X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_x87.c;h=55c9206a90dcbb32b85466f9bbf87abc5b00d34c;hb=ce6161a7e42a48f7422b7babcc64d8ace18e2687;hp=7429188d3ea2811169e23101d2c8f1e14e2ee92d;hpb=302872956d772c0480f371bfa57fe88471cee02c;p=libfirm diff --git a/ir/be/ia32/ia32_x87.c b/ir/be/ia32/ia32_x87.c index 7429188d3..55c9206a9 100644 --- a/ir/be/ia32/ia32_x87.c +++ b/ir/be/ia32/ia32_x87.c @@ -53,18 +53,13 @@ #include "ia32_x87.h" #include "ia32_architecture.h" -#define N_x87_REGS 8 - -/* the unop index */ -#define UNOP_IDX 0 - -#define MASK_TOS(x) ((x) & (N_x87_REGS - 1)) +#define MASK_TOS(x) ((x) & (N_ia32_st_REGS - 1)) /** the debug handle */ DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) /* Forward declaration. */ -typedef struct _x87_simulator x87_simulator; +typedef struct x87_simulator x87_simulator; /** * An exchange template. @@ -73,7 +68,7 @@ typedef struct _x87_simulator x87_simulator; * their opcodes! * Further, x87 supports inverse instructions, so we can handle them. */ -typedef struct _exchange_tmpl { +typedef struct exchange_tmpl { ir_op *normal_op; /**< the normal one */ ir_op *reverse_op; /**< the reverse one if exists */ ir_op *normal_pop_op; /**< the normal one with tos pop */ @@ -83,7 +78,7 @@ typedef struct _exchange_tmpl { /** * An entry on the simulated x87 stack. */ -typedef struct _st_entry { +typedef struct st_entry { int reg_idx; /**< the virtual register index of this stack value */ ir_node *node; /**< the node that produced this value */ } st_entry; @@ -91,11 +86,11 @@ typedef struct _st_entry { /** * The x87 state. */ -typedef struct _x87_state { - st_entry st[N_x87_REGS]; /**< the register stack */ - int depth; /**< the current stack depth */ - int tos; /**< position of the tos */ - x87_simulator *sim; /**< The simulator. */ +typedef struct x87_state { + st_entry st[N_ia32_st_REGS]; /**< the register stack */ + int depth; /**< the current stack depth */ + int tos; /**< position of the tos */ + x87_simulator *sim; /**< The simulator. */ } x87_state; /** An empty state, used for blocks without fp instructions. */ @@ -126,12 +121,12 @@ typedef int (*sim_func)(x87_state *state, ir_node *n); /** * A block state: Every block has a x87 state at the beginning and at the end. */ -typedef struct _blk_state { +typedef struct blk_state { x87_state *begin; /**< state at the begin or NULL if not assigned */ x87_state *end; /**< state at the end or NULL if not assigned */ } blk_state; -#define PTR_TO_BLKSTATE(p) ((blk_state *)(p)) +#define PTR_TO_BLKSTATE(p) ((blk_state *)(p)) /** liveness bitset for vfp registers. */ typedef unsigned char vfp_liveness; @@ -139,7 +134,7 @@ typedef unsigned char vfp_liveness; /** * The x87 simulator. */ -struct _x87_simulator { +struct x87_simulator { struct obstack obst; /**< An obstack for fast allocating. */ pmap *blk_states; /**< Map blocks to states. */ be_lv_t *lv; /**< intrablock liveness. */ @@ -252,7 +247,8 @@ static void x87_fxch(x87_state *state, int pos) state->st[MASK_TOS(state->tos + pos)] = state->st[MASK_TOS(state->tos)]; state->st[MASK_TOS(state->tos)] = entry; - DB((dbg, LEVEL_2, "After FXCH: ")); DEBUG_ONLY(x87_dump_stack(state)); + DB((dbg, LEVEL_2, "After FXCH: ")); + DEBUG_ONLY(x87_dump_stack(state)); } /* x87_fxch */ /** @@ -283,7 +279,7 @@ static int x87_on_stack(const