/* Perform various loop optimizations, including strength reduction. Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001 Free Software Foundation, Inc. This file is part of GNU CC. GNU CC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2, or (at your option) any later version. GNU CC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with GNU CC; see the file COPYING. If not, write to the Free Software Foundation, 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. */ /* This is the loop optimization pass of the compiler. It finds invariant computations within loops and moves them to the beginning of the loop. Then it identifies basic and general induction variables. Strength reduction is applied to the general induction variables, and induction variable elimination is applied to the basic induction variables. It also finds cases where a register is set within the loop by zero-extending a narrower value and changes these to zero the entire register once before the loop and merely copy the low part within the loop. Most of the complexity is in heuristics to decide when it is worth while to do these things. */ #include "config.h" #include "system.h" #include "rtl.h" #include "obstack.h" #include "expr.h" #include "insn-config.h" #include "insn-flags.h" #include "regs.h" #include "hard-reg-set.h" #include "recog.h" #include "flags.h" #include "real.h" #include "loop.h" #include "except.h" #include "toplev.h" /* Vector mapping INSN_UIDs to luids. The luids are like uids but increase monotonically always. We use them to see whether a jump comes from outside a given loop. */ int *uid_luid; /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop number the insn is contained in. */ int *uid_loop_num; /* 1 + largest uid of any insn. */ int max_uid_for_loop; /* 1 + luid of last insn. */ static int max_luid; /* Number of loops detected in current function. Used as index to the next few tables. */ static int max_loop_num; /* Indexed by loop number, contains the first and last insn of each loop. */ static rtx *loop_number_loop_starts, *loop_number_loop_ends; /* Likewise for the continue insn */ static rtx *loop_number_loop_cont; /* The first code_label that is reached in every loop iteration. 0 when not computed yet, initially const0_rtx if a jump couldn't be followed. Also set to 0 when there is no such label before the NOTE_INSN_LOOP_CONT of this loop, or in verify_dominator, if a jump couldn't be followed. */ static rtx *loop_number_cont_dominator; /* For each loop, gives the containing loop number, -1 if none. */ int *loop_outer_loop; #ifdef HAVE_decrement_and_branch_on_count /* Records whether resource in use by inner loop. */ int *loop_used_count_register; #endif /* HAVE_decrement_and_branch_on_count */ /* Indexed by loop number, contains a nonzero value if the "loop" isn't really a loop (an insn outside the loop branches into it). */ static char *loop_invalid; /* Indexed by loop number, links together all LABEL_REFs which refer to code labels outside the loop. Used by routines that need to know all loop exits, such as final_biv_value and final_giv_value. This does not include loop exits due to return instructions. This is because all bivs and givs are pseudos, and hence must be dead after a return, so the presense of a return does not affect any of the optimizations that use this info. It is simpler to just not include return instructions on this list. */ rtx *loop_number_exit_labels; /* Indexed by loop number, counts the number of LABEL_REFs on loop_number_exit_labels for this loop and all loops nested inside it. */ int *loop_number_exit_count; /* Nonzero if there is a subroutine call in the current loop. */ static int loop_has_call; /* Nonzero if there is a volatile memory reference in the current loop. */ static int loop_has_volatile; /* Nonzero if there is a tablejump in the current loop. */ static int loop_has_tablejump; /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the current loop. A continue statement will generate a branch to NEXT_INSN (loop_continue). */ static rtx loop_continue; /* Indexed by register number, contains the number of times the reg is set during the loop being scanned. During code motion, a negative value indicates a reg that has been made a candidate; in particular -2 means that it is an candidate that we know is equal to a constant and -1 means that it is an candidate not known equal to a constant. After code motion, regs moved have 0 (which is accurate now) while the failed candidates have the original number of times set. Therefore, at all times, == 0 indicates an invariant register; < 0 a conditionally invariant one. */ static varray_type set_in_loop; /* Original value of set_in_loop; same except that this value is not set negative for a reg whose sets have been made candidates and not set to 0 for a reg that is moved. */ static varray_type n_times_set; /* Index by register number, 1 indicates that the register cannot be moved or strength reduced. */ static varray_type may_not_optimize; /* Contains the insn in which a register was used if it was used exactly once; contains const0_rtx if it was used more than once. */ static varray_type reg_single_usage; /* Nonzero means reg N has already been moved out of one loop. This reduces the desire to move it out of another. */ static char *moved_once; /* List of MEMs that are stored in this loop. */ static rtx loop_store_mems; /* The insn where the first of these was found. */ static rtx first_loop_store_insn; typedef struct loop_mem_info { rtx mem; /* The MEM itself. */ rtx reg; /* Corresponding pseudo, if any. */ int optimize; /* Nonzero if we can optimize access to this MEM. */ } loop_mem_info; /* Array of MEMs that are used (read or written) in this loop, but cannot be aliased by anything in this loop, except perhaps themselves. In other words, if loop_mems[i] is altered during the loop, it is altered by an expression that is rtx_equal_p to it. */ static loop_mem_info *loop_mems; /* The index of the next available slot in LOOP_MEMS. */ static int loop_mems_idx; /* The number of elements allocated in LOOP_MEMs. */ static int loop_mems_allocated; /* Nonzero if we don't know what MEMs were changed in the current loop. This happens if the loop contains a call (in which case `loop_has_call' will also be set) or if we store into more than NUM_STORES MEMs. */ static int unknown_address_altered; /* Count of movable (i.e. invariant) instructions discovered in the loop. */ static int num_movables; /* Count of memory write instructions discovered in the loop. */ static int num_mem_sets; /* Number of loops contained within the current one, including itself. */ static int loops_enclosed; /* Bound on pseudo register number before loop optimization. A pseudo has valid regscan info if its number is < max_reg_before_loop. */ int max_reg_before_loop; /* This obstack is used in product_cheap_p to allocate its rtl. It may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx. If we used the same obstack that it did, we would be deallocating that array. */ static struct obstack temp_obstack; /* This is where the pointer to the obstack being used for RTL is stored. */ extern struct obstack *rtl_obstack; #define obstack_chunk_alloc xmalloc #define obstack_chunk_free free /* During the analysis of a loop, a chain of `struct movable's is made to record all the movable insns found. Then the entire chain can be scanned to decide which to move. */ struct movable { rtx insn; /* A movable insn */ rtx set_src; /* The expression this reg is set from. */ rtx set_dest; /* The destination of this SET. */ rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST of any registers used within the LIBCALL. */ int consec; /* Number of consecutive following insns that must be moved with this one. */ int regno; /* The register it sets */ short lifetime; /* lifetime of that register; may be adjusted when matching movables that load the same value are found. */ short savings; /* Number of insns we can move for this reg, including other movables that force this or match this one. */ unsigned int cond : 1; /* 1 if only conditionally movable */ unsigned int force : 1; /* 1 means MUST move this insn */ unsigned int global : 1; /* 1 means reg is live outside this loop */ /* If PARTIAL is 1, GLOBAL means something different: that the reg is live outside the range from where it is set to the following label. */ unsigned int done : 1; /* 1 inhibits further processing of this */ unsigned int partial : 1; /* 1 means this reg is used for zero-extending. In particular, moving it does not make it invariant. */ unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to load SRC, rather than copying INSN. */ unsigned int move_insn_first:1;/* Same as above, if this is necessary for the first insn of a consecutive sets group. */ unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */ enum machine_mode savemode; /* Nonzero means it is a mode for a low part that we should avoid changing when clearing the rest of the reg. */ struct movable *match; /* First entry for same value */ struct movable *forces; /* An insn that must be moved if this is */ struct movable *next; }; static struct movable *the_movables; FILE *loop_dump_stream; /* For communicating return values from note_set_pseudo_multiple_uses. */ static int note_set_pseudo_multiple_uses_retval; /* Forward declarations. */ static void verify_dominator PROTO((int)); static void find_and_verify_loops PROTO((rtx)); static void mark_loop_jump PROTO((rtx, int)); static void prescan_loop PROTO((rtx, rtx)); static int reg_in_basic_block_p PROTO((rtx, rtx)); static int consec_sets_invariant_p PROTO((rtx, int, rtx)); static int labels_in_range_p PROTO((rtx, int)); static void count_one_set PROTO((rtx, rtx, varray_type, rtx *)); static void count_loop_regs_set PROTO((rtx, rtx, varray_type, varray_type, int *, int)); static void note_addr_stored PROTO((rtx, rtx)); static void note_set_pseudo_multiple_uses PROTO((rtx, rtx)); static int loop_reg_used_before_p PROTO((rtx, rtx, rtx, rtx, rtx)); static void scan_loop PROTO((rtx, rtx, rtx, int, int)); #if 0 static void replace_call_address PROTO((rtx, rtx, rtx)); #endif static rtx skip_consec_insns PROTO((rtx, int)); static int libcall_benefit PROTO((rtx)); static void ignore_some_movables PROTO((struct movable *)); static void force_movables PROTO((struct movable *)); static void combine_movables PROTO((struct movable *, int)); static int regs_match_p PROTO((rtx, rtx, struct movable *)); static int rtx_equal_for_loop_p PROTO((rtx, rtx, struct movable *)); static void add_label_notes PROTO((rtx, rtx)); static void move_movables PROTO((struct movable *, int, int, rtx, rtx, int)); static int count_nonfixed_reads PROTO((rtx)); static void strength_reduce PROTO((rtx, rtx, rtx, int, rtx, rtx, rtx, int, int)); static void find_single_use_in_loop PROTO((rtx, rtx, varray_type)); static int valid_initial_value_p PROTO((rtx, rtx, int, rtx)); static void find_mem_givs PROTO((rtx, rtx, int, rtx, rtx)); static void record_biv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx *, int, int)); static void check_final_value PROTO((struct induction *, rtx, rtx, unsigned HOST_WIDE_INT)); static void record_giv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx, int, enum g_types, int, rtx *, rtx, rtx)); static void update_giv_derive PROTO((rtx)); static int basic_induction_var PROTO((rtx, enum machine_mode, rtx, rtx, rtx *, rtx *, rtx **)); static rtx simplify_giv_expr PROTO((rtx, int *)); static int general_induction_var PROTO((rtx, rtx *, rtx *, rtx *, int, int *)); static int consec_sets_giv PROTO((int, rtx, rtx, rtx, rtx *, rtx *, rtx *)); static int check_dbra_loop PROTO((rtx, int, rtx, struct loop_info *)); static rtx express_from_1 PROTO((rtx, rtx, rtx)); static rtx combine_givs_p PROTO((struct induction *, struct induction *)); static void combine_givs PROTO((struct iv_class *)); struct recombine_givs_stats; static int find_life_end PROTO((rtx, struct recombine_givs_stats *, rtx, rtx)); static void recombine_givs PROTO((struct iv_class *, rtx, rtx, int)); static int product_cheap_p PROTO((rtx, rtx)); static int maybe_eliminate_biv PROTO((struct iv_class *, rtx, rtx, int, int, int)); static int maybe_eliminate_biv_1 PROTO((rtx, rtx, struct iv_class *, int, rtx)); static int last_use_this_basic_block PROTO((rtx, rtx)); static void record_initial PROTO((rtx, rtx)); static void update_reg_last_use PROTO((rtx, rtx)); static rtx next_insn_in_loop PROTO((rtx, rtx, rtx, rtx)); static void load_mems_and_recount_loop_regs_set PROTO((rtx, rtx, rtx, rtx, int *)); static void load_mems PROTO((rtx, rtx, rtx, rtx)); static int insert_loop_mem PROTO((rtx *, void *)); static int replace_loop_mem PROTO((rtx *, void *)); static int replace_label PROTO((rtx *, void *)); typedef struct rtx_and_int { rtx r; int i; } rtx_and_int; typedef struct rtx_pair { rtx r1; rtx r2; } rtx_pair; /* Nonzero iff INSN is between START and END, inclusive. */ #define INSN_IN_RANGE_P(INSN, START, END) \ (INSN_UID (INSN) < max_uid_for_loop \ && INSN_LUID (INSN) >= INSN_LUID (START) \ && INSN_LUID (INSN) <= INSN_LUID (END)) #ifdef HAVE_decrement_and_branch_on_count /* Test whether BCT applicable and safe. */ static void insert_bct PROTO((rtx, rtx, struct loop_info *)); /* Auxiliary function that inserts the BCT pattern into the loop. */ static void instrument_loop_bct PROTO((rtx, rtx, rtx)); #endif /* HAVE_decrement_and_branch_on_count */ /* Indirect_jump_in_function is computed once per function. */ int indirect_jump_in_function = 0; static int indirect_jump_in_function_p PROTO((rtx)); static int compute_luids PROTO((rtx, rtx, int)); static int biv_elimination_giv_has_0_offset PROTO((struct induction *, struct induction *, rtx)); /* Relative gain of eliminating various kinds of operations. */ static int add_cost; #if 0 static int shift_cost; static int mult_cost; #endif /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to copy the value of the strength reduced giv to its original register. */ static int copy_cost; /* Cost of using a register, to normalize the benefits of a giv. */ static int reg_address_cost; void init_loop () { char *free_point = (char *) oballoc (1); rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1); add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET); #ifdef ADDRESS_COST reg_address_cost = ADDRESS_COST (reg); #else reg_address_cost = rtx_cost (reg, MEM); #endif /* We multiply by 2 to reconcile the difference in scale between these two ways of computing costs. Otherwise the cost of a copy will be far less than the cost of an add. */ copy_cost = 2 * 2; /* Free the objects we just allocated. */ obfree (free_point); /* Initialize the obstack used for rtl in product_cheap_p. */ gcc_obstack_init (&temp_obstack); } /* Compute the mapping from uids to luids. LUIDs are numbers assigned to insns, like uids, except that luids increase monotonically through the code. Start at insn START and stop just before END. Assign LUIDs starting with PREV_LUID + 1. Return the last assigned LUID + 1. */ static int compute_luids (start, end, prev_luid) rtx start, end; int prev_luid; { int i; rtx insn; for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn)) { if (INSN_UID (insn) >= max_uid_for_loop) continue; /* Don't assign luids to line-number NOTEs, so that the distance in luids between two insns is not affected by -g. */ if (GET_CODE (insn) != NOTE || NOTE_LINE_NUMBER (insn) <= 0) uid_luid[INSN_UID (insn)] = ++i; else /* Give a line number note the same luid as preceding insn. */ uid_luid[INSN_UID (insn)] = i; } return i + 1; } /* Entry point of this file. Perform loop optimization on the current function. F is the first insn of the function and DUMPFILE is a stream for output of a trace of actions taken (or 0 if none should be output). */ void loop_optimize (f, dumpfile, unroll_p, bct_p) /* f is the first instruction of a chain of insns for one function */ rtx f; FILE *dumpfile; int unroll_p, bct_p; { register rtx insn; register int i; loop_dump_stream = dumpfile; init_recog_no_volatile (); max_reg_before_loop = max_reg_num (); moved_once = (char *) alloca (max_reg_before_loop); bzero (moved_once, max_reg_before_loop); regs_may_share = 0; /* Count the number of loops. */ max_loop_num = 0; for (insn = f; insn; insn = NEXT_INSN (insn)) { if (GET_CODE (insn) == NOTE && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG) max_loop_num++; } /* Don't waste time if no loops. */ if (max_loop_num == 0) return; /* Get size to use for tables indexed by uids. Leave some space for labels allocated by find_and_verify_loops. */ max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32; uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int)); uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int)); bzero ((char *) uid_luid, max_uid_for_loop * sizeof (int)); bzero ((char *) uid_loop_num, max_uid_for_loop * sizeof (int)); /* Allocate tables for recording each loop. We set each entry, so they need not be zeroed. */ loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx)); loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx)); loop_number_loop_cont = (rtx *) alloca (max_loop_num * sizeof (rtx)); loop_number_cont_dominator = (rtx *) alloca (max_loop_num * sizeof (rtx)); loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int)); loop_invalid = (char *) alloca (max_loop_num * sizeof (char)); loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx)); loop_number_exit_count = (int *) alloca (max_loop_num * sizeof (int)); #ifdef HAVE_decrement_and_branch_on_count /* Allocate for BCT optimization */ loop_used_count_register = (int *) alloca (max_loop_num * sizeof (int)); bzero ((char *) loop_used_count_register, max_loop_num * sizeof (int)); #endif /* HAVE_decrement_and_branch_on_count */ /* Find and process each loop. First, find them, and record them in order of their beginnings. */ find_and_verify_loops (f); /* Now find all register lifetimes. This must be done after find_and_verify_loops, because it might reorder the insns in the function. */ reg_scan (f, max_reg_num (), 1); /* This must occur after reg_scan so that registers created by gcse will have entries in the register tables. We could have added a call to reg_scan after gcse_main in toplev.c, but moving this call to init_alias_analysis is more efficient. */ init_alias_analysis (); /* See if we went too far. Note that get_max_uid already returns one more that the maximum uid of all insn. */ if (get_max_uid () > max_uid_for_loop) abort (); /* Now reset it to the actual size we need. See above. */ max_uid_for_loop = get_max_uid (); /* find_and_verify_loops has already called compute_luids, but it might have rearranged code afterwards, so we need to recompute the luids now. */ max_luid = compute_luids (f, NULL_RTX, 0); /* Don't leave gaps in uid_luid for insns that have been deleted. It is possible that the first or last insn using some register has been deleted by cross-jumping. Make sure that uid_luid for that former insn's uid points to the general area where that insn used to be. */ for (i = 0; i < max_uid_for_loop; i++) { uid_luid[0] = uid_luid[i]; if (uid_luid[0] != 0) break; } for (i = 0; i < max_uid_for_loop; i++) if (uid_luid[i] == 0) uid_luid[i] = uid_luid[i - 1]; /* Create a mapping from loops to BLOCK tree nodes. */ if (unroll_p && write_symbols != NO_DEBUG) find_loop_tree_blocks (); /* Determine if the function has indirect jump. On some systems this prevents low overhead loop instructions from being used. */ indirect_jump_in_function = indirect_jump_in_function_p (f); /* Now scan the loops, last ones first, since this means inner ones are done before outer ones. */ for (i = max_loop_num-1; i >= 0; i--) if (! loop_invalid[i] && loop_number_loop_ends[i]) scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i], loop_number_loop_cont[i], unroll_p, bct_p); /* If debugging and unrolling loops, we must replicate the tree nodes corresponding to the blocks inside the loop, so that the original one to one mapping will remain. */ if (unroll_p && write_symbols != NO_DEBUG) unroll_block_trees (); end_alias_analysis (); } /* Returns the next insn, in execution order, after INSN. START and END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop, respectively. LOOP_TOP, if non-NULL, is the top of the loop in the insn-stream; it is used with loops that are entered near the bottom. */ static rtx next_insn_in_loop (insn, start, end, loop_top) rtx insn; rtx start; rtx end; rtx loop_top; { insn = NEXT_INSN (insn); if (insn == end) { if (loop_top) /* Go to the top of the loop, and continue there. */ insn = loop_top; else /* We're done. */ insn = NULL_RTX; } if (insn == start) /* We're done. */ insn = NULL_RTX; return insn; } /* Optimize one loop whose start is LOOP_START and end is END. LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching NOTE_INSN_LOOP_END. LOOP_CONT is the NOTE_INSN_LOOP_CONT. */ /* ??? Could also move memory writes out of loops if the destination address is invariant, the source is invariant, the memory write is not volatile, and if we can prove that no read inside the loop can read this address before the write occurs. If there is a read of this address after the write, then we can also mark the memory read as invariant. */ static void scan_loop (loop_start, end, loop_cont, unroll_p, bct_p) rtx loop_start, end, loop_cont; int unroll_p, bct_p; { register int i; rtx p; /* 1 if we are scanning insns that could be executed zero times. */ int maybe_never = 0; /* 1 if we are scanning insns that might never be executed due to a subroutine call which might exit before they are reached. */ int call_passed = 0; /* For a rotated loop that is entered near the bottom, this is the label at the top. Otherwise it is zero. */ rtx loop_top = 0; /* Jump insn that enters the loop, or 0 if control drops in. */ rtx loop_entry_jump = 0; /* Place in the loop where control enters. */ rtx scan_start; /* Number of insns in the loop. */ int insn_count; int in_libcall = 0; int tem; rtx temp; /* The SET from an insn, if it is the only SET in the insn. */ rtx set, set1; /* Chain describing insns movable in current loop. */ struct movable *movables = 0; /* Last element in `movables' -- so we can add elements at the end. */ struct movable *last_movable = 0; /* Ratio of extra register life span we can justify for saving an instruction. More if loop doesn't call subroutines since in that case saving an insn makes more difference and more registers are available. */ int threshold; /* Nonzero if we are scanning instructions in a sub-loop. */ int loop_depth = 0; int nregs; /* Determine whether this loop starts with a jump down to a test at the end. This will occur for a small number of loops with a test that is too complex to duplicate in front of the loop. We search for the first insn or label in the loop, skipping NOTEs. However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG (because we might have a loop executed only once that contains a loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END (in case we have a degenerate loop). Note that if we mistakenly think that a loop is entered at the top when, in fact, it is entered at the exit test, the only effect will be slightly poorer optimization. Making the opposite error can generate incorrect code. Since very few loops now start with a jump to the exit test, the code here to detect that case is very conservative. */ for (p = NEXT_INSN (loop_start); p != end && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i' && (GET_CODE (p) != NOTE || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END)); p = NEXT_INSN (p)) ; scan_start = p; /* Set up variables describing this loop. */ prescan_loop (loop_start, end); threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs); /* If loop has a jump before the first label, the true entry is the target of that jump. Start scan from there. But record in LOOP_TOP the place where the end-test jumps back to so we can scan that after the end of the loop. */ if (GET_CODE (p) == JUMP_INSN) { loop_entry_jump = p; /* Loop entry must be unconditional jump (and not a RETURN) */ if (simplejump_p (p) && JUMP_LABEL (p) != 0 /* Check to see whether the jump actually jumps out of the loop (meaning it's no loop). This case can happen for things like do {..} while (0). If this label was generated previously by loop, we can't tell anything about it and have to reject the loop. */ && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, end)) { loop_top = next_label (scan_start); scan_start = JUMP_LABEL (p); } } /* If SCAN_START was an insn created by loop, we don't know its luid as required by loop_reg_used_before_p. So skip such loops. (This test may never be true, but it's best to play it safe.) Also, skip loops where we do not start scanning at a label. This test also rejects loops starting with a JUMP_INSN that failed the test above. */ if (INSN_UID (scan_start) >= max_uid_for_loop || GET_CODE (scan_start) != CODE_LABEL) { if (loop_dump_stream) fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n", INSN_UID (loop_start), INSN_UID (end)); return; } /* Count number of times each reg is set during this loop. Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */ /* Allocate extra space for REGS that might be created by load_mems. We allocate a little extra slop as well, in the hopes that even after the moving of movables creates some new registers we won't have to reallocate these arrays. However, we do grow the arrays, if necessary, in load_mems_recount_loop_regs_set. */ nregs = max_reg_num () + loop_mems_idx + 16; VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop"); VARRAY_INT_INIT (n_times_set, nregs, "n_times_set"); VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize"); VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage"); count_loop_regs_set (loop_top ? loop_top : loop_start, end, may_not_optimize, reg_single_usage, &insn_count, nregs); for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) { VARRAY_CHAR (may_not_optimize, i) = 1; VARRAY_INT (set_in_loop, i) = 1; } #ifdef AVOID_CCMODE_COPIES /* Don't try to move insns which set CC registers if we should not create CCmode register copies. */ for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--) if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC) VARRAY_CHAR (may_not_optimize, i) = 1; #endif bcopy ((char *) &set_in_loop->data, (char *) &n_times_set->data, nregs * sizeof (int)); if (loop_dump_stream) { fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n", INSN_UID (loop_start), INSN_UID (end), insn_count); if (loop_continue) fprintf (loop_dump_stream, "Continue at insn %d.\n", INSN_UID (loop_continue)); } /* Scan through the loop finding insns that are safe to move. Set set_in_loop negative for the reg being set, so that this reg will be considered invariant for subsequent insns. We consider whether subsequent insns use the reg in deciding whether it is worth actually moving. MAYBE_NEVER is nonzero if we have passed a conditional jump insn and therefore it is possible that the insns we are scanning would never be executed. At such times, we must make sure that it is safe to execute the insn once instead of zero times. When MAYBE_NEVER is 0, all insns will be executed at least once so that is not a problem. */ for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top); p != NULL_RTX; p = next_insn_in_loop (p, scan_start, end, loop_top)) { if (GET_RTX_CLASS (GET_CODE (p)) == 'i' && find_reg_note (p, REG_LIBCALL, NULL_RTX)) in_libcall = 1; else if (GET_RTX_CLASS (GET_CODE (p)) == 'i' && find_reg_note (p, REG_RETVAL, NULL_RTX)) in_libcall = 0; if (GET_CODE (p) == INSN && (set = single_set (p)) && GET_CODE (SET_DEST (set)) == REG && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set)))) { int tem1 = 0; int tem2 = 0; int move_insn = 0; rtx src = SET_SRC (set); rtx dependencies = 0; /* Figure out what to use as a source of this insn. If a REG_EQUIV note is given or if a REG_EQUAL note with a constant operand is specified, use it as the source and mark that we should move this insn by calling emit_move_insn rather that duplicating the insn. Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note is present. */ temp = find_reg_note (p, REG_EQUIV, NULL_RTX); if (temp) src = XEXP (temp, 0), move_insn = 1; else { temp = find_reg_note (p, REG_EQUAL, NULL_RTX); if (temp && CONSTANT_P (XEXP (temp, 0))) src = XEXP (temp, 0), move_insn = 1; if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX)) { src = XEXP (temp, 0); /* A libcall block can use regs that don't appear in the equivalent expression. To move the libcall, we must move those regs too. */ dependencies = libcall_other_reg (p, src); } } /* Don't try to optimize a register that was made by loop-optimization for an inner loop. We don't know its life-span, so we can't compute the benefit. */ if (REGNO (SET_DEST (set)) >= max_reg_before_loop) ; else if (/* The register is used in basic blocks other than the one where it is set (meaning that something after this point in the loop might depend on its value before the set). */ ! reg_in_basic_block_p (p, SET_DEST (set)) /* And the set is not guaranteed to be executed one the loop starts, or the value before the set is needed before the set occurs... ??? Note we have quadratic behaviour here, mitigated by the fact that the previous test will often fail for large loops. Rather than re-scanning the entire loop each time for register usage, we should build tables of the register usage and use them here instead. */ && (maybe_never || loop_reg_used_before_p (set, p, loop_start, scan_start, end))) /* It is unsafe to move the set. This code used to consider it OK to move a set of a variable which was not created by the user and not used in an exit test. That behavior is incorrect and was removed. */ ; else if ((tem = invariant_p (src)) && (dependencies == 0 || (tem2 = invariant_p (dependencies)) != 0) && (VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))) == 1 || (tem1 = consec_sets_invariant_p (SET_DEST (set), VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))), p))) /* If the insn can cause a trap (such as divide by zero), can't move it unless it's guaranteed to be executed once loop is entered. Even a function call might prevent the trap insn from being reached (since it might exit!) */ && ! ((maybe_never || call_passed) && may_trap_p (src))) { register struct movable *m; register int regno = REGNO (SET_DEST (set)); /* A potential lossage is where we have a case where two insns can be combined as long as they are both in the loop, but we move one of them outside the loop. For large loops, this can lose. The most common case of this is the address of a function being called. Therefore, if this register is marked as being used exactly once if we are in a loop with calls (a "large loop"), see if we can replace the usage of this register with the source of this SET. If we can, delete this insn. Don't do this if P has a REG_RETVAL note or if we have SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */ if (loop_has_call && VARRAY_RTX (reg_single_usage, regno) != 0 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx && REGNO_FIRST_UID (regno) == INSN_UID (p) && (REGNO_LAST_UID (regno) == INSN_UID (VARRAY_RTX (reg_single_usage, regno))) && VARRAY_INT (set_in_loop, regno) == 1 && ! side_effects_p (SET_SRC (set)) && ! find_reg_note (p, REG_RETVAL, NULL_RTX) && ! find_reg_note (p, REG_LABEL, NULL_RTX) && (! SMALL_REGISTER_CLASSES || (! (GET_CODE (SET_SRC (set)) == REG && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER))) /* This test is not redundant; SET_SRC (set) might be a call-clobbered register and the life of REGNO might span a call. */ && ! modified_between_p (SET_SRC (set), p, VARRAY_RTX (reg_single_usage, regno)) && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno)) && validate_replace_rtx (SET_DEST (set), SET_SRC (set), VARRAY_RTX (reg_single_usage, regno))) { /* Replace any usage in a REG_EQUAL note. Must copy the new source, so that we don't get rtx sharing between the SET_SOURCE and REG_NOTES of insn p. */ REG_NOTES (VARRAY_RTX (reg_single_usage, regno)) = replace_rtx (REG_NOTES (VARRAY_RTX (reg_single_usage, regno)), SET_DEST (set), copy_rtx (SET_SRC (set))); PUT_CODE (p, NOTE); NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED; NOTE_SOURCE_FILE (p) = 0; VARRAY_INT (set_in_loop, regno) = 0; continue; } m = (struct movable *) alloca (sizeof (struct movable)); m->next = 0; m->insn = p; m->set_src = src; m->dependencies = dependencies; m->set_dest = SET_DEST (set); m->force = 0; m->consec = VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))) - 1; m->done = 0; m->forces = 0; m->partial = 0; m->move_insn = move_insn; m->move_insn_first = 0; m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0); m->savemode = VOIDmode; m->regno = regno; /* Set M->cond if either invariant_p or consec_sets_invariant_p returned 2 (only conditionally invariant). */ m->cond = ((tem | tem1 | tem2) > 1); m->global = (uid_luid[REGNO_LAST_UID (regno)] > INSN_LUID (end) || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start)); m->match = 0; m->lifetime = (uid_luid[REGNO_LAST_UID (regno)] - uid_luid[REGNO_FIRST_UID (regno)]); m->savings = VARRAY_INT (n_times_set, regno); if (find_reg_note (p, REG_RETVAL, NULL_RTX)) m->savings += libcall_benefit (p); VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1; /* Add M to the end of the chain MOVABLES. */ if (movables == 0) movables = m; else last_movable->next = m; last_movable = m; if (m->consec > 0) { /* It is possible for the first instruction to have a REG_EQUAL note but a non-invariant SET_SRC, so we must remember the status of the first instruction in case the last instruction doesn't have a REG_EQUAL note. */ m->move_insn_first = m->move_insn; /* Skip this insn, not checking REG_LIBCALL notes. */ p = next_nonnote_insn (p); /* Skip the consecutive insns, if there are any. */ p = skip_consec_insns (p, m->consec); /* Back