+ /* integer psi */
+ 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 */
+ FP_USED(cg);
+
+ if (USE_SSE2(cg)) {
+ /* SSE FPU */
+ set_func = new_rd_ia32_xCmpSet;
+ cmov_func = new_rd_ia32_xCmpCMov;
+ }
+ else {
+ /* x87 FPU */
+ 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_CmpSet;
+ cmov_func = new_rd_ia32_CmpCMov;
+ }
+
+ /* create the nodes */
+
+ /* 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 {
+ 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));
+ }
+ }
+ }
+
+ return new_op;
+}
+
+
+/**
+ * Following conversion rules apply:
+ *
+ * INT -> INT
+ * ============
+ * 1) n bit -> m bit n > m (downscale)
+ * a) target is signed: movsx
+ * b) target is unsigned: and with lower bits sets
+ * 2) n bit -> m bit n == m (sign change)
+ * always ignored
+ * 3) n bit -> m bit n < m (upscale)
+ * a) source is signed: movsx
+ * b) source is unsigned: and with lower bits sets
+ *
+ * INT -> FLOAT
+ * ==============
+ * SSE(1/2) convert to float or double (cvtsi2ss/sd)
+ *
+ * FLOAT -> INT
+ * ==============
+ * SSE(1/2) convert from float or double to 32bit int (cvtss/sd2si)
+ * if target mode < 32bit: additional INT -> INT conversion (see above)
+ *
+ * FLOAT -> FLOAT
+ * ================
+ * SSE(1/2) convert from float or double to double or float (cvtss/sd2sd/ss)
+ * x87 is mode_E internally, conversions happen only at load and store
+ * in non-strict semantic
+ */
+
+/**
+ * Create a conversion from x87 state register to general purpose.
+ */
+static ir_node *gen_x87_fp_to_gp(ia32_transform_env_t *env, ir_mode *tgt_mode) {
+ ia32_code_gen_t *cg = env->cg;
+ entity *ent = cg->fp_to_gp;
+ ir_graph *irg = env->irg;
+ ir_node *block = env->block;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *op = get_Conv_op(env->irn);
+ ir_node *fist, *mem, *load;
+
+ if (! ent) {
+ int size = get_mode_size_bytes(ia32_reg_classes[CLASS_ia32_vfp].mode);
+ ent = cg->fp_to_gp =
+ frame_alloc_area(get_irg_frame_type(env->irg), size, 16, 0);
+ }
+
+ /* do a fist */
+ fist = new_rd_ia32_vfist(env->dbg, irg, block, get_irg_frame(irg), noreg, op, get_irg_no_mem(irg));
+
+ set_ia32_frame_ent(fist, ent);
+ set_ia32_use_frame(fist);
+ 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_F);
+
+ mem = new_r_Proj(irg, block, fist, mode_M, pn_ia32_vfist_M);
+
+ /* do a Load */
+ load = new_rd_ia32_Load(env->dbg, irg, block, get_irg_frame(irg), noreg, mem);
+
+ set_ia32_frame_ent(load, ent);
+ set_ia32_use_frame(load);
+ set_ia32_am_support(load, ia32_am_Source);
+ set_ia32_op_type(load, ia32_AddrModeS);
+ set_ia32_am_flavour(load, ia32_B);
+ set_ia32_ls_mode(load, tgt_mode);
+
+ return new_r_Proj(irg, block, load, tgt_mode, pn_ia32_Load_res);
+}
+
+/**
+ * Create a conversion from x87 state register to general purpose.
