X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Ffunccall.c;h=8b0554f1efb5b0e8c5d5ace1c0ed38bdbfd1d40b;hb=abbf9492df1993d211412ba520feb81de0688c2a;hp=9941e222670c0544114bbc0bdb69d2e6433089e7;hpb=ff3ff6bed9ba9dadcd44250ade97aa19d49545db;p=libfirm diff --git a/ir/opt/funccall.c b/ir/opt/funccall.c index 9941e2226..8b0554f1e 100644 --- a/ir/opt/funccall.c +++ b/ir/opt/funccall.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -23,11 +23,11 @@ * @author Michael Beck * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif -#include "funccall_t.h" +#include + +#include "opt_init.h" #include "irnode_t.h" #include "irgraph_t.h" @@ -36,32 +36,56 @@ #include "irvrfy.h" #include "dbginfo_t.h" #include "irflag_t.h" +#include "irloop_t.h" #include "ircons.h" +#include "iredges_t.h" +#include "irpass_t.h" +#include "iroptimize.h" +#include "analyze_irg_args.h" #include "irhooks.h" #include "debug.h" DEBUG_ONLY(static firm_dbg_module_t *dbg;) /** - * The walker environment for rem_mem_from_const_fkt_calls + * The walker environment for updating function calls. */ typedef struct _env_t { unsigned n_calls_SymConst; unsigned n_calls_Sel; - ir_node *const_call_list; /**< The list of all const function calls that will be changed. */ - ir_node *pure_call_list; /**< The list of all pure function calls that will be changed. */ - ir_node *proj_list; /**< The list of all potential Proj nodes that must be fixed. */ + ir_node *float_const_call_list; /**< The list of all floating const function calls that will be changed. */ + ir_node *nonfloat_const_call_list; /**< The list of all non-floating const function calls that will be changed. */ + ir_node *pure_call_list; /**< The list of all pure function calls that will be changed. */ + ir_node *nothrow_call_list; /**< The list of all nothrow function calls that will be changed. */ + ir_node *proj_list; /**< The list of all potential Proj nodes that must be fixed. */ } env_t; +/** If non-null, evaluates entities for being a heap alloc. */ +static check_alloc_entity_func is_alloc_entity = NULL; + +/** Ready IRG's are marked in the ready set. */ +static unsigned *ready_set; + +/** IRG's that are in progress are marked here. */ +static unsigned *busy_set; + /** - * Collect all calls to const and pure functions + * We misuse the mtp_property_inherited flag as temporary here. + * The is ok, as we cannot set or get it anyway using the + * get_addtional_properties API. + */ +#define mtp_temporary mtp_property_inherited + +/** + * Walker: Collect all calls to const and pure functions * to lists. Collect all Proj(Call) nodes into a Proj list. */ -static void collect_calls(ir_node *node, void *env) { - env_t *ctx = env; - ir_node *call, *ptr; +static void collect_const_and_pure_calls(ir_node *node, void *env) +{ + env_t *ctx = env; + ir_node *call, *ptr; ir_entity *ent; - unsigned mode; + unsigned and_prop, or_prop, prop; if (is_Call(node)) { call = node; @@ -69,11 +93,11 @@ static void collect_calls(ir_node *node, void *env) { /* set the link to NULL for all non-const/pure calls */ set_irn_link(call, NULL); ptr = get_Call_ptr(call); - if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) { - ent = get_SymConst_entity(ptr); + if (is_Global(ptr)) { + ent = get_Global_entity(ptr); - mode = get_entity_additional_properties(ent); - if ((mode & (mtp_property_const|mtp_property_pure)) == 0) + prop = get_entity_additional_properties(ent); + if ((prop & (mtp_property_const|mtp_property_pure)) == 0) return; ++ctx->n_calls_SymConst; } else if (get_opt_closed_world() && @@ -89,28 +113,37 @@ static void collect_calls(ir_node *node, void *env) { } /* note that const function are a subset of pure ones */ - mode = mtp_property_const | mtp_property_pure; + and_prop = mtp_property_const | mtp_property_pure; + or_prop = 0; for (i = 0; i < n_callees; ++i) { ent = get_Call_callee(call, i); if (ent == unknown_entity) { /* we don't know which entity is called here */ return; } - mode &= get_entity_additional_properties(ent); - if (mode == 0) + prop = get_entity_additional_properties(ent); + and_prop &= prop; + or_prop &= prop; + if (and_prop == mtp_no_property) return; } + prop = and_prop | (or_prop & mtp_property_has_loop); ++ctx->n_calls_Sel; } else return; /* ok, if we get here we found a call to a const or a pure function */ - if (mode & mtp_property_pure) { + if (prop & mtp_property_pure) { set_irn_link(call, ctx->pure_call_list); ctx->pure_call_list = call; } else { - set_irn_link(call, ctx->const_call_list); - ctx->const_call_list = call; + if (prop & mtp_property_has_loop) { + set_irn_link(call, ctx->nonfloat_const_call_list); + ctx->nonfloat_const_call_list = call; + } else { + set_irn_link(call, ctx->float_const_call_list); + ctx->float_const_call_list = call; + } } } else if (is_Proj(node)) { /* @@ -123,9 +156,9 @@ static void collect_calls(ir_node *node, void *env) { /* collect the Proj's in the Proj list */ switch (get_Proj_proj(node)) { - case pn_Call_M_regular: + case pn_Call_M: case pn_Call_X_except: - case pn_Call_M_except: + case pn_Call_X_regular: set_irn_link(node, ctx->proj_list); ctx->proj_list = node; break; @@ -133,25 +166,26 @@ static void collect_calls(ir_node *node, void *env) { break; } } -} /* collect_calls */ +} /* collect_const_and_pure_calls */ /** * Fix the list of collected Calls. * - * @param irg the graph that contained calls to pure functions - * @param call_list the list of all call sites of const functions - * @param proj_list the list of all memory/exception Proj's of this call sites + * @param irg the graph that contained calls to pure functions + * @param ctx context */ -static void fix_const_call_list(ir_graph *irg, ir_node *call_list, ir_node *proj_list) { +static void fix_const_call_lists(ir_graph *irg, env_t *ctx) +{ ir_node *call, *next, *mem, *proj; int exc_changed = 0; ir_graph *rem = current_ir_graph; current_ir_graph = irg; - /* First step: fix all calls by removing it's memory input. - It's original memory input is preserved in their link fields. */ - for (call = call_list; call; call = next) { + /* First step: fix all calls by removing their memory input and let + * them floating. + * The original memory input is preserved in their link fields. */ + for (call = ctx->float_const_call_list; call != NULL; call = next) { next = get_irn_link(call); mem = get_Call_mem(call); @@ -159,7 +193,7 @@ static void fix_const_call_list(ir_graph *irg, ir_node *call_list, ir_node *proj set_Call_mem(call, get_irg_no_mem(irg)); /* - * Sorrily we cannot simply set the node to 'float'. + * Unfortunately we cannot simply set the node to 'float'. * There is a reason for that: * * - The call might be inside a loop/if that is NOT entered @@ -173,33 +207,39 @@ static void fix_const_call_list(ir_graph *irg, ir_node *call_list, ir_node *proj * observable states... */ - /* finally, this call can float - set_irn_pinned(call, op_pin_state_floats); */ + /* finally, this call can float */ + set_irn_pinned(call, op_pin_state_floats); hook_func_call(irg, call); } - /* Second step: fix all Proj's */ - for (proj = proj_list; proj; proj = next) { + /* Last step: fix all Proj's */ + for (proj = ctx->proj_list; proj != NULL; proj = next) { next = get_irn_link(proj); call = get_Proj_pred(proj); mem = get_irn_link(call); /* beware of calls in the pure call list */ - if (! mem || get_irn_op(mem) == op_Call) + if (!mem || is_Call(mem)) continue; assert(get_irn_mode(mem) == mode_M); switch (get_Proj_proj(proj)) { - case pn_Call_M_regular: { + case pn_Call_M: { /* in dead code there might be cycles where proj == mem */ if (proj != mem) exchange(proj, mem); - } break; + break; + } case pn_Call_X_except: - case pn_Call_M_except: exc_changed = 1; exchange(proj, get_irg_bad(irg)); break; + case pn_Call_X_regular: { + ir_node *block = get_nodes_block(call); + exc_changed = 1; + exchange(proj, new_r_Jmp(block)); + break; + } default: ; } @@ -214,40 +254,193 @@ static void fix_const_call_list(ir_graph *irg, ir_node *call_list, ir_node *proj set_irg_doms_inconsistent(irg); } current_ir_graph = rem; -} /* fix_call_list */ +} /* fix_const_call_list */ + +/** + * Walker: Collect all calls to nothrow functions + * to lists. Collect all Proj(Call) nodes into a Proj list. + */ +static void collect_nothrow_calls(ir_node *node, void *env) +{ + env_t *ctx = env; + ir_node *call, *ptr; + ir_entity *ent; + unsigned prop; + + if (is_Call(node)) { + call = node; + + /* set the link to NULL for all non-const/pure calls */ + set_irn_link(call, NULL); + ptr = get_Call_ptr(call); + if (is_Global(ptr)) { + ent = get_Global_entity(ptr); -/* a marker */ -static char _mark; -#define MARK &_mark + prop = get_entity_additional_properties(ent); + if ((prop & mtp_property_nothrow) == 0) + return; + ++ctx->n_calls_SymConst; + } else