X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fana%2Firscc.c;h=e4a7595e39fc8d1aa4bf27819bb45db0347c8bcb;hb=64eae41a44856b8b13b1ea6edce58188f1de7e3e;hp=ae3ed653feb47083c38450636c3b166cb2a3e5e8;hpb=f364f32a3809b57418a0cbf04c37b3ed643158ae;p=libfirm diff --git a/ir/ana/irscc.c b/ir/ana/irscc.c index ae3ed653f..e4a7595e3 100644 --- a/ir/ana/irscc.c +++ b/ir/ana/irscc.c @@ -14,22 +14,33 @@ */ #ifdef HAVE_CONFIG_H -#include "config.h" +# include "config.h" #endif -#include +#ifdef HAVE_STRING_H +# include +#endif +#ifdef HAVE_STDLIB_H +# include +#endif #include "irloop_t.h" -#include "irnode.h" + +#include "irprog_t.h" #include "irgraph_t.h" +#include "irnode_t.h" +#include "irgwalk.h" #include "array.h" #include "pmap.h" -#include "irgwalk.h" -#include "irprog_t.h" + #include "irdump.h" -ir_graph *outermost_ir_graph; /* The outermost graph the scc is computed - for */ +/* A variant of the loop tree that avoids loops without head. + This reduces the depth of the loop tree. */ +#define NO_LOOPS_WITHOUT_HEAD 1 + +static ir_graph *outermost_ir_graph; /* The outermost graph the scc is computed + for */ static ir_loop *current_loop; /* Current loop construction is working on. */ static int loop_node_cnt = 0; /* Counts the number of allocated loop nodes. @@ -38,22 +49,25 @@ static int loop_node_cnt = 0; /* Counts the number of allocated loop nodes. static int current_dfn = 1; /* Counter to generate depth first numbering of visited nodes. */ +static int max_loop_depth = 0; + +void link_to_reg_end (ir_node *n, void *env); +void set_projx_link(ir_node *cb_projx, ir_node *end_projx); +ir_node *get_projx_link(ir_node *cb_projx); + /**********************************************************************/ /* Node attributes **/ /**********************************************************************/ -/* A map to get from irnodes to loop nodes. */ -static pmap *node_loop_map = NULL; - /**********************************************************************/ /* Node attributes needed for the construction. **/ /**********************************************************************/ typedef struct scc_info { - bool in_stack; /* Marks whether node is on the stack. */ - int dfn; /* Depth first search number. */ - int uplink; /* dfn number of ancestor. */ - // ir_loop *loop; /* Refers to the containing loop. */ + int in_stack; /**< Marks whether node is on the stack. */ + int dfn; /**< Depth first search number. */ + int uplink; /**< dfn number of ancestor. */ + /* ir_loop *loop; *//* Refers to the containing loop. */ /* struct section *section; xset def; @@ -69,70 +83,66 @@ static INLINE scc_info* new_scc_info(void) { static INLINE void mark_irn_in_stack (ir_node *n) { - assert(get_irn_link(n)); - ((scc_info *)get_irn_link(n))->in_stack = true; + scc_info *scc = get_irn_link(n); + assert(scc); + scc->in_stack = 1; } static INLINE void mark_irn_not_in_stack (ir_node *n) { - assert(get_irn_link(n)); - ((scc_info *)get_irn_link(n))->in_stack = false; + scc_info *scc = get_irn_link(n); + assert(scc); + scc->in_stack = 0; } -static INLINE bool +static INLINE int irn_is_in_stack (ir_node *n) { - assert(get_irn_link(n)); - return ((scc_info *)get_irn_link(n))->in_stack; + scc_info *scc = get_irn_link(n); + assert(scc); + return scc->in_stack; } static INLINE void set_irn_uplink (ir_node *n, int uplink) { - assert(get_irn_link(n)); - ((scc_info *)get_irn_link(n))->uplink = uplink; + scc_info *scc = get_irn_link(n); + assert(scc); + scc->uplink = uplink; } -static INLINE int +int get_irn_uplink (ir_node *n) { - assert(get_irn_link(n)); - return ((scc_info *)get_irn_link(n))->uplink; + scc_info *scc = get_irn_link(n); + assert(scc); + return scc->uplink; } static INLINE void set_irn_dfn (ir_node *n, int dfn) { - if (! get_irn_link(n)) { DDMN(n); DDME(get_irg_ent(current_ir_graph));} - assert(get_irn_link(n)); - ((scc_info *)get_irn_link(n))->dfn = dfn; + scc_info *scc = get_irn_link(n); + assert(scc); + scc->dfn = dfn; } -static INLINE int +int get_irn_dfn (ir_node *n) { - assert(get_irn_link(n)); - return ((scc_info *)get_irn_link(n))->dfn; + scc_info *scc = get_irn_link(n); + assert(scc); + return scc->dfn; } -/* Uses temporary information to set the loop */ -static INLINE void -set_irn_loop (ir_node *n, ir_loop* loop) { - //assert(get_irn_link(n)); - //((scc_info *)get_irn_link(n))->loop = loop; - assert(node_loop_map && "not initialized!"); - pmap_insert(node_loop_map, (void *)n, (void *)loop); + +void +set_irn_loop (ir_node *n, ir_loop *loop) { + n->loop = loop; } /* Uses temporary information to get the loop */ -INLINE ir_loop * +ir_loop * get_irn_loop (ir_node *n) { - ir_loop *res = NULL; - //assert(get_irn_link(n)); - //return ((scc_info *)get_irn_link(n))->loop; - assert(node_loop_map && "not initialized!"); - - if (pmap_contains(node_loop_map, (void *)n)) - res = (ir_loop *) pmap_get(node_loop_map, (void *)n); - - return res; + return n->loop; } + #if 0 static ir_loop *find_nodes_loop (ir_node *n, ir_loop *l) { int i; @@ -168,6 +178,9 @@ ir_loop * get_irn_loop(ir_node *n) { static ir_node **stack = NULL; static int tos = 0; /* top of stack */ +/** + * initializes the stack + */ static INLINE void init_stack(void) { if (stack) { ARR_RESIZE (ir_node *, stack, 1000); @@ -185,6 +198,11 @@ static INLINE void free_stack(void) { } #endif +/** + * push a node onto the stack + * + * @param n The node to push + */ static INLINE void push (ir_node *n) { @@ -198,6 +216,11 @@ push (ir_node *n) mark_irn_in_stack(n); } +/** + * pop a node from the stack + * + * @return The topmost node + */ static INLINE ir_node * pop (void) { @@ -206,55 +229,60 @@ pop (void) return n; } -/* The nodes up to n belong to the current loop. - Removes them from the stack and adds them to the current loop. */ +/** + * The nodes up to n belong to the current loop. + * Removes them from the stack and adds them to the current loop. + */ static INLINE void pop_scc_to_loop (ir_node *n) { ir_node *m; int i = 0; - /*for (;;) {*/ - do - { + do { m = pop(); + + //printf(" dfn: %d, upl %d upl-new %d ", get_irn_dfn(m), get_irn_uplink(m), loop_node_cnt+1); DDMN(m); + loop_node_cnt++; set_irn_dfn(m, loop_node_cnt); add_loop_node(current_loop, m); set_irn_loop(m, current_loop); i++; + /* if (m==n) break;*/ - } while(m != n); + } while(m != n); - if(i > 1) - printf("Mehr als eine Iteration!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n"); + /* i might be bigger than 1 for dead (and that's why bad) loops */ + /* if(i > 1) + printf("Mehr als eine Iteration!!!!!!!!!!!!!!!!!!!!!!!!!!!!11111\n"); + */ } /* GL ??? my last son is my grandson??? Removes loops with no ir_nodes in them. Such loops have only another loop as son. (Why can't they have two loops as sons? Does it never get that far? ) */ -void close_loop (ir_loop *l) +static void close_loop (ir_loop *l) { int last = get_loop_n_elements(l) - 1; loop_element lelement = get_loop_element(l, last); ir_loop *last_son = lelement.son; if (get_kind(last_son) == k_ir_loop && - get_loop_n_elements(last_son) == 1) - { - ir_loop *gson; - - lelement = get_loop_element(last_son, 0); - gson = lelement.son; - if(get_kind(gson) == k_ir_loop) - { - loop_element new_last_son; - - gson -> outer_loop = l; - new_last_son.son = gson; - l -> children[last] = new_last_son; - } + get_loop_n_elements(last_son) == 1) { + ir_loop *gson; + + lelement = get_loop_element(last_son, 0); + gson = lelement.son; + + if (get_kind(gson) == k_ir_loop) { + loop_element new_last_son; + + gson->outer_loop = l; + new_last_son.son = gson; + l->children[last] = new_last_son; } + } current_loop = l; } @@ -278,7 +306,7 @@ pop_scc_unmark_visit (ir_node *n) /* Allocates a new loop as son of current_loop. Sets current_loop to the new loop and returns the father. */ -static ir_loop *new_loop (void) { +ir_loop *new_loop (void) { ir_loop *father, *son; father = current_loop; @@ -293,6 +321,7 @@ static ir_loop *new_loop (void) { son->outer_loop = father; add_loop_son(father, son); son->depth = father->depth+1; + if (son->depth > max_loop_depth) max_loop_depth = son->depth; } else { /* The root loop */ son->outer_loop = son; son->depth = 0; @@ -341,7 +370,7 @@ int get_loop_n_sons (ir_loop *loop) { /* Returns the pos`th loop_node-child * * TODO: This method isn`t very efficient ! * - * Returns NULL if there isnt`t a pos`th loop_node */ + * Returns NULL if there isn`t a pos`th loop_node */ ir_loop *get_loop_son (ir_loop *loop, int pos) { int child_nr = 0, loop_nr = -1; @@ -360,7 +389,7 @@ ir_loop *get_loop_son (ir_loop *loop, int pos) { /* Use EXCLUSIVELY this function to add sons, otherwise the loop->n_sons is invalid! */ -static INLINE void +void add_loop_son(ir_loop *loop, ir_loop *son) { loop_element lson; lson.son = son; @@ -379,7 +408,7 @@ int get_loop_n_nodes (ir_loop *loop) { /* Returns the pos`th ir_node-child * * TODO: This method isn`t very efficient ! * - * Returns NULL if there isnt`t a pos`th ir_node */ + * Returns NULL if there isn`t a pos`th ir_node */ ir_node *get_loop_node (ir_loop *loop, int pos) { int child_nr, node_nr = -1; @@ -392,6 +421,8 @@ ir_node *get_loop_node (ir_loop *loop, int pos) { if(node_nr == pos) return(loop -> children[child_nr].node); } + DDML(loop); + printf("pos: %d\n", pos); assert(0 && "no child at pos found"); return NULL; } @@ -399,12 +430,12 @@ ir_node *get_loop_node (ir_loop *loop, int pos) { /* Use EXCLUSIVELY this function to add nodes, otherwise the loop->n_nodes is invalid! */ -static INLINE void +void add_loop_node(ir_loop *loop, ir_node *n) { loop_element ln; ln.node = n; assert(loop && loop->kind == k_ir_loop); - assert(get_kind(n) == k_ir_node); + assert(get_kind(n) == k_ir_node || get_kind(n) == k_ir_graph); /* used in callgraph.c */ ARR_APP1 (loop_element, loop->children, ln); loop->n_nodes++; } @@ -419,7 +450,7 @@ int get_loop_n_elements (ir_loop *loop) { Returns the pos`th loop element. This may be a loop_node or a ir_node. The caller of this function has to check the *(loop_element.kind) field for "k_ir_node" or "k_ir_loop" - and then select the apropriate "loop_element.node" or "loop_element.son". + and then select the appropriate "loop_element.node" or "loop_element.son". */ loop_element get_loop_element (ir_loop *loop, int pos) { @@ -464,14 +495,18 @@ void *get_loop_link (const ir_loop *loop) { #endif } +int (is_ir_loop)(const void *thing) { + return _is_ir_loop(thing); +} + /* The outermost loop is remarked in the surrounding graph. */ -void set_irg_loop(ir_graph *irg, ir_loop *loop) { - assert(irg); - irg->loop = loop; +void (set_irg_loop)(ir_graph *irg, ir_loop *loop) { + _set_irg_loop(irg, loop); } -ir_loop *get_irg_loop(ir_graph *irg) { - assert(irg); - return irg->loop; + +/* Returns the root loop info (if exists) for an irg. */ +ir_loop *(get_irg_loop)(ir_graph *irg) { + return _get_irg_loop(irg); } @@ -485,39 +520,12 @@ static INLINE void init_node (ir_node *n, void *env) { set_irn_link (n, new_scc_info()); clear_backedges(n); -#if 0 - /* Also init nodes not visible in intraproc_view. */ - /* @@@ init_node is called for too many nodes -- this wastes memory!. - The mem is not lost as its on the obstack. */ - if (get_irn_op(n) == op_Filter) { - for (i = 0; i < get_Filter_n_cg_preds(n); i++) - init_node(get_Filter_cg_pred(n, i), NULL); - } - if (get_irn_op(n) == op_Block) { - for (i = 0; i < get_Block_cg_n_cfgpreds(n); i++) { - init_node(get_Block_cg_cfgpred(n, i), NULL); - } - } - /* The following pattern matches only after a call from above pattern. */ - if ((get_irn_op(n) == op_Proj) /*&& (get_Proj_proj(n) == 0)*/) { - /* @@@ init_node is called for every proj -- this wastes memory!. - The mem is not lost as its on the obstack. */ - ir_node *cb = get_Proj_pred(n); - if ((get_irn_op(cb) == op_CallBegin) || - (get_irn_op(cb) == op_EndReg) || - (get_irn_op(cb) == op_EndExcept)) { - init_node(cb, NULL); - init_node(get_nodes_Block(cb), NULL); - } - } -#endif } static INLINE void init_scc_common (void) { current_dfn = 1; loop_node_cnt = 0; - if (!node_loop_map) node_loop_map = pmap_create(); init_stack(); } @@ -534,17 +542,21 @@ static INLINE void init_ip_scc (void) { init_scc_common(); cg_walk (init_node, NULL, NULL); + +#if EXPERIMENTAL_LOOP_TREE + cg_walk (link_to_reg_end, NULL, NULL); +#endif } /* Condition for breaking the recursion. */ -static bool is_outermost_Start(ir_node *n) { +static int is_outermost_Start(ir_node *n) { /* Test whether this is the outermost Start node. If so recursion must end. */ if ((get_irn_op(n) == op_Block) && (get_Block_n_cfgpreds(n) == 1) && (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Start) && - (get_nodes_Block(skip_Proj(get_Block_cfgpred(n, 0))) == n)) { - return true; + (get_nodes_block(skip_Proj(get_Block_cfgpred(n, 0))) == n)) { + return 1; } #if 0 /* @@@ Bad condition: @@ -553,88 +565,52 @@ static bool is_outermost_Start(ir_node *n) { Besides current_ir_graph is not set properly. */ if ((get_irn_op(n) == op_Block) && (n == get_irg_start_block(current_ir_graph))) { - if ((!interprocedural_view) || - (current_ir_graph == outermost_ir_graph)) - return true; + if ((!get_interprocedural_view()) || + (current_ir_graph == outermost_ir_graph)) + return 1; } #endif - return false; + return 0; } -/* Don't walk from nodes to blocks except for Control flow operations. */ +/* When to walk from nodes to blocks. Only for Control flow operations? */ static INLINE int get_start_index(ir_node *n) { - if (is_cfop(n) || is_fragile_op(n) || get_irn_op(n) == op_Start) - return -1; - else +#undef BLOCK_BEFORE_NODE +#define BLOCK_BEFORE_NODE 1 + +#if BLOCK_BEFORE_NODE + + /* This version assures, that all nodes are ordered absolutely. This allows + to undef all nodes in the heap analysis if the block is false, which means + not reachable. + I.e., with this code, the order on the loop tree is correct. But a (single) + test showed the loop tree is deeper. */ + if (get_irn_op(n) == op_Phi || + get_irn_op(n) == op_Block || + (get_irn_op(n) == op_Filter && get_interprocedural_view()) || + (get_irg_pinned(get_irn_irg(n)) == op_pin_state_floats && + get_irn_pinned(n) == op_pin_state_floats)) + // Here we could test for backedge at -1 which is illegal return 0; -} - -/* Returns current_ir_graph and set it to the irg of predecessor index - of node n. */ -static INLINE ir_graph * -switch_irg (ir_node *n, int index) { - ir_graph *old_current = current_ir_graph; - - if (interprocedural_view) { - /* Only Filter and Block nodes can have predecessors in other graphs. */ - if (get_irn_op(n) == op_Filter) - n = get_nodes_Block(n); - if (get_irn_op(n) == op_Block) { - ir_node *cfop = skip_Proj(get_Block_cfgpred(n, index)); - if (is_ip_cfop(cfop)) { - current_ir_graph = get_irn_irg(cfop); - set_irg_visited(current_ir_graph, get_max_irg_visited()); - } - } - } - - return old_current; -} - -#if 0 -/* Walks up the stack passing n and then finding the node - where we walked into the irg n is contained in. - Here we switch the irg. */ -static ir_graph * -find_irg_on_stack (ir_node *n) { - ir_node *m; - ir_graph *old_current = current_ir_graph; - int i; + else + return -1; - if (interprocedural_view) { - for (i = tos; i >= 0; i--) { - if (stack[i] == n) break; - } - if (i < 0) i = tos; +#else - assert (i >= 0); - for (; i >= 0; i--) { - m = stack[i]; - /*printf(" Visiting %d ", i); DDMN(m);*/ - if (is_ip_cfop(m)) { - current_ir_graph = get_irn_irg(m); - break; - } - if (get_irn_op(m) == op_Filter) { - /* Find the corresponding ip_cfop */ - ir_node *pred = stack[i+1]; - int j; - for (j = 0; j < get_Filter_n_cg_preds(m); j++) - if (get_Filter_cg_pred(m, j) == pred) break; - if (j >= get_Filter_n_cg_preds(m)) - /* It is a filter we didn't pass as the predecessors are marked. */ - continue; - assert(get_Filter_cg_pred(m, j) == pred); - switch_irg(m, j); - break; - } - } - } + /* This version causes deeper loop trees (at least we verified this + for Polymor). + But it guarantees that Blocks are analysed before nodes contained in the + block. If so, we can set the value to undef if the block is not \ + executed. */ + if (is_cfop(n) || is_fragile_op(n) || get_irn_op(n) == op_Start) + return -1; + else + return 0; - return old_current; -} #endif +} + #if 0 static void test(ir_node *pred, ir_node *root, ir_node *this) { @@ -656,42 +632,83 @@ static void test(ir_node *pred, ir_node *root, ir_node *this) { #endif /* Test for legal loop header: Block, Phi, ... */ -INLINE static bool is_possible_loop_head(ir_node *n) { - return ((get_irn_op(n) == op_Block) || - (get_irn_op(n) == op_Phi) || - ((get_irn_op(n) == op_Filter) && interprocedural_view)); +static INLINE int is_possible_loop_head(ir_node *n) { + ir_op *op = get_irn_op(n); + return ((op == op_Block) || + (op == op_Phi) || + ((op == op_Filter) && get_interprocedural_view())); } -/* Returns true if n is a loop header, i.e., it is a Block, Phi +/* Returns non-zero if n is a loop header, i.e., it is a Block, Phi or Filter node and has predecessors within the loop and out of the loop. @arg root: only needed for assertion. */ -static bool +static int is_head (ir_node *n, ir_node *root) { - int i; + int i, arity; int some_outof_loop = 0, some_in_loop = 0; /* Test for legal loop header: Block, Phi, ... */ if (!is_possible_loop_head(n)) - return false; + return 0; if (!is_outermost_Start(n)) { - for (i = get_start_index(n); i < get_irn_arity(n); i++) { + arity = get_irn_arity(n); + for (i = get_start_index(n); i < arity; i++) { ir_node *pred = get_irn_n(n, i); assert(pred); if (is_backedge(n, i)) continue; if (!irn_is_in_stack(pred)) { - some_outof_loop = 1; + some_outof_loop = 1; } else { - assert(get_irn_uplink(pred) >= get_irn_uplink(root)); - some_in_loop = 1; + if(get_irn_uplink(pred) < get_irn_uplink(root)) { + DDMN(n); DDMN(pred); DDMN(root); + assert(get_irn_uplink(pred) >= get_irn_uplink(root)); + } + some_in_loop = 1; } } } return some_outof_loop && some_in_loop; } +/** + * Returns non-zero if n is possible loop head of an endless loop. + * I.e., it is a Block, Phi or Filter node and has only predecessors + * within the loop. + * @param root: only needed for assertion. + */ +static int +is_endless_head (ir_node *n, ir_node *root) +{ + int i, arity; + int some_outof_loop = 0, some_in_loop = 0; + + /* Test for legal loop header: Block, Phi, ... */ + if (!is_possible_loop_head(n)) + return 0; + + if (!is_outermost_Start(n)) { + arity = get_irn_arity(n); + for (i = get_start_index(n); i < arity; i++) { + ir_node *pred = get_irn_n(n, i); + assert(pred); + if (is_backedge(n, i)) { continue; } + if (!irn_is_in_stack(pred)) { + some_outof_loop = 1; //printf(" some out of loop "); + } else { + if(get_irn_uplink(pred) < get_irn_uplink(root)) { + DDMN(pred); DDMN(root); + assert(get_irn_uplink(pred) >= get_irn_uplink(root)); + } + some_in_loop = 1; + } + } + } + return !some_outof_loop && some_in_loop; +} + /* Returns index of the predecessor with the smallest dfn number greater-equal than limit. */ static int @@ -700,13 +717,14 @@ smallest_dfn_pred (ir_node *n, int limit) int i, index = -2, min = -1; if (!is_outermost_Start(n)) { - for (i = get_start_index(n); i < get_irn_arity(n); i++) { + int arity = get_irn_arity(n); + for (i = get_start_index(n); i < arity; i++) { ir_node *pred = get_irn_n(n, i); assert(pred); if (is_backedge(n, i) || !irn_is_in_stack(pred)) continue; if (get_irn_dfn(pred) >= limit && (min == -1 || get_irn_dfn(pred) < min)) { - index = i; - min = get_irn_dfn(pred); + index = i; + min = get_irn_dfn(pred); } } } @@ -720,23 +738,27 @@ largest_dfn_pred (ir_node *n) int i, index = -2, max = -1; if (!is_outermost_Start(n)) { - for (i = get_start_index(n); i < get_irn_arity(n); i++) { + int arity = get_irn_arity(n); + for (i = get_start_index(n); i < arity; i++) { ir_node *pred = get_irn_n(n, i); if (is_backedge (n, i) || !irn_is_in_stack(pred)) continue; if (get_irn_dfn(pred) > max) { - index = i; - max = get_irn_dfn(pred); + index = i; + max = get_irn_dfn(pred); } } } return index; } -/* Searches the stack for possible loop heads. Tests these for backedges. - If it finds a head with an unmarked backedge it marks this edge and - returns the tail of the loop. - If it finds no backedge returns NULL. - ("disable_backedge" in fiasco) */ +/** Searches the stack for possible loop heads. Tests these for backedges. + If it finds a head with an unmarked backedge it marks