/**
- * This file implements the IR transformation from firm into
- * ia32-Firm.
- *
+ * This file implements the IR transformation from firm into ia32-Firm.
+ * @author Christian Wuerdig
* $Id$
*/
#include "dbginfo.h"
#include "irprintf.h"
#include "debug.h"
+#include "irdom.h"
+#include "archop.h" /* we need this for Min and Max nodes */
#include "../benode_t.h"
#include "../besched.h"
#include "../beabi.h"
#include "bearch_ia32_t.h"
-
#include "ia32_nodes_attr.h"
-#include "../arch/archop.h" /* we need this for Min and Max nodes */
#include "ia32_transform.h"
#include "ia32_new_nodes.h"
#include "ia32_map_regs.h"
#include "ia32_dbg_stat.h"
+#include "ia32_optimize.h"
#include "gen_ia32_regalloc_if.h"
return NULL;
}
+/**
+ * SSE convert of an integer node into a floating point node.
+ */
+static ir_node *gen_sse_conv_int2float(ia32_code_gen_t *cg, dbg_info *dbg, ir_graph *irg, ir_node *block,
+ ir_node *in, ir_node *old_node, ir_mode *tgt_mode)
+{
+ ir_node *noreg = ia32_new_NoReg_gp(cg);
+ ir_node *nomem = new_rd_NoMem(irg);
+
+ ir_node *conv = new_rd_ia32_Conv_I2FP(dbg, irg, block, noreg, noreg, in, nomem);
+ set_ia32_src_mode(conv, get_irn_mode(in));
+ set_ia32_tgt_mode(conv, tgt_mode);
+ set_ia32_am_support(conv, ia32_am_Source);
+ SET_IA32_ORIG_NODE(conv, ia32_get_old_node_name(cg, old_node));
+
+ return new_rd_Proj(dbg, irg, block, conv, tgt_mode, pn_ia32_Conv_I2FP_res);
+}
+
/* Generates an entity for a known FP const (used for FP Neg + Abs) */
static ident *gen_fp_known_const(ir_mode *mode, ia32_known_const_t kct) {
static const struct {
/* Check if immediate optimization is on and */
/* if it's an operation with immediate. */
- if (! (env->cg->opt & IA32_OPT_IMMOPS)) {
+ /* MulS and Mulh don't support immediates */
+ if (! (env->cg->opt & IA32_OPT_IMMOPS) ||
+ func == new_rd_ia32_Mulh ||
+ func == new_rd_ia32_MulS)
+ {
expr_op = op1;
imm_op = NULL;
}
if (tv) {
tv = tarval_mod(tv, new_tarval_from_long(32, mode_Iu));
+ set_ia32_Immop_tarval(imm_op, tv);
}
else {
imm_op = NULL;
if (get_irn_op(irn) == op_Div) {
set_Proj_proj(proj, pn_DivMod_res_div);
- in_keep[0] = new_rd_Proj(dbg, irg, block, res, mode_Is, pn_DivMod_res_mod);
+ in_keep[0] = new_rd_Proj(dbg, irg, block, res, mode, pn_DivMod_res_mod);
}
else {
set_Proj_proj(proj, pn_DivMod_res_mod);
- in_keep[0] = new_rd_Proj(dbg, irg, block, res, mode_Is, pn_DivMod_res_div);
+ in_keep[0] = new_rd_Proj(dbg, irg, block, res, mode, pn_DivMod_res_div);
}
be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1, in_keep);
SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env->cg, env->irn));
- set_ia32_res_mode(res, mode_Is);
+ set_ia32_res_mode(res, mode);
return res;
}
* @return the created ia32 Load node
*/
static ir_node *gen_Load(ia32_transform_env_t *env) {
- ir_node *node = env->irn;
- ir_node *noreg = ia32_new_NoReg_gp(env->cg);
- ir_node *ptr = get_Load_ptr(node);
- ir_node *lptr = ptr;
- ir_mode *mode = get_Load_mode(node);
- int is_imm = 0;
+ ir_node *node = env->irn;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *ptr = get_Load_ptr(node);
+ ir_node *lptr = ptr;
+ ir_mode *mode = get_Load_mode(node);
+ int is_imm = 0;
ir_node *new_op;
ia32_am_flavour_t am_flav = ia32_B;
set_ia32_am_flavour(new_op, am_flav);
set_ia32_ls_mode(new_op, mode);
+ /*
+ check for special case: the loaded value might not be used (optimized, volatile, ...)