if (get_opt_closed_world() && + is_Sel(ptr) && + get_irg_callee_info_state(current_ir_graph) == irg_callee_info_consistent) { + /* If all possible callees are nothrow functions, we can remove the exception edge. */ + int i, n_callees = get_Call_n_callees(call); + if (n_callees == 0) { + /* This is kind of strange: dying code or a Call that will raise an exception + when executed as there is no implementation to call. So better not + optimize. */ + return; + } -#define UNMARK_IRG(irg) set_irg_link((irg), NULL) -#define MARK_IRG(irg) set_irg_link((irg), MARK) -#define IS_IRG_MARKED(irg) (get_irg_link(irg) == MARK) + /* note that const function are a subset of pure ones */ + prop = mtp_property_nothrow; + for (i = 0; i < n_callees; ++i) { + ent = get_Call_callee(call, i); + if (ent == unknown_entity) { + /* we don't know which entity is called here */ + return; + } + prop &= get_entity_additional_properties(ent); + if (prop == mtp_no_property) + return; + } + ++ctx->n_calls_Sel; + } else + return; + + /* ok, if we get here we found a call to a nothrow function */ + set_irn_link(call, ctx->nothrow_call_list); + ctx->nothrow_call_list = call; + } else if (is_Proj(node)) { + /* + * Collect all memory and exception Proj's from + * calls. + */ + call = get_Proj_pred(node); + if (! is_Call(call)) + return; + + /* collect the Proj's in the Proj list */ + switch (get_Proj_proj(node)) { + case pn_Call_M: + case pn_Call_X_except: + case pn_Call_X_regular: + set_irn_link(node, ctx->proj_list); + ctx->proj_list = node; + break; + default: + break; + } + } +} /* collect_nothrow_calls */ + +/** + * Fix the list of collected nothrow Calls. + * + * @param irg the graph that contained calls to pure functions + * @param call_list the list of all call sites of const functions + * @param proj_list the list of all memory/exception Proj's of this call sites + */ +static void fix_nothrow_call_list(ir_graph *irg, ir_node *call_list, ir_node *proj_list) +{ + ir_node *call, *next, *proj; + int exc_changed = 0; + ir_graph *rem = current_ir_graph; + + current_ir_graph = irg; + + /* First step: go through the list of calls and mark them. */ + for (call = call_list; call; call = next) { + next = get_irn_link(call); + + /* current_ir_graph is in memory anyway, so it's a good marker */ + set_irn_link(call, ¤t_ir_graph); + hook_func_call(irg, call); + } + + /* Second step: Remove all exception Proj's */ + for (proj = proj_list; proj; proj = next) { + next = get_irn_link(proj); + call = get_Proj_pred(proj); + + /* handle only marked calls */ + if (get_irn_link(call) != ¤t_ir_graph) + continue; + + /* kill any exception flow */ + switch (get_Proj_proj(proj)) { + case pn_Call_X_except: + exc_changed = 1; + exchange(proj, get_irg_bad(irg)); + break; + case pn_Call_X_regular: { + ir_node *block = get_nodes_block(call); + exc_changed = 1; + exchange(proj, new_r_Jmp(block)); + break; + } + default: + ; + } + } + + /* changes were done ... */ + set_irg_outs_inconsistent(irg); + set_irg_loopinfo_state(irg, loopinfo_cf_inconsistent); + + if (exc_changed) { + /* ... including exception edges */ + set_irg_doms_inconsistent(irg); + } + current_ir_graph = rem; +} /* fix_nothrow_call_list */ + +/* marking */ +#define SET_IRG_READY(irg) rbitset_set(ready_set, get_irg_idx(irg)) +#define IS_IRG_READY(irg) rbitset_is_set(ready_set, get_irg_idx(irg)) +#define SET_IRG_BUSY(irg) rbitset_set(busy_set, get_irg_idx(irg)) +#define CLEAR_IRG_BUSY(irg) rbitset_clear(busy_set, get_irg_idx(irg)) +#define IS_IRG_BUSY(irg) rbitset_is_set(busy_set, get_irg_idx(irg)) /* forward */ -static unsigned check_const_or_pure_function(ir_graph *irg); +static unsigned check_const_or_pure_function(ir_graph *irg, int top); +static unsigned check_nothrow_or_malloc(ir_graph *irg, int top); + +/** + * Calculate the bigger property of two. Handle the temporary flag right. + */ +static unsigned max_property(unsigned a, unsigned b) +{ + unsigned r, t = (a | b) & mtp_temporary; + a &= ~mtp_temporary; + b &= ~mtp_temporary; -#define UMAX(a,b) (a) > (b) ? (a) : (b) + if (a == mtp_no_property || b == mtp_no_property) + return mtp_no_property; + r = a > b ? a : b; + return r | t; +} /* max_property */ /** * Follow the memory chain starting at node and determine * the mtp_property. * * @return mtp_property_const if only calls of const functions are detected - * mtp_property_pure if only Loads and const/pure - * calls detected - * bad_property else + * mtp_property_pure if only Loads and const/pure calls detected + * mtp_no_property else */ -static unsigned _follow_mem(ir_node *node) { +static unsigned _follow_mem(ir_node *node) +{ unsigned m, mode = mtp_property_const; ir_node *ptr; int i; for (;;) { - if (irn_visited(node)) - return mode; + if (mode == mtp_no_property) + return mtp_no_property; - mark_irn_visited(node); + if (irn_visited_else_mark(node)) + return mode; switch (get_irn_opcode(node)) { case iro_Proj: @@ -260,117 +453,122 @@ static unsigned _follow_mem(ir_node *node) { case iro_Phi: case iro_Sync: + /* do a dfs search */ for (i = get_irn_arity(node) - 1; i >= 0; --i) { - mode &= _follow_mem(get_irn_n(node, i)); + m = _follow_mem(get_irn_n(node, i)); + mode = max_property(mode, m); + if (mode == mtp_no_property) + return mtp_no_property; } - break; + return mode; case iro_Load: - /* Beware volatile Loads are NOT allowed in pure functions */ + /* Beware volatile Loads are NOT allowed in pure functions. */ if (get_Load_volatility(node) == volatility_is_volatile) - return 0; - mode = mtp_property_pure; + return mtp_no_property; + mode = max_property(mode, mtp_property_pure); node = get_Load_mem(node); break; case iro_Call: - /* a call is only tolerable if its either constant or pure */ + /* A call is only tolerable if its either constant or pure. */ ptr = get_Call_ptr(node); - if (get_irn_op(ptr) == op_SymConst && - get_SymConst_kind(ptr) == symconst_addr_ent) { + if (is_SymConst_addr_ent(ptr)) { ir_entity *ent = get_SymConst_entity(ptr); ir_graph *irg = get_entity_irg(ent); if (irg == current_ir_graph) { - /* A recursive call. The did not mode depend on this call */ + /* A self-recursive call. The property did not depend on this call. */ } else if (irg == NULL) { m = get_entity_additional_properties(ent) & (mtp_property_const|mtp_property_pure); - if (! m) - return 0; - mode = UMAX(mode, m); + mode = max_property(mode, m); } else if (irg != NULL) { - /* we have a graph. Check if it is already analyzed */ - if (IS_IRG_MARKED(irg)) - (void)check_const_or_pure_function(irg); - - m = get_irg_additional_properties(irg) & (mtp_property_const|mtp_property_pure); - if (! m) - return 0; - mode = UMAX(mode, m); + /* we have a graph, analyze it. */ + m = check_const_or_pure_function(irg, /*top=*/0); + mode = max_property(mode, m); } } else - return 0; + return mtp_no_property; node = get_Call_mem(node); break; default: - return 0; + return mtp_no_property; } } -} /* follow_mem */ +} /* _follow_mem */ /** * Follow the memory chain starting at node and determine * the mtp_property. * * @return mtp_property_const if only calls of const functions are detected - * mtp_property_pure if only Loads and const/pure - * calls detected - * 0 else + * mtp_property_pure if only Loads and const/pure calls detected + * mtp_no_property else */ -static unsigned follow_mem(ir_graph *irg, ir_node *node, unsigned mode) { +static unsigned follow_mem(ir_node *node, unsigned mode) +{ unsigned m; - inc_irg_visited(irg); - /* mark the initial mem: recursion stops here */ - mark_irn_visited(get_irg_initial_mem(irg)); m = _follow_mem(node); - if (! m) - return 0; - return UMAX(mode, m); -} /* follow_mwm */ + return max_property(mode, m); +} /* follow_mem */ -/* - * Check if a graph represents a const function. +/** + * Check if a graph represents a const or a pure function. * - * @param irg the graph + * @param irg the graph to check + * @param top if set, this is the top call */ -static unsigned check_const_or_pure_function(ir_graph *irg) { +static unsigned check_const_or_pure_function(ir_graph *irg, int top) +{ ir_node *end, *endbl; int j; - unsigned mode = get_irg_additional_properties(irg); + unsigned prop = get_irg_additional_properties(irg); ir_graph *rem = current_ir_graph; - if (mode & mtp_property_const) { + if (prop & mtp_property_const) { /* already marked as a const function */ return mtp_property_const; } - if (mode & mtp_property_pure) { + if (prop & mtp_property_pure) { /* already marked as a pure function */ return mtp_property_pure; } - if (! IS_IRG_MARKED(irg)) - return 0; - UNMARK_IRG(irg); + if (IS_IRG_READY(irg)) { + /* already checked */ + return mtp_no_property; + } + if (IS_IRG_BUSY(irg)) { + /* we are still evaluate this method. Be optimistic, + return the best