this edge and + returns the tail of the loop. + If it finds no backedge returns NULL. + ("disable_backedge" in fiasco) +* +* @param n A node where uplink == dfn. +**/ static ir_node * find_tail (ir_node *n) { @@ -746,24 +768,62 @@ find_tail (ir_node *n) { /* if (!icfg && rm_cyclic_phis && remove_cyclic_phis (n)) return NULL; */ - m = stack[tos-1]; /* tos = top of stack */ if (is_head (m, n)) { res_index = smallest_dfn_pred(m, 0); if ((res_index == -2) && /* no smallest dfn pred found. */ - (n == m)) + (n == m)) return NULL; } else { - if (m == n) return NULL; - for (i = tos-2; ; --i) { + if (m == n) return NULL; // Is this to catch Phi - self loops? + for (i = tos-2; i >= 0; --i) { m = stack[i]; + if (is_head (m, n)) { - res_index = smallest_dfn_pred (m, get_irn_dfn(m) + 1); - if (res_index == -2) /* no smallest dfn pred found. */ - res_index = largest_dfn_pred (m); - break; + res_index = smallest_dfn_pred (m, get_irn_dfn(m) + 1); + if (res_index == -2) /* no smallest dfn pred found. */ + res_index = largest_dfn_pred (m); + + if ((m == n) && (res_index == -2)) { /* dont walk past loop head. */ + i = -1; + } + break; + } + + /* We should not walk past our selves on the stack: The upcoming nodes + are not in this loop. We assume a loop not reachable from Start. */ + if (m == n) { + i = -1; + break; } + } + + if (i < 0) { + /* A dead loop not reachable from Start. */ + for (i = tos-2; i >= 0; --i) { + m = stack[i]; + if (is_endless_head (m, n)) { + res_index = smallest_dfn_pred (m, get_irn_dfn(m) + 1); + if (res_index == -2) /* no smallest dfn pred found. */ + res_index = largest_dfn_pred (m); + break; + } + if (m == n) { break; } /* It's not an unreachable loop, either. */ + } + //assert(0 && "no head found on stack"); + } + + } + if (res_index <= -2) { + /* It's a completely bad loop: without Phi/Block nodes that can + be a head. I.e., the code is "dying". We break the loop by + setting Bad nodes. */ + int arity = get_irn_arity(n); + for (i = -1; i < arity; ++i) { + set_irn_n(n, i, get_irg_bad(get_irn_irg(n))); + } + return NULL; } assert (res_index > -2); @@ -772,12 +832,95 @@ find_tail (ir_node *n) { } -/* The core algorithm. *****************************************/ +#if EXPERIMENTAL_LOOP_TREE + +/* ---------------------------------------------------------------- + AS: This is experimental code to build loop trees suitable for + the heap analysis. Does not work correctly right now... :-( + + + Search in stack for the corresponding first Call-End-ProjX that + corresponds to one of the control flow predecessors of the given + block, that is the possible callers. + returns: the control predecessor to chose\ + or -1 if no corresponding Call-End-Node could be found + on the stack. + - -------------------------------------------------------------- */ + +int search_endproj_in_stack(ir_node *start_block) +{ + int i, j; + assert(is_Block(start_block)); + for(i = tos - 1; i >= 0; --i) + { + DDMN(stack[i]); + if(get_irn_op(stack[i]) == op_Proj && get_irn_mode(stack[i]) == mode_X && + get_irn_op(get_irn_n(stack[i], 0)) == op_EndReg) + { + printf("FOUND PROJ!!!!!!!!!!!!!!!!!!!!!!!!!!\n"); + ir_node *end_projx = stack[i]; + + int arity = get_irn_arity(start_block); + for(j = 0; j < arity; j++) + { + ir_node *begin_projx = get_Block_cfgpred(get_irg_start_block(get_irn_irg(end_projx)), + get_Proj_proj(end_projx)); + DDMN(begin_projx); + if(get_irn_n(start_block, j) == begin_projx) + { + printf("FOUND IT!!!!!!!!!!!!!!!!!!\n"); + return(j); + } + } + } + } + return(-1); +} + + +static pmap *projx_link = NULL; + +void link_to_reg_end (ir_node *n, void *env) { + if(get_irn_op(n) == op_Proj && + get_irn_mode(n) == mode_X && + get_irn_op(get_irn_n(n, 0)) == op_EndReg) { + /* Reg End Projx -> Find the CallBegin Projx and hash it */ + ir_node *end_projx = n; + ir_node *begin_projx = get_Block_cfgpred(get_irg_start_block(get_irn_irg(end_projx)), + get_Proj_proj(end_projx)); + printf("Linked the following ProjxNodes:\n"); + DDMN(begin_projx); + DDMN(end_projx); + set_projx_link(begin_projx, end_projx); + } +} + +void set_projx_link(ir_node *cb_projx, ir_node *end_projx) +{ + if(projx_link == NULL) + projx_link = pmap_create(); + pmap_insert(projx_link, (void *)cb_projx, (void *)end_projx); +} + +ir_node *get_projx_link(ir_node *cb_projx) +{ + return((ir_node *) pmap_get(projx_link, (void *)cb_projx)); +} + +#endif + +static INLINE int +is_outermost_loop(ir_loop *l) { + return l == get_loop_outer_loop(l); +} + + +/*-----------------------------------------------------------* + * The core algorithm. * + *-----------------------------------------------------------*/ static void scc (ir_node *n) { int i; - // GL @@@ remove experimental stuff ir_graph *rem; - if (irn_visited(n)) return; mark_irn_visited(n); @@ -786,11 +929,6 @@ static void scc (ir_node *n) { set_irn_uplink(n, current_dfn); /* ... is default uplink. */ set_irn_loop(n, NULL); current_dfn ++; - - /* What's this good for? - n->ana.scc.section = NULL; - */ - push(n); /* AS: get_start_index might return -1 for Control Flow Nodes, and thus a negative @@ -798,48 +936,159 @@ static void scc (ir_node *n) { so is_backedge does not access array[-1] but correctly returns false! */ if (!is_outermost_Start(n)) { - for (i = get_start_index(n); i < get_irn_arity(n); i++) { + int arity = get_irn_arity(n); + + for (i = get_start_index(n); i < arity; i++) { ir_node *m; if (is_backedge(n, i)) continue; - - m = get_irn_n(n, i); /*get_irn_ip_pred(n, i);*/ - assert(m); - //if ((!m) || (get_irn_op(m) == op_Unknown)) continue; + m = get_irn_n(n, i); /* get_irn_ip_pred(n, i); */ + /* if ((!m) || (get_irn_op(m) == op_Unknown)) continue; */ scc (m); - // GL @@@ remove experimental stuff /*return_recur(n, i);*/ + if (irn_is_in_stack(m)) { + /* Uplink of m is smaller if n->m is a backedge. + Propagate the uplink to mark the loop. */ + if (get_irn_uplink(m) < get_irn_uplink(n)) + set_irn_uplink(n, get_irn_uplink(m)); + } + } + } + + if (get_irn_dfn(n) == get_irn_uplink(n)) { + /* This condition holds for + 1) the node with the incoming backedge. + That is: We found a loop! + 2) Straight line code, because no uplink has been propagated, so the + uplink still is the same as the dfn. + + But n might not be a proper loop head for the analysis. Proper loop + heads are Block and Phi nodes. find_tail searches the stack for + Block's and Phi's and takes those nodes as loop heads for the current + loop instead and marks the incoming edge as backedge. */ + + ir_node *tail = find_tail(n); + if (tail) { + /* We have a loop, that is no straight line code, + because we found a loop head! + Next actions: Open a new loop on the loop tree and + try to find inner loops */ +#if NO_LOOPS_WITHOUT_HEAD + /* This is an adaption of the algorithm from fiasco / optscc to + * avoid loops without Block or Phi as first node. This should + * severely reduce the number of evaluations of nodes to detect + * a fixpoint in the heap analysis. + * Further it avoids loops without firm nodes that cause errors + * in the heap analyses. + * But attention: don't do it for the outermost loop: This loop + * is not iterated. A first block can be a loop head in case of + * an endless recursion. */ + + ir_loop *l; + int close; + if ((get_loop_n_elements(current_loop) > 0) || (is_outermost_loop(current_loop))) { + l = new_loop(); + close = 1; + } else { + l = current_loop; + close = 0; + } +#else + ir_loop *l = new_loop(); +#endif + + /* Remove the loop from the stack ... */ + pop_scc_unmark_visit (n); + + /* The current backedge has been marked, that is temporarily eliminated, + by find tail. Start the scc algorithm + anew on the subgraph that is left (the current loop without the backedge) + in order to find more inner loops. */ + scc (tail); + + assert (irn_visited(n)); +#if NO_LOOPS_WITHOUT_HEAD + if (close) +#endif + close_loop(l); + } + else + { + /* No loop head was found, that is we have straightline code. + Pop all nodes from the stack to the current loop. */ + pop_scc_to_loop(n); + } + } +} + +static void my_scc (ir_node *n) { + int i; + if (irn_visited(n)) return; + mark_irn_visited(n); + + /* Initialize the node */ + set_irn_dfn(n, current_dfn); /* Depth first number for this node */ + set_irn_uplink(n, current_dfn); /* ... is default uplink. */ + set_irn_loop(n, NULL); + current_dfn ++; + push(n); + + /* AS: get_start_index might return -1 for Control Flow Nodes, and thus a negative + array index would be passed to is_backedge(). But CFG Nodes dont't have a backedge array, + so is_backedge does not access array[-1] but correctly returns false! */ + + if (!is_outermost_Start(n)) { + int arity = get_irn_arity(n); + + for (i = get_start_index(n); i < arity; i++) { + ir_node *m; + if (is_backedge(n, i)) continue; + m = get_irn_n(n, i); /* get_irn_ip_pred(n, i); */ + /* if ((!m) || (get_irn_op(m) == op_Unknown)) continue; */ + my_scc (m); if (irn_is_in_stack(m)) { - /* Uplink of m is smaller if n->m is a backedge. - Propagate the uplink to mark the loop. */ - if (get_irn_uplink(m) < get_irn_uplink(n)) - set_irn_uplink(n, get_irn_uplink(m)); + /* Uplink of m is smaller if n->m is a backedge. + Propagate the uplink to mark the loop. */ + if (get_irn_uplink(m) < get_irn_uplink(n)) + set_irn_uplink(n, get_irn_uplink(m)); } } } if (get_irn_dfn(n) == get_irn_uplink(n)) { - /* This condition holds for the node with the incoming backedge. - AS: That is: For the loop head. */ + /* This condition holds for + 1) the node with the incoming backedge. + That is: We found a loop! + 2) Straight line code, because no uplink has been propagated, so the + uplink still is the same as the dfn. + + But n might not be a proper loop head for the analysis. Proper loop + heads are Block and Phi nodes. find_tail searches the stack for + Block's and Phi's and takes those nodes as loop heads for the current + loop instead and marks the incoming edge as backedge. */ + ir_node *tail = find_tail(n); if (tail) { - /* We found a new inner loop! */ + /* We have a loop, that is no straight line code, + because we found a loop head! + Next actions: Open a new loop on the loop tree and + try to find inner loops */ +#if