+ */
+static ir_node *gen_x87_gp_to_fp(ia32_transform_env_t *env, ir_mode *src_mode) {
+ ia32_code_gen_t *cg = env->cg;
+ entity *ent = cg->gp_to_fp;
+ ir_graph *irg = env->irg;
+ ir_node *block = env->block;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *nomem = get_irg_no_mem(irg);
+ ir_node *op = get_Conv_op(env->irn);
+ ir_node *fild, *store, *mem;
+ int src_bits;
+
+ if (! ent) {
+ int size = get_mode_size_bytes(ia32_reg_classes[CLASS_ia32_gp].mode);
+ ent = cg->gp_to_fp =
+ frame_alloc_area(get_irg_frame_type(env->irg), size, size, 0);
+ }
+
+ /* first convert to 32 bit */
+ src_bits = get_mode_size_bits(src_mode);
+ if (src_bits == 8) {
+ op = new_rd_ia32_Conv_I2I8Bit(env->dbg, irg, block, noreg, noreg, op, nomem);
+ op = new_r_Proj(irg, block, op, mode_Is, 0);
+ }
+ else if (src_bits < 32) {
+ op = new_rd_ia32_Conv_I2I(env->dbg, irg, block, noreg, noreg, op, nomem);
+ op = new_r_Proj(irg, block, op, mode_Is, 0);
+ }
+
+ /* do a store */
+ store = new_rd_ia32_Store(env->dbg, irg, block, get_irg_frame(irg), noreg, op, nomem);
+
+ set_ia32_frame_ent(store, ent);
+ set_ia32_use_frame(store);
+
+ set_ia32_am_support(store, ia32_am_Dest);
+ set_ia32_op_type(store, ia32_AddrModeD);
+ set_ia32_am_flavour(store, ia32_B);
+ set_ia32_ls_mode(store, mode_Is);
+
+ mem = new_r_Proj(irg, block, store, mode_M, 0);
+
+ /* do a fild */
+ fild = new_rd_ia32_vfild(env->dbg, irg, block, get_irg_frame(irg), noreg, mem);
+
+ set_ia32_frame_ent(fild, ent);
+ set_ia32_use_frame(fild);
+ 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_F);
+
+ return new_r_Proj(irg, block, fild, mode_F, 0);
+}
+
+/**
+ * Transforms a Conv node.
+ *
+ * @param env The transformation environment
+ * @return The created ia32 Conv node
+ */
+static ir_node *gen_Conv(ia32_transform_env_t *env) {
+ dbg_info *dbg = env->dbg;
+ ir_graph *irg = env->irg;
+ ir_node *op = get_Conv_op(env->irn);
+ ir_mode *src_mode = get_irn_mode(op);
+ ir_mode *tgt_mode = env->mode;
+ int src_bits = get_mode_size_bits(src_mode);
+ int tgt_bits = get_mode_size_bits(tgt_mode);
+ int pn = -1;
+ int kill = 0;
+ ir_node *block = env->block;
+ ir_node *new_op = NULL;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *nomem = new_rd_NoMem(irg);
+ ir_node *proj;
+ DEBUG_ONLY(firm_dbg_module_t *mod = env->mod;)
+
+ if (src_mode == tgt_mode) {
+ /* this can happen when changing mode_P to mode_Is */
+ DB((mod, LEVEL_1, "killed Conv(mode, mode) ..."));
+ edges_reroute(env->irn, op, irg);
+ }
+ else if (mode_is_float(src_mode)) {
+ /* we convert from float ... */
+ if (mode_is_float(tgt_mode)) {
+ /* ... to float */
+ if (USE_SSE2(env->cg)) {
+ DB((mod, LEVEL_1, "create Conv(float, float) ..."));
+ new_op = new_rd_ia32_Conv_FP2FP(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_FP2FP_res;
+ }
+ else {
+ DB((mod, LEVEL_1, "killed Conv(float, float) ..."));
+ /*
+ remark: we create a intermediate conv here, so modes will be spread correctly
+ these convs will be killed later
+ */
+ new_op = new_rd_ia32_Conv_FP2FP(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_FP2FP_res;
+ kill = 1;
+ }
+ }
+ else {
+ /* ... to int */
+ DB((mod, LEVEL_1, "create Conv(float, int) ..."));
+ if (USE_SSE2(env->cg)) {
+ new_op = new_rd_ia32_Conv_FP2I(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_FP2I_res;
+ }
+ else
+ return gen_x87_fp_to_gp(env, tgt_mode);
+
+ /* if target mode is not int: add an additional downscale convert */