+ we add a Proj + Keep for volatile loads and ignore all other cases
+ */
+ if (! get_proj_for_pn(node, pn_Load_res) && get_Load_volatility(node) == volatility_is_volatile) {
+ /* add a result proj and a Keep to produce a pseudo use */
+ ir_node *proj = new_r_Proj(env->irg, env->block, new_op, mode, pn_ia32_Load_res);
+ be_new_Keep(arch_get_irn_reg_class(env->cg->arch_env, proj, -1), env->irg, env->block, 1, &proj);
+ }
+
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
return new_op;
ir_node *ptr = get_Store_ptr(node);
ir_node *sptr = ptr;
ir_node *mem = get_Store_mem(node);
- ir_mode *mode = get_irn_mode(val);
+ ir_mode *mode = get_irn_link(node);
ir_node *sval = val;
int is_imm = 0;
ir_node *new_op;
set_ia32_am_support(new_op, ia32_am_Dest);
set_ia32_op_type(new_op, ia32_AddrModeD);
set_ia32_am_flavour(new_op, am_flav);
- set_ia32_ls_mode(new_op, get_irn_mode(val));
+ set_ia32_ls_mode(new_op, mode);
set_ia32_immop_type(new_op, immop);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
pn_Cmp pnc = get_Proj_proj(sel);
if ((pnc == pn_Cmp_Eq || pnc == pn_Cmp_Lg) && mode_is_int(get_irn_mode(expr))) {
- if (classify_tarval(get_ia32_Immop_tarval(cnst)) == TV_CLASSIFY_NULL) {
+ if (get_ia32_op_type(cnst) == ia32_Const &&
+ classify_tarval(get_ia32_Immop_tarval(cnst)) == TV_CLASSIFY_NULL)
+ {
/* a Cmp A =/!= 0 */
ir_node *op1 = expr;
ir_node *op2 = expr;
return NULL;
}
-typedef ir_node *set_func_t(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *cmp_a, ir_node *cmp_b, ir_mode *mode);
typedef ir_node *cmov_func_t(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *cmp_a, ir_node *cmp_b, \
ir_node *psi_true, ir_node *psi_default, ir_mode *mode);
ir_node *psi_default = get_Psi_default(node);
ir_node *noreg = ia32_new_NoReg_gp(cg);
ir_node *nomem = new_rd_NoMem(irg);
- ir_node *cmp, *cmp_a, *cmp_b, *and1, *and2, *new_op, *c1, *c2 = NULL;
+ ir_node *cmp, *cmp_a, *cmp_b, *and1, *and2, *new_op = NULL;
int pnc;
-
assert(get_irn_mode(cmp_proj) == mode_b && "Condition for Psi must have mode_b");
cmp = get_Proj_pred(cmp_proj);
/* in case the compare operands are int, we move them into xmm register */
if (! mode_is_float(get_irn_mode(cmp_a))) {
- c1 = new_rd_ia32_Conv_I2FP(dbg, irg, block, noreg, noreg, cmp_a, nomem);
- set_ia32_src_mode(c1, get_irn_mode(cmp_a));
- set_ia32_tgt_mode(c1, mode_D);
- set_ia32_am_support(c1, ia32_am_Source);
- SET_IA32_ORIG_NODE(c1, ia32_get_old_node_name(cg, node));
- c2 = new_rd_ia32_Conv_I2FP(dbg, irg, block, noreg, noreg, cmp_b, nomem);
- set_ia32_src_mode(c2, get_irn_mode(cmp_b));
- set_ia32_tgt_mode(c2, mode_D);
- set_ia32_am_support(c2, ia32_am_Source);
- SET_IA32_ORIG_NODE(c2, ia32_get_old_node_name(cg, node));
-
- cmp_a = new_rd_Proj(dbg, irg, block, c1, mode_D, pn_ia32_Conv_I2FP_res);
- cmp_b = new_rd_Proj(dbg, irg, block, c2, mode_D, pn_ia32_Conv_I2FP_res);
-
- pnc += pn_Cmp_Uo; /* transform integer compare to fp compare */
+ cmp_a = gen_sse_conv_int2float(cg, dbg, irg, block, cmp_a, node, mode_D);
+ cmp_b = gen_sse_conv_int2float(cg, dbg, irg, block, cmp_b, node, mode_D);
+
+ pnc |= 8; /* transform integer compare to fp compare */
}
new_op = new_rd_ia32_xCmp(dbg, irg, block, noreg, noreg, cmp_a, cmp_b, nomem);
SET_IA32_ORIG_NODE(and1, ia32_get_old_node_name(cg, node));
and1 = new_rd_Proj(dbg, irg, block, and1, mode, pn_ia32_xAnd_res);
- and2 = new_rd_ia32_xAndNot(dbg, irg, block, noreg, noreg, psi_default, new_op, nomem);
+ and2 = new_rd_ia32_xAndNot(dbg, irg, block, noreg, noreg, new_op, psi_default, nomem);
set_ia32_am_support(and2, ia32_am_Source);
set_ia32_res_mode(and2, mode);
SET_IA32_ORIG_NODE(and2, ia32_get_old_node_name(cg, node));
}
else {
/* integer psi */
- set_func_t *set_func = NULL;
- cmov_func_t *cmov_func = NULL;
+ construct_binop_func *set_func = NULL;