NO_LOOPS_WITHOUT_HEAD /* This is an adaption of the algorithm from fiasco / optscc to * avoid loops without Block or Phi as first node. This should * severely reduce the number of evaluations of nodes to detect - * a fixpoint in the heap analyses. + * a fixpoint in the heap analysis. * Further it avoids loops without firm nodes that cause errors * in the heap analyses. */ -#define NO_LOOPS_WITHOUT_HEAD 1 -#if NO_LOOPS_WITHOUT_HEAD + ir_loop *l; int close; - if (get_loop_n_elements(current_loop) > 0) { - l = new_loop(); - close = 1; + if ((get_loop_n_elements(current_loop) > 0) || (is_outermost_loop(current_loop))) { + l = new_loop(); + close = 1; } else { - l = current_loop; - close = 0; + l = current_loop; + close = 0; } #else ir_loop *l = new_loop(); @@ -847,21 +1096,23 @@ static void scc (ir_node *n) { /* Remove the loop from the stack ... */ pop_scc_unmark_visit (n); - /* and recompute it in a better order; and so that it goes into - the new loop. */ - // GL @@@ remove experimental stuff rem = find_irg_on_stack(tail); - scc (tail); - // GL @@@ remove experimental stuff current_ir_graph = rem; + /* The current backedge has been marked, that is temporarily eliminated, + by find tail. Start the scc algorithm + anew on the subgraph that is left (the current loop without the backedge) + in order to find more inner loops. */ + my_scc (tail); assert (irn_visited(n)); #if NO_LOOPS_WITHOUT_HEAD if (close) #endif - close_loop(l); - } else { - /* AS: No inner loop was found. Pop all nodes from the stack - to the current loop. */ + close_loop(l); + } + else + { + /* No loop head was found, that is we have straightline code. + Pop all nodes from the stack to the current loop. */ pop_scc_to_loop(n); } } @@ -869,14 +1120,15 @@ static void scc (ir_node *n) { /* Constructs backedge information for irg. In interprocedural view constructs backedges for all methods called by irg, too. */ -void construct_backedges(ir_graph *irg) { +int construct_backedges(ir_graph *irg) { ir_graph *rem = current_ir_graph; ir_loop *head_rem; - assert(!interprocedural_view && - "not implemented, use construct_ip_backedges"); + assert(!get_interprocedural_view() && + "not implemented, use construct_ip_backedges"); - current_ir_graph = irg; + max_loop_depth = 0; + current_ir_graph = irg; outermost_ir_graph = irg; init_scc(current_ir_graph); @@ -885,19 +1137,13 @@ void construct_backedges(ir_graph *irg) { new_loop(); /* sets current_loop */ head_rem = current_loop; /* Just for assertion */ - if (interprocedural_view) { - set_irg_visited(current_ir_graph, inc_max_irg_visited()); - init_ip_walk (); - } else { - inc_irg_visited(current_ir_graph); - } + inc_irg_visited(current_ir_graph); scc(get_irg_end(current_ir_graph)); - if (interprocedural_view) finish_ip_walk(); - assert(head_rem == current_loop); set_irg_loop(current_ir_graph, current_loop); + set_irg_loopinfo_state(current_ir_graph, loopinfo_consistent); assert(get_irg_loop(current_ir_graph)->kind == k_ir_loop); /* irg->loops = current_loop; @@ -909,14 +1155,18 @@ void construct_backedges(ir_graph *irg) { } */ current_ir_graph = rem; + + return max_loop_depth; } -#if 0 -void construct_ip_backedges (void) { +int construct_ip_backedges (void) { ir_graph *rem = current_ir_graph; - int rem_ipv = interprocedural_view; - int i, j; + int rem_ipv = get_interprocedural_view(); + int i; + + max_loop_depth = 0; + assert(get_irp_ip_view_state() == ip_view_valid); outermost_ir_graph = get_irp_main_irg(); @@ -924,46 +1174,76 @@ void construct_ip_backedges (void) { current_loop = NULL; new_loop(); /* sets current_loop */ - interprocedural_view = 1; + set_interprocedural_view(1); inc_max_irg_visited(); for (i = 0; i < get_irp_n_irgs(); i++) set_irg_visited(get_irp_irg(i), get_max_irg_visited()); + /** We have to start the walk at the same nodes as cg_walk. **/ + /* Walk starting at unreachable procedures. Only these + * have End blocks visible in interprocedural view. */ for (i = 0; i < get_irp_n_irgs(); i++) { ir_node *sb; current_ir_graph = get_irp_irg(i); - /* Find real entry points */ + sb = get_irg_start_block(current_ir_graph); + if ((get_Block_n_cfgpreds(sb) > 1) || - (get_nodes_Block(get_Block_cfgpred(sb, 0)) != sb)) continue; - /* Compute scc for this graph */ - outermost_ir_graph = current_ir_graph; - set_irg_visited(outermost_ir_graph, get_max_irg_visited()); + (get_nodes_block(get_Block_cfgpred(sb, 0)) != sb)) continue; + scc(get_irg_end(current_ir_graph)); - for (j = 0; j < get_End_n_keepalives(get_irg_end(outermost_ir_graph)); j++) - scc(get_End_keepalive(get_irg_end(outermost_ir_graph), j)); + } + + /* Check whether we walked all procedures: there could be procedures + with cyclic calls but no call from the outside. */ + for (i = 0; i < get_irp_n_irgs(); i++) { + ir_node *sb; + current_ir_graph = get_irp_irg(i); + + /* Test start block: if inner procedure end and end block are not + * visible and therefore not marked. */ + sb = get_irg_start_block(current_ir_graph); + if (get_irn_visited(sb) < get_irg_visited(current_ir_graph)) scc(sb); + } + + /* Walk all endless loops