possible so far but mark the result as temporary. */ + return mtp_temporary | mtp_property_const; + } + SET_IRG_BUSY(irg); end = get_irg_end(irg); endbl = get_nodes_block(end); - mode = mtp_property_const; + prop = mtp_property_const; current_ir_graph = irg; + ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED); + inc_irg_visited(irg); + /* mark the initial mem: recursion of follow_mem() stops here */ + mark_irn_visited(get_irg_initial_mem(irg)); + /* visit every Return */ for (j = get_Block_n_cfgpreds(endbl) - 1; j >= 0; --j) { - ir_node *node = get_Block_cfgpred(endbl, j); - ir_op *op = get_irn_op(node); - ir_node *mem; + ir_node *node = get_Block_cfgpred(endbl, j); + ir_opcode code = get_irn_opcode(node); + ir_node *mem; /* Bad nodes usually do NOT produce anything, so it's ok */ - if (op == op_Bad) + if (code == iro_Bad) continue; - if (op == op_Return) { + if (code == iro_Return) { mem = get_Return_mem(node); /* Bad nodes usually do NOT produce anything, so it's ok */ @@ -378,40 +576,59 @@ static unsigned check_const_or_pure_function(ir_graph *irg) { continue; if (mem != get_irg_initial_mem(irg)) - mode = follow_mem(irg, mem, mode); + prop = max_property(prop, follow_mem(mem, prop)); } else { - /* exception found. */ - mode = follow_mem(irg, node, mode); + /* Exception found. Cannot be const or pure. */ + prop = mtp_no_property; break; } - if (mode == 0) + if (prop == mtp_no_property) break; } - if (mode != 0) { + if (prop != mtp_no_property) { /* check, if a keep-alive exists */ for (j = get_End_n_keepalives(end) - 1; j >= 0; --j) { - ir_node *mem = get_End_keepalive(end, j); + ir_node *kept = get_End_keepalive(end, j); + + if (is_Block(kept)) { + prop = mtp_no_property; + break; + } - if (mode_M != get_irn_mode(mem)) + if (mode_M != get_irn_mode(kept)) continue; - mode = follow_mem(irg, mem, mode); - if (mode == 0) + prop = max_property(prop, follow_mem(kept, prop)); + if (prop == mtp_no_property) break; } } - if (mode) - set_irg_additional_property(irg, mode); + if (prop != mtp_no_property) { + if (top || (prop & mtp_temporary) == 0) { + /* We use the temporary flag here to mark optimistic result. + Set the property only if we are sure that it does NOT base on + temporary results OR if we are at top-level. */ + set_irg_additional_property(irg, prop & ~mtp_temporary); + SET_IRG_READY(irg); + } + } + if (top) + SET_IRG_READY(irg); + CLEAR_IRG_BUSY(irg); + ir_free_resources(irg, IR_RESOURCE_IRN_VISITED); current_ir_graph = rem; - return mode; + return prop; } /* check_const_or_pure_function */ /** * Handle calls to const functions. + * + * @param ctx context */ -static void handle_const_Calls(env_t *ctx) { +static void handle_const_Calls(env_t *ctx) +{ int i; ctx->n_calls_SymConst = 0; @@ -421,42 +638,434 @@ static void handle_const_Calls(env_t *ctx) { for (i = get_irp_n_irgs() - 1; i >= 0; --i) { ir_graph *irg = get_irp_irg(i); - ctx->const_call_list = NULL; - ctx->pure_call_list = NULL; - ctx->proj_list = NULL; - irg_walk_graph(irg, NULL, collect_calls, ctx); + ctx->float_const_call_list = NULL; + ctx->nonfloat_const_call_list = NULL; + ctx->pure_call_list = NULL; + ctx->proj_list = NULL; - if (ctx->const_call_list) - fix_const_call_list(irg, ctx->const_call_list, ctx->proj_list); + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); + irg_walk_graph(irg, NULL, collect_const_and_pure_calls, ctx); + + if (ctx->float_const_call_list != NULL) + fix_const_call_lists(irg, ctx); + ir_free_resources(irg, IR_RESOURCE_IRN_LINK); } } /* handle_const_Calls */ +/** + * Handle calls to nothrow functions. + * + * @param ctx context + */ +static void handle_nothrow_Calls(env_t *ctx) +{ + int i; + + ctx->n_calls_SymConst = 0; + ctx->n_calls_Sel = 0; + + /* all calls of const functions can be transformed */ + for (i = get_irp_n_irgs() - 1; i >= 0; --i) { + ir_graph *irg = get_irp_irg(i); + + ctx->nothrow_call_list = NULL; + ctx->proj_list = NULL; + + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); + irg_walk_graph(irg, NULL, collect_nothrow_calls, ctx); + + if (ctx->nothrow_call_list) + fix_nothrow_call_list(irg, ctx->nothrow_call_list, ctx->proj_list); + ir_free_resources(irg, IR_RESOURCE_IRN_LINK); + } +} + +/** + * Check, whether a given node represents a return value of + * a malloc like function (ie, new heap allocated memory). + * + * @param node the node to check + */ +static int