+ cmov_func_t *cmov_func = NULL;
if (mode_is_float(get_irn_mode(cmp_a))) {
/* 1st case: compare operands are floats */
set_func = new_rd_ia32_vfCmpSet;
cmov_func = new_rd_ia32_vfCmpCMov;
}
+
+ pnc &= 7; /* fp compare -> int compare */
}
else {
/* 2nd case: compare operand are integer too */
- set_func = new_rd_ia32_Set;
- cmov_func = new_rd_ia32_CMov;
+ set_func = new_rd_ia32_CmpSet;
+ cmov_func = new_rd_ia32_CmpCMov;
}
/* create the nodes */
- if (is_ia32_Const_1(psi_true) && is_ia32_Const_0(psi_default)) {
- /* first case for SETcc: default is 0, set to 1 iff condition is true */
- new_op = set_func(dbg, irg, block, cmp_a, cmp_b, mode);
- set_ia32_pncode(new_op, pnc);
- }
- else if (is_ia32_Const_0(psi_true) && is_ia32_Const_1(psi_default)) {
- /* second case for SETcc: default is 1, set to 0 iff condition is true: */
- /* we invert condition and set default to 0 */
- new_op = set_func(dbg, irg, block, cmp_a, cmp_b, mode);
- set_ia32_pncode(new_op, get_negated_pnc(pnc, mode));
+ /* check for special case first: And/Or -- Cmp with 0 -- Psi */
+ if (is_ia32_Const_0(cmp_b) && is_Proj(cmp_a) && (is_ia32_And(get_Proj_pred(cmp_a)) || is_ia32_Or(get_Proj_pred(cmp_a)))) {
+ if (is_ia32_Const_1(psi_true) && is_ia32_Const_0(psi_default)) {
+ /* first case for SETcc: default is 0, set to 1 iff condition is true */
+ new_op = new_rd_ia32_PsiCondSet(dbg, irg, block, cmp_a, mode);
+ set_ia32_pncode(new_op, pnc);
+ }
+ else if (is_ia32_Const_0(psi_true) && is_ia32_Const_1(psi_default)) {
+ /* second case for SETcc: default is 1, set to 0 iff condition is true: */
+ /* we invert condition and set default to 0 */
+ new_op = new_rd_ia32_PsiCondSet(dbg, irg, block, cmp_a, mode);
+ set_ia32_pncode(new_op, get_inversed_pnc(pnc));
+ }
+ else {
+ /* otherwise: use CMOVcc */
+ new_op = new_rd_ia32_PsiCondCMov(dbg, irg, block, cmp_a, psi_true, psi_default, mode);
+ set_ia32_pncode(new_op, pnc);
+ }
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(cg, node));
}
else {
- /* otherwise: use CMOVcc */
- new_op = cmov_func(dbg, irg, block, cmp_a, cmp_b, psi_true, psi_default, mode);
- set_ia32_pncode(new_op, pnc);
+ env->irn = cmp;
+ if (is_ia32_Const_1(psi_true) && is_ia32_Const_0(psi_default)) {
+ /* first case for SETcc: default is 0, set to 1 iff condition is true */
+ new_op = gen_binop(env, cmp_a, cmp_b, set_func);
+ set_ia32_pncode(get_Proj_pred(new_op), pnc);
+ set_ia32_am_support(get_Proj_pred(new_op), ia32_am_Source);
+ }
+ else if (is_ia32_Const_0(psi_true) && is_ia32_Const_1(psi_default)) {
+ /* second case for SETcc: default is 1, set to 0 iff condition is true: */
+ /* we invert condition and set default to 0 */
+ new_op = gen_binop(env, cmp_a, cmp_b, set_func);
+ set_ia32_pncode(get_Proj_pred(new_op), get_inversed_pnc(pnc));
+ set_ia32_am_support(get_Proj_pred(new_op), ia32_am_Source);
+ }
+ else {
+ /* otherwise: use CMOVcc */
+ new_op = cmov_func(dbg, irg, block, cmp_a, cmp_b, psi_true, psi_default, mode);
+ set_ia32_pncode(new_op, pnc);
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(cg, node));
+ }
}
-
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(cg, node));
}
return new_op;
set_ia32_am_support(fist, ia32_am_Dest);
set_ia32_op_type(fist, ia32_AddrModeD);
set_ia32_am_flavour(fist, ia32_B);
- set_ia32_ls_mode(fist, mode_E);
+ set_ia32_ls_mode(fist, mode_F);
mem = new_r_Proj(irg, block, fist, mode_M, pn_ia32_vfist_M);
set_ia32_am_support(fild, ia32_am_Source);
set_ia32_op_type(fild, ia32_AddrModeS);
set_ia32_am_flavour(fild, ia32_B);
- set_ia32_ls_mode(fild, mode_E);
+ set_ia32_ls_mode(fild, mode_F);
- return new_r_Proj(irg, block, fild, mode_E, 0);
+ return new_r_Proj(irg, block, fild, mode_F, 0);
}
/**
*
********************************************/
+ /**
+ * Decides in which block the transformed StackParam should be placed.