up to the last insn of the consecutive group. */ p = prev_nonnote_insn (p); /* We must now reset m->move_insn, m->is_equiv, and possibly m->set_src to correspond to the effects of all the insns. */ temp = find_reg_note (p, REG_EQUIV, NULL_RTX); if (temp) m->set_src = XEXP (temp, 0), m->move_insn = 1; else { temp = find_reg_note (p, REG_EQUAL, NULL_RTX); if (temp && CONSTANT_P (XEXP (temp, 0))) m->set_src = XEXP (temp, 0), m->move_insn = 1; else m->move_insn = 0; } m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0); } } /* If this register is always set within a STRICT_LOW_PART or set to zero, then its high bytes are constant. So clear them outside the loop and within the loop just load the low bytes. We must check that the machine has an instruction to do so. Also, if the value loaded into the register depends on the same register, this cannot be done. */ else if (SET_SRC (set) == const0_rtx && GET_CODE (NEXT_INSN (p)) == INSN && (set1 = single_set (NEXT_INSN (p))) && GET_CODE (set1) == SET && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART) && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG) && (SUBREG_REG (XEXP (SET_DEST (set1), 0)) == SET_DEST (set)) && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1))) { register int regno = REGNO (SET_DEST (set)); if (VARRAY_INT (set_in_loop, regno) == 2) { register struct movable *m; m = (struct movable *) alloca (sizeof (struct movable)); m->next = 0; m->insn = p; m->set_dest = SET_DEST (set); m->dependencies = 0; m->force = 0; m->consec = 0; m->done = 0; m->forces = 0; m->move_insn = 0; m->move_insn_first = 0; m->partial = 1; /* If the insn may not be executed on some cycles, we can't clear the whole reg; clear just high part. Not even if the reg is used only within this loop. Consider this: while (1) while (s != t) { if (foo ()) x = *s; use (x); } Clearing x before the inner loop could clobber a value being saved from the last time around the outer loop. However, if the reg is not used outside this loop and all uses of the register are in the same basic block as the store, there is no problem. If this insn was made by loop, we don't know its INSN_LUID and hence must make a conservative assumption. */ m->global = (INSN_UID (p) >= max_uid_for_loop || (uid_luid[REGNO_LAST_UID (regno)] > INSN_LUID (end)) || (uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (p)) || (labels_in_range_p (p, uid_luid[REGNO_FIRST_UID (regno)]))); if (maybe_never && m->global) m->savemode = GET_MODE (SET_SRC (set1)); else m->savemode = VOIDmode; m->regno = regno; m->cond = 0; m->match = 0; m->lifetime = (uid_luid[REGNO_LAST_UID (regno)] - uid_luid[REGNO_FIRST_UID (regno)]); m->savings = 1; VARRAY_INT (set_in_loop, regno) = -1; /* Add M to the end of the chain MOVABLES. */ if (movables == 0) movables = m; else last_movable->next = m; last_movable = m; } } } /* Past a call insn, we get to insns which might not be executed because the call might exit. This matters for insns that trap. Call insns inside a REG_LIBCALL/REG_RETVAL block always return, so they don't count. */ else if (GET_CODE (p) == CALL_INSN && ! in_libcall) call_passed = 1; /* Past a label or a jump, we get to insns for which we can't count on whether or how many times they will be executed during each iteration. Therefore, we can only move out sets of trivial variables (those not used after the loop). */ /* Similar code appears twice in strength_reduce. */ else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN) /* If we enter the loop in the middle, and scan around to the beginning, don't set maybe_never for that. This must be an unconditional jump, otherwise the code at the top of the loop might never be executed. Unconditional jumps are followed a by barrier then loop end. */ && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top && NEXT_INSN (NEXT_INSN (p)) == end && simplejump_p (p))) maybe_never = 1; else if (GET_CODE (p) == NOTE) { /* At the virtual top of a converted loop, insns are again known to be executed: logically, the loop begins here even though the exit code has been duplicated. */ if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0) maybe_never = call_passed = 0; else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG) loop_depth++; else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END) loop_depth--; } } /* If one movable subsumes another, ignore that other. */ ignore_some_movables (movables); /* For each movable insn, see if the reg that it loads leads when it dies right into another conditionally movable insn. If so, record that the second insn "forces" the first one, since the second can be moved only if the first is. */ force_movables (movables); /* See if there are multiple movable insns that load the same value. If there are, make all but the first point at the first one through the `match' field, and add the priorities of them all together as the priority of the first. */ combine_movables (movables, nregs); /* Now consider each movable insn to decide whether it is worth moving. Store 0 in set_in_loop for each reg that is moved. Generally this increases code size, so do not move moveables when optimizing for code size. */ if (! optimize_size) move_movables (movables, threshold, insn_count, loop_start, end, nregs); /* Now candidates that still are negative are those not moved. Change set_in_loop to indicate that those are not actually invariant. */ for (i = 0; i < nregs; i++) if (VARRAY_INT (set_in_loop, i) < 0) VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i); /* Now that we've moved some things out of the loop, we might be able to hoist even more memory references. */ load_mems_and_recount_loop_regs_set (scan_start, end, loop_top, loop_start, &insn_count); if (flag_strength_reduce) { the_movables = movables; strength_reduce (scan_start, end, loop_top, insn_count, loop_start, end, loop_cont, unroll_p, bct_p); } VARRAY_FREE (reg_single_usage); VARRAY_FREE (set_in_loop); VARRAY_FREE (n_times_set); VARRAY_FREE (may_not_optimize); } /* Add elements to *OUTPUT to record all the pseudo-regs mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */ void record_excess_regs (in_this, not_in_this, output) rtx in_this, not_in_this; rtx *output; { enum rtx_code code; char *fmt; int i; code = GET_CODE (in_this); switch (code) { case PC: case CC0: case CONST_INT: case CONST_DOUBLE: case CONST: case SYMBOL_REF: case LABEL_REF: return; case REG: if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER && ! reg_mentioned_p (in_this, not_in_this)) *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output); return; default: break; } fmt = GET_RTX_FORMAT (code); for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { int j; switch (fmt[i]) { case 'E': for (j = 0; j < XVECLEN (in_this, i); j++) record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output); break; case 'e': record_excess_regs (XEXP (in_this, i), not_in_this, output); break; } } } /* Check what regs are referred to in the libcall block ending with INSN, aside from those mentioned in the equivalent value. If there are none, return 0. If there are one or more, return an EXPR_LIST containing all of them. */ rtx libcall_other_reg (insn, equiv) rtx insn, equiv; { rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX); rtx p = XEXP (note, 0); rtx output = 0; /* First, find all the regs used in the libcall block that are not mentioned as inputs to the result. */ while (p != insn) { if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN || GET_CODE (p) == CALL_INSN) record_excess_regs (PATTERN (p), equiv, &output); p = NEXT_INSN (p); } return output; } /* Return 1 if all uses of REG are between INSN and the end of the basic block. */ static int reg_in_basic_block_p (insn, reg) rtx insn, reg; { int regno = REGNO (reg); rtx p; if (REGNO_FIRST_UID (regno) != INSN_UID (insn)) return 0; /* Search this basic block for the already recorded last use of the reg. */ for (p = insn; p; p = NEXT_INSN (p)) { switch (GET_CODE (p)) { case NOTE: break; case INSN: case CALL_INSN: /* Ordinary insn: if this is the last use, we win. */ if (REGNO_LAST_UID (regno) == INSN_UID (p)) return 1; break; case JUMP_INSN: /* Jump insn: if this is the last use, we win. */ if (REGNO_LAST_UID (regno) == INSN_UID (p)) return 1; /* Otherwise, it's the end of the basic block, so we lose. */ return 0; case CODE_LABEL: case BARRIER: /* It's the end of the basic block, so we lose. */ return 0; default: break; } } /* The "last use" doesn't follow the "first use"?? */ abort (); } /* Compute the benefit of eliminating the insns in the block whose last insn is LAST. This may be a group of insns used to compute a value directly or can contain a library call. */ static int libcall_benefit (last) rtx last; { rtx insn; int benefit = 0; for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0); insn != last; insn = NEXT_INSN (insn)) { if (GET_CODE (insn) == CALL_INSN) benefit += 10; /* Assume at least this many insns in a library routine. */ else if (GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) != USE && GET_CODE (PATTERN (insn)) != CLOBBER) benefit++; } return benefit; } /* Skip COUNT insns from INSN, counting library calls as 1 insn. */ static rtx skip_consec_insns (insn, count) rtx insn; int count; { for (; count > 0; count--) { rtx temp; /* If first insn of libcall sequence, skip to end. */ /* Do this at start of loop, since INSN is guaranteed to be an insn here. */ if (GET_CODE (insn) != NOTE && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX))) insn = XEXP (temp, 0); do insn = NEXT_INSN (insn); while (GET_CODE (insn) == NOTE); } return insn; } /* Ignore any movable whose insn falls within a libcall which is part of another movable. We make use of the fact that the movable for the libcall value was made later and so appears later on the chain. */ static void ignore_some_movables (movables) struct movable *movables; { register struct movable *m, *m1; for (m = movables; m; m = m->next) { /* Is this a movable for the value of a libcall? */ rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX); if (note) { rtx insn; /* Check for earlier movables inside that range, and mark them invalid. We cannot use LUIDs here because insns created by loop.c for prior loops don't have LUIDs. Rather than reject all such insns from movables, we just explicitly check each insn in the libcall (since invariant libcalls aren't that common). */ for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn)) for (m1 = movables; m1 != m; m1 = m1->next) if (m1->insn == insn) m1->done = 1; } } } /* For each movable insn, see if the reg that it loads leads when it dies right into another conditionally movable insn. If so, record that the second insn "forces" the first one, since the second can be moved only if the first is. */ static void force_movables (movables) struct movable *movables; { register struct movable *m, *m1; for (m1 = movables; m1; m1 = m1->next) /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */ if (!m1->partial && !m1->done) { int regno = m1->regno; for (m = m1->next; m; m = m->next) /* ??? Could this be a bug? What if CSE caused the register of M1 to be used after this insn? Since CSE does not update regno_last_uid, this insn M->insn might not be where it dies. But very likely this doesn't matter; what matters is that M's reg is computed from M1's reg. */ if (INSN_UID (m->insn) == REGNO_LAST_UID (regno) && !m->done) break; if (m != 0 && m->set_src == m1->set_dest /* If m->consec, m->set_src isn't valid. */ && m->consec == 0) m = 0; /* Increase the priority of the moving the first insn since it permits the second to be moved as well. */ if (m != 0) { m->forces = m1; m1->lifetime += m->lifetime; m1->savings += m->savings; } } } /* Find invariant expressions that are equal and can be combined into one register. */ static void combine_movables (movables, nregs) struct movable *movables; int nregs; { register struct movable *m; char *matched_regs = (char *) alloca (nregs); enum machine_mode mode; /* Regs that are set more than once are not allowed to match or be matched. I'm no longer sure why not. */ /* Perhaps testing m->consec_sets would be more appropriate here? */ for (m = movables; m; m = m->next) if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial) { register struct movable *m1; int regno = m->regno; bzero (matched_regs, nregs); matched_regs[regno] = 1; /* We want later insns to match the first one. Don't make the first one match any later ones. So start this loop at m->next. */ for (m1 = m->next; m1; m1 = m1->next) if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1 /* A reg used outside the loop mustn't be eliminated. */ && !m1->global /* A reg used for zero-extending mustn't be eliminated. */ && !m1->partial && (matched_regs[m1->regno] || ( /* Can combine regs with different modes loaded from the same constant only if the modes are the same or if both are integer modes with M wider or the same width as M1. The check for integer is redundant, but safe, since the only case of differing destination modes with equal sources is when both sources are VOIDmode, i.e., CONST_INT. For 2.95, don't do this if the mode of M1 is Pmode. This prevents us from substituting SUBREGs for REGs in memory accesses; not all targets are prepared to handle this properly. */ (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest) || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT && GET_MODE (m1->set_dest) != Pmode && (GET_MODE_BITSIZE (GET_MODE (m->set_dest)) >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest))))) /* See if the source of M1 says it matches M. */ && ((GET_CODE (m1->set_src) == REG && matched_regs[REGNO (m1->set_src)]) || rtx_equal_for_loop_p (m->set_src, m1->set_src, movables)))) && ((m->dependencies == m1->dependencies) || rtx_equal_p (m->dependencies, m1->dependencies))) { m->lifetime += m1->lifetime; m->savings += m1->savings; m1->done = 1; m1->match = m; matched_regs[m1->regno] = 1; } } /* Now combine the regs used for zero-extension. This can be done for those not marked `global' provided their lives don't overlap. */ for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode; mode = GET_MODE_WIDER_MODE (mode)) { register struct movable *m0 = 0; /* Combine all the registers for extension from mode MODE. Don't combine any that are used outside this loop. */ for (m = movables; m; m = m->next) if (m->partial && ! m->global && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn))))) { register struct movable *m1; int first = uid_luid[REGNO_FIRST_UID (m->regno)]; int last = uid_luid[REGNO_LAST_UID (m->regno)]; if (m0 == 0) { /* First one: don't check for overlap, just record it. */ m0 = m; continue; } /* Make sure they extend to the same mode. (Almost always true.) */ if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest)) continue; /* We already have one: check for overlap with those already combined together. */ for (m1 = movables; m1 != m; m1 = m1->next) if (m1 == m0 || (m1->partial && m1->match == m0)) if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last || uid_luid[REGNO_LAST_UID (m1->regno)] < first)) goto overlap; /* No overlap: we can combine this with the others. */ m0->lifetime += m->lifetime; m0->savings += m->savings; m->done = 1; m->match = m0; overlap: ; } } } /* Return 1 if regs X and Y will become the same if moved. */ static int regs_match_p (x, y, movables) rtx x, y; struct movable *movables; { int xn = REGNO (x); int yn = REGNO (y); struct movable *mx, *my; for (mx = movables; mx; mx = mx->next) if (mx->regno == xn) break; for (my = movables; my; my = my->next) if (my->regno == yn) break; return (mx && my && ((mx->match == my->match && mx->match != 0) || mx->match == my || mx == my->match)); } /* Return 1 if X and Y are identical-looking rtx's. This is the Lisp function EQUAL for rtx arguments. If two registers are matching movables or a movable register and an equivalent constant, consider them equal. */ static int rtx_equal_for_loop_p (x, y, movables) rtx x, y; struct movable *movables; { register int i; register int j; register struct movable *m; register enum rtx_code code; register char *fmt; if (x == y) return 1; if (x == 0 || y == 0) return 0; code = GET_CODE (x); /* If we have a register and a constant, they may sometimes be equal. */ if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2 && CONSTANT_P (y)) { for (m = movables; m; m = m->next) if (m->move_insn && m->regno == REGNO (x) && rtx_equal_p (m->set_src, y)) return 1; } else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2 && CONSTANT_P (x)) { for (m = movables; m; m = m->next) if (m->move_insn && m->regno == REGNO (y) && rtx_equal_p (m->set_src, x)) return 1; } /* Otherwise, rtx's of different codes cannot be equal. */ if (code != GET_CODE (y)) return 0; /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. (REG:SI x) and (REG:HI x) are NOT equivalent. */ if (GET_MODE (x) != GET_MODE (y)) return 0; /* These three types of rtx's can be compared nonrecursively. */ if (code == REG) return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables)); if (code == LABEL_REF) return XEXP (x, 0) == XEXP (y, 0); if (code == SYMBOL_REF) return XSTR (x, 0) == XSTR (y, 0); /* Compare the elements. If any pair of corresponding elements fail to match, return 0 for the whole things. */ fmt = GET_RTX_FORMAT (code); for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { switch (fmt[i]) { case 'w': if (XWINT (x, i) != XWINT (y, i)) return 0; break; case 'i': if (XINT (x, i) != XINT (y, i)) return 0; break; case 'E': /* Two vectors must have the same length. */ if (XVECLEN (x, i) != XVECLEN (y, i)) return 0; /* And the corresponding elements must match. */ for (j = 0; j < XVECLEN (x, i); j++) if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0) return 0; break; case 'e': if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0) return 0; break; case 's': if (strcmp (XSTR (x, i), XSTR (y, i))) return 0; break; case 'u': /* These are just backpointers, so they don't matter. */ break; case '0': break; /* It is believed that rtx's at this level will never contain anything but integers and other rtx's, except for within LABEL_REFs and SYMBOL_REFs. */ default: abort (); } } return 1; } /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all insns in INSNS which use thet reference. */ static void add_label_notes (x, insns) rtx x; rtx insns; { enum rtx_code code = GET_CODE (x); int i, j; char *fmt; rtx insn; if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x)) { /* This code used to ignore labels that referred to dispatch tables to avoid flow generating (slighly) worse code. We no longer ignore such label references (see LABEL_REF handling in mark_jump_label for additional information). */ for (insn = insns; insn; insn = NEXT_INSN (insn)) if (reg_mentioned_p (XEXP (x, 0), insn)) REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0), REG_NOTES (insn)); } fmt = GET_RTX_FORMAT (code); for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { if (fmt[i] == 'e') add_label_notes (XEXP (x, i), insns); else if (fmt[i] == 'E') for (j = XVECLEN (x, i) - 1; j >= 0; j--) add_label_notes (XVECEXP (x, i, j), insns); } } /* Scan MOVABLES, and move the insns that deserve to be moved. If two matching movables are combined, replace one reg with the other throughout. */ static void move_movables (movables, threshold, insn_count, loop_start, end, nregs) struct movable *movables; int threshold; int insn_count; rtx loop_start; rtx end; int nregs; { rtx new_start = 0; register struct movable *m; register rtx p; /* Map of pseudo-register replacements to handle combining when we move several insns that load the same value into different pseudo-registers. */ rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx)); char *already_moved = (char *) alloca (nregs); bzero (already_moved, nregs); bzero ((char *) reg_map, nregs * sizeof (rtx)); num_movables = 0; for (m = movables; m; m = m->next) { /* Describe this movable insn. */ if (loop_dump_stream) { fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ", INSN_UID (m->insn), m->regno, m->lifetime); if (m->consec > 0) fprintf (loop_dump_stream, "consec %d, ", m->consec); if (m->cond) fprintf (loop_dump_stream, "cond "); if (m->force) fprintf (loop_dump_stream, "force "); if (m->global) fprintf (loop_dump_stream, "global "); if (m->done) fprintf (loop_dump_stream, "done "); if (m->move_insn) fprintf (loop_dump_stream, "move-insn "); if (m->match) fprintf (loop_dump_stream, "matches %d ", INSN_UID (m->match->insn)); if (m->forces) fprintf (loop_dump_stream, "forces %d ", INSN_UID (m->forces->insn)); } /* Count movables. Value used in heuristics in strength_reduce. */ num_movables++; /* Ignore the insn if it's already done (it matched something else). Otherwise, see if it is now safe to move. */ if (!m->done && (! m->cond || (1 == invariant_p (m->set_src) && (m->dependencies == 0 || 1 == invariant_p (m->dependencies)) && (m->consec == 0 || 1 == consec_sets_invariant_p (m->set_dest, m->consec + 1, m->insn)))) && (! m->forces || m->forces->done)) { register int regno; register rtx p; int savings = m->savings; /* We have an insn that is safe to move. Compute its desirability. */ p = m->insn; regno = m->regno; if (loop_dump_stream) fprintf (loop_dump_stream, "savings %d ", savings); if (moved_once[regno] && loop_dump_stream) fprintf (loop_dump_stream, "halved since already moved "); /* An insn MUST be moved if we already moved something else which is safe only if this one is moved too: that is, if already_moved[REGNO] is nonzero. */ /* An insn is desirable to move if the new lifetime of the register is no more than THRESHOLD times the old lifetime. If it's not desirable, it means the loop is so big that moving won't speed things up much, and it is liable to make register usage worse. */ /* It is also desirable to move if it can be moved at no extra cost because something else was already moved. */ if (already_moved[regno] || flag_move_all_movables || (threshold * savings * m->lifetime) >= (moved_once[regno] ? insn_count * 2 : insn_count) || (m->forces && m->forces->done && VARRAY_INT (n_times_set, m->forces->regno) == 1)) { int count; register struct movable *m1; rtx first; /* Now move the insns that set the reg. */ if (m->partial && m->match) { rtx newpat, i1; rtx r1, r2; /* Find the end of this chain of matching regs. Thus, we load each reg in the chain from that one reg. And that reg is loaded with 0 directly, since it has ->match == 0. */ for (m1 = m; m1->match; m1 = m1->match); newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)), SET_DEST (PATTERN (m1->insn))); i1 = emit_insn_before (newpat, loop_start); /* Mark the moved, invariant reg as being allowed to share a hard reg with the other matching invariant. */ REG_NOTES (i1) = REG_NOTES (m->insn); r1 = SET_DEST (PATTERN (m->insn)); r2 = SET_DEST (PATTERN (m1->insn)); regs_may_share = gen_rtx_EXPR_LIST (VOIDmode, r1, gen_rtx_EXPR_LIST (VOIDmode, r2, regs_may_share)); delete_insn (m->insn); if (new_start == 0) new_start = i1; if (loop_dump_stream) fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1)); } /* If we are to re-generate the item being moved with a new move insn, first delete what we have and then emit the move insn before the loop. */ else if (m->move_insn) { rtx i1, temp; for (count = m->consec; count >= 0; count--) { /* If this is the first insn of a library call sequence, skip to the end. */ if (GET_CODE (p) != NOTE && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX))) p = XEXP (temp, 0); /* If this is the last insn of a libcall sequence, then delete every insn in the sequence except the last. The last insn is handled in the normal manner. */ if (GET_CODE (p) != NOTE && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX))) { temp = XEXP (temp, 0); while (temp != p) temp = delete_insn (temp); } temp = p; p = delete_insn (p); /* simplify_giv_expr expects that it can walk the insns at m->insn forwards and see this old sequence we are tossing here. delete_insn does preserve the next pointers, but when we skip over a NOTE we must fix it up. Otherwise that code walks into the non-deleted insn stream. */ while (p && GET_CODE (p) == NOTE) p = NEXT_INSN (temp) = NEXT_INSN (p); } start_sequence (); emit_move_insn (m->set_dest, m->set_src); temp = get_insns (); end_sequence (); add_label_notes (m->set_src, temp); i1 = emit_insns_before (temp, loop_start); if (! find_reg_note (i1, REG_EQUAL, NULL_RTX)) REG_NOTES (i1) = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL, m->set_src, REG_NOTES (i1)); if (loop_dump_stream) fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1)); /* The more regs we move, the less we like moving them. */ threshold -= 3; } else { for (count = m->consec; count >= 0; count--) { rtx i1, temp; /* If first insn of libcall sequence, skip to end. */ /* Do this at start of loop, since p is guaranteed to be an insn here. */ if (GET_CODE (p) != NOTE && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX))) p = XEXP (temp, 0); /* If last insn of libcall sequence, move all insns except the last before the loop. The last insn is handled in the normal manner. */ if (GET_CODE (p) != NOTE && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX))) { rtx fn_address = 0; rtx fn_reg = 0; rtx fn_address_insn = 0; first = 0; for (temp = XEXP (temp, 0); temp != p; temp = NEXT_INSN (temp)) { rtx body; rtx n; rtx next; if (GET_CODE (temp) == NOTE) continue; body = PATTERN (temp); /* Find the next insn after TEMP, not counting USE or NOTE insns. */ for (next = NEXT_INSN (temp); next != p; next = NEXT_INSN (next)) if (! (GET_CODE (next) == INSN && GET_CODE (PATTERN (next)) == USE) && GET_CODE (next) != NOTE) break; /* If that is the call, this may be the insn that loads the function address. Extract the function address from the insn that loads it into a register. If this insn was cse'd, we get incorrect code. So emit a new move insn that copies the function address into the register that the call insn will use. flow.c will delete any redundant stores that we have created. */ if (GET_CODE (next) == CALL_INSN && GET_CODE (body) == SET && GET_CODE (SET_DEST (body)) == REG && (n = find_reg_note (temp, REG_EQUAL, NULL_RTX))) { fn_reg = SET_SRC (body); if (GET_CODE (fn_reg) != REG) fn_reg = SET_DEST (body); fn_address = XEXP (n, 0); fn_address_insn = temp; } /* We have the call insn. If it uses the register we suspect it might, load it with the correct address directly. */ if (GET_CODE (temp) == CALL_INSN && fn_address != 0 && reg_referenced_p (fn_reg, body)) emit_insn_after (gen_move_insn (fn_reg, fn_address), fn_address_insn); if (GET_CODE (temp) == CALL_INSN) { i1 = emit_call_insn_before (body, loop_start); /* Because the USAGE information potentially contains objects other than hard registers we need to copy it. */ if (CALL_INSN_FUNCTION_USAGE (temp)) CALL_INSN_FUNCTION_USAGE (i1) = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp)); } else i1 = emit_insn_before (body, loop_start); if (first == 0) first = i1; if (temp == fn_address_insn) fn_address_insn = i1; REG_NOTES (i1) = REG_NOTES (temp); delete_insn (temp); } if (new_start == 0) new_start = first; } if (m->savemode != VOIDmode) { /* P sets REG to zero; but we should clear only the bits that are not covered by the mode m->savemode. */ rtx reg = m->set_dest; rtx sequence; rtx tem; start_sequence (); tem = expand_binop (GET_MODE (reg), and_optab, reg, GEN_INT ((((HOST_WIDE_INT) 1 << GET_MODE_BITSIZE (m->savemode))) - 1), reg, 1, OPTAB_LIB_WIDEN); if (tem == 0) abort (); if (tem != reg) emit_move_insn (reg, tem); sequence = gen_sequence (); end_sequence (); i1 = emit_insn_before (sequence, loop_start); } else if (GET_CODE (p) == CALL_INSN) { i1 = emit_call_insn_before (PATTERN (p), loop_start); /* Because the USAGE information potentially contains objects other than hard registers we need to copy it. */ if (CALL_INSN_FUNCTION_USAGE (p)) CALL_INSN_FUNCTION_USAGE (i1) = copy_rtx (CALL_INSN_FUNCTION_USAGE (p)); } else if (count == m->consec && m->move_insn_first) { /* The SET_SRC might not be invariant, so we must use the REG_EQUAL note. */ start_sequence (); emit_move_insn (m->set_dest, m->set_src); temp = get_insns (); end_sequence (); add_label_notes (m->set_src, temp); i1 = emit_insns_before (temp, loop_start); if (! find_reg_note (i1, REG_EQUAL, NULL_RTX)) REG_NOTES (i1) = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV : REG_EQUAL), m->set_src, REG_NOTES (i1)); } else i1 = emit_insn_before (PATTERN (p), loop_start); if (REG_NOTES (i1) == 0) { REG_NOTES (i1) = REG_NOTES (p); /* If there is a REG_EQUAL note present whose value is not loop invariant, then delete it, since it may cause problems with later optimization passes. It is possible for cse to create such notes like this as a result of record_jump_cond. */ if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX)) && ! invariant_p (XEXP (temp, 0))) remove_note (i1, temp); } if (new_start == 0) new_start = i1; if (loop_dump_stream) fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1)); /* If library call, now fix the REG_NOTES that contain insn pointers, namely REG_LIBCALL on FIRST and REG_RETVAL on I1. */ if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX))) { XEXP (temp, 0) = first; temp = find_reg_note (first, REG_LIBCALL, NULL_RTX); XEXP (temp, 0) = i1; } temp = p; delete_insn (p); p = NEXT_INSN (p); /* simplify_giv_expr expects that it can walk the insns at m->insn forwards and see this old sequence we are tossing here. delete_insn does preserve the next pointers, but when we skip over a NOTE we must fix it up. Otherwise that code walks into the non-deleted insn stream. */ while (p && GET_CODE (p) == NOTE) p = NEXT_INSN (temp) = NEXT_INSN (p); } /* The more regs we move, the less we like moving them. */ threshold -= 3; } /* Any other movable that loads the same register MUST be moved. */ already_moved[regno] = 1; /* This reg has been moved out of one loop. */ moved_once[regno] = 1; /* The reg set here is now invariant. */ if (! m->partial) VARRAY_INT (set_in_loop, regno) = 0; m->done = 1; /* Change the length-of-life info for the register to say it lives at least the full length of this loop. This will help guide optimizations in outer loops. */ if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start)) /* This is the old insn before all the moved insns. We can't use the moved insn because it is out of range in uid_luid. Only the old insns have luids. */ REGNO_FIRST_UID (regno) = INSN_UID (loop_start); if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (end)) REGNO_LAST_UID (regno) = INSN_UID (end); /* Combine with this moved insn any other matching movables. */ if (! m->partial) for (m1 = movables; m1; m1 = m1->next) if (m1->match == m) { rtx temp; /* Schedule the reg loaded by M1 for replacement so that shares the reg of M. If the modes differ (only possible in restricted circumstances, make a SUBREG. */ if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)) reg_map[m1->regno] = m->set_dest; else reg_map[m1->regno] = gen_lowpart_common (GET_MODE (m1->set_dest), m->set_dest); /* Get rid of the matching insn and prevent further processing of it. */ m1->done = 1; /* if library call, delete all insn except last, which is deleted below */ if ((temp = find_reg_note (m1->insn, REG_RETVAL, NULL_RTX))) { for (temp = XEXP (temp, 0); temp != m1->insn; temp = NEXT_INSN (temp)) delete_insn (temp); } delete_insn (m1->insn); /* Any other movable that loads the same register MUST be moved. */ already_moved[m1->regno] = 1; /* The reg merged here is now invariant, if the reg it matches is invariant. */ if (! m->partial) VARRAY_INT (set_in_loop, m1->regno) = 0; } } else if (loop_dump_stream) fprintf (loop_dump_stream, "not desirable"); } else if (loop_dump_stream && !m->match) fprintf (loop_dump_stream, "not safe"); if (loop_dump_stream) fprintf (loop_dump_stream, "\n"); } if (new_start == 0) new_start = loop_start; /* Go through all the instructions in the loop, making all the register substitutions scheduled in REG_MAP. */ for (p = new_start; p != end; p = NEXT_INSN (p)) if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN || GET_CODE (p) == CALL_INSN) { replace_regs (PATTERN (p), reg_map, nregs, 0); replace_regs (REG_NOTES (p), reg_map, nregs, 0); INSN_CODE (p) = -1; } } #if 0 /* Scan X and replace the address of any MEM in it with ADDR. REG is the address that MEM should have before the replacement. */ static void replace_call_address (x, reg, addr) rtx x, reg, addr; { register enum rtx_code code; register int i; register char *fmt; if (x == 0) return; code = GET_CODE (x); switch (code) { case PC: case CC0: case CONST_INT: case CONST_DOUBLE: case CONST: case SYMBOL_REF: case LABEL_REF: case REG: return; case SET: /* Short cut for very common case. */ replace_call_address (XEXP (x, 1), reg, addr); return; case CALL: /* Short cut for very common case. */ replace_call_address (XEXP (x, 0), reg, addr); return; case MEM: /* If this MEM uses a reg other than the one we expected, something is wrong. */ if (XEXP (x, 0) != reg) abort (); XEXP (x, 0) = addr; return; default: break; } fmt = GET_RTX_FORMAT (code); for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { if (fmt[i] == 'e') replace_call_address (XEXP (x, i), reg, addr); if (fmt[i] == 'E') { register int j; for (j = 0; j < XVECLEN (x, i); j++) replace_call_address (XVECEXP (x, i, j), reg, addr); } } } #endif /* Return the number of memory refs to addresses that vary in the rtx X. */ static int count_nonfixed_reads (x) rtx x; { register enum rtx_code code; register int i; register char *fmt; int value; if (x == 0) return 0; code = GET_CODE (x); switch (code) { case PC: case CC0: case CONST_INT: case CONST_DOUBLE: case CONST: case SYMBOL_REF: case LABEL_REF: case REG: return 0; case MEM: return ((invariant_p (XEXP (x, 0)) != 1) + count_nonfixed_reads (XEXP (x, 0))); default: break; } value = 0; fmt = GET_RTX_FORMAT (code); for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { if (fmt[i] == 'e') value += count_nonfixed_reads (XEXP (x, i)); if (fmt[i] == 'E') { register int j; for (j = 0; j < XVECLEN (x, i); j++) value += count_nonfixed_reads (XVECEXP (x, i, j)); } } return value; } #if 0 /* P is an instruction that sets a register to the result of a ZERO_EXTEND. Replace it with an instruction to load just the low bytes if the machine supports such an instruction, and insert above LOOP_START an instruction to clear the register. */ static void constant_high_bytes (p, loop_start) rtx p, loop_start; { register rtx new; register int insn_code_number; /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...))) to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */ new = gen_rtx_SET (VOIDmode, gen_rtx_STRICT_LOW_PART (VOIDmode, gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)), SET_DEST (PATTERN (p)), 0)), XEXP (SET_SRC (PATTERN (p)), 0)); insn_code_number = recog (new, p); if (insn_code_number) { register int i; /* Clear destination register before the loop. */ emit_insn_before (gen_rtx_SET (VOIDmode, SET_DEST (PATTERN (p)), const0_rtx), loop_start); /* Inside the loop, just load the low part. */ PATTERN (p) = new; } } #endif /* Scan a loop setting the variables `unknown_address_altered', `num_mem_sets', `loop_continue', `loops_enclosed', `loop_has_call', `loop_has_volatile', and `loop_has_tablejump'. Also, fill in the array `loop_mems' and the list `loop_store_mems'. */ static void prescan_loop (start, end) rtx start, end; { register int level = 1; rtx insn; int loop_has_multiple_exit_targets = 0; /* The label after END. Jumping here is just like falling off the end of the loop. We use next_nonnote_insn instead of next_label as a hedge against the (pathological) case where some actual insn might end up between the two. */ rtx exit_target = next_nonnote_insn (end); if (exit_target == NULL_RTX || GET_CODE (exit_target) != CODE_LABEL) loop_has_multiple_exit_targets = 1; unknown_address_altered = 0; loop_has_call = 0; loop_has_volatile = 0; loop_has_tablejump = 0; loop_store_mems = NULL_RTX; first_loop_store_insn = NULL_RTX; loop_mems_idx = 0; num_mem_sets = 0; loops_enclosed = 1; loop_continue = 0; for (insn = NEXT_INSN (start); insn != NEXT_INSN (end); insn = NEXT_INSN (insn)) { if (GET_CODE (insn) == NOTE) { if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG) { ++level; /* Count number of loops contained in this one. */ loops_enclosed++; } else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END) { --level; if (level == 0) { end = insn; break; } } else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT) { if (level == 1) loop_continue = insn; } } else if (GET_CODE (insn) == CALL_INSN) { if (! CONST_CALL_P (insn)) unknown_address_altered = 1; loop_has_call = 1; } else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN) { rtx label1 = NULL_RTX; rtx label2 = NULL_RTX; if (volatile_refs_p (PATTERN (insn))) loop_has_volatile = 1; if (GET_CODE (insn) == JUMP_INSN && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC || GET_CODE (PATTERN (insn)) == ADDR_VEC)) loop_has_tablejump = 1; note_stores (PATTERN (insn), note_addr_stored); if (! first_loop_store_insn && loop_store_mems) first_loop_store_insn = insn; if (! loop_has_multiple_exit_targets && GET_CODE (insn) == JUMP_INSN && GET_CODE (PATTERN (insn)) == SET && SET_DEST (PATTERN (insn)) == pc_rtx) { if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE) { label1 = XEXP (SET_SRC (PATTERN (insn)), 1); label2 = XEXP (SET_SRC (PATTERN (insn)), 2); } else { label1 = SET_SRC (PATTERN (insn)); } do { if (label1 && label1 != pc_rtx) { if (GET_CODE (label1) != LABEL_REF) { /* Something tricky. */ loop_has_multiple_exit_targets = 1; break; } else if (XEXP (label1, 0) != exit_target && LABEL_OUTSIDE_LOOP_P (label1)) { /* A jump outside the current loop. */ loop_has_multiple_exit_targets = 1; break; } } label1 = label2; label2 = NULL_RTX; } while (label1); } } else if (GET_CODE (insn) == RETURN) loop_has_multiple_exit_targets = 1; } /* Now, rescan the loop, setting up the LOOP_MEMS array. */ if (/* We can't tell what MEMs are aliased by what. */ !unknown_address_altered /* An exception thrown by a called function might land us anywhere. */ && !loop_has_call /* We don't want loads for MEMs moved to a location before the one at which their stack memory becomes allocated. (Note that this is not a problem for malloc, etc., since those require actual function calls. */ && !current_function_calls_alloca /* There are ways to leave the loop other than falling off the end. */ && !loop_has_multiple_exit_targets) for (insn = NEXT_INSN (start); insn != NEXT_INSN (end); insn = NEXT_INSN (insn)) for_each_rtx (&insn, insert_loop_mem, 0); } /* LOOP_NUMBER_CONT_DOMINATOR is now the last label between the loop start and the continue note that is a the destination of a (cond)jump after the continue note. If there is any (cond)jump between the loop start and what we have so far as LOOP_NUMBER_CONT_DOMINATOR that has a target between LOOP_DOMINATOR and the continue note, move LOOP_NUMBER_CONT_DOMINATOR forward to that label; if a jump's destination cannot be determined, clear LOOP_NUMBER_CONT_DOMINATOR. */ static void verify_dominator (loop_number) int loop_number; { rtx insn; if (! loop_number_cont_dominator[loop_number]) /* This can happen for an empty loop, e.g. in gcc.c-torture/compile/920410-2.c */ return; if (loop_number_cont_dominator[loop_number] == const0_rtx) { loop_number_cont_dominator[loop_number] = 0; return; } for (insn = loop_number_loop_starts[loop_number]; insn != loop_number_cont_dominator[loop_number]; insn = NEXT_INSN (insn)) { if (GET_CODE (insn) == JUMP_INSN && GET_CODE (PATTERN (insn)) != RETURN) { rtx label = JUMP_LABEL (insn); int label_luid; /* If it is not a jump we can easily understand or for which we do not have jump target information in the JUMP_LABEL field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear LOOP_NUMBER_CONT_DOMINATOR. */ if ((! condjump_p (insn) && ! condjump_in_parallel_p (insn)) || label == NULL_RTX) { loop_number_cont_dominator[loop_number] = NULL_RTX; return; } label_luid = INSN_LUID (label); if (label_luid < INSN_LUID (loop_number_loop_cont[loop_number]) && (label_luid > INSN_LUID (loop_number_cont_dominator[loop_number]))) loop_number_cont_dominator[loop_number] = label; } } } /* Scan the function looking for loops. Record the start and end of each loop. Also mark as invalid loops any loops that contain a setjmp or are branched to from outside the loop. */ static void find_and_verify_loops (f) rtx f; { rtx insn, label; int current_loop = -1; int next_loop = -1; int loop; compute_luids (f, NULL_RTX, 0); /* If there are jumps to undefined labels, treat them as jumps out of any/all loops. This also avoids writing past end of tables when there are no loops. */ uid_loop_num[0] = -1; /* Find boundaries of loops, mark which loops are contained within loops, and invalidate loops that have setjmp. */ for (insn = f; insn; insn = NEXT_INSN (insn)) { if (GET_CODE (insn) == NOTE) switch (NOTE_LINE_NUMBER (insn)) { case NOTE_INSN_LOOP_BEG: loop_number_loop_starts[++next_loop] = insn; loop_number_loop_ends[next_loop] = 0; loop_number_loop_cont[next_loop] = 0; loop_number_cont_dominator[next_loop] = 0; loop_outer_loop[next_loop] = current_loop; loop_invalid[next_loop] = 0; loop_number_exit_labels[next_loop] = 0; loop_number_exit_count[next_loop] = 0; current_loop = next_loop; break; case NOTE_INSN_SETJMP: /* In this case, we must invalidate our current loop and any enclosing loop. */ for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop]) { loop_invalid[loop] = 1; if (loop_dump_stream) fprintf (loop_dump_stream, "\nLoop at %d ignored due to setjmp.\n", INSN_UID (loop_number_loop_starts[loop])); } break; case NOTE_INSN_LOOP_CONT: loop_number_loop_cont[current_loop] = insn; break; case NOTE_INSN_LOOP_END: if (current_loop == -1) abort (); loop_number_loop_ends[current_loop] = insn; verify_dominator (current_loop); current_loop = loop_outer_loop[current_loop]; break; default: break; } /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT and NOTE_INSN_LOOP_END notes, update loop_number_loop_dominator. */ else if (GET_CODE (insn) == JUMP_INSN && GET_CODE (PATTERN (insn)) != RETURN && current_loop >= 0) { int this_loop; rtx label = JUMP_LABEL (insn); if (! condjump_p (insn) && ! condjump_in_parallel_p (insn)) label = NULL_RTX; this_loop = current_loop; do { /* First see if we care about this loop. */ if (loop_number_loop_cont[this_loop] && loop_number_cont_dominator[this_loop] != const0_rtx) { /* If the jump destination is not known, invalidate loop_number_const_dominator. */ if (! label) loop_number_cont_dominator[this_loop] = const0_rtx; else /* Check if the destination is between loop start and cont. */ if ((INSN_LUID (label) < INSN_LUID (loop_number_loop_cont[this_loop])) && (INSN_LUID (label) > INSN_LUID (loop_number_loop_starts[this_loop])) /* And if there is no later destination already recorded. */ && (! loop_number_cont_dominator[this_loop] || (INSN_LUID (label) > INSN_LUID (loop_number_cont_dominator [this_loop])))) loop_number_cont_dominator[this_loop] = label; } this_loop = loop_outer_loop[this_loop]; } while (this_loop >= 0); } /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the enclosing loop, but this doesn't matter. */ uid_loop_num[INSN_UID (insn)] = current_loop; } /* Any loop containing a label used in an initializer must be invalidated, because it can be jumped into from anywhere. */ for (label = forced_labels; label; label = XEXP (label, 1)) { int loop_num; for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))]; loop_num != -1; loop_num = loop_outer_loop[loop_num]) loop_invalid[loop_num] = 1; } /* Any loop containing a label used for an exception handler must be invalidated, because it can be jumped into from anywhere. */ for (label = exception_handler_labels; label; label = XEXP (label, 1)) { int loop_num; for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))]; loop_num != -1; loop_num = loop_outer_loop[loop_num]) loop_invalid[loop_num] = 1; } /* Now scan all insn's in the function. If any JUMP_INSN branches into a loop that it is not contained within, that loop is marked invalid. If any INSN or CALL_INSN uses a label's address, then the loop containing that label is marked invalid, because it could be jumped into from anywhere. Also look for blocks of code ending in an unconditional branch that exits the loop. If such a block is surrounded by a conditional branch around the block, move the block elsewhere (see below) and invert the jump to point to the code block. This may eliminate a label in our loop and will simplify processing by both us and a possible second cse pass. */ for (insn = f; insn; insn = NEXT_INSN (insn)) if (GET_RTX_CLASS (GET_CODE (insn)) == 'i') { int this_loop_num = uid_loop_num[INSN_UID (insn)]; if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN) { rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX); if (note) { int loop_num; for (loop_num = uid_loop_num[INSN_UID (XEXP (note, 0))]; loop_num != -1; loop_num = loop_outer_loop[loop_num]) loop_invalid[loop_num] = 1; } } if (GET_CODE (insn) != JUMP_INSN) continue; mark_loop_jump (PATTERN (insn), this_loop_num); /* See if this is an unconditional branch outside the loop. */ if (this_loop_num != -1 && (GET_CODE (PATTERN (insn)) == RETURN || (simplejump_p (insn) && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))] != this_loop_num))) && get_max_uid () < max_uid_for_loop) { rtx p; rtx our_next = next_real_insn (insn); rtx last_insn_to_move = NEXT_INSN (insn); int dest_loop; int outer_loop = -1; /* Go backwards until we reach the start of the loop, a label, or a JUMP_INSN. */ for (p = PREV_INSN (insn); GET_CODE (p) != CODE_LABEL && ! (GET_CODE (p) == NOTE && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG) && GET_CODE (p) != JUMP_INSN; p = PREV_INSN (p)) ; /* Check for the case where we have a jump to an inner nested loop, and do not perform the optimization in that case. */ if (JUMP_LABEL (insn)) { dest_loop = uid_loop_num[INSN_UID (JUMP_LABEL (insn))]; if (dest_loop != -1) { for (outer_loop = dest_loop; outer_loop != -1; outer_loop = loop_outer_loop[outer_loop]) if (outer_loop == this_loop_num) break; } } /* Make sure that the target of P is within the current loop. */ if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) && uid_loop_num[INSN_UID (JUMP_LABEL (p))] != this_loop_num) outer_loop = this_loop_num; /* If we stopped on a JUMP_INSN to the next insn after INSN, we have a block of code to try to move. We look backward and then forward from the target of INSN to find a BARRIER at the same loop depth as the target. If we find such a BARRIER, we make a new label for the start of the block, invert the jump in P and point it to that label, and move the block of code to the spot we found. */ if (outer_loop == -1 && GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) != 0 /* Just ignore jumps to labels that were never emitted. These always indicate compilation errors. */ && INSN_UID (JUMP_LABEL (p)) != 0 && condjump_p (p) && ! simplejump_p (p) && next_real_insn (JUMP_LABEL (p)) == our_next /* If it's not safe to move the sequence, then we mustn't try. */ && insns_safe_to_move_p (p, NEXT_INSN (insn), &last_insn_to_move)) { rtx target = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn (); int target_loop_num = uid_loop_num[INSN_UID (target)]; rtx loc, loc2; for (loc = target; loc; loc = PREV_INSN (loc)) if (GET_CODE (loc) == BARRIER /* Don't move things inside a tablejump. */ && ((loc2 = next_nonnote_insn (loc)) == 0 || GET_CODE (loc2) != CODE_LABEL || (loc2 = next_nonnote_insn (loc2)) == 0 || GET_CODE (loc2) != JUMP_INSN || (GET_CODE (PATTERN (loc2)) != ADDR_VEC && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC)) && uid_loop_num[INSN_UID (loc)] == target_loop_num) break; if (loc == 0) for (loc = target; loc; loc = NEXT_INSN (loc)) if (GET_CODE (loc) == BARRIER /* Don't move things inside a tablejump. */ && ((loc2 = next_nonnote_insn (loc)) == 0 || GET_CODE (loc2) != CODE_LABEL || (loc2 = next_nonnote_insn (loc2)) == 0 || GET_CODE (loc2) != JUMP_INSN || (GET_CODE (PATTERN (loc2)) != ADDR_VEC && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC)) && uid_loop_num[INSN_UID (loc)] == target_loop_num) break; if (loc) { rtx cond_label = JUMP_LABEL (p); rtx new_label = get_label_after (p); /* Ensure our label doesn't go away. */ LABEL_NUSES (cond_label)++; /* Verify that uid_loop_num is large enough and that we can invert P. */ if (invert_jump (p, new_label)) { rtx q, r; /* If no suitable BARRIER was found, create a suitable one before TARGET. Since TARGET is a fall through path, we'll need to insert an jump around our block and a add a BARRIER before TARGET. This creates an extra unconditional jump outside the loop. However, the benefits of removing rarely executed instructions from inside the loop usually outweighs the cost of the extra unconditional jump outside the loop. */ if (loc == 0) { rtx temp; temp = gen_jump (JUMP_LABEL (insn)); temp = emit_jump_insn_before (temp, target); JUMP_LABEL (temp) = JUMP_LABEL (insn); LABEL_NUSES (JUMP_LABEL (insn))++; loc = emit_barrier_before (target); } /* Include the BARRIER after INSN and copy the block after LOC. */ new_label = squeeze_notes (new_label, last_insn_to_move); reorder_insns (new_label, last_insn_to_move, loc); /* All those insns are now in TARGET_LOOP_NUM. */ for (q = new_label; q != NEXT_INSN (last_insn_to_move); q = NEXT_INSN (q)) uid_loop_num[INSN_UID (q)] = target_loop_num; /* The label jumped to by INSN is no longer a loop exit. Unless INSN does not have a label (e.g., it is a RETURN insn), search loop_number_exit_labels to find its label_ref, and remove it. Also turn off LABEL_OUTSIDE_LOOP_P bit. */ if (JUMP_LABEL (insn)) { int loop_num; for (q = 0, r = loop_number_exit_labels[this_loop_num]; r; q = r, r = LABEL_NEXTREF (r)) if (XEXP (r, 0) == JUMP_LABEL (insn)) { LABEL_OUTSIDE_LOOP_P (r) = 0; if (q) LABEL_NEXTREF (q) = LABEL_NEXTREF (r); else loop_number_exit_labels[this_loop_num] = LABEL_NEXTREF (r); break; } for (loop_num = this_loop_num; loop_num != -1 && loop_num != target_loop_num; loop_num = loop_outer_loop[loop_num]) loop_number_exit_count[loop_num]--; /* If we didn't find it, then something is wrong. */ if (! r) abort (); } /* P is now a jump outside the loop, so it must be put in loop_number_exit_labels, and marked as such. The easiest way to do this is to just call mark_loop_jump again for P. */ mark_loop_jump (PATTERN (p), this_loop_num); /* If INSN now jumps to the insn after it, delete INSN. */ if (JUMP_LABEL (insn) != 0 && (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn))) delete_insn (insn); } /* Continue the loop after where the conditional branch used to jump, since the only branch insn in the block (if it still remains) is an inter-loop branch and hence needs no processing. */ insn = NEXT_INSN (cond_label); if (--LABEL_NUSES (cond_label) == 0) delete_insn (cond_label); /* This loop will be continued with NEXT_INSN (insn). */ insn = PREV_INSN (insn); } } } } } /* If any label in X jumps to a loop different from LOOP_NUM and any of the loops it is contained in, mark the target loop invalid. For speed, we assume that X is part of a pattern of a JUMP_INSN. */ static void mark_loop_jump (x, loop_num) rtx x; int loop_num; { int dest_loop; int outer_loop; int i; switch (GET_CODE (x)) { case PC: case USE: case CLOBBER: case REG: case MEM: case CONST_INT: case CONST_DOUBLE: case RETURN: return; case CONST: /* There could be a label reference in here. */ mark_loop_jump (XEXP (x, 0), loop_num); return; case PLUS: case MINUS: case MULT: mark_loop_jump (XEXP (x, 0), loop_num); mark_loop_jump (XEXP (x, 1), loop_num); return; case LO_SUM: /* This may refer to a LABEL_REF or SYMBOL_REF. */ mark_loop_jump (XEXP (x, 1), loop_num); return; case SIGN_EXTEND: case ZERO_EXTEND: mark_loop_jump (XEXP (x, 0), loop_num); return; case LABEL_REF: dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))]; /* Link together all labels that branch outside the loop. This is used by final_[bg]iv_value and the loop unrolling code. Also mark this LABEL_REF so we know that this branch should predict false. */ /* A check to make sure the label is not in an inner nested loop, since this does not count as a loop exit. */ if (dest_loop != -1) { for (outer_loop = dest_loop; outer_loop != -1; outer_loop = loop_outer_loop[outer_loop]) if (outer_loop == loop_num) break; } else outer_loop = -1; if (loop_num != -1 && outer_loop == -1) { LABEL_OUTSIDE_LOOP_P (x) = 1; LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num]; loop_number_exit_labels[loop_num] = x; for (outer_loop = loop_num; outer_loop != -1 && outer_loop != dest_loop; outer_loop = loop_outer_loop[outer_loop]) loop_number_exit_count[outer_loop]++; } /* If this is inside a loop, but not in the current loop or one enclosed by it, it invalidates at least one loop. */ if (dest_loop == -1) return; /* We must invalidate every nested loop containing the target of this label, except those that also contain the jump insn. */ for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop]) { /* Stop when we reach a loop that also contains the jump insn. */ for (outer_loop = loop_num; outer_loop != -1; outer_loop = loop_outer_loop[outer_loop]) if (dest_loop == outer_loop) return; /* If we get here, we know we need to invalidate a loop. */ if (loop_dump_stream && ! loop_invalid[dest_loop]) fprintf (loop_dump_stream, "\nLoop at %d ignored due to multiple entry points.\n", INSN_UID (loop_number_loop_starts[dest_loop])); loop_invalid[dest_loop] = 1; } return; case SET: /* If this is not setting pc, ignore. */ if (SET_DEST (x) == pc_rtx) mark_loop_jump (SET_SRC (x), loop_num); return; case IF_THEN_ELSE: mark_loop_jump (XEXP (x, 1), loop_num); mark_loop_jump (XEXP (x, 2), loop_num); return; case PARALLEL: case ADDR_VEC: for (i = 0; i < XVECLEN (x, 0); i++) mark_loop_jump (XVECEXP (x, 0, i), loop_num); return; case ADDR_DIFF_VEC: for (i = 0; i < XVECLEN (x, 1); i++) mark_loop_jump (XVECEXP (x, 1, i), loop_num); return; default: /* Strictly speaking this is not a jump into the loop, only a possible jump out of the loop. However, we have no way to link the destination of this jump onto the list of exit labels. To be safe we mark this loop and any containing loops as invalid. */ if (loop_num != -1) { for (outer_loop = loop_num; outer_loop != -1; outer_loop = loop_outer_loop[outer_loop]) { if (loop_dump_stream && ! loop_invalid[outer_loop]) fprintf (loop_dump_stream, "\nLoop at %d ignored due to unknown exit jump.\n", INSN_UID (loop_number_loop_starts[outer_loop])); loop_invalid[outer_loop] = 1; } } return; } } /* Return nonzero if there is a label in the range from insn INSN to and including the insn whose luid is END INSN must have an assigned luid (i.e., it must not have been previously created by loop.c). */ static int labels_in_range_p (insn, end) rtx insn; int end; { while (insn && INSN_LUID (insn) <= end) { if (GET_CODE (insn) == CODE_LABEL) return 1; insn = NEXT_INSN (insn); } return 0; } /* Record that a memory reference X is being set. */ static void note_addr_stored (x, y) rtx x; rtx y ATTRIBUTE_UNUSED; { if (x == 0 || GET_CODE (x) != MEM) return; /* Count number of memory writes. This affects heuristics in strength_reduce. */ num_mem_sets++; /* BLKmode MEM means all memory is clobbered. */ if (GET_MODE (x) == BLKmode) unknown_address_altered = 1; if (unknown_address_altered) return; loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems); } /* X is a value modified by an INSN that references a biv inside a loop exit test (ie, X is somehow related to the value of the biv). If X is a pseudo that is used more than once, then the biv is (effectively) used more than once. */ static void note_set_pseudo_multiple_uses (x, y) rtx x; rtx y ATTRIBUTE_UNUSED; { if (x == 0) return; while (GET_CODE (x) == STRICT_LOW_PART || GET_CODE (x) == SIGN_EXTRACT || GET_CODE (x) == ZERO_EXTRACT || GET_CODE (x) == SUBREG) x = XEXP (x, 0); if (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER) return; /* If we do not have usage information, or if we know the register is used more than once, note that fact for check_dbra_loop. */ if (REGNO (x) >= max_reg_before_loop || ! VARRAY_RTX (reg_single_usage, REGNO (x)) || VARRAY_RTX (reg_single_usage, REGNO (x)) == const0_rtx) note_set_pseudo_multiple_uses_retval = 1; } /* Return nonzero if the rtx X is invariant over the current loop. The value is 2 if we refer to something only conditionally invariant. If `unknown_address_altered' is nonzero, no memory ref is invariant. Otherwise, a memory ref is invariant if it does not conflict with anything stored in `loop_store_mems'. */ int invariant_p (x) register rtx x; { register int i; register enum rtx_code code; register char *fmt; int conditional = 0; rtx mem_list_entry; if (x == 0) return 1; code = GET_CODE (x); switch (code) { case CONST_INT: case CONST_DOUBLE: case SYMBOL_REF: case CONST: return 1; case LABEL_REF: /* A LABEL_REF is normally invariant, however, if we are unrolling loops, and this label is inside the loop, then it isn't invariant. This is because each unrolled copy of the loop body will have a copy of this label. If this was invariant, then an insn loading the address of this label into a register might get moved outside the loop, and then each loop body would end up using the same label. We don't know the loop bounds here though, so just fail for all labels. */ if (flag_unroll_loops) return 0; else return 1; case PC: case CC0: case UNSPEC_VOLATILE: return 0; case REG: /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid since the reg might be set by initialization within the loop. */ if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx || x == arg_pointer_rtx) && ! current_function_has_nonlocal_goto) return 1; if (loop_has_call && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)]) return 0; if (VARRAY_INT (set_in_loop, REGNO (x)) < 0) return 2; return VARRAY_INT (set_in_loop, REGNO (x)) == 0; case MEM: /* Volatile memory references must be rejected. Do this before checking for read-only items, so that volatile read-only items will be rejected also. */ if (MEM_VOLATILE_P (x)) return 0; /* Read-only items (such as constants in a constant pool) are invariant if their address is. */ if (RTX_UNCHANGING_P (x)) break; /* If we had a subroutine call, any location in memory could have been clobbered. */ if (unknown_address_altered) return 0; /* See if there is any dependence between a store and this load. */ mem_list_entry = loop_store_mems; while (mem_list_entry) { if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode, x, rtx_varies_p)) return 0; mem_list_entry = XEXP (mem_list_entry, 1); } /* It's not invalidated by a store in memory but we must still verify the address is invariant. */ break; case ASM_OPERANDS: /* Don't mess with insns declared volatile. */ if (MEM_VOLATILE_P (x)) return 0; break; default: break; } fmt = GET_RTX_FORMAT (code); for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { if (fmt[i] == 'e') { int tem = invariant_p (XEXP (x, i)); if (tem == 0) return 0; if (tem == 2) conditional = 1; } else if (fmt[i] == 'E') { register int j; for (j = 0; j < XVECLEN (x, i); j++) { int tem = invariant_p (XVECEXP (x, i, j)); if (tem == 0) return 0; if (tem == 2) conditional = 1; } } } return 1 + conditional; } /* Return nonzero if all the insns in the loop that set REG are INSN and the immediately following insns, and if each of those insns sets REG in an invariant way (not counting uses of REG in them). The value is 2 if some of these insns are only conditionally invariant. We assume that INSN itself is the first set of REG and that its source is invariant. */ static int consec_sets_invariant_p (reg, n_sets, insn) int n_sets; rtx reg, insn; { register rtx p = insn; register int regno = REGNO (reg); rtx temp; /* Number of sets we have to insist on finding after INSN. */ int count = n_sets - 1; int old = VARRAY_INT (set_in_loop, regno); int value = 0; int this; /* If N_SETS hit the limit, we can't rely on its value. */ if (n_sets == 127) return 0; VARRAY_INT (set_in_loop, regno) = 0; while (count > 0) { register enum rtx_code code; rtx set; p = NEXT_INSN (p); code = GET_CODE (p); /* If library call, skip to end of it. */ if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX))) p = XEXP (temp, 0); this = 0; if (code == INSN && (set = single_set (p)) && GET_CODE (SET_DEST (set)) == REG && REGNO (SET_DEST (set)) == regno) { this = invariant_p (SET_SRC (set)); if (this != 0) value |= this; else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))) { /* If this is a libcall, then any invariant REG_EQUAL note is OK. If this is an ordinary insn, then only CONSTANT_P REG_EQUAL notes are OK. */ this = (CONSTANT_P (XEXP (temp, 0)) || (find_reg_note (p, REG_RETVAL, NULL_RTX) && invariant_p (XEXP (temp, 0)))); if (this != 0) value |= this; } } if (this != 0) count--; else if (code != NOTE) { VARRAY_INT (set_in_loop, regno) = old; return 0; } } VARRAY_INT (set_in_loop, regno) = old; /* If invariant_p ever returned 2, we return 2. */ return 1 + (value & 2); } #if 0 /* I don't think this condition is sufficient to allow INSN to be moved, so we no longer test it. */ /* Return 1 if all insns in the basic block of INSN and following INSN that set REG are invariant according to TABLE. */ static int all_sets_invariant_p (reg, insn, table) rtx reg, insn; short *table; { register rtx p = insn; register int regno = REGNO (reg); while (1) { register enum rtx_code code; p = NEXT_INSN (p); code = GET_CODE (p); if (code == CODE_LABEL || code == JUMP_INSN) return 1; if (code == INSN && GET_CODE (PATTERN (p)) == SET && GET_CODE (SET_DEST (PATTERN (p))) == REG && REGNO (SET_DEST (PATTERN (p))) == regno) { if (!invariant_p (SET_SRC (PATTERN (p)), table)) return 0; } } } #endif /* 0 */ /* Look at all uses (not sets) of registers in X. For each, if it is the single use, set USAGE[REGNO] to INSN; if there was a previous use in a different insn, set USAGE[REGNO] to const0_rtx. */ static void find_single_use_in_loop (insn, x, usage) rtx insn; rtx x; varray_type usage; { enum rtx_code code = GET_CODE (x); char *fmt = GET_RTX_FORMAT (code); int i, j; if (code == REG) VARRAY_RTX (usage, REGNO (x)) = (VARRAY_RTX (usage, REGNO (x)) != 0 && VARRAY_RTX (usage, REGNO (x)) != insn) ? const0_rtx : insn; else if (code == SET) { /* Don't count SET_DEST if it is a REG; otherwise count things in SET_DEST because if a register is partially modified, it won't show up as a potential movable so we don't care how USAGE is set for it. */ if (GET_CODE (SET_DEST (x)) != REG) find_single_use_in_loop (insn, SET_DEST (x), usage); find_single_use_in_loop (insn, SET_SRC (x), usage); } else for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { if (fmt[i] == 'e' && XEXP (x, i) != 0) find_single_use_in_loop (insn, XEXP (x, i), usage); else if (fmt[i] == 'E') for (j = XVECLEN (x, i) - 1; j >= 0; j--) find_single_use_in_loop (insn, XVECEXP (x, i, j), usage); } } /* Count and record any set in X which is contained in INSN. Update MAY_NOT_MOVE and LAST_SET for any register set in X. */ static void count_one_set (insn, x, may_not_move, last_set) rtx insn, x; varray_type may_not_move; rtx *last_set; { if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG) /* Don't move a reg that has an explicit clobber. It's not worth the pain to try to do it correctly. */ VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1; if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER) { rtx dest = SET_DEST (x); while (GET_CODE (dest) == SUBREG || GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT || GET_CODE (dest) == STRICT_LOW_PART) dest = XEXP (dest, 0); if (GET_CODE (dest) == REG) { register int regno = REGNO (dest); /* If this is the first setting of this reg in current basic block, and it was set before, it must be set in two basic blocks, so it cannot be moved out of the loop. */ if (VARRAY_INT (set_in_loop, regno) > 0 && last_set[regno] == 0) VARRAY_CHAR (may_not_move, regno) = 1; /* If this is not first setting in current basic block, see if reg was used in between previous one and this. If so, neither one can be moved. */ if (last_set[regno] != 0 && reg_used_between_p (dest, last_set[regno], insn)) VARRAY_CHAR (may_not_move, regno) = 1; if (VARRAY_INT (set_in_loop, regno) < 127) ++VARRAY_INT (set_in_loop, regno); last_set[regno] = insn; } } } /* Increment SET_IN_LOOP at the index of each register that is modified by an insn between FROM and TO. If the value of an element of SET_IN_LOOP becomes 127 or more, stop incrementing it, to avoid overflow. Store in SINGLE_USAGE[I] the single insn in which register I is used, if it is only used once. Otherwise, it is set to 0 (for no uses) or const0_rtx for more than one use. This parameter may be zero, in which case this processing is not done. Store in *COUNT_PTR the number of actual instruction in the loop. We use this to decide what is worth moving out. */ /* last_set[n] is nonzero iff reg n has been set in the current basic block. In that case, it is the insn that last set reg n. */ static void count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs) register rtx from, to; varray_type may_not_move; varray_type single_usage; int *count_ptr; int nregs; { register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx)); register rtx insn; register int count = 0; bzero ((char *) last_set, nregs * sizeof (rtx)); for (insn = from; insn != to; insn = NEXT_INSN (insn)) { if (GET_RTX_CLASS (GET_CODE (insn)) == 'i') { ++count; /* Record registers that have exactly one use. */ find_single_use_in_loop (insn, PATTERN (insn), single_usage); /* Include uses in REG_EQUAL notes. */ if (REG_NOTES (insn)) find_single_use_in_loop (insn, REG_NOTES (insn), single_usage); if (GET_CODE (PATTERN (insn)) == SET || GET_CODE (PATTERN (insn)) == CLOBBER) count_one_set (insn, PATTERN (insn), may_not_move, last_set); else if (GET_CODE (PATTERN (insn)) == PARALLEL) { register int i; for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--) count_one_set (insn, XVECEXP (PATTERN (insn), 0, i), may_not_move, last_set); } } if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN) bzero ((char *) last_set, nregs * sizeof (rtx)); } *count_ptr = count; } /* Given a loop that is bounded by LOOP_START and LOOP_END and that is entered at SCAN_START, return 1 if the register set in SET contained in insn INSN is used by any insn that precedes INSN in cyclic order starting from the loop entry point. We don't want to use INSN_LUID here because if we restrict INSN to those that have a valid INSN_LUID, it means we cannot move an invariant out from an inner loop past two loops. */ static int loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end) rtx set, insn, loop_start, scan_start, loop_end; { rtx reg = SET_DEST (set); rtx p; /* Scan forward checking for register usage. If we hit INSN, we are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */ for (p = scan_start; p != insn; p = NEXT_INSN (p)) { if (GET_RTX_CLASS (GET_CODE (p)) == 'i' && reg_overlap_mentioned_p (reg, PATTERN (p))) return 1; if (p == loop_end) p = loop_start; } return 0; } /* A "basic induction variable" or biv is a pseudo reg that is set (within this loop) only by incrementing or decrementing it. */ /* A "general induction variable" or giv is a pseudo reg whose value is a linear function of a biv. */ /* Bivs are recognized by `basic_induction_var'; Givs by `general_induction_var'. */ /* Indexed by register number, indicates whether or not register is an induction variable, and if so what type. */ varray_type reg_iv_type; /* Indexed by register number, contains pointer to `struct induction' if register is an induction variable. This holds general info for all induction variables. */ varray_type reg_iv_info; /* Indexed by register number, contains pointer to `struct iv_class' if register is a basic induction variable. This holds info