+ if (tgt_bits < 32) {
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+ set_ia32_am_support(new_op, ia32_am_Source);
+ set_ia32_tgt_mode(new_op, tgt_mode);
+ set_ia32_src_mode(new_op, src_mode);
+
+ proj = new_rd_Proj(dbg, irg, block, new_op, mode_Is, pn_ia32_Conv_FP2I_res);
+
+ if (tgt_bits == 8 || src_bits == 8) {
+ new_op = new_rd_ia32_Conv_I2I8Bit(dbg, irg, block, noreg, noreg, proj, nomem);
+ pn = pn_ia32_Conv_I2I8Bit_res;
+ }
+ else {
+ new_op = new_rd_ia32_Conv_I2I(dbg, irg, block, noreg, noreg, proj, nomem);
+ pn = pn_ia32_Conv_I2I_res;
+ }
+ }
+ }
+ }
+ else {
+ /* we convert from int ... */
+ if (mode_is_float(tgt_mode)) {
+ FP_USED(env->cg);
+ /* ... to float */
+ DB((mod, LEVEL_1, "create Conv(int, float) ..."));
+ if (USE_SSE2(env->cg)) {
+ new_op = new_rd_ia32_Conv_I2FP(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_I2FP_res;
+ }
+ else
+ return gen_x87_gp_to_fp(env, src_mode);
+ }
+ else {
+ /* ... to int */
+ if (get_mode_size_bits(src_mode) == tgt_bits) {
+ DB((mod, LEVEL_1, "omitting equal size Conv(%+F, %+F) ...", src_mode, tgt_mode));
+ /*
+ remark: we create a intermediate conv here, so modes will be spread correctly
+ these convs will be killed later
+ */
+ new_op = new_rd_ia32_Conv_I2I(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_I2I_res;
+ kill = 1;
+ }
+ else {
+ DB((mod, LEVEL_1, "create Conv(int, int) ...", src_mode, tgt_mode));
+ if (tgt_bits == 8 || src_bits == 8) {
+ new_op = new_rd_ia32_Conv_I2I8Bit(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_I2I8Bit_res;
+ }
+ else {
+ new_op = new_rd_ia32_Conv_I2I(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_I2I_res;
+ }
+ }
+ }
+ }
+
+ if (new_op) {
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+ set_ia32_tgt_mode(new_op, tgt_mode);
+ set_ia32_src_mode(new_op, src_mode);
+
+ set_ia32_am_support(new_op, ia32_am_Source);
+
+ new_op = new_rd_Proj(dbg, irg, block, new_op, tgt_mode, pn);
+
+ if (kill)
+ nodeset_insert(env->cg->kill_conv, new_op);
+ }
+
+ return new_op;
+}
+
+
+
+/********************************************
+ * _ _
+ * | | | |
+ * | |__ ___ _ __ ___ __| | ___ ___
+ * | '_ \ / _ \ '_ \ / _ \ / _` |/ _ \/ __|
+ * | |_) | __/ | | | (_) | (_| | __/\__ \
+ * |_.__/ \___|_| |_|\___/ \__,_|\___||___/
+ *
+ ********************************************/
+
+ /**
+ * 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;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *mem = new_rd_NoMem(env->irg);
+ ir_node *ptr = get_irn_n(node, 0);
+ entity *ent = arch_get_frame_entity(env->cg->arch_env, node);
+ ir_mode *mode = env->mode;
+
+ /* 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);
+ if (USE_SSE2(env->cg))
+ new_op = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, mem);
+ else
+ new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem);
+ }
+ else {
+ new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem);
+ }
+
+ set_ia32_frame_ent(new_op, ent);
+ set_ia32_use_frame(new_op);
+
+ set_ia32_am_support(new_op, ia32_am_Source);
+ set_ia32_op_type(new_op, ia32_AddrModeS);
+ set_ia32_am_flavour(new_op, ia32_B);
+ set_ia32_ls_mode(new_op, mode);
+ set_ia32_flags(new_op, get_ia32_flags(new_op) | arch_irn_flags_rematerializable);
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+
+ return new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, pn_ia32_Load_res);
+}
+
+/**
+ * Transforms a FrameAddr into an ia32 Add.