+ * If the StackParam has more than one user, the dominator block of
+ * the users will be returned. In case of only one user, this is either
+ * the user block or, in case of a Phi, the predecessor block of the Phi.
+ */
+ static ir_node *get_block_transformed_stack_param(ir_node *irn) {
+ ir_node *dom_bl = NULL;
+
+ if (get_irn_n_edges(irn) == 1) {
+ ir_node *src = get_edge_src_irn(get_irn_out_edge_first(irn));
+
+ if (! is_Phi(src)) {
+ dom_bl = get_nodes_block(src);
+ }
+ else {
+ /* Determine on which in position of the Phi the irn is */
+ /* and get the corresponding cfg predecessor block. */
+
+ int i = get_irn_pred_pos(src, irn);
+ assert(i >= 0 && "kaputt");
+ dom_bl = get_Block_cfgpred_block(get_nodes_block(src), i);
+ }
+ }
+ else {
+ dom_bl = node_users_smallest_common_dominator(irn, 1);
+ }
+
+ assert(dom_bl && "dominator block not found");
+
+ return dom_bl;
+ }
+
static ir_node *gen_be_StackParam(ia32_transform_env_t *env) {
ir_node *new_op = NULL;
ir_node *node = env->irn;
entity *ent = be_get_frame_entity(node);
ir_mode *mode = env->mode;
-// /* If the StackParam has only one user -> */
-// /* put it in the Block where the user resides */
-// if (get_irn_n_edges(node) == 1) {
-// env->block = get_nodes_block(get_edge_src_irn(get_irn_out_edge_first(node)));
-// }
+ /* choose the block where to place the load */
+ env->block = get_block_transformed_stack_param(node);
if (mode_is_float(mode)) {
FP_USED(env->cg);
return NULL;
}
+/**********************************************************************
+ * _ _ _
+ * | | | | | |
+ * | | _____ _____ _ __ ___ __| | _ __ ___ __| | ___ ___
+ * | |/ _ \ \ /\ / / _ \ '__/ _ \/ _` | | '_ \ / _ \ / _` |/ _ \/ __|
+ * | | (_) \ V V / __/ | | __/ (_| | | | | | (_) | (_| | __/\__ \
+ * |_|\___/ \_/\_/ \___|_| \___|\__,_| |_| |_|\___/ \__,_|\___||___/
+ *
+ **********************************************************************/
+
+/* These nodes are created in intrinsic lowering (64bit -> 32bit) */
+
+typedef ir_node *construct_load_func(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, \
+ ir_node *mem);
+
+typedef ir_node *construct_store_func(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, \
+ ir_node *val, ir_node *mem);
+
+/**
+ * Transforms a lowered Load into a "real" one.
+ */
+static ir_node *gen_lowered_Load(ia32_transform_env_t *env, construct_load_func func, char fp_unit) {
+ ir_node *node = env->irn;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_mode *mode = get_ia32_ls_mode(node);
+ ir_node *new_op;
+ char *am_offs;
+ ia32_am_flavour_t am_flav = ia32_B;
+
+ /*
+ Could be that we have SSE2 unit, but due to 64Bit Div/Conv
+ lowering we have x87 nodes, so we need to enforce simulation.
+ */
+ if (mode_is_float(mode)) {
+ FP_USED(env->cg);
+ if (fp_unit == fp_x87)
+ FORCE_x87(env->cg);
+ }
+
+ new_op = func(env->dbg, env->irg, env->block, get_irn_n(node, 0), noreg, get_irn_n(node, 1));
+ am_offs = get_ia32_am_offs(node);
+
+ if (am_offs) {
+ am_flav |= ia32_O;
+ add_ia32_am_offs(new_op, am_offs);
+ }
+
+ set_ia32_am_support(new_op, ia32_am_Source);
+ set_ia32_op_type(new_op, ia32_AddrModeS);
+ set_ia32_am_flavour(new_op, am_flav);
+ set_ia32_ls_mode(new_op, mode);
+ set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
+ set_ia32_use_frame(new_op);
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, node));
+
+ return new_op;
+}
+
+/**
+* Transforms a lowered Store into a "real" one.
+*/
+static ir_node *gen_lowered_Store(ia32_transform_env_t *env, construct_store_func func, char fp_unit) {
+ ir_node *node = env->irn;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_mode *mode = get_ia32_ls_mode(node);
+ ir_node *new_op;
+ char *am_offs;
+ ia32_am_flavour_t am_flav = ia32_B;
+
+ /*
+ Could be that we have SSE2 unit, but due to 64Bit Div/Conv
+ lowering we have x87 nodes, so we need to enforce simulation.