describing the class (a related group) of induction variables that the biv belongs to. */ struct iv_class **reg_biv_class; /* The head of a list which links together (via the next field) every iv class for the current loop. */ struct iv_class *loop_iv_list; /* Givs made from biv increments are always splittable for loop unrolling. Since there is no regscan info for them, we have to keep track of them separately. */ int first_increment_giv, last_increment_giv; /* Communication with routines called via `note_stores'. */ static rtx note_insn; /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */ static rtx addr_placeholder; /* ??? Unfinished optimizations, and possible future optimizations, for the strength reduction code. */ /* ??? The interaction of biv elimination, and recognition of 'constant' bivs, may cause problems. */ /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause performance problems. Perhaps don't eliminate things that can be combined with an addressing mode. Find all givs that have the same biv, mult_val, and add_val; then for each giv, check to see if its only use dies in a following memory address. If so, generate a new memory address and check to see if it is valid. If it is valid, then store the modified memory address, otherwise, mark the giv as not done so that it will get its own iv. */ /* ??? Could try to optimize branches when it is known that a biv is always positive. */ /* ??? When replace a biv in a compare insn, we should replace with closest giv so that an optimized branch can still be recognized by the combiner, e.g. the VAX acb insn. */ /* ??? Many of the checks involving uid_luid could be simplified if regscan was rerun in loop_optimize whenever a register was added or moved. Also, some of the optimizations could be a little less conservative. */ /* Perform strength reduction and induction variable elimination. Pseudo registers created during this function will be beyond the last valid index in several tables including n_times_set and regno_last_uid. This does not cause a problem here, because the added registers cannot be givs outside of their loop, and hence will never be reconsidered. But scan_loop must check regnos to make sure they are in bounds. SCAN_START is the first instruction in the loop, as the loop would actually be executed. END is the NOTE_INSN_LOOP_END. LOOP_TOP is the first instruction in the loop, as it is layed out in the instruction stream. LOOP_START is the NOTE_INSN_LOOP_BEG. LOOP_CONT is the NOTE_INSN_LOOP_CONT. */ static void strength_reduce (scan_start, end, loop_top, insn_count, loop_start, loop_end, loop_cont, unroll_p, bct_p) rtx scan_start; rtx end; rtx loop_top; int insn_count; rtx loop_start; rtx loop_end; rtx loop_cont; int unroll_p, bct_p ATTRIBUTE_UNUSED; { rtx p; rtx set; rtx inc_val; rtx mult_val; rtx dest_reg; rtx *location; /* This is 1 if current insn is not executed at least once for every loop iteration. */ int not_every_iteration = 0; /* This is 1 if current insn may be executed more than once for every loop iteration. */ int maybe_multiple = 0; /* This is 1 if we have past a branch back to the top of the loop (aka a loop latch). */ int past_loop_latch = 0; /* Temporary list pointers for traversing loop_iv_list. */ struct iv_class *bl, **backbl; /* Ratio of extra register life span we can justify for saving an instruction. More if loop doesn't call subroutines since in that case saving an insn makes more difference and more registers are available. */ /* ??? could set this to last value of threshold in move_movables */ int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs); /* Map of pseudo-register replacements. */ rtx *reg_map; int reg_map_size; int call_seen; rtx test; rtx end_insert_before; int loop_depth = 0; int n_extra_increment; struct loop_info loop_iteration_info; struct loop_info *loop_info = &loop_iteration_info; /* If scan_start points to the loop exit test, we have to be wary of subversive use of gotos inside expression statements. */ if (prev_nonnote_insn (scan_start) != prev_nonnote_insn (loop_start)) maybe_multiple = back_branch_in_range_p (scan_start, loop_start, loop_end); VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type"); VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info"); reg_biv_class = (struct iv_class **) alloca (max_reg_before_loop * sizeof (struct iv_class *)); bzero ((char *) reg_biv_class, (max_reg_before_loop * sizeof (struct iv_class *))); loop_iv_list = 0; addr_placeholder = gen_reg_rtx (Pmode); /* Save insn immediately after the loop_end. Insns inserted after loop_end must be put before this insn, so that they will appear in the right order (i.e. loop order). If loop_end is the end of the current function, then emit a NOTE_INSN_DELETED after loop_end and set end_insert_before to the dummy note insn. */ if (NEXT_INSN (loop_end) != 0) end_insert_before = NEXT_INSN (loop_end); else end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end); /* Scan through loop to find all possible bivs. */ for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top); p != NULL_RTX; p = next_insn_in_loop (p, scan_start, end, loop_top)) { if (GET_CODE (p) == INSN && (set = single_set (p)) && GET_CODE (SET_DEST (set)) == REG) { dest_reg = SET_DEST (set); if (REGNO (dest_reg) < max_reg_before_loop && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT) { if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)), dest_reg, p, &inc_val, &mult_val, &location)) { /* It is a possible basic induction variable. Create and initialize an induction structure for it. */ struct induction *v = (struct induction *) alloca (sizeof (struct induction)); record_biv (v, p, dest_reg, inc_val, mult_val, location, not_every_iteration, maybe_multiple); REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT; } else if (REGNO (dest_reg) < max_reg_before_loop) REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT; } } /* Past CODE_LABEL, we get to insns that may be executed multiple times. The only way we can be sure that they can't is if every jump insn between here and the end of the loop either returns, exits the loop, is a jump to a location that is still behind the label, or is a jump to the loop start. */ if (GET_CODE (p) == CODE_LABEL) { rtx insn = p; maybe_multiple = 0; while (1) { insn = NEXT_INSN (insn); if (insn == scan_start) break; if (insn == end) { if (loop_top != 0) insn = loop_top; else break; if (insn == scan_start) break; } if (GET_CODE (insn) == JUMP_INSN && GET_CODE (PATTERN (insn)) != RETURN && (! condjump_p (insn) || (JUMP_LABEL (insn) != 0 && JUMP_LABEL (insn) != scan_start && ! loop_insn_first_p (p, JUMP_LABEL (insn))))) { maybe_multiple = 1; break; } } } /* Past a jump, we get to insns for which we can't count on whether they will be executed during each iteration. */ /* This code appears twice in strength_reduce. There is also similar code in scan_loop. */ if (GET_CODE (p) == JUMP_INSN /* If we enter the loop in the middle, and scan around to the beginning, don't set not_every_iteration for that. This can be any kind of jump, since we want to know if insns will be executed if the loop is executed. */ && ! (JUMP_LABEL (p) == loop_top && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p)) || (NEXT_INSN (p) == loop_end && condjump_p (p))))) { rtx label = 0; /* If this is a jump outside the loop, then it also doesn't matter. Check to see if the target of this branch is on the loop_number_exits_labels list. */ for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]; label; label = LABEL_NEXTREF (label)) if (XEXP (label, 0) == JUMP_LABEL (p)) break; if (! label) not_every_iteration = 1; } else if (GET_CODE (p) == NOTE) { /* At the virtual top of a converted loop, insns are again known to be executed each iteration: logically, the loop begins here even though the exit code has been duplicated. Insns are also again known to be executed each iteration at the LOOP_CONT note. */ if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT) && loop_depth == 0) not_every_iteration = 0; else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG) loop_depth++; else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END) loop_depth--; } /* Note if we pass a loop latch. If we do, then we can not clear NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in a loop since a jump before the last CODE_LABEL may have started a new loop iteration. Note that LOOP_TOP is only set for rotated loops and we need this check for all loops, so compare against the CODE_LABEL which immediately follows LOOP_START. */ if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == NEXT_INSN (loop_start)) past_loop_latch = 1; /* Unlike in the code motion pass where MAYBE_NEVER indicates that an insn may never be executed, NOT_EVERY_ITERATION indicates whether or not an insn is known to be executed each iteration of the loop, whether or not any iterations are known to occur. Therefore, if we have just passed a label and have no more labels between here and the test insn of the loop, and we have not passed a jump to the top of the loop, then we know these insns will be executed each iteration. */ if (not_every_iteration && ! past_loop_latch && GET_CODE (p) == CODE_LABEL && no_labels_between_p (p, loop_end) && loop_insn_first_p (p, loop_cont)) not_every_iteration = 0; } /* Scan loop_iv_list to remove all regs that proved not to be bivs. Make a sanity check against n_times_set. */ for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next) { if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT /* Above happens if register modified by subreg, etc. */ /* Make sure it is not recognized as a basic induction var: */ || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count /* If never incremented, it is invariant that we decided not to move. So leave it alone. */ || ! bl->incremented) { if (loop_dump_stream) fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n", bl->regno, (REG_IV_TYPE (bl->regno) != BASIC_INDUCT ? "not induction variable" : (! bl->incremented ? "never incremented" : "count error"))); REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT; *backbl = bl->next; } else { backbl = &bl->next; if (loop_dump_stream) fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno); } } /* Exit if there are no bivs. */ if (! loop_iv_list) { /* Can still unroll the loop anyways, but indicate that there is no strength reduction info available. */ if (unroll_p) unroll_loop (loop_end, insn_count, loop_start, end_insert_before, loop_info, 0); return; } /* Find initial value for each biv by searching backwards from loop_start, halting at first label. Also record any test condition. */ call_seen = 0; for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p)) { note_insn = p; if (GET_CODE (p) == CALL_INSN) call_seen = 1; if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN || GET_CODE (p) == CALL_INSN) note_stores (PATTERN (p), record_initial); /* Record any test of a biv that branches around the loop if no store between it and the start of loop. We only care about tests with constants and registers and only certain of those. */ if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) != 0 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end) && (test = get_condition_for_loop (p)) != 0 && GET_CODE (XEXP (test, 0)) == REG && REGNO (XEXP (test, 0)) < max_reg_before_loop && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start) && bl->init_insn == 0) { /* If an NE test, we have an initial value! */ if (GET_CODE (test) == NE) { bl->init_insn = p; bl->init_set = gen_rtx_SET (VOIDmode, XEXP (test, 0), XEXP (test, 1)); } else bl->initial_test = test; } } /* Look at the each biv and see if we can say anything better about its initial value from any initializing insns set up above. (This is done in two passes to avoid missing SETs in a PARALLEL.) */ for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next) { rtx src; rtx note; if (! bl->init_insn) continue; /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value is a constant, use the value of that. */ if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL && CONSTANT_P (XEXP (note, 0))) || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL && CONSTANT_P (XEXP (note, 0)))) src = XEXP (note, 0); else src = SET_SRC (bl->init_set); if (loop_dump_stream) fprintf (loop_dump_stream, "Biv %d initialized at insn %d: initial value ", bl->regno, INSN_UID (bl->init_insn)); if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno]) || GET_MODE (src) == VOIDmode) && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start)) { bl->initial_value = src; if (loop_dump_stream) { if (GET_CODE (src) == CONST_INT) { fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src)); fputc ('\n', loop_dump_stream); } else { print_rtl (loop_dump_stream, src); fprintf (loop_dump_stream, "\n"); } } } else { struct iv_class *bl2 = 0; rtx increment; /* Biv initial value is not a simple move. If it is the sum of another biv and a constant, check if both bivs are incremented in lockstep. Then we are actually looking at a giv. For simplicity, we only handle the case where there is but a single increment, and the register is not used elsewhere. */ if (bl->biv_count == 1 && bl->regno < max_reg_before_loop && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end) && GET_CODE (src) == PLUS && GET_CODE (XEXP (src, 0)) == REG && CONSTANT_P (XEXP (src, 1)) && ((increment = biv_total_increment (bl, loop_start, loop_end)) != NULL_RTX)) { int regno = REGNO (XEXP (src, 0)); for (bl2 = loop_iv_list; bl2; bl2 = bl2->next) if (bl2->regno == regno) break; } /* Now, can we transform this biv into a giv? */ if (bl2 && bl2->biv_count == 1 && rtx_equal_p (increment, biv_total_increment (bl2, loop_start, loop_end)) /* init_insn is only set to insns that are before loop_start without any intervening labels. */ && ! reg_set_between_p (bl2->biv->src_reg, PREV_INSN (bl->init_insn), loop_start) /* The register from BL2 must be set before the register from BL is set, or we must be able to move the latter set after the former set. Currently there can't be any labels in-between when biv_toal_increment returns nonzero both times but we test it here in case some day some real cfg analysis gets used to set always_computable. */ && ((loop_insn_first_p (bl2->biv->insn, bl->biv->insn) && no_labels_between_p (bl2->biv->insn, bl->biv->insn)) || (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn, bl2->biv->insn) && no_jumps_between_p (bl->biv->insn, bl2->biv->insn))) && validate_change (bl->biv->insn, &SET_SRC (single_set (bl->biv->insn)), copy_rtx (src), 0)) { int loop_num = uid_loop_num[INSN_UID (loop_start)]; rtx dominator = loop_number_cont_dominator[loop_num]; rtx giv = bl->biv->src_reg; rtx giv_insn = bl->biv->insn; rtx after_giv = NEXT_INSN (giv_insn); if (loop_dump_stream) fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno); /* Let this giv be discovered by the generic code. */ REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT; /* We can get better optimization if we can move the giv setting before the first giv use. */ if (dominator && ! loop_insn_first_p (dominator, scan_start) && ! reg_set_between_p (bl2->biv->src_reg, loop_start, dominator) && ! reg_used_between_p (giv, loop_start, dominator) && ! reg_used_between_p (giv, giv_insn, loop_end)) { rtx p; rtx next; for (next = NEXT_INSN (dominator); ; next = NEXT_INSN (next)) { if ((GET_RTX_CLASS (GET_CODE (next)) == 'i' && (reg_mentioned_p (giv, PATTERN (next)) || reg_set_p (bl2->biv->src_reg, next))) || GET_CODE (next) == JUMP_INSN) break; #ifdef HAVE_cc0 if (GET_RTX_CLASS (GET_CODE (next)) != 'i' || ! sets_cc0_p (PATTERN (next))) #endif dominator = next; } if (loop_dump_stream) fprintf (loop_dump_stream, "move after insn %d\n", INSN_UID (dominator)); /* Avoid problems with luids by actually moving the insn and adjusting all luids in the range. */ reorder_insns (giv_insn, giv_insn, dominator); for (p = dominator; INSN_UID (p) >= max_uid_for_loop; ) p = PREV_INSN (p); compute_luids (giv_insn, after_giv, INSN_LUID (p)); /* If the only purpose of the init insn is to initialize this giv, delete it. */ if (single_set (bl->init_insn) && ! reg_used_between_p (giv, bl->init_insn, loop_start)) delete_insn (bl->init_insn); } else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn)) { rtx p = PREV_INSN (giv_insn); while (INSN_UID (p) >= max_uid_for_loop) p = PREV_INSN (p); reorder_insns (giv_insn, giv_insn, bl2->biv->insn); compute_luids (after_giv, NEXT_INSN (giv_insn), INSN_LUID (p)); } /* Remove this biv from the chain. */ if (bl->next) *bl = *bl->next; else { *backbl = 0; break; } } /* If we can't make it a giv, let biv keep initial value of "itself". */ else if (loop_dump_stream) fprintf (loop_dump_stream, "is complex\n"); } } /* If a biv is unconditionally incremented several times in a row, convert all but the last increment into a giv. */ /* Get an upper bound for the number of registers we might have after all bivs have been processed. */ first_increment_giv = max_reg_num (); for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next) n_extra_increment += bl->biv_count - 1; /* If the loop contains volatile memory references do not allow any replacements to take place, since this could loose the volatile markers. Disabled for the gcc-2.95 release. There are still some problems with giv recombination. We have a patch from Joern which should fix those problems. But the patch is fairly complex and not really suitable for the gcc-2.95 branch at this stage. */ if (0 && n_extra_increment && ! loop_has_volatile) { int nregs = first_increment_giv + n_extra_increment; /* Reallocate reg_iv_type and reg_iv_info. */ VARRAY_GROW (reg_iv_type, nregs); VARRAY_GROW (reg_iv_info, nregs); for (bl = loop_iv_list; bl; bl = bl->next) { struct induction **vp, *v, *next; int biv_dead_after_loop = 0; /* The biv increments lists are in reverse order. Fix this first. */ for (v = bl->biv, bl->biv = 0; v; v = next) { next = v->next_iv; v->next_iv = bl->biv; bl->biv = v; } /* We must guard against the case that an early exit between v->insn and next->insn leaves the biv live after the loop, since that would mean that we'd be missing an increment for the final value. The following test to set biv_dead_after_loop is like the first part of the test to set bl->eliminable. We don't check here if we can calculate the final value, since this can't succeed if we already know that there is a jump between v->insn and next->insn, yet next->always_executed is set and next->maybe_multiple is cleared. Such a combination implies that the jump destination is outside the loop. If we want to make this check more sophisticated, we should check each branch between v->insn and next->insn individually to see if the biv is dead at its destination. */ if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end) && bl->init_insn && INSN_UID (bl->init_insn) < max_uid_for_loop && (uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)) #ifdef HAVE_decrement_and_branch_until_zero && ! bl->nonneg #endif && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set))) biv_dead_after_loop = 1; for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;) { HOST_WIDE_INT offset; rtx set, add_val, old_reg, dest_reg, last_use_insn; int old_regno, new_regno; if (! v->always_executed || v->maybe_multiple || GET_CODE (v->add_val) != CONST_INT || ! next->always_executed || next->maybe_multiple || ! CONSTANT_P (next->add_val) || v->mult_val != const1_rtx || next->mult_val != const1_rtx || ! (biv_dead_after_loop || no_jumps_between_p (v->insn, next->insn))) { vp = &v->next_iv; continue; } offset = INTVAL (v->add_val); set = single_set (v->insn); add_val = plus_constant (next->add_val, offset); old_reg = v->dest_reg; dest_reg = gen_reg_rtx (v->mode); /* Unlike reg_iv_type / reg_iv_info, the other three arrays have been allocated with some slop space, so we may not actually need to reallocate them. If we do, the following if statement will be executed just once in this loop. */ if ((unsigned) max_reg_num () > n_times_set->num_elements) { /* Grow all the remaining arrays. */ VARRAY_GROW (set_in_loop, nregs); VARRAY_GROW (n_times_set, nregs); VARRAY_GROW (may_not_optimize, nregs); VARRAY_GROW (reg_single_usage, nregs); } if (! validate_change (next->insn, next->location, add_val, 0)) { vp = &v->next_iv; continue; } /* Here we can try to eliminate the increment by combining it into the uses. */ /* Set last_use_insn so that we can check against it. */ for (last_use_insn = v->insn, p = NEXT_INSN (v->insn); p != next->insn; p = next_insn_in_loop (p, scan_start, end, loop_top)) { if (GET_RTX_CLASS (GET_CODE (p)) != 'i') continue; if (reg_mentioned_p (old_reg, PATTERN (p))) { last_use_insn = p; } } /* If we can't get the LUIDs for the insns, we can't calculate the lifetime. This is likely from unrolling of an inner loop, so there is little point in making this a DEST_REG giv anyways. */ if (INSN_UID (v->insn) >= max_uid_for_loop || INSN_UID (last_use_insn) >= max_uid_for_loop || ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0)) { /* Change the increment at NEXT back to what it was. */ if (! validate_change (next->insn, next->location, next->add_val, 0)) abort (); vp = &v->next_iv; continue; } next->add_val = add_val; v->dest_reg = dest_reg; v->giv_type = DEST_REG; v->location = &SET_SRC (set); v->cant_derive = 0; v->combined_with = 0; v->maybe_dead = 0; v->derive_adjustment = 0; v->same = 0; v->ignore = 0; v->new_reg = 0; v->final_value = 0; v->same_insn = 0; v->auto_inc_opt = 0; v->unrolled = 0; v->shared = 0; v->derived_from = 0; v->always_computable = 1; v->always_executed = 1; v->replaceable = 1; v->no_const_addval = 0; old_regno = REGNO (old_reg); new_regno = REGNO (dest_reg); VARRAY_INT (set_in_loop, old_regno)--; VARRAY_INT (set_in_loop, new_regno) = 1; VARRAY_INT (n_times_set, old_regno)--; VARRAY_INT (n_times_set, new_regno) = 1; VARRAY_CHAR (may_not_optimize, new_regno) = 0; REG_IV_TYPE (new_regno) = GENERAL_INDUCT; REG_IV_INFO (new_regno) = v; /* Remove the increment from the list of biv increments, and record it as a giv. */ *vp = next; bl->biv_count--; v->next_iv = bl->giv; bl->giv = v; bl->giv_count++; v->benefit = rtx_cost (SET_SRC (set), SET); bl->total_benefit += v->benefit; /* Now replace the biv with DEST_REG in all insns between the replaced increment and the next increment, and remember the last insn that needed a replacement. */ for (last_use_insn = v->insn, p = NEXT_INSN (v->insn); p != next->insn; p = next_insn_in_loop (p, scan_start, end, loop_top)) { rtx note; if (GET_RTX_CLASS (GET_CODE (p)) != 'i') continue; if (reg_mentioned_p (old_reg, PATTERN (p))) { last_use_insn = p; if (! validate_replace_rtx (old_reg, dest_reg, p)) abort (); } for (note = REG_NOTES (p); note; note = XEXP (note, 1)) { if (GET_CODE (note) == EXPR_LIST) XEXP (note, 0) = replace_rtx (XEXP (note, 0), old_reg, dest_reg); } } v->last_use = last_use_insn; v->lifetime = INSN_LUID (v->insn) - INSN_LUID (last_use_insn); /* If the lifetime is zero, it means that this register is really a dead store. So mark this as a giv that can be ignored. This will not prevent the biv from being eliminated. */ if (v->lifetime == 0) v->ignore = 1; if (loop_dump_stream) fprintf (loop_dump_stream, "Increment %d of biv %d converted to giv %d.\n\n", INSN_UID (v->insn), old_regno, new_regno); } } } last_increment_giv = max_reg_num () - 1; /* Search the loop for general induction variables. */ /* A register is a giv if: it is only set once, it is a function of a biv and a constant (or invariant), and it is not a biv. */ not_every_iteration = 0; loop_depth = 0; p = scan_start; while (1) { p = NEXT_INSN (p); /* At end of a straight-in loop, we are done. At end of a loop entered at the bottom, scan the top. */ if (p == scan_start) break; if (p == end) { if (loop_top != 0) p = loop_top; else break; if (p == scan_start) break; } /* Look for a general induction variable in a register. */ if (GET_CODE (p) == INSN && (set = single_set (p)) && GET_CODE (SET_DEST (set)) == REG && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set)))) { rtx src_reg; rtx add_val; rtx mult_val; int benefit; rtx regnote = 0; rtx last_consec_insn; dest_reg = SET_DEST (set); if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER) continue; if (/* SET_SRC is a giv. */ (general_induction_var (SET_SRC (set), &src_reg, &add_val, &mult_val, 0, &benefit) /* Equivalent expression is a giv. */ || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX)) && general_induction_var (XEXP (regnote, 0), &src_reg, &add_val, &mult_val, 0, &benefit))) /* Don't try to handle any regs made by loop optimization. We have nothing on them in regno_first_uid, etc. */ && REGNO (dest_reg) < max_reg_before_loop /* Don't recognize a BASIC_INDUCT_VAR here. */ && dest_reg != src_reg /* This must be the only place where the register is set. */ && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1 /* or all sets must be consecutive and make a giv. */ || (benefit = consec_sets_giv (benefit, p, src_reg, dest_reg, &add_val, &mult_val, &last_consec_insn)))) { struct induction *v = (struct induction *) alloca (sizeof (struct induction)); /* If this is a library call, increase benefit. */ if (find_reg_note (p, REG_RETVAL, NULL_RTX)) benefit += libcall_benefit (p); /* Skip the consecutive insns, if there are any. */ if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1) p = last_consec_insn; record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit, DEST_REG, not_every_iteration, NULL_PTR, loop_start, loop_end); } } #ifndef DONT_REDUCE_ADDR /* Look for givs which are memory addresses. */ /* This resulted in worse code on a VAX 8600. I wonder if it still does. */ if (GET_CODE (p) == INSN) find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start, loop_end); #endif /* Update the status of whether giv can derive other givs. This can change when we pass a label or an insn that updates a biv. */ if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN || GET_CODE (p) == CODE_LABEL) update_giv_derive (p); /* Past a jump, we get to insns for which we can't count on whether they will be executed during each iteration. */ /* This code appears twice in strength_reduce. There is also similar code in scan_loop. */ if (GET_CODE (p) == JUMP_INSN /* If we enter the loop in the middle, and scan around to the beginning, don't set not_every_iteration for that. This can be any kind of jump, since we want to know if insns will be executed if the loop is executed. */ && ! (JUMP_LABEL (p) == loop_top && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p)) || (NEXT_INSN (p) == loop_end && condjump_p (p))))) { rtx label = 0; /* If this is a jump outside the loop, then it also doesn't matter. Check to see if the target of this branch is on the loop_number_exits_labels list. */ for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]; label; label = LABEL_NEXTREF (label)) if (XEXP (label, 0) == JUMP_LABEL (p)) break; if (! label) not_every_iteration = 1; } else if (GET_CODE (p) == NOTE) { /* At the virtual top of a converted loop, insns are again known to be executed each iteration: logically, the loop begins here even though the exit code has been duplicated. Insns are also again known to be executed each iteration at the LOOP_CONT note. */ if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT) && loop_depth == 0) not_every_iteration = 0; else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG) loop_depth++; else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END) loop_depth--; } /* Unlike in the code motion pass where MAYBE_NEVER indicates that an insn may never be executed, NOT_EVERY_ITERATION indicates whether or not an insn is known to be executed each iteration of the loop, whether or not any iterations are known to occur. Therefore, if we have just passed a label and have no more labels between here and the test insn of the loop, we know these insns will be executed each iteration. */ if (not_every_iteration && GET_CODE (p) == CODE_LABEL && no_labels_between_p (p, loop_end) && loop_insn_first_p (p, loop_cont)) not_every_iteration = 0; } /* Try to calculate and save the number of loop iterations. This is set to zero if the actual number can not be calculated. This must be called after all giv's have been identified, since otherwise it may fail if the iteration variable is a giv. */ loop_iterations (loop_start, loop_end, loop_info); /* Now for each giv for which we still don't know whether or not it is replaceable, check to see if it is replaceable because its final value can be calculated. This must be done after loop_iterations is called, so that final_giv_value will work correctly. */ for (bl = loop_iv_list; bl; bl = bl->next) { struct induction *v; for (v = bl->giv; v; v = v->next_iv) if (! v->replaceable && ! v->not_replaceable) check_final_value (v, loop_start, loop_end, loop_info->n_iterations); } /* Try to prove that the loop counter variable (if any) is always nonnegative; if so, record that fact with a REG_NONNEG note so that "decrement and branch until zero" insn can be used. */ check_dbra_loop (loop_end, insn_count, loop_start, loop_info); /* Create reg_map to hold substitutions for replaceable giv regs. Some givs might have been made from biv increments, so look at reg_iv_type for a suitable size. */ reg_map_size = reg_iv_type->num_elements; reg_map = (rtx *) alloca (reg_map_size * sizeof (rtx)); bzero ((char *) reg_map, reg_map_size * sizeof (rtx)); /* Examine each iv class for feasibility of strength reduction/induction variable elimination. */ for (bl = loop_iv_list; bl; bl = bl->next) { struct induction *v; int benefit; int all_reduced; rtx final_value = 0; unsigned nregs; /* Test whether it will be possible to eliminate this biv provided all givs are reduced. This is possible if either the reg is not used outside the loop, or we can compute what its final value will be. For architectures with a decrement_and_branch_until_zero insn, don't do this if we put a REG_NONNEG note on the endtest for this biv. */ /* Compare against bl->init_insn rather than loop_start. We aren't concerned with any uses of the biv between init_insn and loop_start since these won't be affected by the value of the biv elsewhere in the function, so long as init_insn doesn't use the biv itself. March 14, 1989 -- self@bayes.arc.nasa.gov */ if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end) && bl->init_insn && INSN_UID (bl->init_insn) < max_uid_for_loop && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn) #ifdef HAVE_decrement_and_branch_until_zero && ! bl->nonneg #endif && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set))) || ((final_value = final_biv_value (bl, loop_start, loop_end, loop_info->n_iterations)) #ifdef HAVE_decrement_and_branch_until_zero && ! bl->nonneg #endif )) bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0, threshold, insn_count); else { if (loop_dump_stream) { fprintf (loop_dump_stream, "Cannot eliminate biv %d.\n", bl->regno); fprintf (loop_dump_stream, "First use: insn %d, last use: insn %d.\n", REGNO_FIRST_UID (bl->regno), REGNO_LAST_UID (bl->regno)); } } /* Combine all giv's for this iv_class. */ combine_givs (bl); /* This will be true at the end, if all givs which depend on this biv have been strength reduced. We can't (currently) eliminate the biv unless this is so. */ all_reduced = 1; /* Check each giv in this class to see if we will benefit by reducing it. Skip giv's combined with others. */ for (v = bl->giv; v; v = v->next_iv) { struct induction *tv; if (v->ignore || v->same) continue; benefit = v->benefit; /* Reduce benefit if not replaceable, since we will insert a move-insn to replace the insn that calculates this giv. Don't do this unless the giv is a user variable, since it will often be marked non-replaceable because of the duplication of the exit code outside the loop. In such a case, the copies we insert are dead and will be deleted. So they don't have a cost. Similar situations exist. */ /* ??? The new final_[bg]iv_value code does a much better job of finding replaceable giv's, and hence this code may no longer be necessary. */ if (! v->replaceable && ! bl->eliminable && REG_USERVAR_P (v->dest_reg)) benefit -= copy_cost; /* Decrease the benefit to count the add-insns that we will insert to increment the reduced reg for the giv. */ benefit -= add_cost * bl->biv_count; /* Decide whether to strength-reduce this giv or to leave the code unchanged (recompute it from the biv each time it is used). This decision can be made independently for each giv. */ #ifdef AUTO_INC_DEC /* Attempt to guess whether autoincrement will handle some of the new add insns; if so, increase BENEFIT (undo the subtraction of add_cost that was done above). */ if (v->giv_type == DEST_ADDR && GET_CODE (v->mult_val) == CONST_INT) { if (HAVE_POST_INCREMENT && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode)) benefit += add_cost * bl->biv_count; else if (HAVE_PRE_INCREMENT && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode)) benefit += add_cost * bl->biv_count; else if (HAVE_POST_DECREMENT && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode)) benefit += add_cost * bl->biv_count; else if (HAVE_PRE_DECREMENT && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode)) benefit += add_cost * bl->biv_count; } #endif /* If an insn is not to be strength reduced, then set its ignore flag, and clear all_reduced. */ /* A giv that depends on a reversed biv must be reduced if it is used after the loop exit, otherwise, it would have the wrong value after the loop exit. To make it simple, just reduce all of such giv's whether or not we know they are used after the loop exit. */ if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count && ! bl->reversed ) { if (loop_dump_stream) fprintf (loop_dump_stream, "giv of insn %d not worth while, %d vs %d.