x87_state *state, int reg_idx) */ static void x87_push_dbl(x87_state *state, int reg_idx, ir_node *node) { - assert(state->depth < N_x87_REGS && "stack overrun"); + assert(state->depth < N_ia32_st_REGS && "stack overrun"); ++state->depth; state->tos = MASK_TOS(state->tos - 1); @@ -347,7 +343,7 @@ static blk_state *x87_get_bl_state(x87_simulator *sim, ir_node *block) pmap_entry *entry = pmap_find(sim->blk_states, block); if (! entry) { - blk_state *bl_state = obstack_alloc(&sim->obst, sizeof(*bl_state)); + blk_state *bl_state = OALLOC(&sim->obst, blk_state); bl_state->begin = NULL; bl_state->end = NULL; @@ -367,7 +363,7 @@ static blk_state *x87_get_bl_state(x87_simulator *sim, ir_node *block) */ static x87_state *x87_alloc_state(x87_simulator *sim) { - x87_state *res = obstack_alloc(&sim->obst, sizeof(*res)); + x87_state *res = OALLOC(&sim->obst, x87_state); res->sim = sim; return res; @@ -451,7 +447,7 @@ static inline const arch_register_t *x87_get_irn_register(const ir_node *irn) { const arch_register_t *res = arch_get_irn_register(irn); - assert(res->reg_class->regs == ia32_vfp_regs); + assert(res->reg_class == &ia32_reg_classes[CLASS_ia32_vfp]); return res; } /* x87_get_irn_register */ @@ -460,10 +456,15 @@ static inline const arch_register_t *x87_irn_get_register(const ir_node *irn, { const arch_register_t *res = arch_irn_get_register(irn, pos); - assert(res->reg_class->regs == ia32_vfp_regs); + assert(res->reg_class == &ia32_reg_classes[CLASS_ia32_vfp]); return res; } /* x87_irn_get_register */ +static inline const arch_register_t *get_st_reg(int index) +{ + return &ia32_registers[REG_ST0 + index]; +} + /* -------------- x87 perm --------------- */ /** @@ -485,8 +486,8 @@ static ir_node *x87_fxch_shuffle(x87_state *state, int pos, ir_node *block) fxch = new_bd_ia32_fxch(NULL, block); attr = get_ia32_x87_attr(fxch); - attr->x87[0] = &ia32_st_regs[pos]; - attr->x87[2] = &ia32_st_regs[0]; + attr->x87[0] = get_st_reg(pos); + attr->x87[2] = get_st_reg(0); keep_alive(fxch); @@ -654,8 +655,8 @@ static ir_node *x87_create_fxch(x87_state *state, ir_node *n, int pos) fxch = new_bd_ia32_fxch(NULL, block); attr = get_ia32_x87_attr(fxch); - attr->x87[0] = &ia32_st_regs[pos]; - attr->x87[2] = &ia32_st_regs[0]; + attr->x87[0] = get_st_reg(pos); + attr->x87[2] = get_st_reg(0); keep_alive(fxch); @@ -682,8 +683,8 @@ static void x87_create_fpush(x87_state *state, ir_node *n, int pos, int op_idx) fpush = new_bd_ia32_fpush(NULL, get_nodes_block(n)); attr = get_ia32_x87_attr(fpush); - attr->x87[0] = &ia32_st_regs[pos]; - attr->x87[2] = &ia32_st_regs[0]; + attr->x87[0] = get_st_reg(pos); + attr->x87[2] = get_st_reg(0); keep_alive(fpush); sched_add_before(n, fpush); @@ -713,9 +714,9 @@ static ir_node *x87_create_fpop(x87_state *state, ir_node *n, int num) else fpop = new_bd_ia32_fpop(NULL, get_nodes_block(n)); attr = get_ia32_x87_attr(fpop); - attr->x87[0] = &ia32_st_regs[0]; - attr->x87[1] = &ia32_st_regs[0]; - attr->x87[2] = &ia32_st_regs[0]; + attr->x87[0] = get_st_reg(0); + attr->x87[1] = get_st_reg(0); + attr->x87[2] = get_st_reg(0); keep_alive(fpop); sched_add_before(n, fpop); @@ -802,7 +803,7 @@ static vfp_liveness vfp_liveness_end_of_block(x87_simulator *sim, const ir_node } /* vfp_liveness_end_of_block */ /** get the register mask from an