+ */
+ if (mode_is_float(mode)) {
+ FP_USED(env->cg);
+ if (fp_unit == fp_x87)
+ FORCE_x87(env->cg);
+ }
+
+ new_op = func(env->dbg, env->irg, env->block, get_irn_n(node, 0), noreg, get_irn_n(node, 1), get_irn_n(node, 2));
+
+ if (am_offs = get_ia32_am_offs(node)) {
+ am_flav |= ia32_O;
+ add_ia32_am_offs(new_op, am_offs);
+ }
+
+ set_ia32_am_support(new_op, ia32_am_Dest);
+ set_ia32_op_type(new_op, ia32_AddrModeD);
+ set_ia32_am_flavour(new_op, am_flav);
+ set_ia32_ls_mode(new_op, mode);
+ set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
+ set_ia32_use_frame(new_op);
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, node));
+
+ return new_op;
+}
+
+
+/**
+ * Transforms an ia32_l_XXX into a "real" XXX node
+ *
+ * @param env The transformation environment
+ * @return the created ia32 XXX node
+ */
+#define GEN_LOWERED_OP(op) \
+ static ir_node *gen_ia32_l_##op(ia32_transform_env_t *env) { \
+ if (mode_is_float(env->mode)) \
+ FP_USED(env->cg); \
+ return gen_binop(env, get_binop_left(env->irn), get_binop_right(env->irn), new_rd_ia32_##op); \
+ }
+
+#define GEN_LOWERED_x87_OP(op) \
+ static ir_node *gen_ia32_l_##op(ia32_transform_env_t *env) { \
+ ir_node *new_op; \
+ FORCE_x87(env->cg); \
+ new_op = gen_binop(env, get_binop_left(env->irn), get_binop_right(env->irn), new_rd_ia32_##op); \
+ set_ia32_am_support(get_Proj_pred(new_op), ia32_am_None); \
+ return new_op; \
+ }
+
+#define GEN_LOWERED_UNOP(op) \
+ static ir_node *gen_ia32_l_##op(ia32_transform_env_t *env) { \
+ return gen_unop(env, get_unop_op(env->irn), new_rd_ia32_##op); \
+ }
+
+#define GEN_LOWERED_SHIFT_OP(op) \
+ static ir_node *gen_ia32_l_##op(ia32_transform_env_t *env) { \
+ return gen_shift_binop(env, get_binop_left(env->irn), get_binop_right(env->irn), new_rd_ia32_##op); \
+ }
+
+#define GEN_LOWERED_LOAD(op, fp_unit) \
+ static ir_node *gen_ia32_l_##op(ia32_transform_env_t *env) { \
+ return gen_lowered_Load(env, new_rd_ia32_##op, fp_unit); \
+ }
+
+#define GEN_LOWERED_STORE(op, fp_unit) \
+ static ir_node *gen_ia32_l_##op(ia32_transform_env_t *env) { \
+ return gen_lowered_Store(env, new_rd_ia32_##op, fp_unit); \
+}
+
+GEN_LOWERED_OP(AddC)
+GEN_LOWERED_OP(Add)
+GEN_LOWERED_OP(SubC)
+GEN_LOWERED_OP(Sub)
+GEN_LOWERED_OP(Mul)
+GEN_LOWERED_OP(Eor)
+GEN_LOWERED_x87_OP(vfdiv)
+GEN_LOWERED_x87_OP(vfmul)
+GEN_LOWERED_x87_OP(vfsub)
+
+GEN_LOWERED_UNOP(Minus)
+
+GEN_LOWERED_LOAD(vfild, fp_x87)
+GEN_LOWERED_LOAD(Load, fp_none)
+GEN_LOWERED_STORE(vfist, fp_x87)
+GEN_LOWERED_STORE(Store, fp_none)
+
+/**
+ * Transforms a l_MulS into a "real" MulS node.
+ *
+ * @param env The transformation environment
+ * @return the created ia32 MulS node
+ */
+static ir_node *gen_ia32_l_MulS(ia32_transform_env_t *env) {
+
+ /* l_MulS is already a mode_T node, so we create the MulS in the normal way */
+ /* and then skip the result Proj, because all needed Projs are already there. */
+
+ ir_node *new_op = gen_binop(env, get_binop_left(env->irn), get_binop_right(env->irn), new_rd_ia32_MulS);
+ ir_node *muls = get_Proj_pred(new_op);
+
+ /* MulS cannot have AM for destination */
+ if (get_ia32_am_support(muls) != ia32_am_None)
+ set_ia32_am_support(muls, ia32_am_Source);
+
+ return muls;
+}
+
+GEN_LOWERED_SHIFT_OP(Shl)
+GEN_LOWERED_SHIFT_OP(Shr)
+GEN_LOWERED_SHIFT_OP(Shrs)
+
+/**
+ * Transforms a l_ShlD/l_ShrD into a ShlD/ShrD. Those nodes have 3 data inputs:
+ * op1 - target to be shifted
+ * op2 - contains bits to be shifted into target
+ * op3 - shift count
+ * Only op3 can be an immediate.
