
}
} else {
return NULL;
}
if (cnst != NULL) {
- offset = get_Const_tarval(cnst);
- if (tarval_is_long(offset)) {
- val = get_tarval_long(offset);
- } else {
- ir_fprintf(stderr, "Optimisation Warning: tarval from %+F is not a "
- "long?\n", cnst);
+ tarval *offset = get_Const_tarval(cnst);
+ if (!tarval_is_long(offset)) {
+ ir_fprintf(stderr, "Optimisation Warning: tarval of %+F is not a long?\n", cnst);
return NULL;
}
+ val = get_tarval_long(offset);
if (!check_immediate_constraint(val, immediate_constraint_type))
return NULL;
}
return NULL;
}
- /* unfortunately the assembler/linker doesn't support -symconst */
- if (symconst_sign)
- return NULL;
-
- if (get_SymConst_kind(symconst) != symconst_addr_ent)
- return NULL;
- symconst_ent = get_SymConst_entity(symconst);
+ symconst_ent = get_Global_entity(symconst);
}
if (cnst == NULL && symconst == NULL)
return NULL;
- if (offset_sign && offset != NULL) {
- offset = tarval_neg(offset);
- }
-
- new_node = ia32_create_Immediate(symconst_ent, symconst_sign, val);
-
+ new_node = ia32_create_Immediate(symconst_ent, 0, val);
return new_node;
}