\n", INSN_UID (v->insn), v->lifetime * threshold * benefit, insn_count); v->ignore = 1; all_reduced = 0; } else { /* Check that we can increment the reduced giv without a multiply insn. If not, reject it. */ for (tv = bl->biv; tv; tv = tv->next_iv) if (tv->mult_val == const1_rtx && ! product_cheap_p (tv->add_val, v->mult_val)) { if (loop_dump_stream) fprintf (loop_dump_stream, "giv of insn %d: would need a multiply.\n", INSN_UID (v->insn)); v->ignore = 1; all_reduced = 0; break; } } } /* Check for givs whose first use is their definition and whose last use is the definition of another giv. If so, it is likely dead and should not be used to derive another giv nor to eliminate a biv. */ for (v = bl->giv; v; v = v->next_iv) { if (v->ignore || (v->same && v->same->ignore)) continue; if (v->last_use) { struct induction *v1; for (v1 = bl->giv; v1; v1 = v1->next_iv) if (v->last_use == v1->insn) v->maybe_dead = 1; } else if (v->giv_type == DEST_REG && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn)) { struct induction *v1; for (v1 = bl->giv; v1; v1 = v1->next_iv) if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn)) v->maybe_dead = 1; } } /* Now that we know which givs will be reduced, try to rearrange the combinations to reduce register pressure. recombine_givs calls find_life_end, which needs reg_iv_type and reg_iv_info to be valid for all pseudos. We do the necessary reallocation here since it allows to check if there are still more bivs to process. */ nregs = max_reg_num (); if (nregs > reg_iv_type->num_elements) { /* If there are still more bivs to process, allocate some slack space so that we're not constantly reallocating these arrays. */ if (bl->next) nregs += nregs / 4; /* Reallocate reg_iv_type and reg_iv_info. */ VARRAY_GROW (reg_iv_type, nregs); VARRAY_GROW (reg_iv_info, nregs); } #if 0 /* Disabled for the gcc-2.95 release. There are still some problems with giv recombination. We have a patch from Joern which should fix those problems. But the patch is fairly complex and not really suitable for the gcc-2.95 branch at this stage. */ recombine_givs (bl, loop_start, loop_end, unroll_p); #endif /* Reduce each giv that we decided to reduce. */ for (v = bl->giv; v; v = v->next_iv) { struct induction *tv; if (! v->ignore && v->same == 0) { int auto_inc_opt = 0; /* If the code for derived givs immediately below has already allocated a new_reg, we must keep it. */ if (! v->new_reg) v->new_reg = gen_reg_rtx (v->mode); if (v->derived_from) { struct induction *d = v->derived_from; /* In case d->dest_reg is not replaceable, we have to replace it in v->insn now. */ if (! d->new_reg) d->new_reg = gen_reg_rtx (d->mode); PATTERN (v->insn) = replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg); PATTERN (v->insn) = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg); if (bl->biv_count != 1) { /* For each place where the biv is incremented, add an insn to set the new, reduced reg for the giv. */ for (tv = bl->biv; tv; tv = tv->next_iv) { /* We always emit reduced giv increments before the biv increment when bl->biv_count != 1. So by emitting the add insns for derived givs after the biv increment, they pick up the updated value of the reduced giv. */ emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn); } } continue; } #ifdef AUTO_INC_DEC /* If the target has auto-increment addressing modes, and this is an address giv, then try to put the increment immediately after its use, so that flow can create an auto-increment addressing mode. */ if (v->giv_type == DEST_ADDR && bl->biv_count == 1 && bl->biv->always_executed && ! bl->biv->maybe_multiple /* We don't handle reversed biv's because bl->biv->insn does not have a valid INSN_LUID. */ && ! bl->reversed && v->always_executed && ! v->maybe_multiple && INSN_UID (v->insn) < max_uid_for_loop) { /* If other giv's have been combined with this one, then this will work only if all uses of the other giv's occur before this giv's insn. This is difficult to check. We simplify this by looking for the common case where there is one DEST_REG giv, and this giv's insn is the last use of the dest_reg of that DEST_REG giv. If the increment occurs after the address giv, then we can perform the optimization. (Otherwise, the increment would have to go before other_giv, and we would not be able to combine it with the address giv to get an auto-inc address.) */ if (v->combined_with) { struct induction *other_giv = 0; for (tv = bl->giv; tv; tv = tv->next_iv) if (tv->same == v) { if (other_giv) break; else other_giv = tv; } if (! tv && other_giv && REGNO (other_giv->dest_reg) < max_reg_before_loop && (REGNO_LAST_UID (REGNO (other_giv->dest_reg)) == INSN_UID (v->insn)) && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn)) auto_inc_opt = 1; } /* Check for case where increment is before the address giv. Do this test in "loop order". */ else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn) && (INSN_LUID (v->insn) < INSN_LUID (scan_start) || (INSN_LUID (bl->biv->insn) > INSN_LUID (scan_start)))) || (INSN_LUID (v->insn) < INSN_LUID (scan_start) && (INSN_LUID (scan_start) < INSN_LUID (bl->biv->insn)))) auto_inc_opt = -1; else auto_inc_opt = 1; #ifdef HAVE_cc0 { rtx prev; /* We can't put an insn immediately after one setting cc0, or immediately before one using cc0. */ if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn))) || (auto_inc_opt == -1 && (prev = prev_nonnote_insn (v->insn)) != 0 && GET_RTX_CLASS (GET_CODE (prev)) == 'i' && sets_cc0_p (PATTERN (prev)))) auto_inc_opt = 0; } #endif if (auto_inc_opt) v->auto_inc_opt = 1; } #endif /* For each place where the biv is incremented, add an insn to increment the new, reduced reg for the giv. */ for (tv = bl->biv; tv; tv = tv->next_iv) { rtx insert_before; if (! auto_inc_opt) insert_before = tv->insn; else if (auto_inc_opt == 1) insert_before = NEXT_INSN (v->insn); else insert_before = v->insn; if (tv->mult_val == const1_rtx) emit_iv_add_mult (tv->add_val, v->mult_val, v->new_reg, v->new_reg, insert_before); else /* tv->mult_val == const0_rtx */ /* A multiply is acceptable here since this is presumed to be seldom executed. */ emit_iv_add_mult (tv->add_val, v->mult_val, v->add_val, v->new_reg, insert_before); } /* Add code at loop start to initialize giv's reduced reg. */ emit_iv_add_mult (bl->initial_value, v->mult_val, v->add_val, v->new_reg, loop_start); } } /* Rescan all givs. If a giv is the same as a giv not reduced, mark it as not reduced. For each giv register that can be reduced now: if replaceable, substitute reduced reg wherever the old giv occurs; else add new move insn "giv_reg = reduced_reg". */ for (v = bl->giv; v; v = v->next_iv) { if (v->same && v->same->ignore) v->ignore = 1; if (v->ignore) continue; /* Update expression if this was combined, in case other giv was replaced. */ if (v->same) v->new_reg = replace_rtx (v->new_reg, v->same->dest_reg, v->same->new_reg); if (v->giv_type == DEST_ADDR) /* Store reduced reg as the address in the memref where we found this giv. */ validate_change (v->insn, v->location, v->new_reg, 0); else if (v->replaceable) { reg_map[REGNO (v->dest_reg)] = v->new_reg; #if 0 /* I can no longer duplicate the original problem. Perhaps this is unnecessary now? */ /* Replaceable; it isn't strictly necessary to delete the old insn and emit a new one, because v->dest_reg is now dead. However, especially when unrolling loops, the special handling for (set REG0 REG1) in the second cse pass may make v->dest_reg live again. To avoid this problem, emit an insn to set the original giv reg from the reduced giv. We can not delete the original insn, since it may be part of a LIBCALL, and the code in flow that eliminates dead libcalls will fail if it is deleted. */ emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg), v->insn); #endif } else { /* Not replaceable; emit an insn to set the original giv reg from the reduced giv, same as above. */ emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg), v->insn); } /* When a loop is reversed, givs which depend on the reversed biv, and which are live outside the loop, must be set to their correct final value. This insn is only needed if the giv is not replaceable. The correct final value is the same as the value that the giv starts the reversed loop with. */ if (bl->reversed && ! v->replaceable) emit_iv_add_mult (bl->initial_value, v->mult_val, v->add_val, v->dest_reg, end_insert_before); else if (v->final_value) { rtx insert_before; /* If the loop has multiple exits, emit the insn before the loop to ensure that it will always be executed no matter how the loop exits. Otherwise, emit the insn after the loop, since this is slightly more efficient. */ if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]) insert_before = loop_start; else insert_before = end_insert_before; emit_insn_before (gen_move_insn (v->dest_reg, v->final_value), insert_before); #if 0 /* If the insn to set the final value of the giv was emitted before the loop, then we must delete the insn inside the loop that sets it. If this is a LIBCALL, then we must delete every insn in the libcall. Note, however, that final_giv_value will only succeed when there are multiple exits if the giv is dead at each exit, hence it does not matter that the original insn remains because it is dead anyways. */ /* Delete the insn inside the loop that sets the giv since the giv is now set before (or after) the loop. */ delete_insn (v->insn); #endif } if (loop_dump_stream) { fprintf (loop_dump_stream, "giv at %d reduced to ", INSN_UID (v->insn)); print_rtl (loop_dump_stream, v->new_reg); fprintf (loop_dump_stream, "\n"); } } /* All the givs based on the biv bl have been reduced if they merit it. */ /* For each giv not marked as maybe dead that has been combined with a second giv, clear any "maybe dead" mark on that second giv. v->new_reg will either be or refer to the register of the giv it combined with. Doing this clearing avoids problems in biv elimination where a giv's new_reg is a complex value that can't be put in the insn but the giv combined with (with a reg as new_reg) is marked maybe_dead. Since the register will be used in either case, we'd prefer it be used from the simpler giv. */ for (v = bl->giv; v; v = v->next_iv) if (! v->maybe_dead && v->same) v->same->maybe_dead = 0; /* Try to eliminate the biv, if it is a candidate. This won't work if ! all_reduced, since the givs we planned to use might not have been reduced. We have to be careful that we didn't initially think we could eliminate this biv because of a giv that we now think may be dead and shouldn't be used as a biv replacement. Also, there is the possibility that we may have a giv that looks like it can be used to eliminate a biv, but the resulting insn isn't valid. This can happen, for example, on the 88k, where a JUMP_INSN can compare a register only with zero. Attempts to replace it with a compare with a constant will fail. Note that in cases where this call fails, we may have replaced some of the occurrences of the biv with a giv, but no harm was done in doing so in the rare cases where it can occur. */ if (all_reduced == 1 && bl->eliminable && maybe_eliminate_biv (bl, loop_start, end, 1, threshold, insn_count)) { /* ?? If we created a new test to bypass the loop entirely, or otherwise drop straight in, based on this test, then we might want to rewrite it also. This way some later pass has more hope of removing the initialization of this biv entirely. */ /* If final_value != 0, then the biv may be used after loop end and we must emit an insn to set it just in case. Reversed bivs already have an insn after the loop setting their value, so we don't need another one. We can't calculate the proper final value for such a biv here anyways. */ if (final_value != 0 && ! bl->reversed) { rtx insert_before; /* If the loop has multiple exits, emit the insn before the loop to ensure that it will always be executed no matter how the loop exits. Otherwise, emit the insn after the loop, since this is slightly more efficient. */ if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]) insert_before = loop_start; else insert_before = end_insert_before; emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value), end_insert_before); } #if 0 /* Delete all of the instructions inside the loop which set the biv, as they are all dead. If is safe to delete them, because an insn setting a biv will never be part of a libcall. */ /* However, deleting them will invalidate the regno_last_uid info, so keeping them around is more convenient. Final_biv_value will only succeed when there are multiple exits if the biv is dead at each exit, hence it does not matter that the original insn remains, because it is dead anyways. */ for (v = bl->biv; v; v = v->next_iv) delete_insn (v->insn); #endif if (loop_dump_stream) fprintf (loop_dump_stream, "Reg %d: biv eliminated\n", bl->regno); } } /* Go through all the instructions in the loop, making all the register substitutions scheduled in REG_MAP. */ for (p = loop_start; p != end; p = NEXT_INSN (p)) if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN || GET_CODE (p) == CALL_INSN) { replace_regs (PATTERN (p), reg_map, reg_map_size, 0); replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0); INSN_CODE (p) = -1; } /* Unroll loops from within strength reduction so that we can use the induction variable information that strength_reduce has already collected. */ if (unroll_p) unroll_loop (loop_end, insn_count, loop_start, end_insert_before, loop_info, 1); #ifdef HAVE_decrement_and_branch_on_count /* Instrument the loop with BCT insn. */ if (HAVE_decrement_and_branch_on_count && bct_p && flag_branch_on_count_reg) insert_bct (loop_start, loop_end, loop_info); #endif /* HAVE_decrement_and_branch_on_count */ if (loop_dump_stream) fprintf (loop_dump_stream, "\n"); VARRAY_FREE (reg_iv_type); VARRAY_FREE (reg_iv_info); } /* Return 1 if X is a valid source for an initial value (or as value being compared against in an initial test). X must be either a register or constant and must not be clobbered between the current insn and the start of the loop. INSN is the insn containing X. */ static int valid_initial_value_p (x, insn, call_seen, loop_start) rtx x; rtx insn; int call_seen; rtx loop_start; { if (CONSTANT_P (x)) return 1; /* Only consider pseudos we know about initialized in insns whose luids we know. */ if (GET_CODE (x) != REG || REGNO (x) >= max_reg_before_loop) return 0; /* Don't use call-clobbered registers across a call which clobbers it. On some machines, don't use any hard registers at all. */ if (REGNO (x) < FIRST_PSEUDO_REGISTER && (SMALL_REGISTER_CLASSES || (call_used_regs[REGNO (x)] && call_seen))) return 0; /* Don't use registers that have been clobbered before the start of the loop. */ if (reg_set_between_p (x, insn, loop_start)) return 0; return 1; } /* Scan X for memory refs and check each memory address as a possible giv. INSN is the insn whose pattern X comes from. NOT_EVERY_ITERATION is 1 if the insn might not be executed during every loop iteration. */ static void find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end) rtx x; rtx insn; int not_every_iteration; rtx loop_start, loop_end; { register int i, j; register enum rtx_code code; register char *fmt; if (x == 0) return; code = GET_CODE (x); switch (code) { case REG: case CONST_INT: case CONST: case CONST_DOUBLE: case SYMBOL_REF: case LABEL_REF: case PC: case CC0: case ADDR_VEC: case ADDR_DIFF_VEC: case USE: case CLOBBER: return; case MEM: { rtx src_reg; rtx add_val; rtx mult_val; int benefit; /* This code used to disable creating GIVs with mult_val == 1 and add_val == 0. However, this leads to lost optimizations when it comes time to combine a set of related DEST_ADDR GIVs, since this one would not be seen. */ if (general_induction_var (XEXP (x, 0), &src_reg, &add_val, &mult_val, 1, &benefit)) { /* Found one; record it. */ struct induction *v = (struct induction *) oballoc (sizeof (struct induction)); record_giv (v, insn, src_reg, addr_placeholder, mult_val, add_val, benefit, DEST_ADDR, not_every_iteration, &XEXP (x, 0), loop_start, loop_end); v->mem_mode = GET_MODE (x); } } return; default: break; } /* Recursively scan the subexpressions for other mem refs. */ fmt = GET_RTX_FORMAT (code); for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) if (fmt[i] == 'e') find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start, loop_end); else if (fmt[i] == 'E') for (j = 0; j < XVECLEN (x, i); j++) find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration, loop_start, loop_end); } /* Fill in the data about one biv update. V is the `struct induction' in which we record the biv. (It is allocated by the caller, with alloca.) INSN is the insn that sets it. DEST_REG is the biv's reg. MULT_VAL is const1_rtx if the biv is being incremented here, in which case INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is being set to INC_VAL. NOT_EVERY_ITERATION is nonzero if this biv update is not know to be executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update can be executed more than once per iteration. If MAYBE_MULTIPLE and NOT_EVERY_ITERATION are both zero, we know that the biv update is executed exactly once per iteration. */ static void record_biv (v, insn, dest_reg, inc_val, mult_val, location, not_every_iteration, maybe_multiple) struct induction *v; rtx insn; rtx dest_reg; rtx inc_val; rtx mult_val; rtx *location; int not_every_iteration; int maybe_multiple; { struct iv_class *bl; v->insn = insn; v->src_reg = dest_reg; v->dest_reg = dest_reg; v->mult_val = mult_val; v->add_val = inc_val; v->location = location; v->mode = GET_MODE (dest_reg); v->always_computable = ! not_every_iteration; v->always_executed = ! not_every_iteration; v->maybe_multiple = maybe_multiple; /* Add this to the reg's iv_class, creating a class if this is the first incrementation of the reg. */ bl = reg_biv_class[REGNO (dest_reg)]; if (bl == 0) { /* Create and initialize new iv_class. */ bl = (struct iv_class *) oballoc (sizeof (struct iv_class)); bl->regno = REGNO (dest_reg); bl->biv = 0; bl->giv = 0; bl->biv_count = 0; bl->giv_count = 0; /* Set initial value to the reg itself. */ bl->initial_value = dest_reg; /* We haven't seen the initializing insn yet */ bl->init_insn = 0; bl->init_set = 0; bl->initial_test = 0; bl->incremented = 0; bl->eliminable = 0; bl->nonneg = 0; bl->reversed = 0; bl->total_benefit = 0; /* Add this class to loop_iv_list. */ bl->next = loop_iv_list; loop_iv_list = bl; /* Put it in the array of biv register classes. */ reg_biv_class[REGNO (dest_reg)] = bl; } /* Update IV_CLASS entry for this biv. */ v->next_iv = bl->biv; bl->biv = v; bl->biv_count++; if (mult_val == const1_rtx) bl->incremented = 1; if (loop_dump_stream) { fprintf (loop_dump_stream, "Insn %d: possible biv, reg %d,", INSN_UID (insn), REGNO (dest_reg)); if (GET_CODE (inc_val) == CONST_INT) { fprintf (loop_dump_stream, " const ="); fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val)); fputc ('\n', loop_dump_stream); } else { fprintf (loop_dump_stream, " const = "); print_rtl (loop_dump_stream, inc_val); fprintf (loop_dump_stream, "\n"); } } } /* Fill in the data about one giv. V is the `struct induction' in which we record the giv. (It is allocated by the caller, with alloca.) INSN is the insn that sets it. BENEFIT estimates the savings from deleting this insn. TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed into a register or is used as a memory address. SRC_REG is the biv reg which the giv is computed from. DEST_REG is the giv's reg (if the giv is stored in a reg). MULT_VAL and ADD_VAL are the coefficients used to compute the giv. LOCATION points to the place where this giv's value appears in INSN. */ static void record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit, type, not_every_iteration, location, loop_start, loop_end) struct induction *v; rtx insn; rtx src_reg; rtx dest_reg; rtx mult_val, add_val; int benefit; enum g_types type; int not_every_iteration; rtx *location; rtx loop_start, loop_end; { struct induction *b; struct iv_class *bl; rtx set = single_set (insn); v->insn = insn; v->src_reg = src_reg; v->giv_type = type; v->dest_reg = dest_reg; v->mult_val = mult_val; v->add_val = add_val; v->benefit = benefit; v->location = location; v->cant_derive = 0; v->combined_with = 0; v->maybe_multiple = 0; v->maybe_dead = 0; v->derive_adjustment = 0; v->same = 0; v->ignore = 0; v->new_reg = 0; v->final_value = 0; v->same_insn = 0; v->auto_inc_opt = 0; v->unrolled = 0; v->shared = 0; v->derived_from = 0; v->last_use = 0; /* The v->always_computable field is used in update_giv_derive, to determine whether a giv can be used to derive another giv. For a DEST_REG giv, INSN computes a new value for the giv, so its value isn't computable if INSN insn't executed every iteration. However, for a DEST_ADDR giv, INSN merely uses the value of the giv; it does not compute a new value. Hence the value is always computable regardless of whether INSN is executed each iteration. */ if (type == DEST_ADDR) v->always_computable = 1; else v->always_computable = ! not_every_iteration; v->always_executed = ! not_every_iteration; if (type == DEST_ADDR) { v->mode = GET_MODE (*location); v->lifetime = 1; } else /* type == DEST_REG */ { v->mode = GET_MODE (SET_DEST (set)); v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))] - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]); /* If the lifetime is zero, it means that this register is really a dead store. So mark this as a giv that can be ignored. This will not prevent the biv from being eliminated. */ if (v->lifetime == 0) v->ignore = 1; REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT; REG_IV_INFO (REGNO (dest_reg)) = v; } /* Add the giv to the class of givs computed from one biv. */ bl = reg_biv_class[REGNO (src_reg)]; if (bl) { v->next_iv = bl->giv; bl->giv = v; /* Don't count DEST_ADDR. This is supposed to count the number of insns that calculate givs. */ if (type == DEST_REG) bl->giv_count++; bl->total_benefit += benefit; } else /* Fatal error, biv missing for this giv? */ abort (); if (type == DEST_ADDR) v->replaceable = 1; else { /* The giv can be replaced outright by the reduced register only if all of the following conditions are true: - the insn that sets the giv is always executed on any iteration on which the giv is used at all (there are two ways to deduce this: either the insn is executed on every iteration, or all uses follow that insn in the same basic block), - the giv is not used outside the loop - no assignments to the biv occur during the giv's lifetime. */ if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn) /* Previous line always fails if INSN was moved by loop opt. */ && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))] < INSN_LUID (loop_end) && (! not_every_iteration || last_use_this_basic_block (dest_reg, insn))) { /* Now check that there are no assignments to the biv within the giv's lifetime. This requires two separate checks. */ /* Check each biv update, and fail if any are between the first and last use of the giv. If this loop contains an inner loop that was unrolled, then the insn modifying the biv may have been emitted by the loop unrolling code, and hence does not have a valid luid. Just mark the biv as not replaceable in this case. It is not very useful as a biv, because it is used in two different loops. It is very unlikely that we would be able to optimize the giv using this biv anyways. */ v->replaceable = 1; for (b = bl->biv; b; b = b->next_iv) { if (INSN_UID (b->insn) >= max_uid_for_loop || ((uid_luid[INSN_UID (b->insn)] >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]) && (uid_luid[INSN_UID (b->insn)] <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]))) { v->replaceable = 0; v->not_replaceable = 1; break; } } /* If there are any backwards branches that go from after the biv update to before it, then this giv is not replaceable. */ if (v->replaceable) for (b = bl->biv; b; b = b->next_iv) if (back_branch_in_range_p (b->insn, loop_start, loop_end)) { v->replaceable = 0; v->not_replaceable = 1; break; } } else { /* May still be replaceable, we don't have enough info here to decide. */ v->replaceable = 0; v->not_replaceable = 0; } } /* Record whether the add_val contains a const_int, for later use by combine_givs. */ { rtx tem = add_val; v->no_const_addval = 1; if (tem == const0_rtx) ; else if (GET_CODE (tem) == CONST_INT) v->no_const_addval = 0; else if (GET_CODE (tem) == PLUS) { while (1) { if (GET_CODE (XEXP (tem, 0)) == PLUS) tem = XEXP (tem, 0); else if (GET_CODE (XEXP (tem, 1)) == PLUS) tem = XEXP (tem, 1); else break; } if (GET_CODE (XEXP (tem, 1)) == CONST_INT) v->no_const_addval = 0; } } if (loop_dump_stream) { if (type == DEST_REG) fprintf (loop_dump_stream, "Insn %d: giv reg %d", INSN_UID (insn), REGNO (dest_reg)); else fprintf (loop_dump_stream, "Insn %d: dest address", INSN_UID (insn)); fprintf (loop_dump_stream, " src reg %d benefit %d", REGNO (src_reg), v->benefit); fprintf (loop_dump_stream, " lifetime %d", v->lifetime); if (v->replaceable) fprintf (loop_dump_stream, " replaceable"); if (v->no_const_addval) fprintf (loop_dump_stream, " ncav"); if (GET_CODE (mult_val) == CONST_INT) { fprintf (loop_dump_stream, " mult "); fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val)); } else { fprintf (loop_dump_stream, " mult "); print_rtl (loop_dump_stream, mult_val); } if (GET_CODE (add_val) == CONST_INT) { fprintf (loop_dump_stream, " add "); fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val)); } else { fprintf (loop_dump_stream, " add "); print_rtl (loop_dump_stream, add_val); } } if (loop_dump_stream) fprintf (loop_dump_stream, "\n"); } /* All this does is determine whether a giv can be made replaceable because its final value can be calculated. This code can not be part of record_giv above, because final_giv_value requires that the number of loop iterations be known, and that can not be accurately calculated until after all givs have been identified. */ static void check_final_value (v, loop_start, loop_end, n_iterations) struct induction *v; rtx loop_start, loop_end; unsigned HOST_WIDE_INT n_iterations; { struct iv_class *bl; rtx final_value = 0; bl = reg_biv_class[REGNO (v->src_reg)]; /* DEST_ADDR givs will never reach here, because they are always marked replaceable above in record_giv. */ /* The giv can be replaced outright by the reduced register only if all of the following conditions are true: - the insn that sets the giv is always executed on any iteration on which the giv is used at all (there are two ways to deduce this: either the insn is executed on every iteration, or all uses follow that insn in the same basic block), - its final value can be calculated (this condition is different than the one above in record_giv) - it's not used before it's set - no assignments to the biv occur during the giv's lifetime. */ #if 0 /* This is only called now when replaceable is known to be false. */ /* Clear replaceable, so that it won't confuse final_giv_value. */ v->replaceable = 0; #endif if ((final_value = final_giv_value (v, loop_start, loop_end, n_iterations)) && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn))) { int biv_increment_seen = 0, before_giv_insn = 0; rtx p = v->insn; rtx last_giv_use; v->replaceable = 1; /* When trying to determine whether or not a biv increment occurs during the lifetime of the giv, we can ignore uses of the variable outside the loop because final_value is true. Hence we can not use regno_last_uid and regno_first_uid as above in record_giv. */ /* Search the loop to determine whether any assignments to the biv occur during the giv's lifetime. Start with the insn that sets the giv, and search around the loop until we come back to that insn again. Also fail if there is a jump within the giv's lifetime that jumps to somewhere outside the lifetime but still within the loop. This catches spaghetti code where the execution order is not linear, and hence the above test fails. Here we assume that the giv lifetime does not extend from one iteration of the loop to the next, so as to make the test easier. Since the lifetime isn't known yet, this requires two loops. See also record_giv above. */ last_giv_use = v->insn; while (1) { p = NEXT_INSN (p); if (p == loop_end) { before_giv_insn = 1; p = NEXT_INSN (loop_start); } if (p == v->insn) break; if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN || GET_CODE (p) == CALL_INSN) { /* It is possible for the BIV increment to use the GIV if we have a cycle. Thus we must be sure to check each insn for both BIV and GIV uses, and we must check for BIV uses first. */ if (! biv_increment_seen && reg_set_p (v->src_reg, PATTERN (p))) biv_increment_seen = 1; if (reg_mentioned_p (v->dest_reg, PATTERN (p))) { if (biv_increment_seen || before_giv_insn) { v->replaceable = 0; v->not_replaceable = 1; break; } last_giv_use = p; } } } /* Now that the lifetime of the giv is known, check for branches from within the lifetime to outside the lifetime if it is still replaceable. */ if (v->replaceable) { p = v->insn; while (1) { p = NEXT_INSN (p); if (p == loop_end) p = NEXT_INSN (loop_start); if (p == last_giv_use) break; if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) && LABEL_NAME (JUMP_LABEL (p)) && ((loop_insn_first_p (JUMP_LABEL (p), v->insn) && loop_insn_first_p (loop_start, JUMP_LABEL (p))) || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p)) && loop_insn_first_p (JUMP_LABEL (p), loop_end)))) { v->replaceable = 0; v->not_replaceable = 1; if (loop_dump_stream) fprintf (loop_dump_stream, "Found branch outside giv lifetime.