2 * This file is part of libFirm.
3 * Copyright (C) 2012 University of Karlsruhe.
8 * @brief Lower small CopyB nodes into a series of Load/Store nodes
9 * @author Michael Beck, Matthias Braun, Manuel Mohr
26 typedef struct entry entry_t;
28 struct list_head list;
33 * Every CopyB is assigned a size category as follows:
34 * - 'small' iff size <= max_small_size,
35 * - 'medium' iff max_small_size < size < min_large_size,
36 * - 'large' iff size >= min_large_size.
38 * The idea is that each backend can apply different optimizations in each
39 * of the three categories.
41 * For small CopyBs, the x86 backend could, e.g., emit a single SSE
42 * instruction to copy 16 bytes. Other backends might just go with a series
43 * of Load/Stores. Therefore, x86 would like to keep the small CopyB nodes
44 * around whereas other backends would not.
45 * For medium-sized CopyBs, the x86 backend might generate a rep-prefixed mov
46 * instruction. Hence, it also wants to keep the CopyBs in these cases. Other
47 * backends might handle this differently.
48 * For large CopyBs, a call to memcpy is worth the call overhead, so large
49 * CopyBs should always be lowered to memcpy calls.
51 * The lowerer performs the following actions if the CopyB is
52 * - 'small': Replace it with a series of Loads/Stores
53 * - 'medium': Nothing.
54 * - 'large': Replace it with a call to memcpy.
56 * max_small_size and min_large_size allow for a flexible configuration.
57 * For example, one backend could specify max_small_size == 0 and
58 * min_large_size == 8192 to keep all CopyB nodes smaller than 8192 and get
59 * memcpy Calls for all others. Here, the set of small CopyBs is empty.
60 * Another backend could specify max_small_size == 63 and min_large_size == 64
61 * to lower all small CopyBs to Loads/Stores and all big CopyBs to memcpy.
62 * Hence, the set of medium-sized CopyBs is empty and this backend never
63 * sees a CopyB node at all.
64 * If memcpy is not available, min_large_size can be set to UINT_MAX to prevent
65 * the creation of calls to memcpy. Note that CopyBs whose size is UINT_MAX
66 * will still be lowered to memcpy calls because we check if the size is greater
67 * *or equal* to min_large_size. However, this should never occur in practice.
70 static unsigned max_small_size; /**< The maximum size of a CopyB node
71 so that it is regarded as 'small'. */
72 static unsigned min_large_size; /**< The minimum size of a CopyB node
73 so that it is regarded as 'large'. */
74 static unsigned native_mode_bytes; /**< The size of the native mode in bytes. */
75 static int allow_misalignments; /**< Whether backend can handle misaligned
78 typedef struct walk_env {
79 struct obstack obst; /**< the obstack where data is allocated
81 struct list_head list; /**< the list of copyb nodes. */
84 static ir_mode *get_ir_mode(unsigned mode_bytes)
87 case 1: return mode_Bu;
88 case 2: return mode_Hu;
89 case 4: return mode_Iu;
90 case 8: return mode_Lu;
91 case 16: return mode_LLu;
93 panic("unexpected mode size requested in copyb lowering");
98 * Turn a small CopyB node into a series of Load/Store nodes.
100 static void lower_small_copyb_node(ir_node *irn)
102 ir_graph *irg = get_irn_irg(irn);
103 ir_node *block = get_nodes_block(irn);
104 ir_type *tp = get_CopyB_type(irn);
105 ir_node *addr_src = get_CopyB_src(irn);
106 ir_node *addr_dst = get_CopyB_dst(irn);
107 ir_node *mem = get_CopyB_mem(irn);
108 ir_mode *addr_mode = get_irn_mode(addr_src);
109 unsigned mode_bytes = allow_misalignments ? native_mode_bytes : tp->align;
110 unsigned size = get_type_size_bytes(tp);
114 while (offset < size) {
115 mode = get_ir_mode(mode_bytes);
116 for (; offset + mode_bytes <= size; offset += mode_bytes) {
117 /* construct offset */
126 addr_const = new_r_Const_long(irg, mode_Iu, offset);
127 add = new_r_Add(block, addr_src, addr_const, addr_mode);
129 load = new_r_Load(block, mem, add, mode, cons_none);
130 load_res = new_r_Proj(load, mode, pn_Load_res);
131 load_mem = new_r_Proj(load, mode_M, pn_Load_M);
133 addr_const = new_r_Const_long(irg, mode_Iu, offset);
134 add = new_r_Add(block, addr_dst, addr_const, addr_mode);
136 store = new_r_Store(block, load_mem, add, load_res, cons_none);
137 store_mem = new_r_Proj(store, mode_M, pn_Store_M);
145 ir_node *const bad = new_r_Bad(irg, mode_X);
146 ir_node *const in[] = {
148 [pn_CopyB_X_regular] = bad,
149 [pn_CopyB_X_except] = bad,
151 turn_into_tuple(irn, ARRAY_SIZE(in), in);
154 static ir_type *get_memcpy_methodtype(void)
156 ir_type *tp = new_type_method(3, 1);
157 ir_mode *size_t_mode = get_ir_mode(native_mode_bytes);
159 set_method_param_type(tp, 0, get_type_for_mode(mode_P));
160 set_method_param_type(tp, 1, get_type_for_mode(mode_P));
161 set_method_param_type(tp, 2, get_type_for_mode(size_t_mode));
162 set_method_res_type (tp, 0, get_type_for_mode(mode_P));
167 static ir_node *get_memcpy_symconst(ir_graph *irg)
169 ident *id = new_id_from_str("memcpy");
170 ir_type *mt = get_memcpy_methodtype();
171 ir_entity *ent = create_compilerlib_entity(id, mt);
175 return new_r_SymConst(irg, mode_P_code, sym, symconst_addr_ent);
179 * Turn a large CopyB node into a memcpy call.
