X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fldstopt.c;h=156293dee344ee9a638a5a31fedbcc577fc3d7fb;hb=eb7ac83b870791219e7a50cc262f2f1ee9c01b39;hp=f5e2cb42834001974ee47d44f0fdb40a6085790b;hpb=e40b97bd28e9d4316aecceb642337847d303d4de;p=libfirm diff --git a/ir/opt/ldstopt.c b/ir/opt/ldstopt.c index f5e2cb428..156293dee 100644 --- a/ir/opt/ldstopt.c +++ b/ir/opt/ldstopt.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -23,9 +23,7 @@ * @author Michael Beck * @version $Id$ */ -#ifdef HAVE_CONFIG_H -# include "config.h" -#endif +#include "config.h" #include @@ -37,33 +35,29 @@ #include "ircons_t.h" #include "irgmod.h" #include "irgwalk.h" -#include "irvrfy.h" #include "tv_t.h" #include "dbginfo_t.h" #include "iropt_dbg.h" #include "irflag_t.h" -#include "array.h" +#include "array_t.h" #include "irhooks.h" #include "iredges.h" -#include "irtools.h" +#include "irpass.h" #include "opt_polymorphy.h" #include "irmemory.h" -#include "xmalloc.h" #include "irphase_t.h" #include "irgopt.h" +#include "set.h" +#include "be.h" #include "debug.h" /** The debug handle. */ DEBUG_ONLY(static firm_dbg_module_t *dbg;) -#ifdef DO_CACHEOPT -#include "cacheopt/cachesim.h" -#endif - #undef IMAX -#define IMAX(a,b) ((a) > (b) ? (a) : (b)) +#define IMAX(a,b) ((a) > (b) ? (a) : (b)) -#define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max) +#define MAX_PROJ IMAX(IMAX((long)pn_Load_max, (long)pn_Store_max), (long)pn_Call_max) enum changes_t { DF_CHANGED = 1, /**< data flow changed */ @@ -73,14 +67,14 @@ enum changes_t { /** * walker environment */ -typedef struct _walk_env_t { +typedef struct walk_env_t { struct obstack obst; /**< list of all stores */ unsigned changes; /**< a bitmask of graph changes */ } walk_env_t; /** A Load/Store info. */ -typedef struct _ldst_info_t { - ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */ +typedef struct ldst_info_t { + ir_node *projs[MAX_PROJ+1]; /**< list of Proj's of this node */ ir_node *exc_block; /**< the exception block if available */ int exc_idx; /**< predecessor index in the exception block */ unsigned visited; /**< visited counter for breaking loops */ @@ -97,7 +91,7 @@ enum block_flags_t { /** * a Block info. */ -typedef struct _block_info_t { +typedef struct block_info_t { unsigned flags; /**< flags for the block */ } block_info_t; @@ -111,12 +105,12 @@ static unsigned master_visited = 0; /** * get the Load/Store info of a node */ -static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) { - ldst_info_t *info = get_irn_link(node); +static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) +{ + ldst_info_t *info = (ldst_info_t*)get_irn_link(node); if (! info) { - info = obstack_alloc(obst, sizeof(*info)); - memset(info, 0, sizeof(*info)); + info = OALLOCZ(obst, ldst_info_t); set_irn_link(node, info); } return info; @@ -125,12 +119,12 @@ static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) { /** * get the Block info of a node */ -static block_info_t *get_block_info(ir_node *node, struct obstack *obst) { - block_info_t *info = get_irn_link(node); +static block_info_t *get_block_info(ir_node *node, struct obstack *obst) +{ + block_info_t *info = (block_info_t*)get_irn_link(node); if (! info) { - info = obstack_alloc(obst, sizeof(*info)); - memset(info, 0, sizeof(*info)); + info = OALLOCZ(obst, block_info_t); set_irn_link(node, info); } return info; @@ -182,10 +176,10 @@ static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos) */ static void collect_nodes(ir_node *node, void *env) { - ir_opcode opcode = get_irn_opcode(node); + walk_env_t *wenv = (walk_env_t *)env; + unsigned opcode = get_irn_opcode(node); ir_node *pred, *blk, *pred_blk; ldst_info_t *ldst_info; - walk_env_t *wenv = env; if (opcode == iro_Proj) { pred = get_Proj_pred(node); @@ -221,7 +215,7 @@ static void collect_nodes(ir_node *node, void *env) if (is_Proj(proj)) { pred = get_Proj_pred(proj); - is_exc = get_Proj_proj(proj) == pn_Generic_X_except; + is_exc = is_x_except_Proj(proj); } /* ignore Bad predecessors, they will be removed later */ @@ -257,10 +251,7 @@ static ir_entity *find_constant_entity(ir_node *ptr) { for (;;) { if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) { - ir_entity *ent = get_SymConst_entity(ptr); - if (variability_constant == get_entity_variability(ent)) - return ent; - return NULL; + return get_SymConst_entity(ptr); } else if (is_Sel(ptr)) { ir_entity *ent = get_Sel_entity(ptr); ir_type *tp = get_entity_owner(ent); @@ -276,10 +267,10 @@ static ir_entity *find_constant_entity(ir_node *ptr) int i, n; for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) { - ir_node *bound; - tarval *tlower, *tupper; - ir_node *index = get_Sel_index(ptr, i); - tarval *tv = computed_value(index); + ir_node *bound; + ir_tarval *tlower, *tupper; + ir_node *index = get_Sel_index(ptr, i); + ir_tarval *tv = computed_value(index); /* check if the index is constant */ if (tv == tarval_bad) @@ -293,16 +284,16 @@ static ir_entity *find_constant_entity(ir_node *ptr) if (tlower == tarval_bad || tupper == tarval_bad) return NULL; - if (tarval_cmp(tv, tlower) & pn_Cmp_Lt) + if (tarval_cmp(tv, tlower) == ir_relation_less) return NULL; - if (tarval_cmp(tupper, tv) & pn_Cmp_Lt) + if (tarval_cmp(tupper, tv) == ir_relation_less) return NULL; /* ok, bounds check finished */ } } - if (variability_constant == get_entity_variability(ent)) + if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) return ent; /* try next */ @@ -325,7 +316,7 @@ static ir_entity *find_constant_entity(ir_node *ptr) ir_node *l = get_Sub_left(ptr); ir_node *r = get_Sub_right(ptr); - if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r)) + if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r)) ptr = l; else return NULL; @@ -340,7 +331,8 @@ static ir_entity *find_constant_entity(ir_node *ptr) /** * Return the Selection index of a Sel node from dimension n */ -static long get_Sel_array_index_long(ir_node *n, int dim) { +static long get_Sel_array_index_long(ir_node *n, int dim) +{ ir_node *index = get_Sel_index(n, dim); assert(is_Const(index)); return get_tarval_long(get_Const_tarval(index)); @@ -354,17 +346,18 @@ static long get_Sel_array_index_long(ir_node *n, int dim) { * @param depth current depth in steps upward from the root * of the address */ -static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) { +static compound_graph_path *rec_get_accessed_path(ir_node *ptr, size_t depth) +{ compound_graph_path *res = NULL; ir_entity *root, *field, *ent; - int path_len, pos, idx; - tarval *tv; + size_t path_len, pos, idx; + ir_tarval *tv; ir_type *tp; if (is_SymConst(ptr)) { /* a SymConst. If the depth is 0, this is an access to a global * entity and we don't need a component path, else we know - * at least it's length. + * at least its length. */ assert(get_SymConst_kind(ptr) == symconst_addr_ent); root = get_SymConst_entity(ptr); @@ -386,19 +379,23 @@ static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) { set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0)); } } else if (is_Add(ptr)) { - ir_node *l = get_Add_left(ptr); - ir_node *r = get_Add_right(ptr); - ir_mode *mode; + ir_mode *mode; + ir_tarval *tmp; - if (is_Const(r)) { - ptr = l; - tv = get_Const_tarval(r); - } else { - ptr = r; - tv = get_Const_tarval(l); + { + ir_node *l = get_Add_left(ptr); + ir_node *r = get_Add_right(ptr); + if (is_Const(r) && get_irn_mode(l) == get_irn_mode(ptr)) { + ptr = l; + tv = get_Const_tarval(r); + } else { + ptr = r; + tv = get_Const_tarval(l); + } } ptr_arith: mode = get_tarval_mode(tv); + tmp = tv; /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */ if (is_Sel(ptr)) { @@ -408,10 +405,9 @@ ptr_arith: } idx = 0; for (ent = field;;) { - unsigned size; - tarval *sz, *tv_index, *tlower, *tupper; - long index; - ir_node *bound; + unsigned size; + ir_tarval *sz, *tv_index, *tlower, *tupper; + ir_node *bound; tp = get_entity_type(ent); if (! is_Array_type(tp)) @@ -420,10 +416,10 @@ ptr_arith: size = get_type_size_bytes(get_entity_type(ent)); sz = new_tarval_from_long(size, mode); - tv_index = tarval_div(tv, sz); - tv = tarval_mod(tv, sz); + tv_index = tarval_div(tmp, sz); + tmp = tarval_mod(tmp, sz); - if (tv_index == tarval_bad || tv == tarval_bad) + if (tv_index == tarval_bad || tmp == tarval_bad) return NULL; assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented"); @@ -435,16 +431,15 @@ ptr_arith: if (tlower == tarval_bad || tupper == tarval_bad) return NULL; - if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt) + if (tarval_cmp(tv_index, tlower) == ir_relation_less) return NULL; - if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt) + if (tarval_cmp(tupper, tv_index) == ir_relation_less) return NULL; /* ok, bounds check finished */ - index = get_tarval_long(tv_index); ++idx; } - if (! tarval_is_null(tv)) { + if (! tarval_is_null(tmp)) { /* access to some struct/union member */ return NULL; } @@ -461,9 +456,9 @@ ptr_arith: pos = path_len - depth - idx; for (ent = field;;) { - unsigned size; - tarval *sz, *tv_index; - long index; + unsigned size; + ir_tarval *sz, *tv_index; + long index; tp = get_entity_type(ent); if (! is_Array_type(tp)) @@ -501,23 +496,26 @@ ptr_arith: * Returns an access path or NULL. The access path is only * valid, if the graph is in phase_high and _no_ address computation is used. */ -static compound_graph_path *get_accessed_path(ir_node *ptr) { - return rec_get_accessed_path(ptr, 0); +static compound_graph_path *get_accessed_path(ir_node *ptr) +{ + compound_graph_path *gr = rec_get_accessed_path(ptr, 0); + return gr; } /* get_accessed_path */ typedef struct path_entry { ir_entity *ent; struct path_entry *next; - long index; + size_t index; } path_entry; -static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) { +static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) +{ path_entry entry, *p; ir_entity *ent, *field; ir_initializer_t *initializer; - tarval *tv; + ir_tarval *tv; ir_type *tp; - unsigned n; + size_t n; entry.next = next; if (is_SymConst(ptr)) { @@ -573,7 +571,7 @@ static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) { assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented"); entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0); } else { - int i, n_members = get_compound_n_members(tp); + size_t i, n_members = get_compound_n_members(tp); for (i = 0; i < n_members; ++i) { if (get_compound_member(tp, i) == field) break; @@ -586,17 +584,19 @@ static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) { } return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry); } else if (is_Add(ptr)) { - ir_node *l = get_Add_left(ptr); - ir_node *r = get_Add_right(ptr); ir_mode *mode; unsigned pos; - if (is_Const(r)) { - ptr = l; - tv = get_Const_tarval(r); - } else { - ptr = r; - tv = get_Const_tarval(l); + { + ir_node *l = get_Add_left(ptr); + ir_node *r = get_Add_right(ptr); + if (is_Const(r)) { + ptr = l; + tv = get_Const_tarval(r); + } else { + ptr = r; + tv = get_Const_tarval(l); + } } ptr_arith: mode = get_tarval_mode(tv); @@ -627,10 +627,10 @@ ptr_arith: /* fill them up */ pos = 0; for (ent = field;;) { - unsigned size; - tarval *sz, *tv_index, *tlower, *tupper; - long index; - ir_node *bound; + unsigned size; + ir_tarval *sz, *tv_index, *tlower, *tupper; + long index; + ir_node *bound; tp = get_entity_type(ent); if (! is_Array_type(tp)) @@ -657,9 +657,9 @@ ptr_arith: if (tlower == tarval_bad || tupper == tarval_bad) return NULL; - if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt) + if (tarval_cmp(tv_index, tlower) == ir_relation_less) return NULL; - if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt) + if (tarval_cmp(tupper, tv_index) == ir_relation_less) return NULL; /* ok, bounds check finished */ @@ -685,7 +685,8 @@ ptr_arith: return NULL; } -static ir_node *find_compound_ent_value(ir_node *ptr) { +static ir_node *find_compound_ent_value(ir_node *ptr) +{ return rec_find_compound_ent_value(ptr, NULL); } @@ -693,10 +694,11 @@ static ir_node *find_compound_ent_value(ir_node *ptr) { static void reduce_adr_usage(ir_node *ptr); /** - * Update a Load that may lost it's usage. + * Update a Load that may have lost its users. */ -static void handle_load_update(ir_node *load) { - ldst_info_t *info = get_irn_link(load); +static void handle_load_update(ir_node *load) +{ + ldst_info_t *info = (ldst_info_t*)get_irn_link(load); /* do NOT touch volatile loads for now */ if (get_Load_volatility(load) == volatility_is_volatile) @@ -706,33 +708,35 @@ static void handle_load_update(ir_node *load) { ir_node *ptr = get_Load_ptr(load); ir_node *mem = get_Load_mem(load); - /* a Load which value is neither used nor exception checked, remove it */ + /* a Load whose value is neither used nor exception checked, remove it */ exchange(info->projs[pn_Load_M], mem); if (info->projs[pn_Load_X_regular]) - exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load))); + exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load))); kill_node(load); reduce_adr_usage(ptr); } } /* handle_load_update */ /** - * A Use of an address node is vanished. Check if this was a Proj + * A use of an address node has vanished. Check if this was a Proj * node and update the counters. */ -static void reduce_adr_usage(ir_node *ptr) { - if (is_Proj(ptr)) { - if (get_irn_n_edges(ptr) <= 0) { - /* this Proj is dead now */ - ir_node *pred = get_Proj_pred(ptr); - - if (is_Load(pred)) { - ldst_info_t *info = get_irn_link(pred); - info->projs[get_Proj_proj(ptr)] = NULL; - - /* this node lost it's result proj, handle that */ - handle_load_update(pred); - } - } +static void reduce_adr_usage(ir_node *ptr) +{ + ir_node *pred; + if (!is_Proj(ptr)) + return; + if (get_irn_n_edges(ptr) > 0) + return; + + /* this Proj is dead now */ + pred = get_Proj_pred(ptr); + if (is_Load(pred)) { + ldst_info_t *info = (ldst_info_t*)get_irn_link(pred); + info->projs[get_Proj_proj(ptr)] = NULL; + + /* this node lost its result proj, handle that */ + handle_load_update(pred); } } /* reduce_adr_usage */ @@ -740,23 +744,34 @@ static void reduce_adr_usage(ir_node *ptr) { * Check, if an already existing value of mode old_mode can be converted * into the needed one new_mode without loss. */ -static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) { +static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) +{ + unsigned old_size; + unsigned new_size; if (old_mode == new_mode) - return 1; + return true; + + old_size = get_mode_size_bits(old_mode); + new_size = get_mode_size_bits(new_mode); /* if both modes are two-complement ones, we can always convert the - Stored value into the needed one. */ - if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) && + Stored value into the needed one. (on big endian