+ */
+static ir_node *gen_lowered_64bit_shifts(ia32_transform_env_t *env, ir_node *op1, ir_node *op2, ir_node *count) {
+ ir_node *new_op = NULL;
+ ir_mode *mode = env->mode;
+ dbg_info *dbg = env->dbg;
+ ir_graph *irg = env->irg;
+ ir_node *block = env->block;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *nomem = new_NoMem();
+ ir_node *imm_op;
+ tarval *tv;
+ DEBUG_ONLY(firm_dbg_module_t *mod = env->mod;)
+
+ assert(! mode_is_float(mode) && "Shift/Rotate with float not supported");
+
+ /* Check if immediate optimization is on and */
+ /* if it's an operation with immediate. */
+ imm_op = (env->cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(NULL, count) : NULL;
+
+ /* Limit imm_op within range imm8 */
+ if (imm_op) {
+ tv = get_ia32_Immop_tarval(imm_op);
+
+ if (tv) {
+ tv = tarval_mod(tv, new_tarval_from_long(32, mode_Iu));
+ set_ia32_Immop_tarval(imm_op, tv);
+ }
+ else {
+ imm_op = NULL;
+ }
+ }
+
+ /* integer operations */
+ if (imm_op) {
+ /* This is ShiftD with const */
+ DB((mod, LEVEL_1, "ShiftD with immediate ..."));
+
+ if (is_ia32_l_ShlD(env->irn))
+ new_op = new_rd_ia32_ShlD(dbg, irg, block, noreg, noreg, op1, op2, noreg, nomem);
+ else
+ new_op = new_rd_ia32_ShrD(dbg, irg, block, noreg, noreg, op1, op2, noreg, nomem);
+ set_ia32_Immop_attr(new_op, imm_op);
+ }
+ else {
+ /* This is a normal ShiftD */
+ DB((mod, LEVEL_1, "ShiftD binop ..."));
+ if (is_ia32_l_ShlD(env->irn))
+ new_op = new_rd_ia32_ShlD(dbg, irg, block, noreg, noreg, op1, op2, count, nomem);
+ else
+ new_op = new_rd_ia32_ShrD(dbg, irg, block, noreg, noreg, op1, op2, count, nomem);
+ }
+
+ /* set AM support */
+ set_ia32_am_support(new_op, ia32_am_Dest);
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+
+ set_ia32_res_mode(new_op, mode);
+ set_ia32_emit_cl(new_op);
+
+ return new_rd_Proj(dbg, irg, block, new_op, mode, 0);
+}
+
+static ir_node *gen_ia32_l_ShlD(ia32_transform_env_t *env) {
+ return gen_lowered_64bit_shifts(env, get_irn_n(env->irn, 0), get_irn_n(env->irn, 1), get_irn_n(env->irn, 2));
+}
+
+static ir_node *gen_ia32_l_ShrD(ia32_transform_env_t *env) {
+ return gen_lowered_64bit_shifts(env, get_irn_n(env->irn, 0), get_irn_n(env->irn, 1), get_irn_n(env->irn, 2));
+}
+
+/**
+ * In case SSE Unit is used, the node is transformed into a vfst + xLoad.
+ */
+static ir_node *gen_ia32_l_X87toSSE(ia32_transform_env_t *env) {
+ ia32_code_gen_t *cg = env->cg;
+ ir_node *res = NULL;
+ ir_node *ptr = get_irn_n(env->irn, 0);
+ ir_node *val = get_irn_n(env->irn, 1);
+ ir_node *mem = get_irn_n(env->irn, 2);
+
+ if (USE_SSE2(cg)) {
+ ir_node *noreg = ia32_new_NoReg_gp(cg);
+
+ /* Store x87 -> MEM */
+ res = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, mem);
+ set_ia32_frame_ent(res, get_ia32_frame_ent(env->irn));
+ set_ia32_use_frame(res);
+ set_ia32_ls_mode(res, get_ia32_ls_mode(env->irn));
+ set_ia32_am_support(res, ia32_am_Dest);
+ set_ia32_am_flavour(res, ia32_B);
+ res = new_rd_Proj(env->dbg, env->irg, env->block, res, mode_M, pn_ia32_vfst_M);
+
+ /* Load MEM -> SSE */
+ res = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, res);
+ set_ia32_frame_ent(res, get_ia32_frame_ent(env->irn));
+ set_ia32_use_frame(res);
+ set_ia32_ls_mode(res, get_ia32_ls_mode(env->irn));
+ set_ia32_am_support(res, ia32_am_Source);
+ set_ia32_am_flavour(res, ia32_B);
+ res = new_rd_Proj(env->dbg, env->irg, env->block, res, get_ia32_ls_mode(env->irn), pn_ia32_xLoad_res);
+ }
+ else {
+ /* SSE unit is not used -> skip this node. */
+ int i;
+
+ edges_reroute(env->irn, val, env->irg);
+ for (i = get_irn_arity(env->irn) - 1; i >= 0; i--)
+ set_irn_n(env->irn, i, get_irg_bad(env->irg));
+ }
+
+ return res;
+}
+
+/**
+ * In case SSE Unit is used, the node is transformed into a xStore + vfld.
