/* Try to unroll loops, and split induction variables. Copyright (C) 1992, 1993 Free Software Foundation, Inc. Contributed by James E. Wilson, Cygnus Support/UC Berkeley. 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, 675 Mass Ave, Cambridge, MA 02139, USA. */ /* Try to unroll a loop, and split induction variables. Loops for which the number of iterations can be calculated exactly are handled specially. If the number of iterations times the insn_count is less than MAX_UNROLLED_INSNS, then the loop is unrolled completely. Otherwise, we try to unroll the loop a number of times modulo the number of iterations, so that only one exit test will be needed. It is unrolled a number of times approximately equal to MAX_UNROLLED_INSNS divided by the insn count. Otherwise, if the number of iterations can be calculated exactly at run time, and the loop is always entered at the top, then we try to precondition the loop. That is, at run time, calculate how many times the loop will execute, and then execute the loop body a few times so that the remaining iterations will be some multiple of 4 (or 2 if the loop is large). Then fall through to a loop unrolled 4 (or 2) times, with only one exit test needed at the end of the loop. Otherwise, if the number of iterations can not be calculated exactly, not even at run time, then we still unroll the loop a number of times approximately equal to MAX_UNROLLED_INSNS divided by the insn count, but there must be an exit test after each copy of the loop body. For each induction variable, which is dead outside the loop (replaceable) or for which we can easily calculate the final value, if we can easily calculate its value at each place where it is set as a function of the current loop unroll count and the variable's value at loop entry, then the induction variable is split into `N' different variables, one for each copy of the loop body. One variable is live across the backward branch, and the others are all calculated as a function of this variable. This helps eliminate data dependencies, and leads to further opportunities for cse. */ /* Possible improvements follow: */ /* ??? Add an extra pass somewhere to determine whether unrolling will give any benefit. E.g. after generating all unrolled insns, compute the cost of all insns and compare against cost of insns in rolled loop. - On traditional architectures, unrolling a non-constant bound loop is a win if there is a giv whose only use is in memory addresses, the memory addresses can be split, and hence giv increments can be eliminated. - It is also a win if the loop is executed many times, and preconditioning can be performed for the loop. Add code to check for these and similar cases. */ /* ??? Improve control of which loops get unrolled. Could use profiling info to only unroll the most commonly executed loops. Perhaps have a user specifyable option to control the amount of code expansion, or the percent of loops to consider for unrolling. Etc. */ /* ??? Look at the register copies inside the loop to see if they form a simple permutation. If so, iterate the permutation until it gets back to the start state. This is how many times we should unroll the loop, for best results, because then all register copies can be eliminated. For example, the lisp nreverse function should be unrolled 3 times while (this) { next = this->cdr; this->cdr = prev; prev = this; this = next; } ??? The number of times to unroll the loop may also be based on data references in the loop. For example, if we have a loop that references x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */ /* ??? Add some simple linear equation solving capability so that we can determine the number of loop iterations for more complex loops. For example, consider this loop from gdb #define SWAP_TARGET_AND_HOST(buffer,len) { char tmp; char *p = (char *) buffer; char *q = ((char *) buffer) + len - 1; int iterations = (len + 1) >> 1; int i; for (p; p < q; p++, q--;) { tmp = *q; *q = *p; *p = tmp; } } Note that: start value = p = &buffer + current_iteration end value = q = &buffer + len - 1 - current_iteration Given the loop exit test of "p < q", then there must be "q - p" iterations, set equal to zero and solve for number of iterations: q - p = len - 1 - 2*current_iteration = 0 current_iteration = (len - 1) / 2 Hence, there are (len - 1) / 2 (rounded up to the nearest integer) iterations of this loop. */ /* ??? Currently, no labels are marked as loop invariant when doing loop unrolling. This is because an insn inside the loop, that loads the address of a label inside the loop into a register, could be moved outside the loop by the invariant code motion pass if labels were invariant. If the loop is subsequently unrolled, the code will be wrong because each unrolled body of the loop will use the same address, whereas each actually needs a different address. A case where this happens is when a loop