is_malloc_call_result(const ir_node *node) +{ + if (is_Alloc(node) && get_Alloc_where(node) == heap_alloc) { + /* Firm style high-level allocation */ + return 1; + } + if (is_alloc_entity != NULL && is_Call(node)) { + ir_node *ptr = get_Call_ptr(node); + + if (is_Global(ptr)) { + ir_entity *ent = get_Global_entity(ptr); + return is_alloc_entity(ent); + } + } + return 0; +} /* is_malloc_call_result */ + +/** + * Update a property depending on a call property. + */ +static unsigned update_property(unsigned orig_prop, unsigned call_prop) +{ + unsigned t = (orig_prop | call_prop) & mtp_temporary; + unsigned r = orig_prop & call_prop; + return r | t; +} /** update_property */ + +/** + * Check if a node is stored. + */ +static int is_stored(const ir_node *n) +{ + const ir_edge_t *edge; + const ir_node *ptr; + + foreach_out_edge(n, edge) { + const ir_node *succ = get_edge_src_irn(edge); + + switch (get_irn_opcode(succ)) { + case iro_Return: + case iro_Load: + case iro_Cmp: + /* ok */ + break; + case iro_Store: + if (get_Store_value(succ) == n) + return 1; + /* ok if its only the address input */ + break; + case iro_Sel: + case iro_Cast: + case iro_Confirm: + if (is_stored(succ)) + return 1; + break; + case iro_Call: + ptr = get_Call_ptr(succ); + if (is_Global(ptr)) { + ir_entity *ent = get_Global_entity(ptr); + int i; + + /* we know the called entity */ + for (i = get_Call_n_params(succ) - 1; i >= 0; --i) { + if (get_Call_param(succ, i) == n) { + /* n is the i'th param of the call */ + if (get_method_param_access(ent, i) & ptr_access_store) { + /* n is store in ent */ + return 1; + } + } + } + } else { + /* unknown call address */ + return 1; + } + break; + default: + /* bad, potential alias */ + return 1; + } + } + return 0; +} /* is_stored */ + +/** + * Check that the return value of an irg is not stored anywhere. + * + * return ~mtp_property_malloc if return values are stored, ~0 else + */ +static unsigned check_stored_result(ir_graph *irg) +{ + ir_node *end_blk = get_irg_end_block(irg); + int i, j; + unsigned res = ~0; + int old_edges = edges_assure_kind(irg, EDGE_KIND_NORMAL); + + for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) { + ir_node *pred = get_Block_cfgpred(end_blk, i); + + if (! is_Return(pred)) + continue; + for (j = get_Return_n_ress(pred) - 1; j >= 0; --j) { + const ir_node *irn = get_Return_res(pred, j); + + if (is_stored(irn)) { + /* bad, might create an alias */ + res = ~mtp_property_malloc; + goto finish; + } + } + } +finish: + if (! old_edges) + edges_deactivate_kind(irg, EDGE_KIND_NORMAL); + return res; +} /* check_stored_result */ + +/** + * Check if a graph represents a nothrow or a malloc function. + * + * @param irg the graph to check + * @param top if set, this is the top call + */ +static unsigned check_nothrow_or_malloc(ir_graph *irg, int top) +{ + ir_node *end_blk = get_irg_end_block(irg); + ir_entity *ent; + ir_type *mtp; + int i, j; + unsigned curr_prop = mtp_property_malloc | mtp_property_nothrow; + + if (IS_IRG_READY(irg)) { + /* already checked */ + return get_irg_additional_properties(irg); + } + if (IS_IRG_BUSY(irg)) { + /* we are still evaluate this method. Be optimistic, + return the best possible so far but mark the result as temporary. */ + return mtp_temporary | mtp_property_malloc | mtp_property_nothrow; + } + SET_IRG_BUSY(irg); + + ent = get_irg_entity(irg); + mtp = get_entity_type(ent); + + if (get_method_n_ress(mtp) <= 0) + curr_prop &= ~mtp_property_malloc; + + for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) { + ir_node *pred = get_Block_cfgpred(end_blk, i); + + if (is_Return(pred)) { + if (curr_prop & mtp_property_malloc) { + /* check, if malloc is called here */ + for (j = get_Return_n_ress(pred) - 1; j >= 0; --j) { + ir_node *res = get_Return_res(pred, j); + + /* skip Confirms and Casts */ + res = skip_HighLevel_ops(res); + /* skip Proj's */ + while (is_Proj(res)) + res = get_Proj_pred(res); + if (is_malloc_call_result(res)) { + /* ok, this is a malloc */ + } else if (is_Call(res)) { + ir_node *ptr = get_Call_ptr(res); + + if (is_Global(ptr)) { + /* a direct call */ + ir_entity *ent = get_Global_entity(ptr); + ir_graph *callee = get_entity_irg(ent); + + if (callee == irg) { + /* A self-recursive call. The property did not depend on this call. */ + } else if (callee != NULL) { + unsigned prop = check_nothrow_or_malloc(callee, /*top=*/0); + curr_prop = update_property(curr_prop, prop); + } else { + curr_prop = update_property(curr_prop, get_entity_additional_properties(ent)); + } + } else if (get_opt_closed_world() && + is_Sel(ptr) && + get_irg_callee_info_state(irg) == irg_callee_info_consistent) { + /* check if all possible callees are malloc functions. */ + int i, n_callees = get_Call_n_callees(res); + if (n_callees == 0) { + /* This is kind of strange: dying code or a Call that will raise an exception + when executed as there is no implementation to call. So better not + optimize. */ + curr_prop &= ~mtp_property_malloc; + continue; + } + + for (i = 0; i < n_callees; ++i) { + ir_entity *ent = get_Call_callee(res, i); + if (ent == unknown_entity) { + /* we don't know which entity is called here */ + curr_prop &= ~mtp_property_malloc; + break; + } + if ((get_entity_additional_properties(ent) & mtp_property_malloc) == 0) { + curr_prop &= ~mtp_property_malloc; + break; + } + } + /* if we pass the for cycle, malloc is still ok */ + } else { + /* unknown call */ + curr_prop &= ~mtp_property_malloc; + } + } else { + /* unknown return value */ + curr_prop &= ~mtp_property_malloc; + } + } + } + } else if (curr_prop & mtp_property_nothrow) { + /* exception flow detected */ + pred = skip_Proj(pred); + + if (is_Call(pred)) { + ir_node *ptr = get_Call_ptr(pred); + + if (is_Global(ptr)) { + /* a direct call */ + ir_entity *ent = get_Global_entity(ptr); + ir_graph *callee = get_entity_irg(ent); + + if (callee == irg) { + /* A self-recursive call. The property did not depend on this call. */ + } else if (callee != NULL) { + /* Note: we check here for nothrow only, so do NOT reset the malloc property */ + unsigned prop = check_nothrow_or_malloc(callee, /*top=*/0) | mtp_property_malloc; + curr_prop = update_property(curr_prop, prop); + } else { + if ((get_entity_additional_properties(ent) & mtp_property_nothrow) == 0) + curr_prop &= ~mtp_property_nothrow; + } + } else if (get_opt_closed_world() && + is_Sel(ptr) && + get_irg_callee_info_state(irg) == irg_callee_info_consistent) { + /* check if all possible callees are nothrow functions. */ + int i, n_callees = get_Call_n_callees(pred); + if (n_callees == 0) { + /* This is kind of strange: dying code or a Call that will raise an exception + when executed as there is no implementation to call. So better not + optimize. */ + curr_prop &= ~mtp_property_nothrow; + continue; + } + + for (i = 0; i < n_callees; ++i) { + ir_entity *ent = get_Call_callee(pred, i); + if (ent == unknown_entity) { + /* we don't know which entity is called here */ + curr_prop &= ~mtp_property_nothrow; + break; + } + if ((get_entity_additional_properties(ent) & mtp_property_nothrow) == 0) { + curr_prop &= ~mtp_property_nothrow; + break; + } + } + /* if we pass the for cycle, nothrow is still ok */ + } else { + /* unknown call */ + curr_prop &= ~mtp_property_nothrow; + } + } else { + /* real exception flow possible. */ + curr_prop &= ~mtp_property_nothrow; + } + } + if ((curr_prop & ~mtp_temporary) == mtp_no_property) { + /* no need to search further */ + break; + } + } + + if (curr_prop & mtp_property_malloc) { + /* + * Note that the malloc property means not only return newly allocated + * memory, but also that this memory is ALIAS FREE. + * To ensure that, we do NOT allow that the returned memory is somewhere + * stored. + */ + curr_prop &= check_stored_result(irg); + } + + if (curr_prop != mtp_no_property) { + if (top || (curr_prop & mtp_temporary) == 0) { + /* We use the temporary flag here to mark an optimistic result. + Set the property only if we are sure that it does NOT base on + temporary results OR if we are at top-level. */ + set_irg_additional_property(irg, curr_prop & ~mtp_temporary); + SET_IRG_READY(irg); + } + } + if (top) + SET_IRG_READY(irg); + CLEAR_IRG_BUSY(irg); + return curr_prop; +} /* check_nothrow_or_malloc */ + +/** + * When a function was detected as "const", it might be moved out of loops. + * This might be dangerous if the graph can contain endless loops. + */ +static void check_for_possible_endless_loops(ir_graph *irg) +{ + ir_loop *root_loop; + assure_cf_loop(irg); + + root_loop = get_irg_loop(irg); + if (root_loop->flags & loop_outer_loop) + set_irg_additional_property(irg, mtp_property_has_loop); +} + /* * optimize function calls by handling const functions */ -void optimize_funccalls(int