181 static void lower_large_copyb_node(ir_node *irn)
183 ir_graph *irg = get_irn_irg(irn);
184 ir_node *block = get_nodes_block(irn);
185 dbg_info *dbgi = get_irn_dbg_info(irn);
186 ir_node *mem = get_CopyB_mem(irn);
187 ir_node *addr_src = get_CopyB_src(irn);
188 ir_node *addr_dst = get_CopyB_dst(irn);
189 ir_type *copyb_tp = get_CopyB_type(irn);
190 unsigned size = get_type_size_bytes(copyb_tp);
192 ir_node *symconst = get_memcpy_symconst(irg);
193 ir_type *call_tp = get_memcpy_methodtype();
194 ir_mode *mode_size_t = get_ir_mode(native_mode_bytes);
201 in[2] = new_r_Const_long(irg, mode_size_t, size);
202 call = new_rd_Call(dbgi, block, mem, symconst, 3, in, call_tp);
203 call_mem = new_r_Proj(call, mode_M, pn_Call_M);
205 ir_node *const tuple_in[] = { call_mem };
206 turn_into_tuple(irn, ARRAY_SIZE(tuple_in), tuple_in);
209 static void lower_copyb_node(ir_node *irn)
211 ir_type *tp = get_CopyB_type(irn);
212 unsigned size = get_type_size_bytes(tp);
214 if (size <= max_small_size)
215 lower_small_copyb_node(irn);
216 else if (size >= min_large_size)
217 lower_large_copyb_node(irn);
219 assert(!"CopyB of invalid size handed to lower_copyb_node");
223 * Post-Walker: find CopyB nodes.
225 static void find_copyb_nodes(ir_node *irn, void *ctx)
227 walk_env_t *env = (walk_env_t*)ctx;
234 ir_node *pred = get_Proj_pred(irn);
236 if (is_CopyB(pred) && get_Proj_proj(irn) != pn_CopyB_M) {
237 /* found an exception Proj: remove it from the list again */
238 entry = (entry_t*)get_irn_link(pred);
239 list_del(&entry->list);
247 tp = get_CopyB_type(irn);
248 if (get_type_state(tp) != layout_fixed)
251 size = get_type_size_bytes(tp);
252 medium_sized = max_small_size < size && size < min_large_size;
254 return; /* Nothing to do for medium-sized CopyBs. */
256 /* Okay, either small or large CopyB, so link it in and lower it later. */
257 entry = OALLOC(&env->obst, entry_t);
259 INIT_LIST_HEAD(&entry->list);
260 set_irn_link(irn, entry);
261 list_add_tail(&entry->list, &env->list);
264 void lower_CopyB(ir_graph *irg, unsigned max_small_sz, unsigned min_large_sz,
267 const backend_params *bparams = be_get_backend_param();
270 assert(max_small_sz < min_large_sz && "CopyB size ranges must not overlap");
272 max_small_size = max_small_sz;
273 min_large_size = min_large_sz;
274 native_mode_bytes = bparams->machine_size / 8;
275 allow_misalignments = allow_misaligns;
277 obstack_init(&env.obst);
278 INIT_LIST_HEAD(&env.list);
279 irg_walk_graph(irg, NULL, find_copyb_nodes, &env);
281 list_for_each_entry(entry_t, entry, &env.list, list) {
282 lower_copyb_node(entry->copyb);
285 obstack_free(&env.obst, NULL);