+ */
+static ir_node *gen_ia32_l_SSEtoX87(ia32_transform_env_t *env) {
+ ia32_code_gen_t *cg = env->cg;
+ ir_node *res = NULL;
+ ir_node *ptr = get_irn_n(env->irn, 0);
+ ir_node *val = get_irn_n(env->irn, 1);
+ ir_node *mem = get_irn_n(env->irn, 2);
+
+ if (USE_SSE2(cg)) {
+ ir_node *noreg = ia32_new_NoReg_gp(cg);
+
+ /* Store SSE -> MEM */
+ res = new_rd_ia32_xStore(env->dbg, env->irg, env->block, ptr, noreg, val, mem);
+ set_ia32_frame_ent(res, get_ia32_frame_ent(env->irn));
+ set_ia32_use_frame(res);
+ set_ia32_ls_mode(res, get_ia32_ls_mode(env->irn));
+ set_ia32_am_support(res, ia32_am_Dest);
+ set_ia32_am_flavour(res, ia32_B);
+ res = new_rd_Proj(env->dbg, env->irg, env->block, res, mode_M, pn_ia32_xStore_M);
+
+ /* Load MEM -> x87 */
+ res = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem);
+ set_ia32_frame_ent(res, get_ia32_frame_ent(env->irn));
+ set_ia32_use_frame(res);
+ set_ia32_ls_mode(res, get_ia32_ls_mode(env->irn));
+ set_ia32_am_support(res, ia32_am_Source);
+ set_ia32_am_flavour(res, ia32_B);
+ res = new_rd_Proj(env->dbg, env->irg, env->block, res, get_ia32_ls_mode(env->irn), pn_ia32_vfld_res);
+ }
+ else {
+ /* SSE unit is not used -> skip this node. */
+ int i;
+
+ edges_reroute(env->irn, val, env->irg);
+ for (i = get_irn_arity(env->irn) - 1; i >= 0; i--)
+ set_irn_n(env->irn, i, get_irg_bad(env->irg));
+ }
+
+ return res;
+}
/*********************************************************
* _ _ _
am_flav = get_ia32_am_flavour(irn);
+ if (get_ia32_am_sc(irn))
+ return;
+
/* only some LEAs can be transformed to an Add */
if (am_flav != ia32_am_B && am_flav != ia32_am_OB && am_flav != ia32_am_OI && am_flav != ia32_am_BI)
return;
GEN(Mux);
GEN(Psi);
+ /* transform ops from intrinsic lowering */
+ GEN(ia32_l_Add);
+ GEN(ia32_l_AddC);
+ GEN(ia32_l_Sub);
+ GEN(ia32_l_SubC);
+ GEN(ia32_l_Minus);
+ GEN(ia32_l_Mul);
+ GEN(ia32_l_Eor);
+ GEN(ia32_l_MulS);
+ GEN(ia32_l_Shl);
+ GEN(ia32_l_Shr);
+ GEN(ia32_l_Shrs);
+ GEN(ia32_l_ShlD);
+ GEN(ia32_l_ShrD);
+ GEN(ia32_l_vfdiv);
+ GEN(ia32_l_vfmul);
+ GEN(ia32_l_vfsub);
+ GEN(ia32_l_vfild);
+ GEN(ia32_l_Load);
+ GEN(ia32_l_vfist);
+ GEN(ia32_l_Store);
+ GEN(ia32_l_X87toSSE);
+ GEN(ia32_l_SSEtoX87);
+
IGN(Call);
IGN(Alloc);
DB((cg->mod, LEVEL_1, "ignored\n"));
}
}
+
+/**
+ * Transforms a psi condition.