+ */
+static ir_node *gen_be_FrameAddr(ia32_transform_env_t *env) {
+ ir_node *new_op = NULL;
+ ir_node *node = env->irn;
+ ir_node *op = get_irn_n(node, 0);
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *nomem = new_rd_NoMem(env->irg);
+
+ new_op = new_rd_ia32_Add(env->dbg, env->irg, env->block, noreg, noreg, op, noreg, nomem);
+ set_ia32_frame_ent(new_op, arch_get_frame_entity(env->cg->arch_env, node));
+ set_ia32_am_support(new_op, ia32_am_Full);
+ set_ia32_use_frame(new_op);
+ set_ia32_immop_type(new_op, ia32_ImmConst);
+ set_ia32_commutative(new_op);
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+
+ return new_rd_Proj(env->dbg, env->irg, env->block, new_op, env->mode, pn_ia32_Add_res);
+}
+
+/**
+ * Transforms a FrameLoad into an ia32 Load.
+ */
+static ir_node *gen_be_FrameLoad(ia32_transform_env_t *env) {
+ ir_node *new_op = NULL;
+ ir_node *node = env->irn;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *mem = get_irn_n(node, 0);
+ ir_node *ptr = get_irn_n(node, 1);
+ entity *ent = arch_get_frame_entity(env->cg->arch_env, node);
+ ir_mode *mode = get_type_mode(get_entity_type(ent));
+
+ if (mode_is_float(mode)) {
+ FP_USED(env->cg);
+ if (USE_SSE2(env->cg))
+ new_op = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, mem);
+ else
+ new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem);
+ }
+ else
+ new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem);
+
+ set_ia32_frame_ent(new_op, ent);
+ set_ia32_use_frame(new_op);
+
+ set_ia32_am_support(new_op, ia32_am_Source);
+ set_ia32_op_type(new_op, ia32_AddrModeS);
+ set_ia32_am_flavour(new_op, ia32_B);
+ set_ia32_ls_mode(new_op, mode);
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+
+ return new_op;
+}
+
+
+/**
+ * Transforms a FrameStore into an ia32 Store.
+ */
+static ir_node *gen_be_FrameStore(ia32_transform_env_t *env) {
+ ir_node *new_op = NULL;
+ ir_node *node = env->irn;
+ ir_node *noreg = ia32_new_NoReg_gp(env->cg);
+ ir_node *mem = get_irn_n(node, 0);
+ ir_node *ptr = get_irn_n(node, 1);
+ ir_node *val = get_irn_n(node, 2);
+ entity *ent = arch_get_frame_entity(env->cg->arch_env, node);
+ ir_mode *mode = get_irn_mode(val);
+
+ if (mode_is_float(mode)) {
+ FP_USED(env->cg);
+ if (USE_SSE2(env->cg))
+ new_op = new_rd_ia32_xStore(env->dbg, env->irg, env->block, ptr, noreg, val, mem);
+ else
+ new_op = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, mem);
+ }
+ else if (get_mode_size_bits(mode) == 8) {
+ new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, mem);
+ }
+ else {
+ new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, mem);
+ }
+
+ set_ia32_frame_ent(new_op, ent);
+ set_ia32_use_frame(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, ia32_B);
+ set_ia32_ls_mode(new_op, mode);
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+
+ return new_op;
+}
+
+/**
+ * In case SSE is used we need to copy the result from FPU TOS.