\n"); break; } } } /* If it is replaceable, then save the final value. */ if (v->replaceable) v->final_value = final_value; } if (loop_dump_stream && v->replaceable) fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n", INSN_UID (v->insn), REGNO (v->dest_reg)); } /* Update the status of whether a giv can derive other givs. We need to do something special if there is or may be an update to the biv between the time the giv is defined and the time it is used to derive another giv. In addition, a giv that is only conditionally set is not allowed to derive another giv once a label has been passed. The cases we look at are when a label or an update to a biv is passed. */ static void update_giv_derive (p) rtx p; { struct iv_class *bl; struct induction *biv, *giv; rtx tem; int dummy; /* Search all IV classes, then all bivs, and finally all givs. There are three cases we are concerned with. First we have the situation of a giv that is only updated conditionally. In that case, it may not derive any givs after a label is passed. The second case is when a biv update occurs, or may occur, after the definition of a giv. For certain biv updates (see below) that are known to occur between the giv definition and use, we can adjust the giv definition. For others, or when the biv update is conditional, we must prevent the giv from deriving any other givs. There are two sub-cases within this case. If this is a label, we are concerned with any biv update that is done conditionally, since it may be done after the giv is defined followed by a branch here (actually, we need to pass both a jump and a label, but this extra tracking doesn't seem worth it). If this is a jump, we are concerned about any biv update that may be executed multiple times. We are actually only concerned about backward jumps, but it is probably not worth performing the test on the jump again here. If this is a biv update, we must adjust the giv status to show that a subsequent biv update was performed. If this adjustment cannot be done, the giv cannot derive further givs. */ for (bl = loop_iv_list; bl; bl = bl->next) for (biv = bl->biv; biv; biv = biv->next_iv) if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN || biv->insn == p) { for (giv = bl->giv; giv; giv = giv->next_iv) { /* If cant_derive is already true, there is no point in checking all of these conditions again. */ if (giv->cant_derive) continue; /* If this giv is conditionally set and we have passed a label, it cannot derive anything. */ if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable) giv->cant_derive = 1; /* Skip givs that have mult_val == 0, since they are really invariants. Also skip those that are replaceable, since we know their lifetime doesn't contain any biv update. */ else if (giv->mult_val == const0_rtx || giv->replaceable) continue; /* The only way we can allow this giv to derive another is if this is a biv increment and we can form the product of biv->add_val and giv->mult_val. In this case, we will be able to compute a compensation. */ else if (biv->insn == p) { tem = 0; if (biv->mult_val == const1_rtx) tem = simplify_giv_expr (gen_rtx_MULT (giv->mode, biv->add_val, giv->mult_val), &dummy); if (tem && giv->derive_adjustment) tem = simplify_giv_expr (gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment), &dummy); if (tem) giv->derive_adjustment = tem; else giv->cant_derive = 1; } else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable) || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple)) giv->cant_derive = 1; } } } /* Check whether an insn is an increment legitimate for a basic induction var. X is the source of insn P, or a part of it. MODE is the mode in which X should be interpreted. DEST_REG is the putative biv, also the destination of the insn. We accept patterns of these forms: REG = REG + INVARIANT (includes REG = REG - CONSTANT) REG = INVARIANT + REG If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX, store the additive term into *INC_VAL, and store the place where we found the additive term into *LOCATION. If X is an assignment of an invariant into DEST_REG, we set *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL. We also want to detect a BIV when it corresponds to a variable whose mode was promoted via PROMOTED_MODE. In that case, an increment of the variable may be a PLUS that adds a SUBREG of that variable to an invariant and then sign- or zero-extends the result of the PLUS into the variable. Most GIVs in such cases will be in the promoted mode, since that is the probably the natural computation mode (and almost certainly the mode used for addresses) on the machine. So we view the pseudo-reg containing the variable as the BIV, as if it were simply incremented. Note that treating the entire pseudo as a BIV will result in making simple increments to any GIVs based on it. However, if the variable overflows in its declared mode but not its promoted mode, the result will be incorrect. This is acceptable if the variable is signed, since overflows in such cases are undefined, but not if it is unsigned, since those overflows are defined. So we only check for SIGN_EXTEND and not ZERO_EXTEND. If we cannot find a biv, we return 0. */ static int basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val, location) register rtx x; enum machine_mode mode; rtx p; rtx dest_reg; rtx *inc_val; rtx *mult_val; rtx **location; { register enum rtx_code code; rtx *argp, arg; rtx insn, set = 0; code = GET_CODE (x); switch (code) { case PLUS: if (rtx_equal_p (XEXP (x, 0), dest_reg) || (GET_CODE (XEXP (x, 0)) == SUBREG && SUBREG_PROMOTED_VAR_P (XEXP (x, 0)) && SUBREG_REG (XEXP (x, 0)) == dest_reg)) { argp = &XEXP (x, 1); } else if (rtx_equal_p (XEXP (x, 1), dest_reg) || (GET_CODE (XEXP (x, 1)) == SUBREG && SUBREG_PROMOTED_VAR_P (XEXP (x, 1)) && SUBREG_REG (XEXP (x, 1)) == dest_reg)) { argp = &XEXP (x, 0); } else return 0; arg = *argp; if (invariant_p (arg) != 1) return 0; *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0); *mult_val = const1_rtx; *location = argp; return 1; case SUBREG: /* If this is a SUBREG for a promoted variable, check the inner value. */ if (SUBREG_PROMOTED_VAR_P (x)) return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)), dest_reg, p, inc_val, mult_val, location); return 0; case REG: /* If this register is assigned in a previous insn, look at its source, but don't go outside the loop or past a label. */ insn = p; while (1) { rtx dest; do { insn = PREV_INSN (insn); } while (insn && GET_CODE (insn) == NOTE && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG); if (!insn) break; set = single_set (insn); if (set == 0) break; dest = SET_DEST (set); if (dest == x || (GET_CODE (dest) == SUBREG && (GET_MODE_SIZE (GET_MODE (dest)) <= UNITS_PER_WORD) && (GET_MODE_CLASS (GET_MODE (dest)) == MODE_INT) && SUBREG_REG (dest) == x)) return basic_induction_var (SET_SRC (set), (GET_MODE (SET_SRC (set)) == VOIDmode ? GET_MODE (x) : GET_MODE (SET_SRC (set))), dest_reg, insn, inc_val, mult_val, location); while (GET_CODE (dest) == SIGN_EXTRACT || GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SUBREG || GET_CODE (dest) == STRICT_LOW_PART) dest = XEXP (dest, 0); if (dest == x) break; } /* ... fall through ... */ /* Can accept constant setting of biv only when inside inner most loop. Otherwise, a biv of an inner loop may be incorrectly recognized as a biv of the outer loop, causing code to be moved INTO the inner loop. */ case MEM: if (invariant_p (x) != 1) return 0; case CONST_INT: case SYMBOL_REF: case CONST: /* convert_modes aborts if we try to convert to or from CCmode, so just exclude that case. It is very unlikely that a condition code value would be a useful iterator anyways. */ if (loops_enclosed == 1 && GET_MODE_CLASS (mode) != MODE_CC && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC) { /* Possible bug here? Perhaps we don't know the mode of X. */ *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0); *mult_val = const0_rtx; return 1; } else return 0; case SIGN_EXTEND: return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)), dest_reg, p, inc_val, mult_val, location); case ASHIFTRT: /* Similar, since this can be a sign extension. */ for (insn = PREV_INSN (p); (insn && GET_CODE (insn) == NOTE && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG); insn = PREV_INSN (insn)) ; if (insn) set = single_set (insn); if (set && SET_DEST (set) == XEXP (x, 0) && GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) >= 0 && GET_CODE (SET_SRC (set)) == ASHIFT && XEXP (x, 1) == XEXP (SET_SRC (set), 1)) return basic_induction_var (XEXP (SET_SRC (set), 0), GET_MODE (XEXP (x, 0)), dest_reg, insn, inc_val, mult_val, location); return 0; default: return 0; } } /* A general induction variable (giv) is any quantity that is a linear function of a basic induction variable, i.e. giv = biv * mult_val + add_val. The coefficients can be any loop invariant quantity. A giv need not be computed directly from the biv; it can be computed by way of other givs. */ /* Determine whether X computes a giv. If it does, return a nonzero value which is the benefit from eliminating the computation of X; set *SRC_REG to the register of the biv that it is computed from; set *ADD_VAL and *MULT_VAL to the coefficients, such that the value of X is biv * mult + add; */ static int general_induction_var (x, src_reg, add_val, mult_val, is_addr, pbenefit) rtx x; rtx *src_reg; rtx *add_val; rtx *mult_val; int is_addr; int *pbenefit; { rtx orig_x = x; char *storage; /* If this is an invariant, forget it, it isn't a giv. */ if (invariant_p (x) == 1) return 0; /* See if the expression could be a giv and get its form. Mark our place on the obstack in case we don't find a giv. */ storage = (char *) oballoc (0); *pbenefit = 0; x = simplify_giv_expr (x, pbenefit); if (x == 0) { obfree (storage); return 0; } switch (GET_CODE (x)) { case USE: case CONST_INT: /* Since this is now an invariant and wasn't before, it must be a giv with MULT_VAL == 0. It doesn't matter which BIV we associate this with. */ *src_reg = loop_iv_list->biv->dest_reg; *mult_val = const0_rtx; *add_val = x; break; case REG: /* This is equivalent to a BIV. */ *src_reg = x; *mult_val = const1_rtx; *add_val = const0_rtx; break; case PLUS: /* Either (plus (biv) (invar)) or (plus (mult (biv) (invar_1)) (invar_2)). */ if (GET_CODE (XEXP (x, 0)) == MULT) { *src_reg = XEXP (XEXP (x, 0), 0); *mult_val = XEXP (XEXP (x, 0), 1); } else { *src_reg = XEXP (x, 0); *mult_val = const1_rtx; } *add_val = XEXP (x, 1); break; case MULT: /* ADD_VAL is zero. */ *src_reg = XEXP (x, 0); *mult_val = XEXP (x, 1); *add_val = const0_rtx; break; default: abort (); } /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be unless they are CONST_INT). */ if (GET_CODE (*add_val) == USE) *add_val = XEXP (*add_val, 0); if (GET_CODE (*mult_val) == USE) *mult_val = XEXP (*mult_val, 0); if (is_addr) { #ifdef ADDRESS_COST *pbenefit += ADDRESS_COST (orig_x) - reg_address_cost; #else *pbenefit += rtx_cost (orig_x, MEM) - reg_address_cost; #endif } else *pbenefit += rtx_cost (orig_x, SET); /* Always return true if this is a giv so it will be detected as such, even if the benefit is zero or negative. This allows elimination of bivs that might otherwise not be eliminated. */ return 1; } /* Given an expression, X, try to form it as a linear function of a biv. We will canonicalize it to be of the form (plus (mult (BIV) (invar_1)) (invar_2)) with possible degeneracies. The invariant expressions must each be of a form that can be used as a machine operand. We surround then with a USE rtx (a hack, but localized and certainly unambiguous!) if not a CONST_INT for simplicity in this routine; it is the caller's responsibility to strip them. If no such canonicalization is possible (i.e., two biv's are used or an expression that is neither invariant nor a biv or giv), this routine returns 0. For a non-zero return, the result will have a code of CONST_INT, USE, REG (for a BIV), PLUS, or MULT. No other codes will occur. *BENEFIT will be incremented by the benefit of any sub-giv encountered. */ static rtx sge_plus PROTO ((enum machine_mode, rtx, rtx)); static rtx sge_plus_constant PROTO ((rtx, rtx)); static rtx simplify_giv_expr (x, benefit) rtx x; int *benefit; { enum machine_mode mode = GET_MODE (x); rtx arg0, arg1; rtx tem; /* If this is not an integer mode, or if we cannot do arithmetic in this mode, this can't be a giv. */ if (mode != VOIDmode && (GET_MODE_CLASS (mode) != MODE_INT || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT)) return NULL_RTX; switch (GET_CODE (x)) { case PLUS: arg0 = simplify_giv_expr (XEXP (x, 0), benefit); arg1 = simplify_giv_expr (XEXP (x, 1), benefit); if (arg0 == 0 || arg1 == 0) return NULL_RTX; /* Put constant last, CONST_INT last if both constant. */ if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT) && ! ((GET_CODE (arg0) == USE && GET_CODE (arg1) == USE) || GET_CODE (arg1) == CONST_INT)) tem = arg0, arg0 = arg1, arg1 = tem; /* Handle addition of zero, then addition of an invariant. */ if (arg1 == const0_rtx) return arg0; else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE) switch (GET_CODE (arg0)) { case CONST_INT: case USE: /* Adding two invariants must result in an invariant, so enclose addition operation inside a USE and return it. */ if (GET_CODE (arg0) == USE) arg0 = XEXP (arg0, 0); if (GET_CODE (arg1) == USE) arg1 = XEXP (arg1, 0); if (GET_CODE (arg0) == CONST_INT) tem = arg0, arg0 = arg1, arg1 = tem; if (GET_CODE (arg1) == CONST_INT) tem = sge_plus_constant (arg0, arg1); else tem = sge_plus (mode, arg0, arg1); if (GET_CODE (tem) != CONST_INT) tem = gen_rtx_USE (mode, tem); return tem; case REG: case MULT: /* biv + invar or mult + invar. Return sum. */ return gen_rtx_PLUS (mode, arg0, arg1); case PLUS: /* (a + invar_1) + invar_2. Associate. */ return simplify_giv_expr ( gen_rtx_PLUS (mode, XEXP (arg0, 0), gen_rtx_PLUS (mode, XEXP (arg0, 1), arg1)), benefit); default: abort (); } /* Each argument must be either REG, PLUS, or MULT. Convert REG to MULT to reduce cases. */ if (GET_CODE (arg0) == REG) arg0 = gen_rtx_MULT (mode, arg0, const1_rtx); if (GET_CODE (arg1) == REG) arg1 = gen_rtx_MULT (mode, arg1, const1_rtx); /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT. Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT. Recurse to associate the second PLUS. */ if (GET_CODE (arg1) == MULT) tem = arg0, arg0 = arg1, arg1 = tem; if (GET_CODE (arg1) == PLUS) return simplify_giv_expr (gen_rtx_PLUS (mode, gen_rtx_PLUS (mode, arg0, XEXP (arg1, 0)), XEXP (arg1, 1)), benefit); /* Now must have MULT + MULT. Distribute if same biv, else not giv. */ if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT) return NULL_RTX; if (!rtx_equal_p (arg0, arg1)) return NULL_RTX; return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (arg0, 0), gen_rtx_PLUS (mode, XEXP (arg0, 1), XEXP (arg1, 1))), benefit); case MINUS: /* Handle "a - b" as "a + b * (-1)". */ return simplify_giv_expr (gen_rtx_PLUS (mode, XEXP (x, 0), gen_rtx_MULT (mode, XEXP (x, 1), constm1_rtx)), benefit); case MULT: arg0 = simplify_giv_expr (XEXP (x, 0), benefit); arg1 = simplify_giv_expr (XEXP (x, 1), benefit); if (arg0 == 0 || arg1 == 0) return NULL_RTX; /* Put constant last, CONST_INT last if both constant. */ if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT) && GET_CODE (arg1) != CONST_INT) tem = arg0, arg0 = arg1, arg1 = tem; /* If second argument is not now constant, not giv. */ if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT) return NULL_RTX; /* Handle multiply by 0 or 1. */ if (arg1 == const0_rtx) return const0_rtx; else if (arg1 == const1_rtx) return arg0; switch (GET_CODE (arg0)) { case REG: /* biv * invar. Done. */ return gen_rtx_MULT (mode, arg0, arg1); case CONST_INT: /* Product of two constants. */ return GEN_INT (INTVAL (arg0) * INTVAL (arg1)); case USE: /* invar * invar. It is a giv, but very few of these will actually pay off, so limit to simple registers. */ if (GET_CODE (arg1) != CONST_INT) return NULL_RTX; arg0 = XEXP (arg0, 0); if (GET_CODE (arg0) == REG) tem = gen_rtx_MULT (mode, arg0, arg1); else if (GET_CODE (arg0) == MULT && GET_CODE (XEXP (arg0, 0)) == REG && GET_CODE (XEXP (arg0, 1)) == CONST_INT) { tem = gen_rtx_MULT (mode, XEXP (arg0, 0), GEN_INT (INTVAL (XEXP (arg0, 1)) * INTVAL (arg1))); } else return NULL_RTX; return gen_rtx_USE (mode, tem); case MULT: /* (a * invar_1) * invar_2. Associate. */ return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (arg0, 0), gen_rtx_MULT (mode, XEXP (arg0, 1), arg1)), benefit); case PLUS: /* (a + invar_1) * invar_2. Distribute. */ return simplify_giv_expr (gen_rtx_PLUS (mode, gen_rtx_MULT (mode, XEXP (arg0, 0), arg1), gen_rtx_MULT (mode, XEXP (arg0, 1), arg1)), benefit); default: abort (); } case ASHIFT: /* Shift by constant is multiply by power of two. */ if (GET_CODE (XEXP (x, 1)) != CONST_INT) return 0; return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (x, 0), GEN_INT ((HOST_WIDE_INT) 1 << INTVAL (XEXP (x, 1)))), benefit); case NEG: /* "-a" is "a * (-1)" */ return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx), benefit); case NOT: /* "~a" is "-a - 1". Silly, but easy. */ return simplify_giv_expr (gen_rtx_MINUS (mode, gen_rtx_NEG (mode, XEXP (x, 0)), const1_rtx), benefit); case USE: /* Already in proper form for invariant. */ return x; case REG: /* If this is a new register, we can't deal with it. */ if (REGNO (x) >= max_reg_before_loop) return 0; /* Check for biv or giv. */ switch (REG_IV_TYPE (REGNO (x))) { case BASIC_INDUCT: return x; case GENERAL_INDUCT: { struct induction *v = REG_IV_INFO (REGNO (x)); /* Form expression from giv and add benefit. Ensure this giv can derive another and subtract any needed adjustment if so. */ *benefit += v->benefit; if (v->cant_derive) return 0; tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode, v->src_reg, v->mult_val), v->add_val); if (v->derive_adjustment) tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment); return simplify_giv_expr (tem, benefit); } default: /* If it isn't an induction variable, and it is invariant, we may be able to simplify things further by looking through the bits we just moved outside the loop. */ if (invariant_p (x) == 1) { struct movable *m; for (m = the_movables; m ; m = m->next) if (rtx_equal_p (x, m->set_dest)) { /* Ok, we found a match. Substitute and simplify. */ /* If we match another movable, we must use that, as this one is going away. */ if (m->match) return simplify_giv_expr (m->match->set_dest, benefit); /* If consec is non-zero, this is a member of a group of instructions that were moved together. We handle this case only to the point of seeking to the last insn and looking for a REG_EQUAL. Fail if we don't find one. */ if (m->consec != 0) { int i = m->consec; tem = m->insn; do { tem = NEXT_INSN (tem); } while (--i > 0); tem = find_reg_note (tem, REG_EQUAL, NULL_RTX); if (tem) tem = XEXP (tem, 0); } else { tem = single_set (m->insn); if (tem) tem = SET_SRC (tem); } if (tem) { /* What we are most interested in is pointer arithmetic on invariants -- only take patterns we may be able to do something with. */ if (GET_CODE (tem) == PLUS || GET_CODE (tem) == MULT || GET_CODE (tem) == ASHIFT || GET_CODE (tem) == CONST_INT || GET_CODE (tem) == SYMBOL_REF) { tem = simplify_giv_expr (tem, benefit); if (tem) return tem; } else if (GET_CODE (tem) == CONST && GET_CODE (XEXP (tem, 0)) == PLUS && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT) { tem = simplify_giv_expr (XEXP (tem, 0), benefit); if (tem) return tem; } } break; } } break; } /* Fall through to general case. */ default: /* If invariant, return as USE (unless CONST_INT). Otherwise, not giv. */ if (GET_CODE (x) == USE) x = XEXP (x, 0); if (invariant_p (x) == 1) { if (GET_CODE (x) == CONST_INT) return x; if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT) x = XEXP (x, 0); return gen_rtx_USE (mode, x); } else return 0; } } /* This routine folds invariants such that there is only ever one CONST_INT in the summation. It is only used by simplify_giv_expr. */ static rtx sge_plus_constant (x, c) rtx x, c; { if (GET_CODE (x) == CONST_INT) return GEN_INT (INTVAL (x) + INTVAL (c)); else if (GET_CODE (x) != PLUS) return gen_rtx_PLUS (GET_MODE (x), x, c); else if (GET_CODE (XEXP (x, 1)) == CONST_INT) { return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0), GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c))); } else if (GET_CODE (XEXP (x, 0)) == PLUS || GET_CODE (XEXP (x, 1)) != PLUS) { return gen_rtx_PLUS (GET_MODE (x), sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1)); } else { return gen_rtx_PLUS (GET_MODE (x), sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0)); } } static rtx sge_plus (mode, x, y) enum machine_mode mode; rtx x, y; { while (GET_CODE (y) == PLUS) { rtx a = XEXP (y, 0); if (GET_CODE (a) == CONST_INT) x = sge_plus_constant (x, a); else x = gen_rtx_PLUS (mode, x, a); y = XEXP (y, 1); } if (GET_CODE (y) == CONST_INT) x = sge_plus_constant (x, y); else x = gen_rtx_PLUS (mode, x, y); return x; } /* Help detect a giv that is calculated by several consecutive insns; for example, giv = biv * M giv = giv + A The caller has already identified the first insn P as having a giv as dest; we check that all other insns that set the same register follow immediately after P, that they alter nothing else, and that the result of the last is still a giv. The value is 0 if the reg set in P is not really a giv. Otherwise, the value is the amount gained by eliminating all the consecutive insns that compute the value. FIRST_BENEFIT is the amount gained by eliminating the first insn, P. SRC_REG is the reg of the biv; DEST_REG is the reg of the giv. The coefficients of the ultimate giv value are stored in *MULT_VAL and *ADD_VAL. */ static int consec_sets_giv (first_benefit, p, src_reg, dest_reg, add_val, mult_val, last_consec_insn) int first_benefit; rtx p; rtx src_reg; rtx dest_reg; rtx *add_val; rtx *mult_val; rtx *last_consec_insn; { int count; enum rtx_code code; int benefit; rtx temp; rtx set; /* Indicate that this is a giv so that we can update the value produced in each insn of the multi-insn sequence. This induction structure will be used only by the call to general_induction_var below, so we can allocate it on our stack. If this is a giv, our caller will replace the induct var entry with a new induction structure. */ struct induction *v = (struct induction *) alloca (sizeof (struct induction)); v->src_reg = src_reg; v->mult_val = *mult_val; v->add_val = *add_val; v->benefit = first_benefit; v->cant_derive = 0; v->derive_adjustment = 0; REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT; REG_IV_INFO (REGNO (dest_reg)) = v; count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1; while (count > 0) { p = NEXT_INSN (p); code = GET_CODE (p); /* If libcall, skip to end of call sequence. */ if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX))) p = XEXP (temp, 0); if (code == INSN && (set = single_set (p)) && GET_CODE (SET_DEST (set)) == REG && SET_DEST (set) == dest_reg && (general_induction_var (SET_SRC (set), &src_reg, add_val, mult_val, 0, &benefit) /* Giv created by equivalent expression. */ || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)) && general_induction_var (XEXP (temp, 0), &src_reg, add_val, mult_val, 0, &benefit))) && src_reg == v->src_reg) { if (find_reg_note (p, REG_RETVAL, NULL_RTX)) benefit += libcall_benefit (p); count--; v->mult_val = *mult_val; v->add_val = *add_val; v->benefit = benefit; } else if (code != NOTE) { /* Allow insns that set something other than this giv to a constant. Such insns are needed on machines which cannot include long constants and should not disqualify a giv. */ if (code == INSN && (set = single_set (p)) && SET_DEST (set) != dest_reg && CONSTANT_P (SET_SRC (set))) continue; REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT; return 0; } } *last_consec_insn = p; return v->benefit; } /* Return an rtx, if any, that expresses giv G2 as a function of the register represented by G1. If no such expression can be found, or it is clear that it cannot possibly be a valid address, 0 is returned. To perform the computation, we note that G1 = x * v + a and G2 = y * v + b where `v' is the biv. So G2 = (y/b) * G1 + (b - a*y/x). Note that MULT = y/x. Update: A and B are now allowed to be additive expressions such that B contains all variables in A. That is, computing B-A will not require subtracting variables. */ static rtx express_from_1 (a, b, mult) rtx a, b, mult; { /* If MULT is zero, then A*MULT is zero, and our expression is B. */ if (mult == const0_rtx) return b; /* If MULT is not 1, we cannot handle A with non-constants, since we would then be required to subtract multiples of the registers in A. This is theoretically possible, and may even apply to some Fortran constructs, but it is a lot of work and we do not attempt it here. */ if (mult != const1_rtx && GET_CODE (a) != CONST_INT) return NULL_RTX; /* In general these structures are sorted top to bottom (down the PLUS chain), but not left to right across the PLUS. If B is a higher order giv than A, we can strip one level and recurse. If A is higher order, we'll eventually bail out, but won't know that until the end. If they are the same, we'll strip one level around this loop. */ while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS) { rtx ra, rb, oa, ob, tmp; ra = XEXP (a, 0), oa = XEXP (a, 1); if (GET_CODE (ra) == PLUS) tmp = ra, ra = oa, oa = tmp; rb = XEXP (b, 0), ob = XEXP (b, 1); if (GET_CODE (rb) == PLUS) tmp = rb, rb = ob, ob = tmp; if (rtx_equal_p (ra, rb)) /* We matched: remove one reg completely. */ a = oa, b = ob; else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob)) /* An alternate match. */ a = oa, b = rb; else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb)) /* An alternate match. */ a = ra, b = ob; else { /* Indicates an extra register in B. Strip one level from B and recurse, hoping B was the higher order expression. */ ob = express_from_1 (a, ob, mult); if (ob == NULL_RTX) return NULL_RTX; return gen_rtx_PLUS (GET_MODE (b), rb, ob); } } /* Here we are at the last level of A, go through the cases hoping to get rid of everything but a constant. */ if (GET_CODE (a) == PLUS) { rtx ra, oa; ra = XEXP (a, 0), oa = XEXP (a, 1); if (rtx_equal_p (oa, b)) oa = ra; else if (!rtx_equal_p (ra, b)) return NULL_RTX; if (GET_CODE (oa) != CONST_INT) return NULL_RTX; return GEN_INT (-INTVAL (oa) * INTVAL (mult)); } else if (GET_CODE (a) == CONST_INT) { return plus_constant (b, -INTVAL (a) * INTVAL (mult)); } else if (GET_CODE (b) == PLUS) { if (rtx_equal_p (a, XEXP (b, 0))) return XEXP (b, 1); else if (rtx_equal_p (a, XEXP (b, 1))) return XEXP (b, 0); else return NULL_RTX; } else if (rtx_equal_p (a, b)) return const0_rtx; return NULL_RTX; } rtx express_from (g1, g2) struct induction *g1, *g2; { rtx mult, add; /* The value that G1 will be multiplied by must be a constant integer. Also, the only chance we have of getting a valid address is if b*c/a (see above for notation) is also an integer. */ if (GET_CODE (g1->mult_val) == CONST_INT && GET_CODE (g2->mult_val) == CONST_INT) { if (g1->mult_val == const0_rtx || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0) return NULL_RTX; mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val)); } else if (rtx_equal_p (g1->mult_val, g2->mult_val)) mult = const1_rtx; else { /* ??? Find out if the one is a multiple of the other? */ return NULL_RTX; } add = express_from_1 (g1->add_val, g2->add_val, mult); if (add == NULL_RTX) return NULL_RTX; /* Form simplified final result. */ if (mult == const0_rtx) return add; else if (mult == const1_rtx) mult = g1->dest_reg; else mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult); if (add == const0_rtx) return mult; else { if (GET_CODE (add) == PLUS && CONSTANT_P (XEXP (add, 1))) { rtx tem = XEXP (add, 1); mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0)); add = tem; } return gen_rtx_PLUS (g2->mode, mult, add); } } /* Return an rtx, if any, that expresses giv G2 as a function of the register represented by G1. This indicates that G2 should be combined with G1 and that G2 can use (either directly or via an address expression) a register used to represent G1. */ static rtx combine_givs_p (g1, g2) struct induction *g1, *g2; { rtx tem = express_from (g1, g2); /* If these givs are identical, they can be combined. We use the results of express_from because the addends are not in a canonical form, so rtx_equal_p is a weaker test. */ /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the combination to be the other way round. */ if (tem == g1->dest_reg && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR)) { return g1->dest_reg; } /* If G2 can be expressed as a function of G1 and that function is valid as an address and no more expensive than using a register for G2, the expression of G2 in terms of G1 can be used. */ if (tem != NULL_RTX && g2->giv_type == DEST_ADDR && memory_address_p (g2->mem_mode, tem) /* ??? Looses, especially with -fforce-addr, where *g2->location will always be a register, and so anything more complicated gets discarded. */ #if 0 #ifdef ADDRESS_COST && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location) #else && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM) #endif #endif ) { return tem; } return NULL_RTX; } struct combine_givs_stats { int giv_number; int total_benefit; }; static int cmp_combine_givs_stats (x, y) struct combine_givs_stats *x, *y; { int d; d = y->total_benefit - x->total_benefit; /* Stabilize the sort. */ if (!d) d = x->giv_number - y->giv_number; return d; } /* Check all pairs of givs for iv_class BL and see if any can be combined with any other. If so, point SAME to the giv combined with and set NEW_REG to be an expression (in terms of the other giv's DEST_REG) equivalent to the giv. Also, update BENEFIT and related fields for cost/benefit analysis. */ static void combine_givs (bl) struct iv_class *bl; { /* Additional benefit to add for being combined multiple times. */ const int extra_benefit = 3; struct induction *g1, *g2, **giv_array; int i, j, k, giv_count; struct combine_givs_stats *stats; rtx *can_combine; /* Count givs, because bl->giv_count is incorrect here. */ giv_count = 0; for (g1 = bl->giv; g1; g1 = g1->next_iv) if (!g1->ignore) giv_count++; giv_array = (struct induction **) alloca (giv_count * sizeof (struct induction *)); i = 0; for (g1 = bl->giv; g1; g1 = g1->next_iv) if (!g1->ignore) giv_array[i++] = g1; stats = (struct combine_givs_stats *) alloca (giv_count * sizeof (*stats)); bzero ((char *) stats, giv_count * sizeof (*stats)); can_combine = (rtx *) alloca (giv_count * giv_count * sizeof(rtx)); bzero ((char *) can_combine, giv_count * giv_count * sizeof(rtx)); for (i = 0; i < giv_count; i++) { int this_benefit; rtx single_use; g1 = giv_array[i]; stats[i].giv_number = i; /* If a DEST_REG GIV is used only once, do not allow it to combine with anything, for in doing so we will gain nothing that cannot be had by simply letting the GIV with which we would have combined to be reduced on its own. The losage shows up in particular with DEST_ADDR targets on hosts with reg+reg addressing, though it can be seen elsewhere as well. */ if (g1->giv_type == DEST_REG && (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg))) && single_use != const0_rtx) continue; this_benefit = g1->benefit; /* Add an additional weight for zero addends. */ if (g1->no_const_addval) this_benefit += 1; for (j = 0; j < giv_count; j++) { rtx this_combine; g2 = giv_array[j]; if (g1 != g2 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX) { can_combine[i*giv_count + j] = this_combine; this_benefit += g2->benefit + extra_benefit; } } stats[i].total_benefit = this_benefit; } /* Iterate, combining until we can't. */ restart: qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats); if (loop_dump_stream) { fprintf (loop_dump_stream, "Sorted combine statistics:\n"); for (k = 0; k < giv_count; k++) { g1 = giv_array[stats[k].giv_number]; if (!g1->combined_with && !g1->same) fprintf (loop_dump_stream, " {%d, %d}", INSN_UID (giv_array[stats[k].giv_number]->insn), stats[k].total_benefit); } putc ('\n', loop_dump_stream); } for (k = 0; k < giv_count; k++) { int g1_add_benefit = 0; i = stats[k].giv_number; g1 = giv_array[i]; /* If it has already been combined, skip. */ if (g1->combined_with || g1->same) continue; for (j = 0; j < giv_count; j++) { g2 = giv_array[j]; if (g1 != g2 && can_combine[i*giv_count + j] /* If it has already been combined, skip. */ && ! g2->same && ! g2->combined_with) { int l; g2->new_reg = can_combine[i*giv_count + j]; g2->same = g1; g1->combined_with++; g1->lifetime += g2->lifetime; g1_add_benefit += g2->benefit; /* ??? The new final_[bg]iv_value code does a much better job of finding replaceable giv's, and hence this code may no longer be necessary. */ if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg)) g1_add_benefit -= copy_cost; /* To help optimize the next set of combinations, remove this giv from the benefits of other potential mates. */ for (l = 0; l < giv_count; ++l) { int m = stats[l].giv_number; if (can_combine[m*giv_count + j]) stats[l].total_benefit -= g2->benefit + extra_benefit; } if (loop_dump_stream) fprintf (loop_dump_stream, "giv at %d combined with giv at %d\n", INSN_UID (g2->insn), INSN_UID (g1->insn)); } } /* To help optimize the next set of combinations, remove this giv from the benefits of other potential mates. */ if (g1->combined_with) { for (j = 0; j < giv_count; ++j) { int m = stats[j].giv_number; if (can_combine[m*giv_count + i]) stats[j].total_benefit -= g1->benefit + extra_benefit; } g1->benefit += g1_add_benefit; /* We've finished with this giv, and everything it touched. Restart the combination so that proper weights for the rest of the givs are properly taken into account. */ /* ??? Ideally we would compact the arrays at this point, so as to not cover old ground. But sanely compacting can_combine is tricky. */ goto restart; } } } struct recombine_givs_stats { int giv_number; int start_luid, end_luid; }; /* Used below as comparison function for qsort. We want a ascending luid when scanning the array starting at the end, thus the arguments are used in reverse. */ static int cmp_recombine_givs_stats (x, y) struct recombine_givs_stats *x, *y; { int d; d = y->start_luid - x->start_luid; /* Stabilize the sort. */ if (!d) d = y->giv_number - x->giv_number; return d; } /* Scan X, which is a part of INSN, for the end of life of a giv. Also look for the start of life of a giv where the start has not been seen yet to unlock the search for the end of its life. Only consider givs that belong to BIV. Return the total number of lifetime ends that have been found. */ static int find_life_end (x, stats, insn, biv) rtx x, insn, biv; struct recombine_givs_stats *stats; { enum rtx_code code; char *fmt; int i, j; int retval; code = GET_CODE (x); switch (code) { case SET: { rtx reg = SET_DEST (x); if (GET_CODE (reg) == REG) { int regno = REGNO (reg); struct induction *v = REG_IV_INFO (regno); if (REG_IV_TYPE (regno) == GENERAL_INDUCT && ! v->ignore && v->src_reg == biv && stats[v->ix].end_luid <= 0) { /* If we see a 0 here for end_luid, it means that we have scanned the entire loop without finding any use at all. We must not predicate this code on a start_luid match since that would make the test fail for givs that have been hoisted out of inner loops. */ if (stats[v->ix].end_luid == 0) { stats[v->ix].end_luid = stats[v->ix].start_luid; return 1 + find_life_end (SET_SRC (x), stats, insn, biv); } else if (stats[v->ix].start_luid == INSN_LUID (insn)) stats[v->ix].end_luid = 0; } return find_life_end (SET_SRC (x), stats, insn, biv); } break; } case REG: { int regno = REGNO (x); struct induction *v = REG_IV_INFO (regno); if (REG_IV_TYPE (regno) == GENERAL_INDUCT && ! v->ignore && v->src_reg == biv && stats[v->ix].end_luid == 0) { while (INSN_UID (insn) >= max_uid_for_loop) insn = NEXT_INSN (insn); stats[v->ix].end_luid = INSN_LUID (insn); return 1; } return 0; } case LABEL_REF: case CONST_DOUBLE: case CONST_INT: case CONST: return 0; default: break; } fmt = GET_RTX_FORMAT (code); retval = 0; for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { if (fmt[i] == 'e') retval += find_life_end (XEXP (x, i), stats, insn, biv); else if (fmt[i] == 'E') for (j = XVECLEN (x, i) - 1; j >= 0; j--) retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv); } return retval; } /* For each giv that has been combined with another, look if we can combine it with the most recently used one instead. This tends to shorten giv lifetimes, and helps the next step: try to derive givs from other givs. */ static void recombine_givs (bl, loop_start, loop_end, unroll_p) struct iv_class *bl; rtx loop_start, loop_end; int unroll_p; { struct induction *v, **giv_array, *last_giv; struct recombine_givs_stats *stats; int giv_count; int i, rescan; int ends_need_computing; for (giv_count = 0, v = bl->giv; v; v = v->next_iv) { if (! v->ignore) giv_count++; } giv_array = (struct induction **) alloca (giv_count * sizeof (struct induction *)); stats = (struct recombine_givs_stats *) alloca (giv_count * sizeof *stats); /* Initialize stats and set up the ix field for each giv in stats to name the corresponding index into stats. */ for (i = 0, v = bl->giv; v; v = v->next_iv) { rtx p; if (v->ignore) continue; giv_array[i] = v; stats[i].giv_number = i; /* If this giv has been hoisted out of an inner loop, use the luid of the previous insn. */ for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; ) p = PREV_INSN (p); stats[i].start_luid = INSN_LUID (p); v->ix = i; i++; } qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats); /* Do the actual most-recently-used recombination. */ for (last_giv = 0, i = giv_count - 1; i >= 0; i--) { v = giv_array[stats[i].giv_number]; if (v->same) { struct induction *old_same = v->same; rtx new_combine; /* combine_givs_p actually says if we can make this transformation. The other tests are here only to avoid keeping a giv alive that could otherwise be eliminated. */ if (last_giv && ((old_same->maybe_dead && ! old_same->combined_with) || ! last_giv->maybe_dead || last_giv->combined_with) && (new_combine = combine_givs_p (last_giv, v))) { old_same->combined_with--; v->new_reg = new_combine; v->same = last_giv; last_giv->combined_with++; /* No need to update lifetimes / benefits here since we have already decided what to reduce. */ if (loop_dump_stream) { fprintf (loop_dump_stream, "giv at %d recombined with giv at %d as ", INSN_UID (v->insn), INSN_UID (last_giv->insn)); print_rtl (loop_dump_stream, v->new_reg); putc ('\n', loop_dump_stream); } continue; } v = v->same; } else if (v->giv_type != DEST_REG) continue; if (! last_giv || (last_giv->maybe_dead && ! last_giv->combined_with) || ! v->maybe_dead || v->combined_with) last_giv = v; } ends_need_computing = 0; /* For each DEST_REG giv, compute lifetime starts, and try to compute lifetime ends from regscan info. */ for (i = 0, v = bl->giv; v; v = v->next_iv) { if (v->ignore) continue; if (v->giv_type == DEST_ADDR) { /* Loop unrolling of an inner loop can even create new DEST_REG givs. */ rtx p; for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; ) p = PREV_INSN (p); stats[i].start_luid = stats[i].end_luid = INSN_LUID (p); if (p != v->insn) stats[i].end_luid++; } else /* v->giv_type == DEST_REG */ { if (v->last_use) { stats[i].start_luid = INSN_LUID (v->insn); stats[i].end_luid = INSN_LUID (v->last_use); } else if (INSN_UID (v->insn) >= max_uid_for_loop) { rtx p; /* This insn has been created by loop optimization on an inner loop. We don't have a proper start_luid that will match when we see the first set. But we do know that there will be no use before the set, so we can set end_luid to 0 so that we'll start looking for the last use right away. */ for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; ) p = PREV_INSN (p); stats[i].start_luid = INSN_LUID (p); stats[i].end_luid = 0; ends_need_computing++; } else { int regno = REGNO (v->dest_reg); int count = VARRAY_INT (n_times_set, regno) - 1; rtx p = v->insn; /* Find the first insn that sets the giv, so that we can verify if this giv's lifetime wraps around the loop. We also need the luid of the first setting insn in order to detect the last use properly. */ while (count) { p = prev_nonnote_insn (p); if (reg_set_p (v->dest_reg, p)) count--; } stats[i].start_luid = INSN_LUID (p); if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)]) { stats[i].end_luid = -1; ends_need_computing++; } else { stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)]; if (stats[i].end_luid > INSN_LUID (loop_end)) { stats[i].end_luid = -1; ends_need_computing++; } } } } i++; } /* If the regscan information was unconclusive for one or more DEST_REG givs, scan the all insn in the loop to find out lifetime ends. */ if (ends_need_computing) { rtx biv = bl->biv->src_reg; rtx p = loop_end; do { if (p == loop_start) p = loop_end; p = PREV_INSN (p); if (GET_RTX_CLASS (GET_CODE (p)) != 'i') continue; ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv); } while (ends_need_computing); } /* Set start_luid back to the last insn that sets the giv. This allows more combinations. */ for (i = 0, v = bl->giv; v; v = v->next_iv) { if (v->ignore) continue; if (INSN_UID (v->insn) < max_uid_for_loop) stats[i].start_luid = INSN_LUID (v->insn); i++; } /* Now adjust lifetime ends by taking combined givs into account. */ for (i = 0, v = bl->giv; v; v = v->next_iv) { unsigned luid; int j; if (v->ignore) continue; if (v->same && ! v->same->ignore) { j = v->same->ix; luid = stats[i].start_luid; /* Use unsigned arithmetic to model loop wrap-around. */ if (luid - stats[j].start_luid > (unsigned) stats[j].end_luid - stats[j].start_luid) stats[j].end_luid = luid; } i++; } qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats); /* Try to derive DEST_REG givs from previous DEST_REG givs with the same mult_val and non-overlapping lifetime. This reduces register pressure. Once we find a DEST_REG giv that is suitable to derive others from, we set last_giv to this giv, and try to derive as many other DEST_REG givs from it without joining overlapping lifetimes. If we then encounter a DEST_REG giv that we can't derive, we set rescan to the index for this giv (unless rescan is already set). When we are finished with the current LAST_GIV (i.e. the inner loop terminates), we start again with rescan, which then becomes the new LAST_GIV. */ for (i = giv_count - 1; i >= 0; i = rescan) { int life_start, life_end; for (last_giv = 0, rescan = -1; i >= 0; i--) { rtx sum; v = giv_array[stats[i].giv_number]; if (v->giv_type != DEST_REG || v->derived_from || v->same) continue; if (! last_giv) { /* Don't use a giv that's likely to be dead to derive others - that would be likely to keep that giv alive. */ if (! v->maybe_dead || v->combined_with) { last_giv = v; life_start = stats[i].start_luid; life_end = stats[i].end_luid; } continue; } /* Use unsigned arithmetic to model loop wrap around. */ if (((unsigned) stats[i].start_luid - life_start >= (unsigned) life_end - life_start) && ((unsigned) stats[i].end_luid - life_start > (unsigned) life_end - life_start) /* Check that the giv insn we're about to use for deriving precedes all uses of that giv. Note that initializing the derived giv would defeat the purpose of reducing register pressure. ??? We could arrange to move the insn. */ && ((unsigned) stats[i].end_luid - INSN_LUID (loop_start) > (unsigned) stats[i].start_luid - INSN_LUID (loop_start)) && rtx_equal_p (last_giv->mult_val, v->mult_val) /* ??? Could handle libcalls, but would need more logic. */ && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX) /* We would really like to know if for any giv that v is combined with, v->insn or any intervening biv increment dominates that combined giv. However, we don't have this detailed control flow information. N.B. since last_giv will be reduced, it is valid anywhere in the loop, so we don't need to check the validity of last_giv. We rely here on the fact that v->always_executed implies that there is no jump to someplace else in the loop before the giv insn, and hence any insn that is executed before the giv insn in the loop will have a lower luid. */ && (v->always_executed || ! v->combined_with) && (sum = express_from (last_giv, v)) /* Make sure we don't make the add more expensive. ADD_COST doesn't take different costs of registers and constants into account, so compare the cost of the actual SET_SRCs. */ && (rtx_cost (sum, SET) <= rtx_cost (SET_SRC (single_set (v->insn)), SET)) /* ??? unroll can't understand anything but reg + const_int sums. It would be cleaner to fix unroll. */ && ((GET_CODE (sum) == PLUS && GET_CODE (XEXP (sum, 0)) == REG && GET_CODE (XEXP (sum, 1)) == CONST_INT) || ! unroll_p) && validate_change (v->insn, &PATTERN (v->insn), gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0)) { v->derived_from = last_giv; life_end = stats[i].end_luid; if (loop_dump_stream) { fprintf (loop_dump_stream, "giv at %d derived from %d as ", INSN_UID (v->insn), INSN_UID (last_giv->insn)); print_rtl (loop_dump_stream, sum); putc ('\n', loop_dump_stream); } } else if (rescan < 0) rescan = i; } } } /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */ void emit_iv_add_mult (b, m, a, reg, insert_before) rtx b; /* initial value of basic induction variable */ rtx m; /* multiplicative constant */ rtx a; /* additive constant */ rtx reg; /* destination register */ rtx insert_before; { rtx seq; rtx result; /* Prevent unexpected sharing of these rtx. */ a = copy_rtx (a); b = copy_rtx (b); /* Increase the lifetime of any invariants moved further in code. */ update_reg_last_use (a, insert_before); update_reg_last_use (b, insert_before); update_reg_last_use (m, insert_before); start_sequence (); result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0); if (reg != result) emit_move_insn (reg, result); seq = gen_sequence (); end_sequence (); emit_insn_before (seq, insert_before); /* It is entirely possible that the expansion created lots of new registers. Iterate over the sequence we just created and record them all. */ if (GET_CODE (seq) == SEQUENCE) { int i; for (i = 0; i < XVECLEN (seq, 0); ++i) { rtx set = single_set (XVECEXP (seq, 0, i)); if (set && GET_CODE (SET_DEST (set)) == REG) record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0); } } else if (GET_CODE (seq) == SET && GET_CODE (SET_DEST (seq)) == REG) record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0); } /* Test whether A * B can be computed without an actual multiply insn. Value is 1 if so. */ static int product_cheap_p (a, b) rtx a; rtx b; { int i; rtx tmp; struct obstack *old_rtl_obstack = rtl_obstack; char *storage = (char *) obstack_alloc (&temp_obstack, 0); int win = 1; /* If only one is constant, make it B. */ if (GET_CODE (a) == CONST_INT) tmp = a, a = b, b = tmp; /* If first constant, both constant, so don't need multiply. */ if (GET_CODE (a) == CONST_INT) return 1; /* If second not constant, neither is constant, so would need multiply. */ if (GET_CODE (b) != CONST_INT) return 0; /* One operand is constant, so might not need multiply insn. Generate the code for the multiply and see if a call or multiply, or long sequence of insns is generated. */ rtl_obstack = &temp_obstack; start_sequence (); expand_mult (GET_MODE (a), a, b, NULL_RTX, 0); tmp = gen_sequence (); end_sequence (); if (GET_CODE (tmp) == SEQUENCE) { if (XVEC (tmp, 0) == 0) win = 1; else if (XVECLEN (tmp, 0) > 3) win = 0; else for (i = 0; i < XVECLEN (tmp, 0); i++) { rtx insn = XVECEXP (tmp, 0, i); if (GET_CODE (insn) != INSN || (GET_CODE (PATTERN (insn)) == SET && GET_CODE (SET_SRC (PATTERN (insn))) == MULT) || (GET_CODE (PATTERN (insn)) == PARALLEL && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT)) { win = 0; break; } } } else if (GET_CODE (tmp) == SET && GET_CODE (SET_SRC (tmp)) == MULT) win = 0; else if (GET_CODE (tmp) == PARALLEL && GET_CODE (XVECEXP (tmp, 0, 0)) == SET && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT) win = 0; /* Free any storage we obtained in generating this multiply and restore rtl allocation to its normal obstack. */ obstack_free (&temp_obstack, storage); rtl_obstack = old_rtl_obstack; return win; } /* Check to see if loop can be terminated by a "decrement and branch until zero" instruction. If so, add a REG_NONNEG note to the branch insn if so. Also try reversing an increment loop to a decrement loop to see if the optimization can be performed. Value is nonzero if optimization was performed. */ /* This is useful even if the architecture doesn't have such an insn, because it might change a loops which increments from 0 to n to a loop which decrements from n to 0. A loop that decrements to zero is usually faster than one that increments from zero. */ /* ??? This could be rewritten to use some of the loop unrolling procedures, such as approx_final_value, biv_total_increment, loop_iterations, and final_[bg]iv_value. */ static int check_dbra_loop (loop_end, insn_count, loop_start, loop_info) rtx loop_end; int insn_count; rtx loop_start; struct loop_info *loop_info; { struct iv_class *bl; rtx reg; rtx jump_label; rtx final_value; rtx start_value; rtx new_add_val; rtx comparison; rtx before_comparison; rtx p; rtx jump; rtx first_compare; int compare_and_branch; /* If last insn is a conditional branch, and the insn before tests a register value, try to optimize it. Otherwise, we can't do anything. */ jump = PREV_INSN (loop_end); comparison = get_condition_for_loop (jump); if (comparison == 0) return 0; /* Try to compute whether the compare/branch at the loop end is one or two instructions. */ get_condition (jump, &first_compare); if (first_compare == jump) compare_and_branch = 1; else if (first_compare == prev_nonnote_insn (jump)) compare_and_branch = 2; else return 0; /* Check all of the bivs to see if the compare uses one of them. Skip biv's set more than once because we can't guarantee that it will be zero on the last iteration. Also skip if the biv is used between its update and the test insn. */ for (bl = loop_iv_list; bl; bl = bl->next) { if (bl->biv_count == 1 && ! bl->biv->maybe_multiple && bl->biv->dest_reg == XEXP (comparison, 0) && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn, first_compare)) break; } if (! bl) return 0; /* Look for the case where the basic induction variable is always nonnegative, and equals zero on the last iteration. In this case, add a reg_note REG_NONNEG, which allows the m68k DBRA instruction to be used. */ if (((GET_CODE (comparison) == GT && GET_CODE (XEXP (comparison, 1)) == CONST_INT && INTVAL (XEXP (comparison, 1)) == -1) || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx)) && GET_CODE (bl->biv->add_val) == CONST_INT && INTVAL (bl->biv->add_val) < 0) { /* Initial value must be greater than 0, init_val % -dec_value == 0 to ensure that it equals zero on the last iteration */ if (GET_CODE (bl->initial_value) == CONST_INT && INTVAL (bl->initial_value) > 0 && (INTVAL (bl->initial_value) % (-INTVAL (bl->biv->add_val))) == 0) { /* register always nonnegative, add REG_NOTE to branch */ REG_NOTES (PREV_INSN (loop_end)) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX, REG_NOTES (PREV_INSN (loop_end))); bl->nonneg = 1; return 1; } /* If the decrement is 1 and the value was tested as >= 0 before the loop, then we can safely optimize. */ for (p = loop_start; p; p = PREV_INSN (p)) { if (GET_CODE (p) == CODE_LABEL) break; if (GET_CODE (p) != JUMP_INSN) continue; before_comparison = get_condition_for_loop (p); if (before_comparison && XEXP (before_comparison, 0) == bl->biv->dest_reg && GET_CODE (before_comparison) == LT && XEXP (before_comparison, 1) == const0_rtx && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start) && INTVAL (bl->biv->add_val) == -1) { REG_NOTES (PREV_INSN (loop_end)) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX, REG_NOTES (PREV_INSN (loop_end))); bl->nonneg = 1; return 1; } } } else if (GET_CODE (bl->biv->add_val) == CONST_INT && INTVAL (bl->biv->add_val) > 0) { /* Try to change inc to dec, so can apply above optimization. */ /* Can do this if: all registers modified are induction variables or invariant, all memory references have non-overlapping addresses (obviously true if only one write) allow 2 insns for the compare/jump at the end of the loop. */ /* Also, we must avoid any instructions which use both the reversed biv and another biv. Such instructions will fail if the loop is reversed. We meet this condition by requiring that either no_use_except_counting is true, or else that there is only one biv. */ int num_nonfixed_reads = 0; /* 1 if the iteration var is used only to count iterations. */ int no_use_except_counting = 0; /* 1 if the loop has no memory store, or it has a single memory store which is reversible. */ int reversible_mem_store = 1; if (bl->giv_count == 0 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]) { rtx bivreg = regno_reg_rtx[bl->regno]; /* If there are no givs for this biv, and the only exit is the fall through at the end of the loop, then see if perhaps there are no uses except to count. */ no_use_except_counting = 1; for (p = loop_start; p != loop_end; p = NEXT_INSN (p)) if (GET_RTX_CLASS (GET_CODE (p)) == 'i') { rtx set = single_set (p); if (set && GET_CODE (SET_DEST (set)) == REG && REGNO (SET_DEST (set)) == bl->regno) /* An insn that sets the biv is okay. */ ; else if ((p == prev_nonnote_insn (prev_nonnote_insn (loop_end)) || p == prev_nonnote_insn (loop_end)) && reg_mentioned_p (bivreg, PATTERN (p))) { /* If either of these insns uses the biv and sets a pseudo that has more than one usage, then the biv has uses other than counting since it's used to derive a value that is used more than one time. */ note_set_pseudo_multiple_uses_retval = 0; note_stores (PATTERN (p), note_set_pseudo_multiple_uses); if (note_set_pseudo_multiple_uses_retval) { no_use_except_counting = 0; break; } } else if (reg_mentioned_p (bivreg, PATTERN (p))) { no_use_except_counting = 0; break; } } } if (no_use_except_counting) ; /* no need to worry about MEMs. */ else if (num_mem_sets <= 1) { for (p = loop_start; p != loop_end; p = NEXT_INSN (p)) if (GET_RTX_CLASS (GET_CODE (p)) == 'i') num_nonfixed_reads += count_nonfixed_reads (PATTERN (p)); /* If the loop has a single store, and the destination address is invariant, then we can't reverse the loop, because this address might then have the wrong value at loop exit. This would work if the source was invariant also, however, in that case, the insn should have been moved out of the loop. */ if (num_mem_sets == 1) { struct induction *v; reversible_mem_store = (! unknown_address_altered && ! invariant_p (XEXP (XEXP (loop_store_mems, 0), 0))); /* If the store depends on a register that is set after the store, it depends on the initial value, and is thus not reversible. */ for (v = bl->giv; reversible_mem_store && v; v = v->next_iv) { if (v->giv_type == DEST_REG && reg_mentioned_p (v->dest_reg, PATTERN (first_loop_store_insn)) && loop_insn_first_p (first_loop_store_insn, v->insn)) reversible_mem_store = 0; } } } else return 0; /* This code only acts for innermost loops. Also it simplifies the memory address check by only reversing loops with zero or one memory access. Two memory accesses could involve parts of the same array, and that can't be reversed. If the biv is used only for counting, than we don't need to worry about all these things. */ if ((num_nonfixed_reads <= 1 && !loop_has_call && !loop_has_volatile && reversible_mem_store && (bl->giv_count + bl->biv_count + num_mem_sets + num_movables + compare_and_branch == insn_count) && (bl == loop_iv_list && bl->next == 0)) || no_use_except_counting) { rtx tem; /* Loop can be reversed. */ if (loop_dump_stream) fprintf (loop_dump_stream, "Can reverse loop\n"); /* Now check other conditions: The increment must be a constant, as must the initial value, and the comparison code must be LT. This test can probably be improved since +/- 1 in the constant can be obtained by changing LT to LE and vice versa; this is confusing. */ if (comparison /* for constants, LE gets turned into LT */ && (GET_CODE (comparison) == LT || (GET_CODE (comparison) == LE && no_use_except_counting))) { HOST_WIDE_INT add_val, add_adjust, comparison_val; rtx initial_value, comparison_value; int nonneg = 0; enum rtx_code cmp_code; int comparison_const_width; unsigned HOST_WIDE_INT comparison_sign_mask; add_val = INTVAL (bl->biv->add_val); comparison_value = XEXP (comparison, 1); if (GET_MODE (comparison_value) == VOIDmode) comparison_const_width = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0))); else comparison_const_width = GET_MODE_BITSIZE (GET_MODE (comparison_value)); if (comparison_const_width > HOST_BITS_PER_WIDE_INT) comparison_const_width = HOST_BITS_PER_WIDE_INT; comparison_sign_mask = (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1); /* If the comparison value is not a loop invariant, then we can not reverse this loop. ??? If the insns which initialize the comparison value as a whole compute an invariant result, then we could move them out of the loop and proceed with loop reversal. */ if (!invariant_p (comparison_value)) return 0; if (GET_CODE (comparison_value) == CONST_INT) comparison_val = INTVAL (comparison_value); initial_value = bl->initial_value; /* Normalize the initial value if it is an integer and has no other use except as a counter. This will allow a few more loops to be reversed. */ if (no_use_except_counting && GET_CODE (comparison_value) == CONST_INT && GET_CODE (initial_value) == CONST_INT) { comparison_val = comparison_val - INTVAL (bl->initial_value); /* The code below requires comparison_val to be a multiple of add_val in order to do the loop reversal, so round up comparison_val to a multiple of add_val. Since comparison_value is constant, we know that the current comparison code is LT. */ comparison_val = comparison_val + add_val - 1; comparison_val -= (unsigned HOST_WIDE_INT) comparison_val % add_val; /* We postpone overflow checks for COMPARISON_VAL here; even if there is an overflow, we might still be able to reverse the loop, if converting the loop exit test to NE is possible. */ initial_value = const0_rtx; } /* First check if we can do a vanilla loop reversal. */ if (initial_value == const0_rtx /* If we have a decrement_and_branch_on_count, prefer the NE test, since this will allow that instruction to be generated. Note that we must use a vanilla loop reversal if the biv is used to calculate a giv or has a non-counting use. */ #if ! defined (HAVE_decrement_and_branch_until_zero) && defined (HAVE_decrement_and_branch_on_count) && (! (add_val == 1 && loop_info->vtop && (bl->biv_count == 0 || no_use_except_counting))) #endif && GET_CODE (comparison_value) == CONST_INT /* Now do postponed overflow checks on COMPARISON_VAL. */ && ! (((comparison_val - add_val) ^ INTVAL (comparison_value)) & comparison_sign_mask)) { /* Register will always be nonnegative, with value 0 on last iteration */ add_adjust = add_val; nonneg = 1; cmp_code = GE; } else if (add_val == 1 && loop_info->vtop && (bl->biv_count == 0 || no_use_except_counting)) { add_adjust = 0; cmp_code = NE; } else return 0; if (GET_CODE (comparison) == LE) add_adjust -= add_val; /* If the initial value is not zero, or if the comparison value is not an exact multiple of the increment, then we can not reverse this loop. */ if (initial_value == const0_rtx && GET_CODE (comparison_value) == CONST_INT) { if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0) return 0; } else { if (! no_use_except_counting || add_val != 1) return 0; } final_value = comparison_value; /* Reset these in case we normalized the initial value and comparison value above. */ if (GET_CODE (comparison_value) == CONST_INT && GET_CODE (initial_value) == CONST_INT) { comparison_value = GEN_INT (comparison_val); final_value = GEN_INT (comparison_val + INTVAL (bl->initial_value)); } bl->initial_value = initial_value; /* Save some info needed to produce the new insns. */ reg = bl->biv->dest_reg; jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1); if (jump_label == pc_rtx) jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2); new_add_val = GEN_INT (- INTVAL (bl->biv->add_val)); /* Set start_value; if this is not a CONST_INT, we need to generate a SUB. Initialize biv to start_value before loop start. The old initializing insn will be deleted as a dead store by flow.c. */ if (initial_value == const0_rtx && GET_CODE (comparison_value) == CONST_INT) { start_value = GEN_INT (comparison_val - add_adjust); emit_insn_before (gen_move_insn (reg, start_value), loop_start); } else if (GET_CODE (initial_value) == CONST_INT) { rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust); enum machine_mode mode = GET_MODE (reg); enum insn_code icode = add_optab->handlers[(int) mode].insn_code; if (! (*insn_operand_predicate[icode][0]) (reg, mode) || ! ((*insn_operand_predicate[icode][1]) (comparison_value, mode)) || ! (*insn_operand_predicate[icode][2]) (offset, mode)) return 0; start_value = gen_rtx_PLUS (mode, comparison_value, offset); emit_insn_before ((GEN_FCN (icode) (reg, comparison_value, offset)), loop_start); if (GET_CODE (comparison) == LE) final_value = gen_rtx_PLUS (mode, comparison_value, GEN_INT (add_val)); } else if (! add_adjust) { enum machine_mode mode = GET_MODE (reg); enum insn_code icode = sub_optab->handlers[(int) mode].insn_code; if (! (*insn_operand_predicate[icode][0]) (reg, mode) || ! ((*insn_operand_predicate[icode][1]) (comparison_value, mode)) || ! ((*insn_operand_predicate[icode][2]) (initial_value, mode))) return 0; start_value = gen_rtx_MINUS (mode, comparison_value, initial_value); emit_insn_before ((GEN_FCN (icode) (reg, comparison_value, initial_value)), loop_start); } else /* We could handle the other cases too, but it'll be better to have a testcase first. */ return 0; /* We may not have a single insn which can increment a reg, so create a sequence to hold all the insns from expand_inc. */ start_sequence (); expand_inc (reg, new_add_val); tem = gen_sequence (); end_sequence (); p = emit_insn_before (tem, bl->biv->insn); delete_insn (bl->biv->insn); /* Update biv info to reflect its new status. */ bl->biv->insn = p; bl->initial_value = start_value; bl->biv->add_val = new_add_val; /* Update loop info. */ loop_info->initial_value = reg; loop_info->initial_equiv_value = reg; loop_info->final_value = const0_rtx; loop_info->final_equiv_value = const0_rtx; loop_info->comparison_value = const0_rtx; loop_info->comparison_code = cmp_code; loop_info->increment = new_add_val; /* Inc LABEL_NUSES so that delete_insn will not delete the label. */ LABEL_NUSES (XEXP (jump_label, 0)) ++; /* Emit an insn after the end of the loop to set the biv's proper exit value if it is used anywhere outside the loop. */ if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare)) || ! bl->init_insn || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn)) emit_insn_after (gen_move_insn (reg, final_value), loop_end); /* Delete compare/branch at end of loop. */ delete_insn (PREV_INSN (loop_end)); if (compare_and_branch == 2) delete_insn (first_compare); /* Add new compare/branch insn at end of loop. */ start_sequence (); emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX, GET_MODE (reg), 0, 0, XEXP (jump_label, 0)); tem = gen_sequence (); end_sequence (); emit_jump_insn_before (tem, loop_end); for (tem = PREV_INSN (loop_end); tem && GET_CODE (tem) != JUMP_INSN; tem = PREV_INSN (tem)) ; if (tem) JUMP_LABEL (tem) = XEXP (jump_label, 0); if (nonneg) { if (tem) { /* Increment of LABEL_NUSES done above. */ /* Register is now always nonnegative, so add REG_NONNEG note to the branch. */ REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX, REG_NOTES (tem)); } bl->nonneg = 1; } /* No insn may reference both the reversed and another biv or it will fail (see comment near the top of the loop reversal code). Earlier on, we have verified that the biv has no use except counting, or it is the only biv in this function. However, the code that computes no_use_except_counting does not verify reg notes. It's possible to have an insn that references another biv, and has a REG_EQUAL note with an expression based on the reversed biv. To avoid this case, remove all REG_EQUAL notes based on the reversed biv here. */ for (p = loop_start; p != loop_end; p = NEXT_INSN (p)) if (GET_RTX_CLASS (GET_CODE (p)) == 'i') { rtx *pnote; rtx set = single_set (p); /* If this is a set of a GIV based on the reversed biv, any REG_EQUAL notes should still be correct. */ if (! set || GET_CODE (SET_DEST (set)) != REG || (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements || REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT || REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg) for (pnote = ®_NOTES (p); *pnote;) { if (REG_NOTE_KIND (*pnote) == REG_EQUAL && reg_mentioned_p (regno_reg_rtx[bl->regno], XEXP (*pnote, 0))) *pnote = XEXP (*pnote, 1); else pnote = &XEXP (*pnote, 1); } } /* Mark that this biv has been reversed. Each giv which depends on this biv, and which is also live past the end of the loop will have to be fixed up. */ bl->reversed = 1; if (loop_dump_stream) { fprintf (loop_dump_stream, "Reversed loop"); if (bl->nonneg) fprintf (loop_dump_stream, " and added reg_nonneg\n"); else fprintf (loop_dump_stream, "\n"); } return 1; } } } return 0; } /* Verify whether the biv BL appears to be eliminable, based on the insns in the loop that refer to it. LOOP_START is the first insn of the loop, and END is the end insn. If ELIMINATE_P is non-zero, actually do the elimination. THRESHOLD and INSN_COUNT are from loop_optimize and are used to determine whether invariant insns should be placed inside or at the start of the loop. */ static int maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count) struct iv_class *bl; rtx loop_start; rtx end; int eliminate_p; int threshold, insn_count; { rtx reg = bl->biv->dest_reg; rtx p; /* Scan all insns in the loop, stopping if we find one that uses the biv in a way that we cannot eliminate. */ for (p = loop_start; p != end; p = NEXT_INSN (p)) { enum rtx_code code = GET_CODE (p); rtx where = threshold >= insn_count ? loop_start : p; /* If this is a libcall that sets a giv, skip ahead to its end. */ if (GET_RTX_CLASS (code) == 'i') { rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX); if (note) { rtx last = XEXP (note, 0); rtx set = single_set (last); if (set && GET_CODE (SET_DEST (set)) == REG) { int regno = REGNO (SET_DEST (set)); if (regno < max_reg_before_loop && REG_IV_TYPE (regno) == GENERAL_INDUCT && REG_IV_INFO (regno)->src_reg == bl->biv->src_reg) p = last; } } } if ((code == INSN || code == JUMP_INSN || code == CALL_INSN) && reg_mentioned_p (reg, PATTERN (p)) && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where)) { if (loop_dump_stream) fprintf (loop_dump_stream, "Cannot eliminate biv %d: biv used in insn %d.\n", bl->regno, INSN_UID (p)); break; } } if (p == end) { if (loop_dump_stream) fprintf (loop_dump_stream, "biv %d %s eliminated.\n", bl->regno, eliminate_p ? "was" : "can be"); return 1; } return 0; } /* INSN and REFERENCE are instructions in the same insn chain. Return non-zero if INSN is first. */ int loop_insn_first_p (insn, reference) rtx insn, reference; { rtx p, q; for (p = insn, q = reference; ;) { /* Start with test for not first so that INSN == REFERENCE yields not first. */ if (q == insn || ! p) return 0; if (p == reference || ! q) return 1; /* Either of P or Q might be a NOTE. Notes have the same LUID as the previous insn, hence the <= comparison below does not work if P is a note. */ if (INSN_UID (p) < max_uid_for_loop && INSN_UID (q) < max_uid_for_loop && GET_CODE (p) != NOTE) return INSN_LUID (p) <= INSN_LUID (q); if (INSN_UID (p) >= max_uid_for_loop || GET_CODE (p) == NOTE) p = NEXT_INSN (p); if (INSN_UID (q) >= max_uid_for_loop) q = NEXT_INSN (q); } } /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if the offset that we have to take into account due to auto-increment / div derivation is zero. */ static int biv_elimination_giv_has_0_offset (biv, giv, insn) struct induction *biv, *giv; rtx insn; { /* If the giv V had the auto-inc address optimization applied to it, and INSN occurs between the giv insn and the biv insn, then we'd have to adjust the value used here. This is rare, so we don't bother to make this possible. */ if (giv->auto_inc_opt && ((loop_insn_first_p (giv->insn, insn) && loop_insn_first_p (insn, biv->insn)) || (loop_insn_first_p (biv->insn, insn) && loop_insn_first_p (insn, giv->insn)))) return 0; /* If the giv V was derived from another giv, and INSN does not occur between the giv insn and the biv insn, then we'd have to adjust the value used here. This is rare, so we don't bother to make this possible. */ if (giv->derived_from && ! (giv->always_executed && loop_insn_first_p (giv->insn, insn) && loop_insn_first_p (insn, biv->insn))) return 0; if (giv->same && giv->same->derived_from && ! (giv->same->always_executed && loop_insn_first_p (giv->same->insn, insn) && loop_insn_first_p (insn, biv->insn))) return 0; return 1; } /* If BL appears in X (part of the pattern of INSN), see if we can eliminate its use. If so, return 1. If not, return 0. If BIV does not appear in X, return 1. If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates where extra insns should be added. Depending on how many items have been moved out of the loop, it will either be before INSN or at the start of the loop. */ static int maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where) rtx x, insn; struct iv_class *bl; int eliminate_p; rtx where; { enum rtx_code code = GET_CODE (x); rtx reg = bl->biv->dest_reg; enum machine_mode mode = GET_MODE (reg); struct induction *v; rtx arg, tem; #ifdef HAVE_cc0 rtx new; #endif int arg_operand; char *fmt; int i, j; switch (code) { case REG: /* If we haven't already been able to do something with this BIV, we can't eliminate it. */ if (x == reg) return 0; return 1; case SET: /* If this sets the BIV, it is not a problem. */ if (SET_DEST (x) == reg) return 1; /* If this is an insn that defines a giv, it is also ok because it will go away when the giv is reduced. */ for (v = bl->giv; v; v = v->next_iv) if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg) return 1; #ifdef HAVE_cc0 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg) { /* Can replace with any giv that was reduced and that has (MULT_VAL != 0) and (ADD_VAL == 0). Require a constant for MULT_VAL, so we know it's nonzero. ??? We disable this optimization to avoid potential overflows. */ for (v = bl->giv; v; v = v->next_iv) if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx && v->add_val == const0_rtx && ! v->ignore && ! v->maybe_dead && v->always_computable && v->mode == mode && 0) { if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn)) continue; if (! eliminate_p) return 1; /* If the giv has the opposite direction of change, then reverse the comparison. */ if (INTVAL (v->mult_val) < 0) new = gen_rtx_COMPARE (GET_MODE (v->new_reg), const0_rtx, v->new_reg); else new = v->new_reg; /* We can probably test that giv's reduced reg. */ if (validate_change (insn, &SET_SRC (x), new, 0)) return 1; } /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0); replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL). Require a constant for MULT_VAL, so we know it's nonzero. ??? Do this only if ADD_VAL is a pointer to avoid a potential overflow problem. */ for (v = bl->giv; v; v = v->next_iv) if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx && ! v->ignore && ! v->maybe_dead && v->always_computable && v->mode == mode && (GET_CODE (v->add_val) == SYMBOL_REF || GET_CODE (v->add_val) == LABEL_REF || GET_CODE (v->add_val) == CONST || (GET_CODE (v->add_val) == REG && REGNO_POINTER_FLAG (REGNO (v->add_val))))) { if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn)) continue; if (! eliminate_p) return 1; /* If the giv has the opposite direction of change, then reverse the comparison. */ if (INTVAL (v->mult_val) < 0) new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val), v->new_reg); else new = gen_rtx_COMPARE (VOIDmode, v->new_reg, copy_rtx (v->add_val)); /* Replace biv with the giv's reduced register. */ update_reg_last_use (v->add_val, insn); if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0)) return 1; /* Insn doesn't support that constant or invariant. Copy it into a register (it