+ */
+static void transform_psi_cond(ir_node *cond, ir_mode *mode, ia32_code_gen_t *cg) {
+ int i;
+
+ /* if the mode is target mode, we have already seen this part of the tree */
+ if (get_irn_mode(cond) == mode)
+ return;
+
+ assert(get_irn_mode(cond) == mode_b && "logical operator for condition must be mode_b");
+
+ set_irn_mode(cond, mode);
+
+ for (i = get_irn_arity(cond) - 1; i >= 0; i--) {
+ ir_node *in = get_irn_n(cond, i);
+
+ /* if in is a compare: transform into Set/xCmp */
+ if (is_Proj(in)) {
+ ir_node *new_op = NULL;
+ ir_node *cmp = get_Proj_pred(in);
+ ir_node *cmp_a = get_Cmp_left(cmp);
+ ir_node *cmp_b = get_Cmp_right(cmp);
+ dbg_info *dbg = get_irn_dbg_info(cmp);
+ ir_graph *irg = get_irn_irg(cmp);
+ ir_node *block = get_nodes_block(cmp);
+ ir_node *noreg = ia32_new_NoReg_gp(cg);
+ ir_node *nomem = new_rd_NoMem(irg);
+ int pnc = get_Proj_proj(in);
+
+ /* this is a compare */
+ if (mode_is_float(mode)) {
+ /* Psi is float, we need a floating point compare */
+
+ if (USE_SSE2(cg)) {
+ /* SSE FPU */
+ if (! mode_is_float(get_irn_mode(cmp_a))) {
+ cmp_a = gen_sse_conv_int2float(cg, dbg, irg, block, cmp_a, cmp_a, mode);
+ cmp_b = gen_sse_conv_int2float(cg, dbg, irg, block, cmp_b, cmp_b, mode);
+ pnc |= 8;
+ }
+
+ new_op = new_rd_ia32_xCmp(dbg, irg, block, noreg, noreg, cmp_a, cmp_b, nomem);
+ set_ia32_pncode(new_op, pnc);
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(cg, cmp));
+ }
+ else {
+ /* x87 FPU */
+ assert(0);
+ }
+ }
+ else {
+ /* integer Psi */
+ ia32_transform_env_t tenv;
+ construct_binop_func *set_func = NULL;
+
+ if (mode_is_float(get_irn_mode(cmp_a))) {
+ /* 1st case: compare operands are floats */
+ FP_USED(cg);
+
+ if (USE_SSE2(cg)) {
+ /* SSE FPU */
+ set_func = new_rd_ia32_xCmpSet;
+ }
+ else {
+ /* x87 FPU */
+ set_func = new_rd_ia32_vfCmpSet;
+ }
+
+ pnc &= 7; /* fp compare -> int compare */
+ }
+ else {
+ /* 2nd case: compare operand are integer too */
+ set_func = new_rd_ia32_CmpSet;
+ }
+
+ tenv.block = block;
+ tenv.cg = cg;
+ tenv.dbg = dbg;
+ tenv.irg = irg;
+ tenv.irn = cmp;
+ tenv.mode = mode;
+ tenv.mod = cg->mod;
+
+ new_op = gen_binop(&tenv, cmp_a, cmp_b, set_func);
+ set_ia32_pncode(get_Proj_pred(new_op), pnc);
+ set_ia32_am_support(get_Proj_pred(new_op), ia32_am_Source);
+ }
+
+ /* the the new compare as in */
+ set_irn_n(cond, i, new_op);
+ }
+ else {
+ /* another complex condition */
+ transform_psi_cond(in, mode, cg);
+ }
+ }
+}
+
+/**
+ * The Psi selector can be a tree of compares combined with "And"s and "Or"s.
+ * We create a Set node, respectively a xCmp in case the Psi is a float, for each
+ * compare, which causes the compare result to be stores in a register. The
+ * "And"s and "Or"s are transformed later, we just have to set their mode right.
+ */
+void ia32_transform_psi_cond_tree(ir_node *node, void *env) {
+ ia32_code_gen_t *cg = (ia32_code_gen_t *)env;
+ ir_node *psi_sel, *new_cmp, *block;
+ ir_graph *irg;
+ ir_mode *mode;
+
+ /* check for Psi */
+ if (get_irn_opcode(node) != iro_Psi)
+ return;
+
+ psi_sel = get_Psi_cond(node, 0);
+
+ /* if psi_cond is a cmp: do nothing, this case is covered by gen_Psi */
+ if (is_Proj(psi_sel))
+ return;
+
+ mode = get_irn_mode(node);
+
+ transform_psi_cond(psi_sel, mode, cg);
+
+ irg = get_irn_irg(node);
+ block = get_nodes_block(node);
+
+ /* we need to compare the evaluated condition tree with 0 */
+
+ /* BEWARE: new_r_Const_long works for floating point as well */
+ new_cmp = new_r_Cmp(irg, block, psi_sel, new_r_Const_long(irg, block, mode, 0));
+ /* transform the const */
+ ia32_place_consts_set_modes(new_cmp, cg);
+ new_cmp = new_r_Proj(irg, block, new_cmp, mode_b, pn_Cmp_Ne + (mode_is_float(mode) ? pn_Cmp_Uo : 0));
+
+ set_Psi_cond(node, 0, new_cmp);
+}
projects / libfirm / blobdiff