+ */
+static ir_node *gen_be_Call(ia32_transform_env_t *env) {
+ ir_node *call_res = get_proj_for_pn(env->irn, pn_be_Call_first_res);
+ ir_node *call_mem = get_proj_for_pn(env->irn, pn_be_Call_M_regular);
+ ir_mode *mode;
+
+ if (! call_res || ! USE_SSE2(env->cg))
+ return NULL;
+
+ mode = get_irn_mode(call_res);
+
+ /* in case there is no memory output: create one to serialize the copy FPU -> SSE */
+ if (! call_mem)
+ call_mem = new_r_Proj(env->irg, env->block, env->irn, mode_M, pn_be_Call_M_regular);
+
+ if (mode_is_float(mode)) {
+ /* store st(0) onto stack */
+ ir_node *frame = get_irg_frame(env->irg);
+ ir_node *fstp = new_rd_ia32_GetST0(env->dbg, env->irg, env->block, frame, call_mem);
+ ir_node *mproj = new_r_Proj(env->irg, env->block, fstp, mode_M, pn_ia32_GetST0_M);
+ entity *ent = frame_alloc_area(get_irg_frame_type(env->irg), get_mode_size_bytes(mode), 16, 0);
+ ir_node *sse_load;
+
+ set_ia32_ls_mode(fstp, mode);
+ set_ia32_op_type(fstp, ia32_AddrModeD);
+ set_ia32_use_frame(fstp);
+ set_ia32_frame_ent(fstp, ent);
+ set_ia32_am_flavour(fstp, ia32_B);
+ set_ia32_am_support(fstp, ia32_am_Dest);
+
+ /* load into SSE register */
+ sse_load = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, frame, ia32_new_NoReg_gp(env->cg), mproj);
+ set_ia32_ls_mode(sse_load, mode);
+ set_ia32_op_type(sse_load, ia32_AddrModeS);
+ set_ia32_use_frame(sse_load);
+ set_ia32_frame_ent(sse_load, ent);
+ set_ia32_am_flavour(sse_load, ia32_B);
+ set_ia32_am_support(sse_load, ia32_am_Source);
+ sse_load = new_r_Proj(env->irg, env->block, sse_load, mode, pn_ia32_xLoad_res);
+
+ /* reroute all users of the result proj to the sse load */
+ edges_reroute(call_res, sse_load, env->irg);
+ }
+
+ return NULL;
+}
+
+/**
+ * In case SSE is used we need to copy the result from XMM0 to FPU TOS before return.
+ */
+static ir_node *gen_be_Return(ia32_transform_env_t *env) {
+ ir_node *ret_val = get_irn_n(env->irn, be_pos_Return_val);
+ ir_node *ret_mem = get_irn_n(env->irn, be_pos_Return_mem);
+ entity *ent = get_irg_entity(get_irn_irg(ret_val));
+ ir_type *tp = get_entity_type(ent);
+
+ if (be_Return_get_n_rets(env->irn) < 1 || ! ret_val || ! USE_SSE2(env->cg))
+ return NULL;
+
+
+ if (get_method_n_ress(tp) == 1) {
+ ir_type *res_type = get_method_res_type(tp, 0);
+ ir_mode *mode;
+
+ if(is_Primitive_type(res_type)) {
+ mode = get_type_mode(res_type);
+ if(mode_is_float(mode)) {
+ ir_node *frame = get_irg_frame(env->irg);
+ entity *ent = frame_alloc_area(get_irg_frame_type(env->irg), get_mode_size_bytes(mode), 16, 0);
+ ir_node *sse_store, *fld, *mproj;
+
+ /* store xmm0 onto stack */
+ sse_store = new_rd_ia32_xStoreSimple(env->dbg, env->irg, env->block, frame, ret_val, ret_mem);
+ set_ia32_ls_mode(sse_store, mode);
+ set_ia32_op_type(sse_store, ia32_AddrModeD);
+ set_ia32_use_frame(sse_store);
+ set_ia32_frame_ent(sse_store, ent);
+ set_ia32_am_flavour(sse_store, ia32_B);
+ set_ia32_am_support(sse_store, ia32_am_Dest);
+ sse_store = new_r_Proj(env->irg, env->block, sse_store, mode_M, pn_ia32_xStore_M);
+
+ /* load into st0 */
+ fld = new_rd_ia32_SetST0(env->dbg, env->irg, env->block, frame, sse_store);
+ set_ia32_ls_mode(fld, mode);
+ set_ia32_op_type(fld, ia32_AddrModeS);
+ set_ia32_use_frame(fld);
+ set_ia32_frame_ent(fld, ent);
+ set_ia32_am_flavour(fld, ia32_B);
+ set_ia32_am_support(fld, ia32_am_Source);
+ mproj = new_r_Proj(env->irg, env->block, fld, mode_M, pn_ia32_SetST0_M);
+ fld = new_r_Proj(env->irg, env->block, fld, mode, pn_ia32_SetST0_res);
+
+ /* set new return value */
+ set_irn_n(env->irn, be_pos_Return_val, fld);
+ set_irn_n(env->irn, be_pos_Return_mem, mproj);
+ }
+ }
+ }
+
+ return NULL;
+}
+
+/**
+ * Transform a be_AddSP into an ia32_AddSP. Eat up const sizes.