force_run) +void optimize_funccalls(int force_run, check_alloc_entity_func callback) { - int i, n; - unsigned num_const = 0; - unsigned num_pure = 0; + int i, last_idx; + unsigned num_const = 0; + unsigned num_pure = 0; + unsigned num_nothrow = 0; + unsigned num_malloc = 0; + + is_alloc_entity = callback; /* prepare: mark all graphs as not analyzed */ - n = get_irp_n_irgs(); - for (i = n - 1; i >= 0; --i) - MARK_IRG(get_irp_irg(i)); + last_idx = get_irp_last_idx(); + ready_set = rbitset_malloc(last_idx); + busy_set = rbitset_malloc(last_idx); - /* first step: detect, which functions are const, i.e. do NOT touch any memory */ - DBG((dbg, LEVEL_2, "Detecting const and pure properties ...\n")); - for (i = n - 1; i >= 0; --i) { + /* first step: detect, which functions are nothrow or malloc */ + DB((dbg, LEVEL_2, "Detecting nothrow and malloc properties ...\n")); + for (i = get_irp_n_irgs() - 1; i >= 0; --i) { ir_graph *irg = get_irp_irg(i); - unsigned mode = check_const_or_pure_function(irg); + unsigned prop = check_nothrow_or_malloc(irg, /*top=*/1); + + if (prop & mtp_property_nothrow) { + ++num_nothrow; + DB((dbg, LEVEL_2, "%+F has the nothrow property\n", irg)); + } else if (prop & mtp_property_malloc) { + ++num_malloc; + DB((dbg, LEVEL_2, "%+F has the malloc property\n", irg)); + } + } - if (mode & mtp_property_const) { + /* second step: remove exception edges: this must be done before the + detection of const and pure functions take place. */ + if (force_run || num_nothrow > 0) { + env_t ctx; + + handle_nothrow_Calls(&ctx); + DB((dbg, LEVEL_1, "Detected %u nothrow graphs, %u malloc graphs.\n", num_nothrow, num_malloc)); + DB((dbg, LEVEL_1, "Optimizes %u(SymConst) + %u(Sel) calls to nothrow functions.\n", + ctx.n_calls_SymConst, ctx.n_calls_Sel)); + } else { + DB((dbg, LEVEL_1, "No graphs without side effects detected\n")); + } + + rbitset_clear_all(ready_set, last_idx); + rbitset_clear_all(busy_set, last_idx); + + /* third step: detect, which functions are const or pure */ + DB((dbg, LEVEL_2, "Detecting const and pure properties ...\n")); + for (i = get_irp_n_irgs() - 1; i >= 0; --i) { + ir_graph *irg = get_irp_irg(i); + unsigned prop = check_const_or_pure_function(irg, /*top=*/1); + + if (prop & mtp_property_const) { ++num_const; - DBG((dbg, LEVEL_2, "%+F has the const property\n", irg)); - } else if (mode & mtp_property_pure) { + DB((dbg, LEVEL_2, "%+F has the const property\n", irg)); + check_for_possible_endless_loops(irg); + } else if (prop & mtp_property_pure) { ++num_pure; - DBG((dbg, LEVEL_2, "%+F has the pure property\n", irg)); + DB((dbg, LEVEL_2, "%+F has the pure property\n", irg)); } } @@ -464,14 +1073,14 @@ void optimize_funccalls(int force_run) env_t ctx; handle_const_Calls(&ctx); - if (get_firm_verbosity()) { - DBG((dbg, LEVEL_1, "Detected %u const graphs, %u pure graphs.\n", num_const, num_pure)); - DBG((dbg, LEVEL_1, "Optimizes %u(SymConst) + %u(Sel) calls to const functions.\n", - ctx.n_calls_SymConst, ctx.n_calls_Sel)); - } + DB((dbg, LEVEL_1, "Detected %u const graphs, %u pure graphs.\n", num_const, num_pure)); + DB((dbg, LEVEL_1, "Optimizes %u(SymConst) + %u(Sel) calls to const functions.\n", + ctx.n_calls_SymConst, ctx.n_calls_Sel)); } else { - DBG((dbg, LEVEL_1, "No graphs without side effects detected\n")); + DB((dbg, LEVEL_1, "No graphs without side effects detected\n")); } + xfree(busy_set); + xfree(ready_set); } /* optimize_funccalls */ /* initialize the funccall optimization */ @@ -479,3 +1088,35 @@ void firm_init_funccalls(void) { FIRM_DBG_REGISTER(dbg, "firm.opt.funccalls"); } /* firm_init_funccalls */ + +struct pass_t { + ir_prog_pass_t pass; + int force_run; + check_alloc_entity_func callback; +}; + +/** + * Wrapper for running optimize_funccalls() as an ir_prog pass. + */ +static int pass_wrapper(ir_prog *irp, void *context) +{ + struct pass_t *pass = context; + + (void)irp; + optimize_funccalls(pass->force_run, pass->callback); + return 0; +} /* pass_wrapper */ + +/* Creates an ir_prog pass for optimize_funccalls. */ +ir_prog_pass_t *optimize_funccalls_pass( + const char *name, + int force_run, check_alloc_entity_func callback) +{ + struct pass_t *pass = XMALLOCZ(struct pass_t); + + pass->force_run = force_run; + pass->callback = callback; + + return def_prog_pass_constructor( + &pass->pass, name ? name : "funccall", pass_wrapper); +} /* optimize_funccalls_pass */