containing a switch statement is unrolled. It would be better to let labels be considered invariant. When we unroll loops here, check to see if any insns using a label local to the loop were moved before the loop. If so, then correct the problem, by moving the insn back into the loop, or perhaps replicate the insn before the loop, one copy for each time the loop is unrolled. */ /* The prime factors looked for when trying to unroll a loop by some number which is modulo the total number of iterations. Just checking for these 4 prime factors will find at least one factor for 75% of all numbers theoretically. Practically speaking, this will succeed almost all of the time since loops are generally a multiple of 2 and/or 5. */ #define NUM_FACTORS 4 struct _factor { int factor, count; } factors[NUM_FACTORS] = { {2, 0}, {3, 0}, {5, 0}, {7, 0}}; /* Describes the different types of loop unrolling performed. */ enum unroll_types { UNROLL_COMPLETELY, UNROLL_MODULO, UNROLL_NAIVE }; #include "config.h" #include "rtl.h" #include "insn-config.h" #include "integrate.h" #include "regs.h" #include "flags.h" #include "expr.h" #include #include "loop.h" /* This controls which loops are unrolled, and by how much we unroll them. */ #ifndef MAX_UNROLLED_INSNS #define MAX_UNROLLED_INSNS 100 #endif /* Indexed by register number, if non-zero, then it contains a pointer to a struct induction for a DEST_REG giv which has been combined with one of more address givs. This is needed because whenever such a DEST_REG giv is modified, we must modify the value of all split address givs that were combined with this DEST_REG giv. */ static struct induction **addr_combined_regs; /* Indexed by register number, if this is a splittable induction variable, then this will hold the current value of the register, which depends on the iteration number. */ static rtx *splittable_regs; /* Indexed by register number, if this is a splittable induction variable, then this will hold the number of instructions in the loop that modify the induction variable. Used to ensure that only the last insn modifying a split iv will update the original iv of the dest. */ static int *splittable_regs_updates; /* Values describing the current loop's iteration variable. These are set up by loop_iterations, and used by precondition_loop_p. */ static rtx loop_iteration_var; static rtx loop_initial_value; static rtx loop_increment; static rtx loop_final_value; /* Forward declarations. */ static void init_reg_map (); static int precondition_loop_p (); static void copy_loop_body (); static void iteration_info (); static rtx approx_final_value (); static int find_splittable_regs (); static int find_splittable_givs (); static rtx fold_rtx_mult_add (); /* Try to unroll one loop and split induction variables in the loop. The loop is described by the arguments LOOP_END, INSN_COUNT, and LOOP_START. END_INSERT_BEFORE indicates where insns should be added which need to be executed when the loop falls through. STRENGTH_REDUCTION_P indicates whether information generated in the strength reduction pass is available. This function is intended to be called from within `strength_reduce' in loop.c. */ void unroll_loop (loop_end, insn_count, loop_start, end_insert_before, strength_reduce_p) rtx loop_end; int insn_count; rtx loop_start; rtx end_insert_before; int strength_reduce_p; { int i, j, temp; int unroll_number = 1; rtx copy_start, copy_end; rtx insn, copy, sequence, pattern, tem; int max_labelno, max_insnno; rtx insert_before; struct inline_remap *map; char *local_label; int maxregnum; int new_maxregnum; rtx exit_label = 0; rtx start_label; struct iv_class *bl; struct induction *v; int splitting_not_safe = 0; enum unroll_types unroll_type; int loop_preconditioned = 0; rtx safety_label; /* This points to the last real insn in the loop, which should be either a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional jumps). */ rtx last_loop_insn; /* Don't bother unrolling huge loops. Since the minimum factor is two, loops greater than one half of MAX_UNROLLED_INSNS will never be unrolled. */ if (insn_count > MAX_UNROLLED_INSNS / 2) { if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling failure: Loop too big.\n"); return; } /* When emitting debugger info, we can't unroll loops with unequal numbers of block_beg and block_end notes, because that would unbalance the block structure of the function. This can happen as a result of the "if (foo) bar; else break;" optimization in jump.c. */ if (write_symbols != NO_DEBUG) { int block_begins = 0; int block_ends = 0; for (insn = loop_start; insn != loop_end; insn = NEXT_INSN (insn)) { if (GET_CODE (insn) == NOTE) { if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_BEG) block_begins++; else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_END) block_ends++; } } if (block_begins != block_ends) { if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling failure: Unbalanced block notes.