+ */
+static ir_node *gen_be_AddSP(ia32_transform_env_t *env) {
+ ir_node *new_op;
+ const ir_edge_t *edge;
+ ir_node *sz = get_irn_n(env->irn, be_pos_AddSP_size);
+ ir_node *sp = get_irn_n(env->irn, be_pos_AddSP_old_sp);
+
+ new_op = new_rd_ia32_AddSP(env->dbg, env->irg, env->block, sp, sz);
+
+ if (is_ia32_Const(sz)) {
+ set_ia32_Immop_attr(new_op, sz);
+ set_irn_n(new_op, 1, ia32_new_NoReg_gp(env->cg));
+ }
+ else if (is_ia32_Load(sz) && get_ia32_am_flavour(sz) == ia32_O) {
+ set_ia32_immop_type(new_op, ia32_ImmSymConst);
+ set_ia32_op_type(new_op, ia32_AddrModeS);
+ set_ia32_am_sc(new_op, get_ia32_am_sc(sz));
+ add_ia32_am_offs(new_op, get_ia32_am_offs(sz));
+ set_irn_n(new_op, 1, ia32_new_NoReg_gp(env->cg));
+ }
+
+ /* fix proj nums */
+ foreach_out_edge(env->irn, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+
+ assert(is_Proj(proj));
+
+ if (get_Proj_proj(proj) == pn_be_AddSP_res) {
+ /* the node is not yet exchanged: we need to set the register manually */
+ ia32_attr_t *attr = get_ia32_attr(new_op);
+ attr->slots[pn_ia32_AddSP_stack] = &ia32_gp_regs[REG_ESP];
+ set_Proj_proj(proj, pn_ia32_AddSP_stack);
+ }
+ else if (get_Proj_proj(proj) == pn_be_AddSP_M) {
+ set_Proj_proj(proj, pn_ia32_AddSP_M);
+ }
+ else {
+ assert(0);
+ }
+ }
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+
+ return new_op;
+}
+
+/**
+ * This function just sets the register for the Unknown node
+ * as this is not done during register allocation because Unknown
+ * is an "ignore" node.
+ */
+static ir_node *gen_Unknown(ia32_transform_env_t *env) {
+ ir_mode *mode = env->mode;
+ ir_node *irn = env->irn;
+
+ if (mode_is_float(mode)) {
+ if (USE_SSE2(env->cg))
+ arch_set_irn_register(env->cg->arch_env, irn, &ia32_xmm_regs[REG_XMM_UKNWN]);
+ else
+ arch_set_irn_register(env->cg->arch_env, irn, &ia32_vfp_regs[REG_VFP_UKNWN]);
+ }
+ else if (mode_is_int(mode) || mode_is_reference(mode)) {
+ arch_set_irn_register(env->cg->arch_env, irn, &ia32_gp_regs[REG_GP_UKNWN]);
+ }
+ else {
+ assert(0 && "unsupported Unknown-Mode");
+ }
+
+ 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)) != NULL) {
+ 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));
+ }
+