machines we currently + only support this for modes of same size) */ + if (old_size >= new_size && get_mode_arithmetic(old_mode) == irma_twos_complement && - get_mode_arithmetic(new_mode) == irma_twos_complement) - return 1; - return 0; -} /* can_use_stored_value */ + get_mode_arithmetic(new_mode) == irma_twos_complement && + (!be_get_backend_param()->byte_order_big_endian + || old_size == new_size)) { + return true; + } + return false; +} /** - * Check whether a Call is at least pure, ie. does only read memory. + * Check whether a Call is at least pure, i.e. does only read memory. */ -static unsigned is_Call_pure(ir_node *call) { +static unsigned is_Call_pure(ir_node *call) +{ ir_type *call_tp = get_Call_type(call); unsigned prop = get_method_additional_properties(call_tp); @@ -774,6 +789,147 @@ static unsigned is_Call_pure(ir_node *call) { return (prop & (mtp_property_const|mtp_property_pure)) != 0; } /* is_Call_pure */ +static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset) +{ + ir_mode *mode = get_irn_mode(ptr); + long offset = 0; + + /* TODO: long might not be enough, we should probably use some tarval thingy... */ + for (;;) { + if (is_Add(ptr)) { + ir_node *l = get_Add_left(ptr); + ir_node *r = get_Add_right(ptr); + + if (get_irn_mode(l) != mode || !is_Const(r)) + break; + + offset += get_tarval_long(get_Const_tarval(r)); + ptr = l; + } else if (is_Sub(ptr)) { + ir_node *l = get_Sub_left(ptr); + ir_node *r = get_Sub_right(ptr); + + if (get_irn_mode(l) != mode || !is_Const(r)) + break; + + offset -= get_tarval_long(get_Const_tarval(r)); + ptr = l; + } else if (is_Sel(ptr)) { + ir_entity *ent = get_Sel_entity(ptr); + ir_type *tp = get_entity_owner(ent); + + if (is_Array_type(tp)) { + int size; + ir_node *index; + + /* only one dimensional arrays yet */ + if (get_Sel_n_indexs(ptr) != 1) + break; + index = get_Sel_index(ptr, 0); + if (! is_Const(index)) + break; + + tp = get_entity_type(ent); + if (get_type_state(tp) != layout_fixed) + break; + + size = get_type_size_bytes(tp); + offset += size * get_tarval_long(get_Const_tarval(index)); + } else { + if (get_type_state(tp) != layout_fixed) + break; + offset += get_entity_offset(ent); + } + ptr = get_Sel_ptr(ptr); + } else + break; + } + + *pOffset = offset; + return ptr; +} + +static int try_load_after_store(ir_node *load, + ir_node *load_base_ptr, long load_offset, ir_node *store) +{ + ldst_info_t *info; + ir_node *store_ptr = get_Store_ptr(store); + long store_offset; + ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset); + ir_node *store_value; + ir_mode *store_mode; + ir_node *load_ptr; + ir_mode *load_mode; + long load_mode_len; + long store_mode_len; + long delta; + int res; + + if (load_base_ptr != store_base_ptr) + return 0; + + load_mode = get_Load_mode(load); + load_mode_len = get_mode_size_bytes(load_mode); + store_mode = get_irn_mode(get_Store_value(store)); + store_mode_len = get_mode_size_bytes(store_mode); + delta = load_offset - store_offset; + store_value = get_Store_value(store); + + if (delta != 0 || store_mode != load_mode) { + /* TODO: implement for big-endian */ + if (delta < 0 || delta + load_mode_len > store_mode_len + || (be_get_backend_param()->byte_order_big_endian + && load_mode_len != store_mode_len)) + return 0; + + if (get_mode_arithmetic(store_mode) != irma_twos_complement || + get_mode_arithmetic(load_mode) != irma_twos_complement) + return 0; + + + /* produce a shift to adjust offset delta */ + if (delta > 0) { + ir_node *cnst; + ir_graph *irg = get_irn_irg(load); + + cnst = new_r_Const_long(irg, mode_Iu, delta * 8); + store_value = new_r_Shr(get_nodes_block(load), + store_value, cnst, store_mode); + } + + /* add an convert if needed */ + if (store_mode != load_mode) { + store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode); + } + } + + DBG_OPT_RAW(load, store_value); + + info = (ldst_info_t*)get_irn_link(load); + if (info->projs[pn_Load_M]) + exchange(info->projs[pn_Load_M], get_Load_mem(load)); + + res = 0; + /* no exception */ + if (info->projs[pn_Load_X_except]) { + ir_graph *irg = get_irn_irg(load); + exchange( info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X)); + res |= CF_CHANGED; + } + if (info->projs[pn_Load_X_regular]) { + exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load))); + res |= CF_CHANGED; + } + + if (info->projs[pn_Load_res]) + exchange(info->projs[pn_Load_res], store_value); + + load_ptr = get_Load_ptr(load); + kill_node(load); + reduce_adr_usage(load_ptr); + return res | DF_CHANGED; +} + /** * Follow the memory chain as long as there are only Loads, * alias free Stores, and constant Calls and try to replace the @@ -784,80 +940,59 @@ static unsigned is_Call_pure(ir_node *call) { * * INC_MASTER() must be called before dive into */ -static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) { - unsigned res = 0; - ldst_info_t *info = get_irn_link(load); - ir_node *pred; - ir_node *ptr = get_Load_ptr(load); - ir_node *mem = get_Load_mem(load); - ir_mode *load_mode = get_Load_mode(load); +static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) +{ + unsigned res = 0; + ldst_info_t *info = (ldst_info_t*)get_irn_link(load); + ir_node *pred; + ir_node *ptr = get_Load_ptr(load); + ir_node *mem = get_Load_mem(load); + ir_mode *load_mode = get_Load_mode(load); for (pred = curr; load != pred; ) { - ldst_info_t *pred_info = get_irn_link(pred); + ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred); /* - * BEWARE: one might think that checking the modes is useless, because - * if the pointers are identical, they refer to the same object. - * This is only true in strong typed languages, not in C were the following - * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ... + * a Load immediately after a Store -- a read after write. + * We may remove the Load, if both Load & Store does not have an + * exception handler OR they are in the same Block. In the latter + * case the Load cannot throw an exception when the previous Store was + * quiet. + * + * Why we need to check for Store Exception? If the Store cannot + * be executed (ROM) the exception handler might simply jump into + * the load Block :-( + * We could make it a little bit better if we would know that the + * exception handler of the Store jumps directly to the end... */ - if (is_Store(pred) && get_Store_ptr(pred) == ptr && - can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) { - /* - * a Load immediately after a Store -- a read after write. - * We may remove the Load, if both Load & Store does not have an exception handler - * OR they are in the same MacroBlock. In the latter case the Load cannot - * throw an exception when the previous Store was quiet. - * - * Why we need to check for Store Exception? If the Store cannot - * be executed (ROM) the exception handler might simply jump into - * the load MacroBlock :-( - * We could make it a little bit better if we would know that the exception - * handler of the Store jumps directly to the end... - */ - if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) || - get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) { - ir_node *value = get_Store_value(pred); - - DBG_OPT_RAW(load, value); - - /* add an convert if needed */ - if (get_irn_mode(get_Store_value(pred)) != load_mode) { - value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode); - } - - if (info->projs[pn_Load_M]) - exchange(info->projs[pn_Load_M], mem); - - /* no exception */ - if (info->projs[pn_Load_X_except]) { - exchange( info->projs[pn_Load_X_except], new_Bad()); - res |= CF_CHANGED; - } - if (info->projs[pn_Load_X_regular]) { - exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load))); - res |= CF_CHANGED; - } - - if (info->projs[pn_Load_res]) - exchange(info->projs[pn_Load_res], value); - - kill_node(load); - reduce_adr_usage(ptr); - return res | DF_CHANGED; - } + if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL + && info->projs[pn_Load_X_except] == NULL) + || get_nodes_block(load) == get_nodes_block(pred))) + { + long load_offset; + ir_node *base_ptr = get_base_and_offset(ptr, &load_offset); + int changes = try_load_after_store(load, base_ptr, load_offset, pred); + + if (changes != 0) + return res | changes; } else if (is_Load(pred) && get_Load_ptr(pred) == ptr && can_use_stored_value(get_Load_mode(pred), load_mode)) { /* * a Load after a Load -- a read after read. - * We may remove the second Load, if it does not have an exception handler - * OR they are in the same MacroBlock. In the later case the Load cannot - * throw an exception when the previous Load was quiet. + * We may remove the second Load, if it does not have an exception + * handler OR they are in the same Block. In the later case + * the Load cannot throw an exception when the previous Load was + * quiet. + * + * Here, there is no need to check if the previous Load has an + * exception hander because they would have exact the same + * exception... * - * Here, there is no need to check if the previous Load has an exception - * hander because they would have exact the same exception... + * TODO: implement load-after-load with different mode for big + * endian */ - if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) { + if (info->projs[pn_Load_X_except] == NULL + || get_nodes_block(load) == get_nodes_block(pred)) { ir_node *value; DBG_OPT_RAR(load, pred); @@ -866,13 +1001,13 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) { if (info->projs[pn_Load_res]) { if (pred_info->projs[pn_Load_res] == NULL) { /* create a new Proj again */ - pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res); + pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res); } value = pred_info->projs[pn_Load_res]; /* add an convert if needed */ if (get_Load_mode(pred) != load_mode) { - value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode); + value = new_r_Conv(get_nodes_block(load), value, load_mode); } exchange(info->projs[pn_Load_res], value); @@ -883,11 +1018,12 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) { /* no exception */ if (info->projs[pn_Load_X_except]) { - exchange(info->projs[pn_Load_X_except], new_Bad()); + ir_graph *irg = get_irn_irg(load); + exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X)); res |= CF_CHANGED; } if (info->projs[pn_Load_X_regular]) { - exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load))); + exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load))); res |= CF_CHANGED; } @@ -900,7 +1036,6 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) { if (is_Store(pred)) { /* check if we can pass through this store */ ir_alias_relation rel = get_alias_relation( - current_ir_graph, get_Store_ptr(pred), get_irn_mode(get_Store_value(pred)), ptr, load_mode); @@ -944,6 +1079,29 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) { return res; } /* follow_Mem_chain */ +/* + * Check if we can replace the load by a given const from + * the const code irg. + */ +ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) +{ + ir_mode *c_mode = get_irn_mode(c); + ir_mode *l_mode = get_Load_mode(load); + ir_node *block = get_nodes_block(load); + dbg_info *dbgi = get_irn_dbg_info(load); + ir_node *res = copy_const_value(dbgi, c, block); + + if (c_mode != l_mode) { + /* check, if the mode matches OR can be easily converted info */ + if (is_reinterpret_cast(c_mode, l_mode)) { + /* copy the value from the const code irg and cast it */ + res = new_rd_Conv(dbgi, block, res, l_mode); + } + return NULL; + } + return res; +} + /** * optimize a Load * @@ -951,10 +1109,11 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) { */ static unsigned optimize_load(ir_node *load) { - ldst_info_t *info = get_irn_link(load); - ir_node *mem, *ptr, *new_node; - ir_entity *ent; - unsigned res = 0; + ldst_info_t *info = (ldst_info_t*)get_irn_link(load); + ir_node *mem, *ptr, *value; + ir_entity *ent; + long dummy; + unsigned res = 0; /* do NOT touch volatile loads for now */ if (get_Load_volatility(load) == volatility_is_volatile) @@ -963,60 +1122,17 @@ static unsigned optimize_load(ir_node *load) /* the address of the load to be optimized */ ptr = get_Load_ptr(load); - /* - * Check if we can remove the exception from a Load: - * This can be done, if the address is from an Sel(Alloc) and - * the Sel type is a subtype of the allocated type. - * - * This optimizes some often used OO constructs, - * like x = new O; x->t; - */ - if (info->projs[pn_Load_X_except]) { - if (is_Sel(ptr)) { - ir_node *mem = get_Sel_mem(ptr); - - /* FIXME: works with the current FE, but better use the base */ - if (is_Alloc(skip_Proj(mem))) { - /* ok, check the types */ - ir_entity *ent = get_Sel_entity(ptr); - ir_type *s_type = get_entity_type(ent); - ir_type *a_type = get_Alloc_type(mem); - - if (is_SubClass_of(s_type, a_type)) { - /* ok, condition met: there can't be an exception because - * Alloc guarantees that enough memory was allocated */ - - exchange(info->projs[pn_Load_X_except], new_Bad()); - info->projs[pn_Load_X_except] = NULL; - exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load))); - info->projs[pn_Load_X_regular] = NULL; - res |= CF_CHANGED; - } - } - } else if (is_Alloc(skip_Proj(skip_Cast(ptr)))) { - /* simple case: a direct load after an Alloc. Firm Alloc throw - * an exception in case of out-of-memory. So, there is no way for an - * exception in this load. - * This code is constructed by the "exception lowering" in the Jack compiler. - */ - exchange(info->projs[pn_Load_X_except], new_Bad()); - info->projs[pn_Load_X_except] = NULL; - exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load))); - info->projs[pn_Load_X_regular] = NULL; - res |= CF_CHANGED; - } - } - /* The mem of the Load. Must still be returned after optimization. */ - mem = get_Load_mem(load); + mem = get_Load_mem(load); - if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) { - /* a Load which value is neither used nor exception checked, remove it */ + if (info->projs[pn_Load_res] == NULL + && info->projs[pn_Load_X_except] == NULL) { + /* the value is never used and we don't care about exceptions, remove */ exchange(info->projs[pn_Load_M], mem); if (info->projs[pn_Load_X_regular]) { /* should not happen, but if it does, remove it */ - exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load))); + exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load))); res |= CF_CHANGED; } kill_node(load); @@ -1026,124 +1142,80 @@ static unsigned optimize_load(ir_node *load) /* Load from a constant polymorphic field, where we can resolve polymorphism. */ - new_node = transform_node_Load(load); - if (new_node != load) { - if (info->projs[pn_Load_M]) { - exchange(info->projs[pn_Load_M], mem); - info->projs[pn_Load_M] = NULL; - } - if (info->projs[pn_Load_X_except]) { - exchange(info->projs[pn_Load_X_except], new_Bad()); - info->projs[pn_Load_X_except] = NULL; - res |= CF_CHANGED; - } - if (info->projs[pn_Load_X_regular]) { - exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load))); - info->projs[pn_Load_X_regular] = NULL; - res |= CF_CHANGED; - } - if (info->projs[pn_Load_res]) - exchange(info->projs[pn_Load_res], new_node); - - kill_node(load); - reduce_adr_usage(ptr); - return res | DF_CHANGED; - } - - /* check if we can determine the entity that will be loaded */ - ent = find_constant_entity(ptr); - if (ent) { - if ((allocation_static == get_entity_allocation(ent)) && - (visibility_external_allocated != get_entity_visibility(ent))) { - /* a static allocation that is not external: there should be NO exception - * when loading. */ + value = transform_polymorph_Load(load); + if (value == load) { + value = NULL; + /* check if we can determine the entity that will be loaded */ + ent = find_constant_entity(ptr); + if (ent != NULL + && get_entity_visibility(ent) != ir_visibility_external) { + /* a static allocation that is not external: there should be NO + * exception when loading even if we cannot replace the load itself. + */ /* no exception, clear the info field as it might be checked later again */ if (info->projs[pn_Load_X_except]) { - exchange(info->projs[pn_Load_X_except], new_Bad()); + ir_graph *irg = get_irn_irg(load); + exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X)); info->projs[pn_Load_X_except] = NULL; res |= CF_CHANGED; } if (info->projs[pn_Load_X_regular]) { - exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load))); + exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load))); info->projs[pn_Load_X_regular] = NULL; res |= CF_CHANGED; } - if (variability_constant == get_entity_variability(ent)) { - if (is_atomic_entity(ent)) { - /* Might not be atomic after - lowering of Sels. In this - case we could also load, but - it's more complicated. */ - /* more simpler case: we load the content of a constant value: - * replace it by the constant itself - */ - - /* no memory */ - if (info->projs[pn_Load_M]) { - exchange(info->projs[pn_Load_M], mem); - res |= DF_CHANGED; - } - /* no result :-) */ - if (info->projs[pn_Load_res]) { - if (is_atomic_entity(ent)) { - ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent)); - - DBG_OPT_RC(load, c); - exchange(info->projs[pn_Load_res], c); - res |= DF_CHANGED; - } - } - kill_node(load); - reduce_adr_usage(ptr); - return res; - } else { - ir_node *c = NULL; - if (ent->has_initializer) { - /* new style initializer */ - c = find_compound_ent_value(ptr); - } else { - /* old style initializer */ - compound_graph_path *path = get_accessed_path(ptr); + if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) { + if (ent->initializer != NULL) { + /* new style initializer */ + value = find_compound_ent_value(ptr); + } else if (entity_has_compound_ent_values(ent)) { + /* old style initializer */ + compound_graph_path *path = get_accessed_path(ptr); - if (path) { - assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1)); + if (path != NULL) { + assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1)); - c = get_compound_ent_value_by_path(ent, path); - free_compound_graph_path(path); - } - } - if (c != NULL) { - if (info->projs[pn_Load_M]) { - exchange(info->projs[pn_Load_M], mem); - res |= DF_CHANGED; - } - if (info->projs[pn_Load_res]) { - exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c)); - res |= DF_CHANGED; - } - kill_node(load); - reduce_adr_usage(ptr); - return res; - } else { - /* We can not determine a correct access path. E.g., in jack, we load - a byte from an object to generate an exception. Happens in test program - Reflectiontest. - printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent), - get_entity_name(get_irg_entity(current_ir_graph))); - ir_printf(" load: %+F\n", load); - ir_printf(" ptr: %+F\n", ptr); - */ + value = get_compound_ent_value_by_path(ent, path); + DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value)); + free_compound_graph_path(path); } } + if (value != NULL) + value = can_replace_load_by_const(load, value); } } } + if (value != NULL) { + /* we completely replace the load by this value */ + if (info->projs[pn_Load_X_except]) { + ir_graph *irg = get_irn_irg(load); + exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X)); + info->projs[pn_Load_X_except] = NULL; + res |= CF_CHANGED; + } + if (info->projs[pn_Load_X_regular]) { + exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load))); + info->projs[pn_Load_X_regular] = NULL; + res |= CF_CHANGED; + } + if (info->projs[pn_Load_M]) { + exchange(info->projs[pn_Load_M], mem); + res |= DF_CHANGED; + } + if (info->projs[pn_Load_res]) { + exchange(info->projs[pn_Load_res], value); + res |= DF_CHANGED; + } + kill_node(load); + reduce_adr_usage(ptr); + return res; + } /* Check, if the address of this load is used more than once. - * If not, this load cannot be removed in any case. */ - if (get_irn_n_uses(ptr) <= 1) + * If not, more load cannot be removed in any case. */ + if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1) return res; /* @@ -1167,48 +1239,94 @@ static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode) return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode); } /* is_completely_overwritten */ +/** + * Check whether small is a part of large (starting at same address). + */ +static int is_partially_same(ir_node *small, ir_node *large) +{ + ir_mode *sm = get_irn_mode(small); + ir_mode *lm = get_irn_mode(large); + + /* FIXME: Check endianness */ + return is_Conv(small) && get_Conv_op(small) == large + && get_mode_size_bytes(sm) < get_mode_size_bytes(lm) + && get_mode_arithmetic(sm) == irma_twos_complement + && get_mode_arithmetic(lm) == irma_twos_complement; +} /* is_partially_same */ + /** * follow the memory chain as long as there are only Loads and alias free Stores. * * INC_MASTER() must be called before dive into */ -static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) { +static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) +{ unsigned res = 0; - ldst_info_t *info = get_irn_link(store); + ldst_info_t *info = (ldst_info_t*)get_irn_link(store); ir_node *pred; ir_node *ptr = get_Store_ptr(store); ir_node *mem = get_Store_mem(store); ir_node *value = get_Store_value(store); ir_mode *mode = get_irn_mode(value); ir_node *block = get_nodes_block(store); - ir_node *mblk = get_Block_MacroBlock(block); for (pred = curr; pred != store;) { - ldst_info_t *pred_info = get_irn_link(pred); + ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred); /* * BEWARE: one might think that checking the modes is useless, because * if the pointers are identical, they refer to the same object. * This is only true in strong typed languages, not is C were the following * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ... - * However, if the mode that is written have a bigger or equal size the the old - * one, the old value is completely overwritten and can be killed ... + * However, if the size of the mode that is written is bigger or equal the + * size of the old one, the old value is completely overwritten and can be + * killed ... */ if (is_Store(pred) && get_Store_ptr(pred) == ptr && - get_nodes_MacroBlock(pred) == mblk && - is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) { + get_nodes_block(pred) == block) { + /* + * a Store after a Store in the same Block -- a write after write. + */ + /* - * a Store after a Store in the same MacroBlock -- a write after write. - * We may remove the first Store, if it does not have an exception handler. + * We may remove the first Store, if the old value is completely + * overwritten or the old value is a part of the new value, + * and if it does not have an exception handler. * * TODO: What, if both have the same exception handler ??? */ - if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) { - DBG_OPT_WAW(pred, store); - exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred)); - kill_node(pred); - reduce_adr_usage(ptr); - return DF_CHANGED; + if (get_Store_volatility(pred) != volatility_is_volatile + && !pred_info->projs[pn_Store_X_except]) { + ir_node *predvalue = get_Store_value(pred); + ir_mode *predmode = get_irn_mode(predvalue); + + if (is_completely_overwritten(predmode, mode) + || is_partially_same(predvalue, value)) { + DBG_OPT_WAW(pred, store); + exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred)); + kill_node(pred); + reduce_adr_usage(ptr); + return DF_CHANGED; + } + } + + /* + * We may remove the Store, if the old value already contains + * the new value, and if it does not have an exception handler. + * + * TODO: What, if both have the same exception handler ??? + */ + if (get_Store_volatility(store) != volatility_is_volatile + && !info->projs[pn_Store_X_except]) { + ir_node *predvalue = get_Store_value(pred); + + if (is_partially_same(value, predvalue)) { + DBG_OPT_WAW(pred, store); + exchange(info->projs[pn_Store_M], mem); + kill_node(store); + reduce_adr_usage(ptr); + return DF_CHANGED; + } } } else if (is_Load(pred) && get_Load_ptr(pred) == ptr && value == pred_info->projs[pn_Load_res]) { @@ -1228,9 +1346,8 @@ static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) { } if (is_Store(pred)) { - /* check if we can pass thru this store */ + /* check if we can pass through this store */ ir_alias_relation rel = get_alias_relation( - current_ir_graph, get_Store_ptr(pred), get_irn_mode(get_Store_value(pred)), ptr, mode); @@ -1240,7 +1357,7 @@ static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) { pred = skip_Proj(get_Store_mem(pred)); } else if (is_Load(pred)) { ir_alias_relation rel = get_alias_relation( - current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred), + get_Load_ptr(pred), get_Load_mode(pred), ptr, mode); if (rel != ir_no_alias) break; @@ -1270,18 +1387,63 @@ static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) { return res; } /* follow_Mem_chain_for_Store */ +/** find entity used as base for an address calculation */ +static ir_entity *find_entity(ir_node *ptr) +{ + switch (get_irn_opcode(ptr)) { + case iro_SymConst: + return get_SymConst_entity(ptr); + case iro_Sel: { + ir_node *pred = get_Sel_ptr(ptr); + if (get_irg_frame(get_irn_irg(ptr)) == pred) + return get_Sel_entity(ptr); + + return find_entity(pred); + } + case iro_Sub: + case iro_Add: { + ir_node *left = get_binop_left(ptr); + ir_node *right; + if (mode_is_reference(get_irn_mode(left))) + return find_entity(left); + right = get_binop_right(ptr); + if (mode_is_reference(get_irn_mode(right))) + return find_entity(right); + return NULL; + } + default: + return NULL; + } +} + /** * optimize a Store * * @param store the Store node */ -static unsigned optimize_store(ir_node *store) { - ir_node *ptr, *mem; +static unsigned optimize_store(ir_node *store) +{ + ir_node *ptr; + ir_node *mem; + ir_entity *entity; if (get_Store_volatility(store) == volatility_is_volatile) return 0; - ptr = get_Store_ptr(store); + ptr = get_Store_ptr(store); + entity = find_entity(ptr); + + /* a store to an entity which is never read is unnecessary */ + if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) { + ldst_info_t *info = (ldst_info_t*)get_irn_link(store); + if (info->projs[pn_Store_X_except] == NULL) { + DB((dbg, LEVEL_1, " Killing useless %+F to never read entity %+F\n", store, entity)); + exchange(info->projs[pn_Store_M], get_Store_mem(store)); + kill_node(store); + reduce_adr_usage(ptr); + return DF_CHANGED; + } + } /* Check, if the address of this Store is used more than once. * If not, this Store cannot be removed in any case. */ @@ -1318,7 +1480,10 @@ static unsigned optimize_store(ir_node *store) { static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) { int i, n; - ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM; + ir_node *store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM; +#ifdef DO_CACHEOPT + ir_node *old_store; +#endif ir_mode *mode; ir_node **inM, **inD, **projMs; int *idx; @@ -1341,19 +1506,17 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) return 0; store = skip_Proj(projM); +#ifdef DO_CACHEOPT old_store = store; +#endif if (!is_Store(store)) return 0; block = get_nodes_block(store); - /* abort on dead blocks */ - if (is_Block_dead(block)) - return 0; - /* check if the block is post dominated by Phi-block and has no exception exit */ - bl_info = get_irn_link(block); + bl_info = (block_info_t*)get_irn_link(block); if (bl_info->flags & BLOCK_HAS_EXC) return 0; @@ -1364,7 +1527,7 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) /* this is the address of the store */ ptr = get_Store_ptr(store); mode = get_irn_mode(get_Store_value(store)); - info = get_irn_link(store); + info = (ldst_info_t*)get_irn_link(store); exc = info->exc_block; for (i = 1; i < n; ++i) { @@ -1380,22 +1543,19 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred))) return 0; - info = get_irn_link(pred); + info = (ldst_info_t*)get_irn_link(pred); /* check, if all stores have the same exception flow */ if (exc != info->exc_block) return 0; - /* abort on dead blocks */ block = get_nodes_block(pred); - if (is_Block_dead(block)) - return 0; /* check if the block is post dominated by Phi-block and has no exception exit. Note that block must be different from Phi-block, else we would move a Store from end End of a block to its Start... */ - bl_info = get_irn_link(block); + bl_info = (block_info_t*)get_irn_link(block); if (bl_info->flags & BLOCK_HAS_EXC) return 0; if (block == phi_block || ! block_postdominates(phi_block, block)) @@ -1435,7 +1595,7 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) assert(is_Proj(projMs[i])); store = get_Proj_pred(projMs[i]); - info = get_irn_link(store); + info = (ldst_info_t*)get_irn_link(store); inM[i] = get_Store_mem(store); inD[i] = get_Store_value(store); @@ -1444,16 +1604,16 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) block = get_nodes_block(phi); /* second step: create a new memory Phi */ - phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M); + phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M); /* third step: create a new data Phi */ - phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode); + phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode); /* rewire memory and kill the node */ for (i = n - 1; i >= 0; --i) { ir_node *proj = projMs[i]; - if(is_Proj(proj)) { + if (is_Proj(proj)) { ir_node *store = get_Proj_pred(proj); exchange(proj, inM[i]); kill_node(store); @@ -1461,19 +1621,19 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) } /* fourth step: create the Store */ - store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD); + store = new_rd_Store(db, block, phiM, ptr, phiD, cons_none); #ifdef DO_CACHEOPT co_set_irn_name(store, co_get_irn_ident(old_store)); #endif - projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M); + projM = new_rd_Proj(NULL, store, mode_M, pn_Store_M); info = get_ldst_info(store, &wenv->obst); info->projs[pn_Store_M] = projM; /* fifths step: repair exception flow */ if (exc) { - ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except); + ir_node *projX = new_rd_Proj(NULL, store, mode_X, pn_Store_X_except); info->projs[pn_Store_X_except] = projX; info->exc_block = exc; @@ -1499,8 +1659,9 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) /** * walker, do the optimizations */ -static void do_load_store_optimize(ir_node *n, void *env) { - walk_env_t *wenv = env; +static void do_load_store_optimize(ir_node *n, void *env) +{ + walk_env_t *wenv = (walk_env_t*)env; switch (get_irn_opcode(n)) { @@ -1517,20 +1678,19 @@ static void do_load_store_optimize(ir_node *n, void *env) { break; default: - ; + break; } } /* do_load_store_optimize */ /** A scc. */ typedef struct scc { - ir_node *head; /**< the head of the list */ + ir_node *head; /**< the head of the list */ } scc; /** A node entry. */ typedef struct node_entry { unsigned DFSnum; /**< the DFS number of this node */ unsigned low; /**< the low number of this node */ - ir_node *header; /**< the header of this node */ int in_stack; /**< flag, set if the node is on the stack */ ir_node *next; /**< link to the next node the the same scc */ scc *pscc; /**< the scc of this node */ @@ -1541,7 +1701,7 @@ typedef struct node_entry { typedef struct loop_env { ir_phase ph; /**< the phase object */ ir_node **stack; /**< the node stack */ - int tos; /**< tos index */ + size_t tos; /**< tos index */ unsigned nextDFSnum; /**< the current DFS number */ unsigned POnum; /**< current post order number */ @@ -1551,12 +1711,13 @@ typedef struct loop_env { /** * Gets the node_entry of a node */ -static node_entry *get_irn_ne(ir_node *irn, loop_env *env) { +static node_entry *get_irn_ne(ir_node *irn, loop_env *env) +{ ir_phase *ph = &env->ph; - node_entry *e = phase_get_irn_data(&env->ph, irn); + node_entry *e = (node_entry*)phase_get_irn_data(&env->ph, irn); if (! e) { - e = phase_alloc(ph, sizeof(*e)); + e = (node_entry*)phase_alloc(ph, sizeof(*e)); memset(e, 0, sizeof(*e)); phase_set_irn_data(ph, irn, e); } @@ -1569,11 +1730,12 @@ static node_entry *get_irn_ne(ir_node *irn, loop_env *env) { * @param env the loop environment * @param n the node to push */ -static void push(loop_env *env, ir_node *n) { +static void push(loop_env *env, ir_node *n) +{ node_entry *e; if (env->tos == ARR_LEN(env->stack)) { - int nlen = ARR_LEN(env->stack) * 2; + size_t nlen = ARR_LEN(env->stack) * 2; ARR_RESIZE(ir_node *, env->stack, nlen); } env->stack[env->tos++] = n; @@ -1588,7 +1750,8 @@ static void push(loop_env *env, ir_node *n) { * * @return The topmost node */ -static ir_node *pop(loop_env *env) { +static ir_node *pop(loop_env *env) +{ ir_node *n = env->stack[--env->tos]; node_entry *e = get_irn_ne(n, env); @@ -1603,7 +1766,8 @@ static ir_node *pop(loop_env *env) { * @param irn the node to check * @param header_block the header block of the induction variable */ -static int is_rc(ir_node *irn, ir_node *header_block) { +static int is_rc(ir_node *irn, ir_node *header_block) +{ ir_node *block = get_nodes_block(irn); return (block != header_block) && block_dominates(block, header_block); @@ -1617,17 +1781,49 @@ struct phi_entry { phi_entry *next; }; +/** + * An entry in the avail set. + */ +typedef struct avail_entry_t { + ir_node *ptr; /**< the address pointer */ + ir_mode *mode; /**< the load mode */ + ir_node *load; /**< the associated Load */ +} avail_entry_t; + +/** + * Compare two avail entries. + */ +static int cmp_avail_entry(const void *elt, const void *key, size_t size) +{ + const avail_entry_t *a = (const avail_entry_t*)elt; + const avail_entry_t *b = (const avail_entry_t*)key; + (void) size; + + return a->ptr != b->ptr || a->mode != b->mode; +} /* cmp_avail_entry */ + +/** + * Calculate the hash value of an avail entry. + */ +static unsigned hash_cache_entry(const avail_entry_t *entry) +{ + return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode); +} /* hash_cache_entry */ + /** * Move loops out of loops if possible. * * @param pscc the loop described by an SCC * @param env the loop environment */ -static void move_loads_out_of_loops(scc *pscc, loop_env *env) { +static void move_loads_out_of_loops(scc *pscc, loop_env *env) +{ ir_node *phi, *load, *next, *other, *next_other; - ir_entity *ent; int j; phi_entry *phi_list = NULL; + set *avail; + + avail = new_set(cmp_avail_entry, 8); /* collect all outer memories */ for (phi = pscc->head; phi != NULL; phi = next) { @@ -1638,7 +1834,7 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) { if (! is_Phi(phi)) continue; - assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi"); + assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi"); for (j = get_irn_arity(phi) - 1; j >= 0; --j) { ir_node *pred = get_irn_n(phi, j); @@ -1646,7 +1842,7 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) { if (pe->pscc != ne->pscc) { /* not in the same SCC, is region const */ - phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe)); + phi_entry *pe = (phi_entry*)phase_alloc(&env->ph, sizeof(*pe)); pe->phi = phi; pe->pos = j; @@ -1658,23 +1854,26 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) { /* no Phis no fun */ assert(phi_list != NULL && "DFS found a loop without Phi"); + /* for now, we cannot handle more than one input (only reducible cf) */ + if (phi_list->next != NULL) + return; + for (load = pscc->head; load; load = next) { ir_mode *load_mode; node_entry *ne = get_irn_ne(load, env); next = ne->next; if (is_Load(load)) { - ldst_info_t *info = get_irn_link(load); + ldst_info_t *info = (ldst_info_t*)get_irn_link(load); ir_node *ptr = get_Load_ptr(load); /* for now, we cannot handle Loads with exceptions */ if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL) continue; - /* for now, we can only handle Load(Global) */ + /* for now, we can only move Load(Global) */ if (! is_Global(ptr)) continue; - ent = get_Global_entity(ptr); load_mode = get_Load_mode(load); for (other = pscc->head; other != NULL; other = next_other) { node_entry *ne = get_irn_ne(other, env); @@ -1682,7 +1881,6 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) { if (is_Store(other)) { ir_alias_relation rel = get_alias_relation( - current_ir_graph, get_Store_ptr(other), get_irn_mode(get_Store_value(other)), ptr, load_mode); @@ -1690,17 +1888,13 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) { if (rel != ir_no_alias) break; } - /* only pure Calls are allowed here, so ignore them */ + /* only Phis and pure Calls are allowed here, so ignore them */ } if (other == NULL) { - ldst_info_t *ninfo; + ldst_info_t *ninfo = NULL; phi_entry *pe; dbg_info *db; - /* for now, we cannot handle more than one input */ - if (phi_list->next != NULL) - return; - /* yep, no aliasing Store found, Load can be moved */ DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load)); @@ -1711,16 +1905,31 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) { ir_node *blk = get_nodes_block(phi); ir_node *pred = get_Block_cfgpred_block(blk, pos); ir_node *irn, *mem; + avail_entry_t entry, *res; - pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode); + entry.ptr = ptr; + entry.mode = load_mode; + res = (avail_entry_t*)set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry)); + if (res != NULL) { + irn = res->load; + } else { + irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, cons_none); + entry.load = irn; + set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry)); + DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred)); + } + pe->load = irn; ninfo = get_ldst_info(irn, phase_obst(&env->ph)); - ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M); - set_Phi_pred(phi, pos, mem); - - ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res); + ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M); + if (res == NULL) { + /* irn is from cache, so do not set phi pred again. + * There might be other Loads between phi and irn already. + */ + set_Phi_pred(phi, pos, mem); + } - DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred)); + ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res); } /* now kill the old Load */ @@ -1731,6 +1940,7 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) { } } } + del_set(avail); } /* move_loads_out_of_loops */ /** @@ -1739,7 +1949,8 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) { * @param pscc the SCC * @param env the loop environment */ -static void process_loop(scc *pscc, loop_env *env) { +static void process_loop(scc *pscc, loop_env *env) +{ ir_node *irn, *next, *header = NULL; node_entry *b, *h = NULL; int j, only_phi, num_outside, process = 0; @@ -1753,13 +1964,12 @@ static void process_loop(scc *pscc, loop_env *env) { next = e->next; b = get_irn_ne(block, env); - if (header) { + if (header != NULL) { if (h->POnum < b->POnum) { header = block; h = b; } - } - else { + } else { header = block; h = b; } @@ -1814,10 +2024,12 @@ static void process_loop(scc *pscc, loop_env *env) { /* not a memory loop */ goto fail; } - if (! out_rc) { + if (out_rc == NULL) { + /* first region constant */ out_rc = pred; ++num_outside; } else if (out_rc != pred) { + /* another region constant */ ++num_outside; } } @@ -1837,22 +2049,20 @@ static void process_loop(scc *pscc, loop_env *env) { for (irn = pscc->head; irn; irn = next) { node_entry *e = get_irn_ne(irn, env); next = e->next; - e->header = NULL; exchange(irn, out_rc); } env->changes |= DF_CHANGED; return; } - /* set the header for every node in this scc */ +#ifdef DEBUG_libfirm for (irn = pscc->head; irn; irn = next) { node_entry *e = get_irn_ne(irn, env); - e->header = header; next = e->next; DB((dbg, LEVEL_2, " %+F,", irn)); } DB((dbg, LEVEL_2, "\n")); - +#endif move_loads_out_of_loops(pscc, env); fail: @@ -1865,7 +2075,8 @@ fail: * @param pscc the SCC * @param env the loop environment */ -static void process_scc(scc *pscc, loop_env *env) { +static void process_scc(scc *pscc, loop_env *env) +{ ir_node *head = pscc->head; node_entry *e = get_irn_ne(head, env); @@ -1915,7 +2126,7 @@ static void dfs(ir_node *irn, loop_env *env) ir_node *pred = get_irn_n(irn, i); node_entry *o = get_irn_ne(pred, env); - if (irn_not_visited(pred)) { + if (!irn_visited(pred)) { dfs(pred, env); node->low = MIN(node->low, o->low); } @@ -1926,7 +2137,7 @@ static void dfs(ir_node *irn, loop_env *env) ir_node *pred = get_fragile_op_mem(irn); node_entry *o = get_irn_ne(pred, env); - if (irn_not_visited(pred)) { + if (!irn_visited(pred)) { dfs(pred, env); node->low = MIN(node->low, o->low); } @@ -1936,7 +2147,7 @@ static void dfs(ir_node *irn, loop_env *env) ir_node *pred = get_Proj_pred(irn); node_entry *o = get_irn_ne(pred, env); - if (irn_not_visited(pred)) { + if (!irn_visited(pred)) { dfs(pred, env); node->low = MIN(node->low, o->low); } @@ -1948,7 +2159,7 @@ static void dfs(ir_node *irn, loop_env *env) } if (node->low == node->DFSnum) { - scc *pscc = phase_alloc(&env->ph, sizeof(*pscc)); + scc *pscc = (scc*)phase_alloc(&env->ph, sizeof(*pscc)); ir_node *x; pscc->head = NULL; @@ -1972,12 +2183,11 @@ static void dfs(ir_node *irn, loop_env *env) * @param irg the graph to process * @param env the loop environment */ -static void do_dfs(ir_graph *irg, loop_env *env) { - ir_graph *rem = current_ir_graph; +static void do_dfs(ir_graph *irg, loop_env *env) +{ ir_node *endblk, *end; int i; - current_ir_graph = irg; inc_irg_visited(irg); /* visit all memory nodes */ @@ -1986,13 +2196,15 @@ static void do_dfs(ir_graph *irg, loop_env *env) { ir_node *pred = get_Block_cfgpred(endblk, i); pred = skip_Proj(pred); - if (is_Return(pred)) + if (is_Return(pred)) { dfs(get_Return_mem(pred), env); - else if (is_Raise(pred)) + } else if (is_Raise(pred)) { dfs(get_Raise_mem(pred), env); - else if (is_fragile_op(pred)) + } else if (is_fragile_op(pred)) { dfs(get_fragile_op_mem(pred), env); - else { + } else if (is_Bad(pred)) { + /* ignore non-optimized block predecessor */ + } else { assert(0 && "Unknown EndBlock predecessor"); } } @@ -2002,28 +2214,18 @@ static void do_dfs(ir_graph *irg, loop_env *env) { for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) { ir_node *ka = get_End_keepalive(end, i); - if (is_Phi(ka) && irn_not_visited(ka)) + if (is_Phi(ka) && !irn_visited(ka)) dfs(ka, env); } - current_ir_graph = rem; } /* do_dfs */ -/** - * Initialize new phase data. We do this always explicit, so return NULL here - */ -static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) { - (void)ph; - (void)irn; - (void)data; - return NULL; -} /* init_loop_data */ - /** * Optimize Loads/Stores in loops. * * @param irg the graph */ -static int optimize_loops(ir_graph *irg) { +static int optimize_loops(ir_graph *irg) +{ loop_env env; env.stack = NEW_ARR_F(ir_node *, 128); @@ -2031,13 +2233,13 @@ static int optimize_loops(ir_graph *irg) { env.nextDFSnum = 0; env.POnum = 0; env.changes = 0; - phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL); + phase_init(&env.ph, irg, phase_irn_init_default); /* calculate the SCC's and drive loop optimization. */ do_dfs(irg, &env); DEL_ARR_F(env.stack); - phase_free(&env.ph); + phase_deinit(&env.ph); return env.changes; } /* optimize_loops */ @@ -2045,7 +2247,8 @@ static int optimize_loops(ir_graph *irg) { /* * do the load store optimization */ -void optimize_load_store(ir_graph *irg) { +int optimize_load_store(ir_graph *irg) +{ walk_env_t env; FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt"); @@ -2063,8 +2266,8 @@ void optimize_load_store(ir_graph *irg) { assure_postdoms(irg); if (get_opt_alias_analysis()) { - assure_irg_address_taken_computed(irg); - assure_irp_globals_address_taken_computed(); + assure_irg_entity_usage_computed(irg); + assure_irp_globals_entity_usage_computed(); } obstack_init(&env.obst); @@ -2083,7 +2286,8 @@ void optimize_load_store(ir_graph *irg) { /* Handle graph state */ if (env.changes) { - set_irg_outs_inconsistent(irg); + set_irg_entity_usage_state(irg, ir_entity_usage_not_computed); + edges_deactivate(irg); } if (env.changes & CF_CHANGED) { @@ -2091,4 +2295,10 @@ void optimize_load_store(ir_graph *irg) { have Bad() predecessors. */ set_irg_doms_inconsistent(irg); } + return env.changes != 0; } /* optimize_load_store */ + +ir_graph_pass_t *optimize_load_store_pass(const char *name) +{ + return def_graph_pass_ret(name ? name : "ldst", optimize_load_store); +} /* optimize_load_store_pass */