\n"); return; } } /* Determine type of unroll to perform. Depends on the number of iterations and the size of the loop. */ /* If there is no strength reduce info, then set loop_n_iterations to zero. This can happen if strength_reduce can't find any bivs in the loop. A value of zero indicates that the number of iterations could not be calculated. */ if (! strength_reduce_p) loop_n_iterations = 0; if (loop_dump_stream && loop_n_iterations > 0) fprintf (loop_dump_stream, "Loop unrolling: %d iterations.\n", loop_n_iterations); /* Find and save a pointer to the last nonnote insn in the loop. */ last_loop_insn = prev_nonnote_insn (loop_end); /* Calculate how many times to unroll the loop. Indicate whether or not the loop is being completely unrolled. */ if (loop_n_iterations == 1) { /* If number of iterations is exactly 1, then eliminate the compare and branch at the end of the loop since they will never be taken. Then return, since no other action is needed here. */ /* If the last instruction is not a BARRIER or a JUMP_INSN, then don't do anything. */ if (GET_CODE (last_loop_insn) == BARRIER) { /* Delete the jump insn. This will delete the barrier also. */ delete_insn (PREV_INSN (last_loop_insn)); } else if (GET_CODE (last_loop_insn) == JUMP_INSN) { #ifdef HAVE_cc0 /* The immediately preceding insn is a compare which must be deleted. */ delete_insn (last_loop_insn); delete_insn (PREV_INSN (last_loop_insn)); #else /* The immediately preceding insn may not be the compare, so don't delete it. */ delete_insn (last_loop_insn); #endif } return; } else if (loop_n_iterations > 0 && loop_n_iterations * insn_count < MAX_UNROLLED_INSNS) { unroll_number = loop_n_iterations; unroll_type = UNROLL_COMPLETELY; } else if (loop_n_iterations > 0) { /* Try to factor the number of iterations. Don't bother with the general case, only using 2, 3, 5, and 7 will get 75% of all numbers theoretically, and almost all in practice. */ for (i = 0; i < NUM_FACTORS; i++) factors[i].count = 0; temp = loop_n_iterations; for (i = NUM_FACTORS - 1; i >= 0; i--) while (temp % factors[i].factor == 0) { factors[i].count++; temp = temp / factors[i].factor; } /* Start with the larger factors first so that we generally get lots of unrolling. */ unroll_number = 1; temp = insn_count; for (i = 3; i >= 0; i--) while (factors[i].count--) { if (temp * factors[i].factor < MAX_UNROLLED_INSNS) { unroll_number *= factors[i].factor; temp *= factors[i].factor; } else break; } /* If we couldn't find any factors, then unroll as in the normal case. */ if (unroll_number == 1) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: No factors found.\n"); } else unroll_type = UNROLL_MODULO; } /* Default case, calculate number of times to unroll loop based on its size. */ if (unroll_number == 1) { if (8 * insn_count < MAX_UNROLLED_INSNS) unroll_number = 8; else if (4 * insn_count < MAX_UNROLLED_INSNS) unroll_number = 4; else unroll_number = 2; unroll_type = UNROLL_NAIVE; } /* Now we know how many times to unroll the loop. */ if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling loop %d times.\n", unroll_number); if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO) { /* Loops of these types should never start with a jump down to the exit condition test. For now, check for this case just to be sure. UNROLL_NAIVE loops can be of this form, this case is handled below. */ insn = loop_start; while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN) insn = NEXT_INSN (insn); if (GET_CODE (insn) == JUMP_INSN) abort (); } if (unroll_type == UNROLL_COMPLETELY) { /* Completely unrolling the loop: Delete the compare and branch at the end (the last two instructions). This delete must done at the very end of loop unrolling, to avoid problems with calls to back_branch_in_range_p, which is called by find_splittable_regs. All increments of splittable bivs/givs are changed to load constant instructions. */ copy_start = loop_start; /* Set insert_before to the instruction immediately after the JUMP_INSN (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of the loop will be correctly handled by copy_loop_body. */ insert_before = NEXT_INSN (last_loop_insn); /* Set copy_end to the insn before the jump at the end of the loop. */ if (GET_CODE (last_loop_insn) == BARRIER) copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); else if (GET_CODE (last_loop_insn) == JUMP_INSN) { #ifdef HAVE_cc0 /* The instruction immediately before the JUMP_INSN is a compare instruction which we do not want to copy. */ copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); #else /* The instruction immediately before the JUMP_INSN may not be the compare, so we must copy it. */ copy_end = PREV_INSN (last_loop_insn); #endif } else { /* We currently can't unroll a loop if it doesn't end with a JUMP_INSN. There would need to be a mechanism that recognizes this case, and then inserts a jump after each loop body, which jumps to after the last loop body. */ if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling failure: loop does not end with a JUMP_INSN.