in inner procedures. + * We recognize an inner procedure if the End node is not visited. */ + for (i = 0; i < get_irp_n_irgs(); i++) { + ir_node *e; + current_ir_graph = get_irp_irg(i); + + e = get_irg_end(current_ir_graph); + if (get_irn_visited(e) < get_irg_visited(current_ir_graph)) { + int j; + /* Don't visit the End node. */ + for (j = 0; j < get_End_n_keepalives(e); j++) scc(get_End_keepalive(e, j)); + } } set_irg_loop(outermost_ir_graph, current_loop); + set_irg_loopinfo_state(current_ir_graph, loopinfo_ip_consistent); assert(get_irg_loop(outermost_ir_graph)->kind == k_ir_loop); current_ir_graph = rem; - interprocedural_view = rem_ipv; + set_interprocedural_view(rem_ipv); + return max_loop_depth; } -#else -void construct_ip_backedges (void) { + +void my_construct_ip_backedges (void) { ir_graph *rem = current_ir_graph; - int rem_ipv = interprocedural_view; + int rem_ipv = get_interprocedural_view(); int i; + assert(get_irp_ip_view_state() == ip_view_valid); + outermost_ir_graph = get_irp_main_irg(); init_ip_scc(); current_loop = NULL; new_loop(); /* sets current_loop */ - interprocedural_view = 1; + set_interprocedural_view(1); inc_max_irg_visited(); for (i = 0; i < get_irp_n_irgs(); i++) @@ -979,9 +1259,9 @@ void construct_ip_backedges (void) { sb = get_irg_start_block(current_ir_graph); if ((get_Block_n_cfgpreds(sb) > 1) || - (get_nodes_block(get_Block_cfgpred(sb, 0)) != sb)) continue; + (get_nodes_block(get_Block_cfgpred(sb, 0)) != sb)) continue; - scc(get_irg_end(current_ir_graph)); + my_scc(get_irg_end(current_ir_graph)); } /* Check whether we walked all procedures: there could be procedures @@ -1011,52 +1291,67 @@ void construct_ip_backedges (void) { } set_irg_loop(outermost_ir_graph, current_loop); + set_irg_loopinfo_state(current_ir_graph, loopinfo_ip_consistent); assert(get_irg_loop(outermost_ir_graph)->kind == k_ir_loop); current_ir_graph = rem; - interprocedural_view = rem_ipv; + set_interprocedural_view(rem_ipv); } -#endif static void reset_backedges(ir_node *n) { if (is_possible_loop_head(n)) { - int rem = interprocedural_view; - interprocedural_view = 1; + int rem = get_interprocedural_view(); + + set_interprocedural_view(1); clear_backedges(n); - interprocedural_view = 0; + set_interprocedural_view(1); clear_backedges(n); - interprocedural_view = rem; + set_interprocedural_view(rem); } } + +/* static void loop_reset_backedges(ir_loop *l) { int i; reset_backedges(get_loop_node(l, 0)); + for (i = 0; i < get_loop_n_nodes(l); ++i) + set_irn_loop(get_loop_node(l, i), NULL); for (i = 0; i < get_loop_n_sons(l); ++i) { loop_reset_backedges(get_loop_son(l, i)); } } +*/ + +static void loop_reset_node(ir_node *n, void *env) { + set_irn_loop(n, NULL); + reset_backedges(n); +} + /** Removes all loop information. Resets all backedges */ void free_loop_information(ir_graph *irg) { - if (get_irg_loop(irg)) - loop_reset_backedges(get_irg_loop(irg)); + /* We can not use this recursion, as the loop might contain + illegal nodes by now. Why else would we throw away the + representation? + if (get_irg_loop(irg)) loop_reset_backedges(get_irg_loop(irg)); + */ + irg_walk_graph(irg, loop_reset_node, NULL, NULL); set_irg_loop(irg, NULL); + set_irg_loopinfo_state(current_ir_graph, loopinfo_none); /* We cannot free the loop nodes, they are on the obstack. */ } void free_all_loop_information (void) { int i; - int rem = interprocedural_view; - interprocedural_view = 1; /* To visit all filter nodes */ + int rem = get_interprocedural_view(); + set_interprocedural_view(1); /* To visit all filter nodes */ for (i = 0; i < get_irp_n_irgs(); i++) { free_loop_information(get_irp_irg(i)); } - pmap_destroy(node_loop_map); - node_loop_map = NULL; - interprocedural_view = rem; + set_interprocedural_view(rem); } @@ -1118,9 +1413,6 @@ static int test_loop_node(ir_loop *l) { dump_loop(l, "-ha"); } - if (get_loop_loop_nr(l) == 11819) - dump_loop(l, "-ha-debug"); - return found_problem; } @@ -1136,3 +1428,37 @@ void find_strange_loop_nodes(ir_loop *l) { if (found_problem) exit(0); } + +/* ------------------------------------------------------------------- */ +/* Simple analyses based on the loop information */ +/* ------------------------------------------------------------------- */ + +int is_loop_variant(ir_loop *l, ir_loop *b) { + int i, n_elems; + + if (l == b) return 1; + + n_elems = get_loop_n_elements(l); + for (i = 0; i < n_elems; ++i) { + loop_element e = get_loop_element(l, i); + if (is_ir_loop(e.kind)) + if (is_loop_variant(e.son, b)) + return 1; + } + + return 0; +} + +/* Test whether a value is loop invariant. + * + * @param n The node to be tested. + * @param block A block node. We pass the block, not the loop as we must + * start off with a block loop to find all proper uses. + * + * Returns non-zero, if the node n is not changed in the loop block + * belongs to or in inner loops of this blocks loop. */ +int is_loop_invariant(ir_node *n, ir_node *block) { + ir_loop *l = get_irn_loop(block); + ir_node *b = (is_Block(n)) ? n : get_nodes_block(n); + return !is_loop_variant(l, get_irn_loop(b)); +}