will be a loop invariant.) */ tem = gen_reg_rtx (GET_MODE (v->new_reg)); emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)), where); /* Substitute the new register for its invariant value in the compare expression. */ XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem; if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0)) return 1; } } #endif break; case COMPARE: case EQ: case NE: case GT: case GE: case GTU: case GEU: case LT: case LE: case LTU: case LEU: /* See if either argument is the biv. */ if (XEXP (x, 0) == reg) arg = XEXP (x, 1), arg_operand = 1; else if (XEXP (x, 1) == reg) arg = XEXP (x, 0), arg_operand = 0; else break; if (CONSTANT_P (arg)) { /* First try to replace with any giv that has constant positive mult_val and constant add_val. We might be able to support negative mult_val, but it seems complex to do it in general. */ for (v = bl->giv; v; v = v->next_iv) if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0 && (GET_CODE (v->add_val) == SYMBOL_REF || GET_CODE (v->add_val) == LABEL_REF || GET_CODE (v->add_val) == CONST || (GET_CODE (v->add_val) == REG && REGNO_POINTER_FLAG (REGNO (v->add_val)))) && ! v->ignore && ! v->maybe_dead && v->always_computable && v->mode == mode) { if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn)) continue; if (! eliminate_p) return 1; /* Replace biv with the giv's reduced reg. */ XEXP (x, 1-arg_operand) = v->new_reg; /* If all constants are actually constant integers and the derived constant can be directly placed in the COMPARE, do so. */ if (GET_CODE (arg) == CONST_INT && GET_CODE (v->mult_val) == CONST_INT && GET_CODE (v->add_val) == CONST_INT && validate_change (insn, &XEXP (x, arg_operand), GEN_INT (INTVAL (arg) * INTVAL (v->mult_val) + INTVAL (v->add_val)), 0)) return 1; /* Otherwise, load it into a register. */ tem = gen_reg_rtx (mode); emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where); if (validate_change (insn, &XEXP (x, arg_operand), tem, 0)) return 1; /* If that failed, put back the change we made above. */ XEXP (x, 1-arg_operand) = reg; } /* Look for giv with positive constant mult_val and nonconst add_val. Insert insns to calculate new compare value. ??? Turn this off due to possible overflow. */ for (v = bl->giv; v; v = v->next_iv) if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0 && ! v->ignore && ! v->maybe_dead && v->always_computable && v->mode == mode && 0) { rtx tem; if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn)) continue; if (! eliminate_p) return 1; tem = gen_reg_rtx (mode); /* Replace biv with giv's reduced register. */ validate_change (insn, &XEXP (x, 1 - arg_operand), v->new_reg, 1); /* Compute value to compare against. */ emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where); /* Use it in this insn. */ validate_change (insn, &XEXP (x, arg_operand), tem, 1); if (apply_change_group ()) return 1; } } else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM) { if (invariant_p (arg) == 1) { /* Look for giv with constant positive mult_val and nonconst add_val. Insert insns to compute new compare value. ??? Turn this off due to possible overflow. */ for (v = bl->giv; v; v = v->next_iv) if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0 && ! v->ignore && ! v->maybe_dead && v->always_computable && v->mode == mode && 0) { rtx tem; if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn)) continue; if (! eliminate_p) return 1; tem = gen_reg_rtx (mode); /* Replace biv with giv's reduced register. */ validate_change (insn, &XEXP (x, 1 - arg_operand), v->new_reg, 1); /* Compute value to compare against. */ emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where); validate_change (insn, &XEXP (x, arg_operand), tem, 1); if (apply_change_group ()) return 1; } } /* This code has problems. Basically, you can't know when seeing if we will eliminate BL, whether a particular giv of ARG will be reduced. If it isn't going to be reduced, we can't eliminate BL. We can try forcing it to be reduced, but that can generate poor code. The problem is that the benefit of reducing TV, below should be increased if BL can actually be eliminated, but this means we might have to do a topological sort of the order in which we try to process biv. It doesn't seem worthwhile to do this sort of thing now. */ #if 0 /* Otherwise the reg compared with had better be a biv. */ if (GET_CODE (arg) != REG || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT) return 0; /* Look for a pair of givs, one for each biv, with identical coefficients. */ for (v = bl->giv; v; v = v->next_iv) { struct induction *tv; if (v->ignore || v->maybe_dead || v->mode != mode) continue; for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv) if (! tv->ignore && ! tv->maybe_dead && rtx_equal_p (tv->mult_val, v->mult_val) && rtx_equal_p (tv->add_val, v->add_val) && tv->mode == mode) { if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn)) continue; if (! eliminate_p) return 1; /* Replace biv with its giv's reduced reg. */ XEXP (x, 1-arg_operand) = v->new_reg; /* Replace other operand with the other giv's reduced reg. */ XEXP (x, arg_operand) = tv->new_reg; return 1; } } #endif } /* If we get here, the biv can't be eliminated. */ return 0; case MEM: /* If this address is a DEST_ADDR giv, it doesn't matter if the biv is used in it, since it will be replaced. */ for (v = bl->giv; v; v = v->next_iv) if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0)) return 1; break; default: break; } /* See if any subexpression fails elimination. */ fmt = GET_RTX_FORMAT (code); for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { switch (fmt[i]) { case 'e': if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl, eliminate_p, where)) return 0; break; case 'E': for (j = XVECLEN (x, i) - 1; j >= 0; j--) if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl, eliminate_p, where)) return 0; break; } } return 1; } /* Return nonzero if the last use of REG is in an insn following INSN in the same basic block. */ static int last_use_this_basic_block (reg, insn) rtx reg; rtx insn; { rtx n; for (n = insn; n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN; n = NEXT_INSN (n)) { if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n)) return 1; } return 0; } /* Called via `note_stores' to record the initial value of a biv. Here we just record the location of the set and process it later. */ static void record_initial (dest, set) rtx dest; rtx set; { struct iv_class *bl; if (GET_CODE (dest) != REG || REGNO (dest) >= max_reg_before_loop || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT) return; bl = reg_biv_class[REGNO (dest)]; /* If this is the first set found, record it. */ if (bl->init_insn == 0) { bl->init_insn = note_insn; bl->init_set = set; } } /* If any of the registers in X are "old" and currently have a last use earlier than INSN, update them to have a last use of INSN. Their actual last use will be the previous insn but it will not have a valid uid_luid so we can't use it. */ static void update_reg_last_use (x, insn) rtx x; rtx insn; { /* Check for the case where INSN does not have a valid luid. In this case, there is no need to modify the regno_last_uid, as this can only happen when code is inserted after the loop_end to set a pseudo's final value, and hence this insn will never be the last use of x. */ if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop && INSN_UID (insn) < max_uid_for_loop && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)]) REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn); else { register int i, j; register char *fmt = GET_RTX_FORMAT (GET_CODE (x)); for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) { if (fmt[i] == 'e') update_reg_last_use (XEXP (x, i), insn); else if (fmt[i] == 'E') for (j = XVECLEN (x, i) - 1; j >= 0; j--) update_reg_last_use (XVECEXP (x, i, j), insn); } } } /* Given a jump insn JUMP, return the condition that will cause it to branch to its JUMP_LABEL. If the condition cannot be understood, or is an inequality floating-point comparison which needs to be reversed, 0 will be returned. If EARLIEST is non-zero, it is a pointer to a place where the earliest insn used in locating the condition was found. If a replacement test of the condition is desired, it should be placed in front of that insn and we will be sure that the inputs are still valid. The condition will be returned in a canonical form to simplify testing by callers. Specifically: (1) The code will always be a comparison operation (EQ, NE, GT, etc.). (2) Both operands will be machine operands; (cc0) will have been replaced. (3) If an operand is a constant, it will be the second operand. (4) (LE x const) will be replaced with (LT x ) and similarly for GE, GEU, and LEU. */ rtx get_condition (jump, earliest) rtx jump; rtx *earliest; { enum rtx_code code; rtx prev = jump; rtx set; rtx tem; rtx op0, op1; int reverse_code = 0; int did_reverse_condition = 0; enum machine_mode mode; /* If this is not a standard conditional jump, we can't parse it. */ if (GET_CODE (jump) != JUMP_INSN || ! condjump_p (jump) || simplejump_p (jump)) return 0; code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0)); mode = GET_MODE (XEXP (SET_SRC (PATTERN (jump)), 0)); op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0); op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1); if (earliest) *earliest = jump; /* If this branches to JUMP_LABEL when the condition is false, reverse the condition. */ if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump)) code = reverse_condition (code), did_reverse_condition ^= 1; /* If we are comparing a register with zero, see if the register is set in the previous insn to a COMPARE or a comparison operation. Perform the same tests as a function of STORE_FLAG_VALUE as find_comparison_args in cse.c */ while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0))) { /* Set non-zero when we find something of interest. */ rtx x = 0; #ifdef HAVE_cc0 /* If comparison with cc0, import actual comparison from compare insn. */ if (op0 == cc0_rtx) { if ((prev = prev_nonnote_insn (prev)) == 0 || GET_CODE (prev) != INSN || (set = single_set (prev)) == 0 || SET_DEST (set) != cc0_rtx) return 0; op0 = SET_SRC (set); op1 = CONST0_RTX (GET_MODE (op0)); if (earliest) *earliest = prev; } #endif /* If this is a COMPARE, pick up the two things being compared. */ if (GET_CODE (op0) == COMPARE) { op1 = XEXP (op0, 1); op0 = XEXP (op0, 0); continue; } else if (GET_CODE (op0) != REG) break; /* Go back to the previous insn. Stop if it is not an INSN. We also stop if it isn't a single set or if it has a REG_INC note because we don't want to bother dealing with it. */ if ((prev = prev_nonnote_insn (prev)) == 0 || GET_CODE (prev) != INSN || FIND_REG_INC_NOTE (prev, 0) || (set = single_set (prev)) == 0) break; /* If this is setting OP0, get what it sets it to if it looks relevant. */ if (rtx_equal_p (SET_DEST (set), op0)) { enum machine_mode inner_mode = GET_MODE (SET_SRC (set)); /* ??? We may not combine comparisons done in a CCmode with comparisons not done in a CCmode. This is to aid targets like Alpha that have an IEEE compliant EQ instruction, and a non-IEEE compliant BEQ instruction. The use of CCmode is actually artificial, simply to prevent the combination, but should not affect other platforms. However, we must allow VOIDmode comparisons to match either CCmode or non-CCmode comparison, because some ports have modeless comparisons inside branch patterns. ??? This mode check should perhaps look more like the mode check in simplify_comparison in combine. */ if ((GET_CODE (SET_SRC (set)) == COMPARE || (((code == NE || (code == LT && GET_MODE_CLASS (inner_mode) == MODE_INT && (GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_WIDE_INT) && (STORE_FLAG_VALUE & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (inner_mode) - 1)))) #ifdef FLOAT_STORE_FLAG_VALUE || (code == LT && GET_MODE_CLASS (inner_mode) == MODE_FLOAT && FLOAT_STORE_FLAG_VALUE < 0) #endif )) && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')) && (((GET_MODE_CLASS (mode) == MODE_CC) == (GET_MODE_CLASS (inner_mode) == MODE_CC)) || mode == VOIDmode || inner_mode == VOIDmode)) x = SET_SRC (set); else if (((code == EQ || (code == GE && (GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_WIDE_INT) && GET_MODE_CLASS (inner_mode) == MODE_INT && (STORE_FLAG_VALUE & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (inner_mode) - 1)))) #ifdef FLOAT_STORE_FLAG_VALUE || (code == GE && GET_MODE_CLASS (inner_mode) == MODE_FLOAT && FLOAT_STORE_FLAG_VALUE < 0) #endif )) && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<' && (((GET_MODE_CLASS (mode) == MODE_CC) == (GET_MODE_CLASS (inner_mode) == MODE_CC)) || mode == VOIDmode || inner_mode == VOIDmode)) { /* We might have reversed a LT to get a GE here. But this wasn't actually the comparison of data, so we don't flag that we have had to reverse the condition. */ did_reverse_condition ^= 1; reverse_code = 1; x = SET_SRC (set); } else break; } else if (reg_set_p (op0, prev)) /* If this sets OP0, but not directly, we have to give up. */ break; if (x) { if (GET_RTX_CLASS (GET_CODE (x)) == '<') code = GET_CODE (x); if (reverse_code) { code = reverse_condition (code); did_reverse_condition ^= 1; reverse_code = 0; } op0 = XEXP (x, 0), op1 = XEXP (x, 1); if (earliest) *earliest = prev; } } /* If constant is first, put it last. */ if (CONSTANT_P (op0)) code = swap_condition (code), tem = op0, op0 = op1, op1 = tem; /* If OP0 is the result of a comparison, we weren't able to find what was really being compared, so fail. */ if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC) return 0; /* Canonicalize any ordered comparison with integers involving equality if we can do computations in the relevant mode and we do not overflow. */ if (GET_CODE (op1) == CONST_INT && GET_MODE (op0) != VOIDmode && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT) { HOST_WIDE_INT const_val = INTVAL (op1); unsigned HOST_WIDE_INT uconst_val = const_val; unsigned HOST_WIDE_INT max_val = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0)); switch (code) { case LE: if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1) code = LT, op1 = GEN_INT (const_val + 1); break; /* When cross-compiling, const_val might be sign-extended from BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */ case GE: if ((HOST_WIDE_INT) (const_val & max_val) != (((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1)))) code = GT, op1 = GEN_INT (const_val - 1); break; case LEU: if (uconst_val < max_val) code = LTU, op1 = GEN_INT (uconst_val + 1); break; case GEU: if (uconst_val != 0) code = GTU, op1 = GEN_INT (uconst_val - 1); break; default: break; } } /* If this was floating-point and we reversed anything other than an EQ or NE, return zero. */ if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT && did_reverse_condition && code != NE && code != EQ && ! flag_fast_math && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT) return 0; #ifdef HAVE_cc0 /* Never return CC0; return zero instead. */ if (op0 == cc0_rtx) return 0; #endif return gen_rtx_fmt_ee (code, VOIDmode, op0, op1); } /* Similar to above routine, except that we also put an invariant last unless both operands are invariants. */ rtx get_condition_for_loop (x) rtx x; { rtx comparison = get_condition (x, NULL_PTR); if (comparison == 0 || ! invariant_p (XEXP (comparison, 0)) || invariant_p (XEXP (comparison, 1))) return comparison; return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode, XEXP (comparison, 1), XEXP (comparison, 0)); } #ifdef HAVE_decrement_and_branch_on_count /* Instrument loop for insertion of bct instruction. We distinguish between loops with compile-time bounds and those with run-time bounds. Information from loop_iterations() is used to compute compile-time bounds. Run-time bounds should use loop preconditioning, but currently ignored. */ static void insert_bct (loop_start, loop_end, loop_info) rtx loop_start, loop_end; struct loop_info *loop_info; { int i; unsigned HOST_WIDE_INT n_iterations; int increment_direction, compare_direction; /* If the loop condition is <= or >=, the number of iteration is 1 more than the range of the bounds of the loop. */ int add_iteration = 0; enum machine_mode loop_var_mode = word_mode; int loop_num = uid_loop_num [INSN_UID (loop_start)]; /* It's impossible to instrument a competely unrolled loop. */ if (loop_info->unroll_number == -1) return; /* Make sure that the count register is not in use. */ if (loop_used_count_register [loop_num]) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: BCT instrumentation failed: count register already in use\n", loop_num); return; } /* Make sure that the function has no indirect jumps. */ if (indirect_jump_in_function) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: BCT instrumentation failed: indirect jump in function\n", loop_num); return; } /* Make sure that the last loop insn is a conditional jump. */ if (GET_CODE (PREV_INSN (loop_end)) != JUMP_INSN || ! condjump_p (PREV_INSN (loop_end)) || simplejump_p (PREV_INSN (loop_end))) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: BCT instrumentation failed: invalid jump at loop end\n", loop_num); return; } /* Make sure that the loop does not contain a function call (the count register might be altered by the called function). */ if (loop_has_call) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: BCT instrumentation failed: function call in loop\n", loop_num); return; } /* Make sure that the loop does not jump via a table. (the count register might be used to perform the branch on table). */ if (loop_has_tablejump) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: BCT instrumentation failed: computed branch in the loop\n", loop_num); return; } /* Account for loop unrolling in instrumented iteration count. */ if (loop_info->unroll_number > 1) n_iterations = loop_info->n_iterations / loop_info->unroll_number; else n_iterations = loop_info->n_iterations; if (n_iterations != 0 && n_iterations < 3) { /* Allow an enclosing outer loop to benefit if possible. */ if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: Too few iterations to benefit from BCT optimization\n", loop_num); return; } /* Try to instrument the loop. */ /* Handle the simpler case, where the bounds are known at compile time. */ if (n_iterations > 0) { /* Mark all enclosing loops that they cannot use count register. */ for (i = loop_num; i != -1; i = loop_outer_loop[i]) loop_used_count_register[i] = 1; instrument_loop_bct (loop_start, loop_end, GEN_INT (n_iterations)); return; } /* Handle the more complex case, that the bounds are NOT known at compile time. In this case we generate run_time calculation of the number of iterations. */ if (loop_info->iteration_var == 0) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: BCT Runtime Instrumentation failed: no loop iteration variable found\n", loop_num); return; } if (GET_MODE_CLASS (GET_MODE (loop_info->iteration_var)) != MODE_INT || GET_MODE_SIZE (GET_MODE (loop_info->iteration_var)) != UNITS_PER_WORD) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: BCT Runtime Instrumentation failed: loop variable not integer\n", loop_num); return; } /* With runtime bounds, if the compare is of the form '!=' we give up */ if (loop_info->comparison_code == NE) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct %d: BCT Runtime Instrumentation failed: runtime bounds with != comparison\n", loop_num); return; } /* Use common loop preconditioning code instead. */ #if 0 else { /* We rely on the existence of run-time guard to ensure that the loop executes at least once. */ rtx sequence; rtx iterations_num_reg; unsigned HOST_WIDE_INT increment_value_abs = INTVAL (increment) * increment_direction; /* make sure that the increment is a power of two, otherwise (an expensive) divide is needed. */ if (exact_log2 (increment_value_abs) == -1) { if (loop_dump_stream) fprintf (loop_dump_stream, "insert_bct: not instrumenting BCT because the increment is not power of 2\n"); return; } /* compute the number of iterations */ start_sequence (); { rtx temp_reg; /* Again, the number of iterations is calculated by: ; ; compare-val - initial-val + (increment -1) + additional-iteration ; num_iterations = ----------------------------------------------------------------- ; increment */ /* ??? Do we have to call copy_rtx here before passing rtx to expand_binop? */ if (compare_direction > 0) { /* <, <= :the loop variable is increasing */ temp_reg = expand_binop (loop_var_mode, sub_optab, comparison_value, initial_value, NULL_RTX, 0, OPTAB_LIB_WIDEN); } else { temp_reg = expand_binop (loop_var_mode, sub_optab, initial_value, comparison_value, NULL_RTX, 0, OPTAB_LIB_WIDEN); } if (increment_value_abs - 1 + add_iteration != 0) temp_reg = expand_binop (loop_var_mode, add_optab, temp_reg, GEN_INT (increment_value_abs - 1 + add_iteration), NULL_RTX, 0, OPTAB_LIB_WIDEN); if (increment_value_abs != 1) { /* ??? This will generate an expensive divide instruction for most targets. The original authors apparently expected this to be a shift, since they test for power-of-2 divisors above, but just naively generating a divide instruction will not give a shift. It happens to work for the PowerPC target because the rs6000.md file has a divide pattern that emits shifts. It will probably not work for any other target. */ iterations_num_reg = expand_binop (loop_var_mode, sdiv_optab, temp_reg, GEN_INT (increment_value_abs), NULL_RTX, 0, OPTAB_LIB_WIDEN); } else iterations_num_reg = temp_reg; } sequence = gen_sequence (); end_sequence (); emit_insn_before (sequence, loop_start); instrument_loop_bct (loop_start, loop_end, iterations_num_reg); } return; #endif /* Complex case */ } /* Instrument loop by inserting a bct in it as follows: 1. A new counter register is created. 2. In the head of the loop the new variable is initialized to the value passed in the loop_num_iterations parameter. 3. At the end of the loop, comparison of the register with 0 is generated. The created comparison follows the pattern defined for the decrement_and_branch_on_count insn, so this insn will be generated. 4. The branch on the old variable are deleted. The compare must remain because it might be used elsewhere. If the loop-variable or condition register are used elsewhere, they will be eliminated by flow. */ static void instrument_loop_bct (loop_start, loop_end, loop_num_iterations) rtx loop_start, loop_end; rtx loop_num_iterations; { rtx counter_reg; rtx start_label; rtx sequence; if (HAVE_decrement_and_branch_on_count) { if (loop_dump_stream) { fputs ("instrument_bct: Inserting BCT (", loop_dump_stream); if (GET_CODE (loop_num_iterations) == CONST_INT) fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (loop_num_iterations)); else fputs ("runtime", loop_dump_stream); fputs (" iterations)", loop_dump_stream); } /* Discard original jump to continue loop. Original compare result may still be live, so it cannot be discarded explicitly. */ delete_insn (PREV_INSN (loop_end)); /* Insert the label which will delimit the start of the loop. */ start_label = gen_label_rtx (); emit_label_after (start_label, loop_start); /* Insert initialization of the count register into the loop header. */ start_sequence (); counter_reg = gen_reg_rtx (word_mode); emit_insn (gen_move_insn (counter_reg, loop_num_iterations)); sequence = gen_sequence (); end_sequence (); emit_insn_before (sequence, loop_start); /* Insert new comparison on the count register instead of the old one, generating the needed BCT pattern (that will be later recognized by assembly generation phase). */ emit_jump_insn_before (gen_decrement_and_branch_on_count (counter_reg, start_label), loop_end); JUMP_LABEL (prev_nonnote_insn (loop_end)) = start_label; LABEL_NUSES (start_label)++; } } #endif /* HAVE_decrement_and_branch_on_count */ /* Scan the function and determine whether it has indirect (computed) jumps. This is taken mostly from flow.c; similar code exists elsewhere in the compiler. It may be useful to put this into rtlanal.c. */ static int indirect_jump_in_function_p (start) rtx start; { rtx insn; for (insn = start; insn; insn = NEXT_INSN (insn)) if (computed_jump_p (insn)) return 1; return 0; } /* Add MEM to the LOOP_MEMS array, if appropriate. See the documentation for LOOP_MEMS for the definition of `appropriate'. This function is called from prescan_loop via for_each_rtx. */ static int insert_loop_mem (mem, data) rtx *mem; void *data ATTRIBUTE_UNUSED; { int i; rtx m = *mem; if (m == NULL_RTX) return 0; switch (GET_CODE (m)) { case MEM: break; case CONST_DOUBLE: /* We're not interested in the MEM associated with a CONST_DOUBLE, so there's no need to traverse into this. */ return -1; default: /* This is not a MEM. */ return 0; } /* See if we've already seen this MEM. */ for (i = 0; i < loop_mems_idx; ++i) if (rtx_equal_p (m, loop_mems[i].mem)) { if (GET_MODE (m) != GET_MODE (loop_mems[i].mem)) /* The modes of the two memory accesses are different. If this happens, something tricky is going on, and we just don't optimize accesses to this MEM. */ loop_mems[i].optimize = 0; return 0; } /* Resize the array, if necessary. */ if (loop_mems_idx == loop_mems_allocated) { if (loop_mems_allocated != 0) loop_mems_allocated *= 2; else loop_mems_allocated = 32; loop_mems = (loop_mem_info*) xrealloc (loop_mems, loop_mems_allocated * sizeof (loop_mem_info)); } /* Actually insert the MEM. */ loop_mems[loop_mems_idx].mem = m; /* We can't hoist this MEM out of the loop if it's a BLKmode MEM because we can't put it in a register. We still store it in the table, though, so that if we see the same address later, but in a non-BLK mode, we'll not think we can optimize it at that point. */ loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode); loop_mems[loop_mems_idx].reg = NULL_RTX; ++loop_mems_idx; return 0; } /* Like load_mems, but also ensures that SET_IN_LOOP, MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct values after load_mems. */ static void load_mems_and_recount_loop_regs_set (scan_start, end, loop_top, start, insn_count) rtx scan_start; rtx end; rtx loop_top; rtx start; int *insn_count; { int nregs = max_reg_num (); load_mems (scan_start, end, loop_top, start); /* Recalculate set_in_loop and friends since load_mems may have created new registers. */ if (max_reg_num () > nregs) { int i; int old_nregs; old_nregs = nregs; nregs = max_reg_num (); if ((unsigned) nregs > set_in_loop->num_elements) { /* Grow all the arrays. */ VARRAY_GROW (set_in_loop, nregs); VARRAY_GROW (n_times_set, nregs); VARRAY_GROW (may_not_optimize, nregs); VARRAY_GROW (reg_single_usage, nregs); } /* Clear the arrays */ bzero ((char *) &set_in_loop->data, nregs * sizeof (int)); bzero ((char *) &may_not_optimize->data, nregs * sizeof (char)); bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx)); count_loop_regs_set (loop_top ? loop_top : start, end, may_not_optimize, reg_single_usage, insn_count, nregs); for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) { VARRAY_CHAR (may_not_optimize, i) = 1; VARRAY_INT (set_in_loop, i) = 1; } #ifdef AVOID_CCMODE_COPIES /* Don't try to move insns which set CC registers if we should not create CCmode register copies. */ for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--) if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC) VARRAY_CHAR (may_not_optimize, i) = 1; #endif /* Set n_times_set for the new registers. */ bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs), (char *) (&n_times_set->data.i[0] + old_nregs), (nregs - old_nregs) * sizeof (int)); } } /* Move MEMs into registers for the duration of the loop. SCAN_START is the first instruction in the loop (as it is executed). The other parameters are as for next_insn_in_loop. */ static void load_mems (scan_start, end, loop_top, start) rtx scan_start; rtx end; rtx loop_top; rtx start; { int maybe_never = 0; int i; rtx p; rtx label = NULL_RTX; rtx end_label; if (loop_mems_idx > 0) { /* Nonzero if the next instruction may never be executed. */ int next_maybe_never = 0; /* Check to see if it's possible that some instructions in the loop are never executed. */ for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top); p != NULL_RTX && !maybe_never; p = next_insn_in_loop (p, scan_start, end, loop_top)) { if (GET_CODE (p) == CODE_LABEL) maybe_never = 1; else if (GET_CODE (p) == JUMP_INSN /* If we enter the loop in the middle, and scan around to the beginning, don't set maybe_never for that. This must be an unconditional jump, otherwise the code at the top of the loop might never be executed. Unconditional jumps are followed a by barrier then loop end. */ && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top && NEXT_INSN (NEXT_INSN (p)) == end && simplejump_p (p))) { if (!condjump_p (p)) /* Something complicated. */ maybe_never = 1; else /* If there are any more instructions in the loop, they might not be reached. */ next_maybe_never = 1; } else if (next_maybe_never) maybe_never = 1; } /* Actually move the MEMs. */ for (i = 0; i < loop_mems_idx; ++i) { int written = 0; rtx reg; rtx mem = loop_mems[i].mem; rtx mem_list_entry; if (MEM_VOLATILE_P (mem) || invariant_p (XEXP (mem, 0)) != 1) /* There's no telling whether or not MEM is modified. */ loop_mems[i].optimize = 0; /* Go through the MEMs written to in the loop to see if this one is aliased by one of them. */ mem_list_entry = loop_store_mems; while (mem_list_entry) { if (rtx_equal_p (mem, XEXP (mem_list_entry, 0))) written = 1; else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode, mem, rtx_varies_p)) { /* MEM is indeed aliased by this store. */ loop_mems[i].optimize = 0; break; } mem_list_entry = XEXP (mem_list_entry, 1); } if (flag_float_store && written && GET_MODE_CLASS (GET_MODE (mem)) == MODE_FLOAT) loop_mems[i].optimize = 0; /* If this MEM is written to, we must be sure that there are no reads from another MEM that aliases this one. */ if (loop_mems[i].optimize && written) { int j; for (j = 0; j < loop_mems_idx; ++j) { if (j == i) continue; else if (true_dependence (mem, VOIDmode, loop_mems[j].mem, rtx_varies_p)) { /* It's not safe to hoist loop_mems[i] out of the loop because writes to it might not be seen by reads from loop_mems[j]. */ loop_mems[i].optimize = 0; break; } } } if (maybe_never && may_trap_p (mem)) /* We can't access the MEM outside the loop; it might cause a trap that wouldn't have happened otherwise. */ loop_mems[i].optimize = 0; if (!loop_mems[i].optimize) /* We thought we were going to lift this MEM out of the loop, but later discovered that we could not. */ continue; /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in order to keep scan_loop from moving stores to this MEM out of the loop just because this REG is neither a user-variable nor used in the loop test. */ reg = gen_reg_rtx (GET_MODE (mem)); REG_USERVAR_P (reg) = 1; loop_mems[i].reg = reg; /* Now, replace all references to the MEM with the corresponding pesudos. */ for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top); p != NULL_RTX; p = next_insn_in_loop (p, scan_start, end, loop_top)) { rtx_and_int ri; ri.r = p; ri.i = i; for_each_rtx (&p, replace_loop_mem, &ri); } if (!apply_change_group ()) /* We couldn't replace all occurrences of the MEM. */ loop_mems[i].optimize = 0; else { rtx set; /* Load the memory immediately before START, which is the NOTE_LOOP_BEG. */ set = gen_move_insn (reg, mem); emit_insn_before (set, start); if (written) { if (label == NULL_RTX) { /* We must compute the former right-after-the-end label before we insert the new one. */ end_label = next_label (end); label = gen_label_rtx (); emit_label_after (label, end); } /* Store the memory immediately after END, which is the NOTE_LOOP_END. */ set = gen_move_insn (copy_rtx (mem), reg); emit_insn_after (set, label); } if (loop_dump_stream) { fprintf (loop_dump_stream, "Hoisted regno %d %s from ", REGNO (reg), (written ? "r/w" : "r/o")); print_rtl (loop_dump_stream, mem); fputc ('\n', loop_dump_stream); } } } } if (label != NULL_RTX) { /* Now, we need to replace all references to the previous exit label with the new one. */ rtx_pair rr; rr.r1 = end_label; rr.r2 = label; for (p = start; p != end; p = NEXT_INSN (p)) { for_each_rtx (&p, replace_label, &rr); /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL field. This is not handled by for_each_rtx because it doesn't handle unprinted ('0') fields. We need to update JUMP_LABEL because the immediately following unroll pass will use it. replace_label would not work anyways, because that only handles LABEL_REFs. */ if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label) JUMP_LABEL (p) = label; } } } /* Replace MEM with its associated pseudo register. This function is called from load_mems via for_each_rtx. DATA is actually an rtx_and_int * describing the instruction currently being scanned and the MEM we are currently replacing. */ static int replace_loop_mem (mem, data) rtx *mem; void *data; { rtx_and_int *ri; rtx insn; int i; rtx m = *mem; if (m == NULL_RTX) return 0; switch (GET_CODE (m)) { case MEM: break; case CONST_DOUBLE: /* We're not interested in the MEM associated with a CONST_DOUBLE, so there's no need to traverse into one. */ return -1; default: /* This is not a MEM. */ return 0; } ri = (rtx_and_int*) data; i = ri->i; if (!rtx_equal_p (loop_mems[i].mem, m)) /* This is not the MEM we are currently replacing. */ return 0; insn = ri->r; /* Actually replace the MEM. */ validate_change (insn, mem, loop_mems[i].reg, 1); return 0; } /* Replace occurrences of the old exit label for the loop with the new one. DATA is an rtx_pair containing the old and new labels, respectively. */ static int replace_label (x, data) rtx *x; void *data; { rtx l = *x; rtx old_label = ((rtx_pair*) data)->r1; rtx new_label = ((rtx_pair*) data)->r2; if (l == NULL_RTX) return 0; if (GET_CODE (l) != LABEL_REF) return 0; if (XEXP (l, 0) != old_label) return 0; XEXP (l, 0) = new_label; ++LABEL_NUSES (new_label); --LABEL_NUSES (old_label); return 0; }