arch_register */ -#define REGMASK(reg) (1 << (arch_register_get_index(reg))) +#define REGMASK(reg) (1 << (arch_register_get_index(reg))) /** * Return a bitset of argument registers which are live at the end of a node. @@ -917,18 +918,18 @@ static int sim_binop(x87_state *state, ir_node *n, const exchange_tmpl *tmpl) op1_idx = x87_on_stack(state, reg_index_1); assert(op1_idx >= 0); - op1_live_after = is_vfp_live(arch_register_get_index(op1_reg), live); + op1_live_after = is_vfp_live(reg_index_1, live); attr = get_ia32_x87_attr(n); permuted = attr->attr.data.ins_permuted; - if (reg_index_2 != REG_VFP_NOREG) { + if (reg_index_2 != REG_VFP_VFP_NOREG) { assert(!permuted); /* second operand is a vfp register */ op2_idx = x87_on_stack(state, reg_index_2); assert(op2_idx >= 0); - op2_live_after = is_vfp_live(arch_register_get_index(op2_reg), live); + op2_live_after = is_vfp_live(reg_index_2, live); if (op2_live_after) { /* Second operand is live. */ @@ -1034,13 +1035,13 @@ static int sim_binop(x87_state *state, ir_node *n, const exchange_tmpl *tmpl) } /* patch the operation */ - attr->x87[0] = op1_reg = &ia32_st_regs[op1_idx]; - if (reg_index_2 != REG_VFP_NOREG) { - attr->x87[1] = op2_reg = &ia32_st_regs[op2_idx]; + attr->x87[0] = op1_reg = get_st_reg(op1_idx); + if (reg_index_2 != REG_VFP_VFP_NOREG) { + attr->x87[1] = op2_reg = get_st_reg(op2_idx); } - attr->x87[2] = out = &ia32_st_regs[out_idx]; + attr->x87[2] = out = get_st_reg(out_idx); - if (reg_index_2 != REG_VFP_NOREG) { + if (reg_index_2 != REG_VFP_VFP_NOREG) { DB((dbg, LEVEL_1, "<<< %s %s, %s -> %s\n", get_irn_opname(n), arch_register_get_name(op1_reg), arch_register_get_name(op2_reg), arch_register_get_name(out))); @@ -1066,7 +1067,7 @@ static int sim_unop(x87_state *state, ir_node *n, ir_op *op) { int op1_idx; x87_simulator *sim = state->sim; - const arch_register_t *op1 = x87_get_irn_register(get_irn_n(n, UNOP_IDX)); + const arch_register_t *op1 = x87_get_irn_register(get_irn_n(n, 0)); const arch_register_t *out = x87_get_irn_register(n); ia32_x87_attr_t *attr; unsigned live = vfp_live_args_after(sim, n, REGMASK(out)); @@ -1078,7 +1079,7 @@ static int sim_unop(x87_state *state, ir_node *n, ir_op *op) if (is_vfp_live(arch_register_get_index(op1), live)) { /* push the operand here */ - x87_create_fpush(state, n, op1_idx, UNOP_IDX); + x87_create_fpush(state, n, op1_idx, 0); op1_idx = 0; } else { @@ -1091,8 +1092,8 @@ static int sim_unop(x87_state *state, ir_node *n, ir_op *op) x87_set_tos(state, arch_register_get_index(out), x87_patch_insn(n, op)); attr = get_ia32_x87_attr(n); - attr->x87[0] = op1 = &ia32_st_regs[0]; - attr->x87[2] = out = &ia32_st_regs[0]; + attr->x87[0] = op1 = get_st_reg(0); + attr->x87[2] = out = get_st_reg(0); DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), out->name)); return NO_NODE_ADDED; @@ -1116,7 +1117,7 @@ static int sim_load(x87_state *state, ir_node *n, ir_op *op, int res_pos) x87_push(state, arch_register_get_index(out), x87_patch_insn(n, op)); assert(out == x87_irn_get_register(n, res_pos)); attr = get_ia32_x87_attr(n); - attr->x87[2] = out = &ia32_st_regs[0]; + attr->x87[2] = out = get_st_reg(0); DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), arch_register_get_name(out))); return NO_NODE_ADDED; @@ -1189,21 +1190,21 @@ static int sim_store(x87_state *state, ir_node *n, ir_op *op, ir_op *op_p) Note that we cannot test on mode_E, because floats might be 96bit ... */ if (get_mode_size_bits(mode) > 64 || (mode_is_int(mode) && get_mode_size_bits(mode) > 32)) { - if (depth < N_x87_REGS) { + if (depth < N_ia32_st_REGS) { /* ok, we have a free register: push + fstp */ x87_create_fpush(state, n, op2_idx, n_ia32_vfst_val); x87_pop(state); x87_patch_insn(n, op_p); } else { ir_node *vfld, *mem, *block, *rproj, *mproj; - ir_graph *irg; + ir_graph *irg = get_irn_irg(n); /* stack full here: need fstp + load */ x87_pop(state); x87_patch_insn(n, op_p); block = get_nodes_block(n); - vfld = new_bd_ia32_vfld(NULL, block, get_irn_n(n, 0), get_irn_n(n, 1), new_NoMem(), get_ia32_ls_mode(n)); + vfld = new_bd_ia32_vfld(NULL, block, get_irn_n(n, 0), get_irn_n(n, 1), new_r_NoMem(irg), get_ia32_ls_mode(n)); /* copy all attributes */ set_ia32_frame_ent(vfld, get_ia32_frame_ent(n)); @@ -1223,7 +1224,6 @@ static int sim_store(x87_state *state, ir_node *n, ir_op *op, ir_op *op_p) arch_set_irn_register(rproj, op2); /* reroute all former users of the store memory to the load memory */ - irg = get_irn_irg(n); edges_reroute(mem, mproj, irg); /* set the memory input of the load to the store memory */ set_irn_n(vfld, n_ia32_vfld_mem, mem); @@ -1254,7 +1254,7 @@ static int sim_store(x87_state *state, ir_node *n, ir_op *op, ir_op *op_p) } attr = get_ia32_x87_attr(n); - attr->x87[1] = op2 = &ia32_st_regs[0]; + attr->x87[1] = op2 = get_st_reg(0); DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2))); return insn; @@ -1267,7 +1267,7 @@ static int sim_##op(x87_state *state, ir_node *n) { \ } #define GEN_BINOP(op) _GEN_BINOP(op, op) -#define GEN_BINOPR(op) _GEN_BINOP(op, op##r) +#define GEN_BINOPR(op) _GEN_BINOP(op, op##r) #define GEN_LOAD(op) \ static int sim_##op(x87_state *state, ir_node *n) { \ @@ -1334,7 +1334,7 @@ static int sim_fisttp(x87_state *state, ir_node *n) x87_patch_insn(n, op_ia32_fisttp); attr = get_ia32_x87_attr(n); - attr->x87[1] = op2 = &ia32_st_regs[0]; + attr->x87[1] = op2 = get_st_reg(0); DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2))); return NO_NODE_ADDED; @@ -1372,7 +1372,7 @@ static int sim_FtstFnstsw(x87_state *state, ir_node *n) /* patch the operation */ x87_patch_insn(n, op_ia32_FtstFnstsw); - reg1 = &ia32_st_regs[op1_idx]; + reg1 = get_st_reg(op1_idx); attr->x87[0] = reg1; attr->x87[1] = NULL; attr->x87[2] = NULL; @@ -1419,7 +1419,7 @@ static int sim_Fucom(x87_state *state, ir_node *n) assert(op1_idx >= 0); /* BEWARE: check for comp a,a cases, they might happen */ - if (reg_index_2 != REG_VFP_NOREG) { + if (reg_index_2 != REG_VFP_VFP_NOREG) { /* second operand is a vfp register */ op2_idx = x87_on_stack(state, reg_index_2); assert(op2_idx >= 0); @@ -1602,10 +1602,10 @@ static int sim_Fucom(x87_state *state, ir_node *n) op2_idx = tmp; } - op1 = &ia32_st_regs[op1_idx]; + op1 = get_st_reg(op1_idx); attr->x87[0] = op1; if (op2_idx >= 0) { - op2 = &ia32_st_regs[op2_idx]; + op2 = get_st_reg(op2_idx); attr->x87[1] = op2; } attr->x87[2] = NULL; @@ -1733,7 +1733,7 @@ static ir_node *create_Copy(x87_state *state, ir_node *n) x87_push(state, arch_register_get_index(out), res); attr = get_ia32_x87_attr(res); - attr->x87[2] = &ia32_st_regs[0]; + attr->x87[2] = get_st_reg(0); } else { int op1_idx = x87_on_stack(state, arch_register_get_index(op1)); @@ -1742,8 +1742,8 @@ static ir_node *create_Copy(x87_state *state, ir_node *n) x87_push(state, arch_register_get_index(out), res); attr = get_ia32_x87_attr(res); - attr->x87[0] = &ia32_st_regs[op1_idx]; - attr->x87[2] = &ia32_st_regs[0]; + attr->x87[0] = get_st_reg(op1_idx); + attr->x87[2] = get_st_reg(0); } arch_set_irn_register(res, out); @@ -1769,7 +1769,7 @@ static int sim_Copy(x87_state *state, ir_node *n) unsigned live; cls = arch_get_irn_reg_class_out(n); - if (cls->regs != ia32_vfp_regs) + if (cls != &ia32_reg_classes[CLASS_ia32_vfp]) return 0; pred = get_irn_n(n, 0); @@ -1820,7 +1820,7 @@ static int sim_Copy(x87_state *state, ir_node *n) /* best case, simple remove and rename */ x87_patch_insn(n, op_ia32_Pop); attr = get_ia32_x87_attr(n); - attr->x87[0] = op1 = &ia32_st_regs[0]; + attr->x87[0] = op1 = get_st_reg(0); x87_pop(state); x87_set_st(state, arch_register_get_index(out), n, op1_idx - 1); @@ -1833,7 +1833,7 @@ static int sim_Copy(x87_state *state, ir_node *n) } x87_patch_insn(n, op_ia32_Pop); attr = get_ia32_x87_attr(n); - attr->x87[0] = op1 = &ia32_st_regs[out_idx]; + attr->x87[0] = op1 = get_st_reg(out_idx); x87_pop(state); x87_set_st(state, arch_register_get_index(out), n, out_idx - 1); @@ -1953,7 +1953,7 @@ static int sim_Return(x87_state *state, ir_node *n) return NO_NODE_ADDED; } /* sim_Return */ -typedef struct _perm_data_t { +typedef struct perm_data_t { const arch_register_t *in; const arch_register_t *out; } perm_data_t; @@ -2209,7 +2209,7 @@ static void x87_init_simulator(x87_simulator *sim, ir_graph *irg) obstack_init(&sim->obst); sim->blk_states = pmap_create(); sim->n_idx = get_irg_last_idx(irg); - sim->live = obstack_alloc(&sim->obst, sizeof(*sim->live) * sim->n_idx); + sim->live = OALLOCN(&sim->obst, vfp_liveness, sim->n_idx); DB((dbg, LEVEL_1, "--------------------------------\n" "x87 Simulator started for %+F\n", irg)); @@ -2259,7 +2259,7 @@ static void x87_destroy_simulator(x87_simulator *sim) */ static void update_liveness_walker(ir_node *block, void *data) { - x87_simulator *sim = data; + x87_simulator *sim = (x87_simulator*)data; update_liveness(sim, block); } /* update_liveness_walker */ @@ -2268,14 +2268,13 @@ static void update_liveness_walker(ir_node *block, void *data) * Replaces all virtual floating point instructions and registers * by real ones. */ -void x87_simulate_graph(be_irg_t *birg) +void x87_simulate_graph(ir_graph *irg) { /* TODO improve code quality (less executed fxch) by using execfreqs */ ir_node *block, *start_block; blk_state *bl_state; x87_simulator sim; - ir_graph *irg = be_get_birg_irg(birg); /* create the simulator */ x87_init_simulator(&sim, irg); @@ -2290,8 +2289,8 @@ void x87_simulate_graph(be_irg_t *birg) sim.worklist = new_waitq(); waitq_put(sim.worklist, start_block); - be_assure_liveness(birg); - sim.lv = be_get_birg_liveness(birg); + be_assure_liveness(irg); + sim.lv = be_get_irg_liveness(irg); be_liveness_assure_sets(sim.lv); /* Calculate the liveness for all nodes. We must precalculate this info, @@ -2304,7 +2303,7 @@ void x87_simulate_graph(be_irg_t *birg) /* iterate */ do { - block = waitq_get(sim.worklist); + block = (ir_node*)waitq_get(sim.worklist); x87_simulate_block(&sim, block); } while (! waitq_empty(sim.worklist));