\n"); return; } } else if (unroll_type == UNROLL_MODULO) { /* Partially unrolling the loop: The compare and branch at the end (the last two instructions) must remain. Don't copy the compare and branch instructions at the end of the loop. Insert the unrolled code immediately before the compare/branch at the end so that the code will fall through to them as before. */ copy_start = loop_start; /* Set insert_before to the jump insn at the end of the loop. Set copy_end to before the jump insn at the end of the loop. */ if (GET_CODE (last_loop_insn) == BARRIER) { insert_before = PREV_INSN (last_loop_insn); copy_end = PREV_INSN (insert_before); } else if (GET_CODE (last_loop_insn) == JUMP_INSN) { #ifdef HAVE_cc0 /* The instruction immediately before the JUMP_INSN is a compare instruction which we do not want to copy or delete. */ insert_before = PREV_INSN (last_loop_insn); copy_end = PREV_INSN (insert_before); #else /* The instruction immediately before the JUMP_INSN may not be the compare, so we must copy it. */ insert_before = last_loop_insn; copy_end = PREV_INSN (last_loop_insn); #endif } else { /* We currently can't unroll a loop if it doesn't end with a JUMP_INSN. There would need to be a mechanism that recognizes this case, and then inserts a jump after each loop body, which jumps to after the last loop body. */ if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling failure: loop does not end with a JUMP_INSN.\n"); return; } } else { /* Normal case: Must copy the compare and branch instructions at the end of the loop. */ if (GET_CODE (last_loop_insn) == BARRIER) { /* Loop ends with an unconditional jump and a barrier. Handle this like above, don't copy jump and barrier. This is not strictly necessary, but doing so prevents generating unconditional jumps to an immediately following label. This will be corrected below if the target of this jump is not the start_label. */ insert_before = PREV_INSN (last_loop_insn); copy_end = PREV_INSN (insert_before); } else if (GET_CODE (last_loop_insn) == JUMP_INSN) { /* Set insert_before to immediately after the JUMP_INSN, so that NOTEs at the end of the loop will be correctly handled by copy_loop_body. */ insert_before = NEXT_INSN (last_loop_insn); copy_end = last_loop_insn; } else { /* We currently can't unroll a loop if it doesn't end with a JUMP_INSN. There would need to be a mechanism that recognizes this case, and then inserts a jump after each loop body, which jumps to after the last loop body. */ if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling failure: loop does not end with a JUMP_INSN.\n"); return; } /* If copying exit test branches because they can not be eliminated, then must convert the fall through case of the branch to a jump past the end of the loop. Create a label to emit after the loop and save it for later use. Do not use the label after the loop, if any, since it might be used by insns outside the loop, or there might be insns added before it later by final_[bg]iv_value which must be after the real exit label. */ exit_label = gen_label_rtx (); insn = loop_start; while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN) insn = NEXT_INSN (insn); if (GET_CODE (insn) == JUMP_INSN) { /* The loop starts with a jump down to the exit condition test. Start copying the loop after the barrier following this jump insn. */ copy_start = NEXT_INSN (insn); /* Splitting induction variables doesn't work when the loop is entered via a jump to the bottom, because then we end up doing a comparison against a new register for a split variable, but we did not execute the set insn for the new register because it was skipped over. */ splitting_not_safe = 1; if (loop_dump_stream) fprintf (loop_dump_stream, "Splitting not safe, because loop not entered at top.\n"); } else copy_start = loop_start; } /* This should always be the first label in the loop. */ start_label = NEXT_INSN (copy_start); /* There may be a line number note and/or a loop continue note here. */ while (GET_CODE (start_label) == NOTE) start_label = NEXT_INSN (start_label); if (GET_CODE (start_label) != CODE_LABEL) { /* This can happen as a result of jump threading. If the first insns in the loop test the same condition as the loop's backward jump, or the opposite condition, then the backward jump will be modified to point to elsewhere, and the loop's start label is deleted. This case currently can not be handled by the loop unrolling code. */ if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling failure: unknown insns between BEG note and loop label.\n"); return; } if (LABEL_NAME (start_label)) { /* The jump optimization pass must have combined the original start label with a named label for a goto. We can't unroll this case because jumps which go to the named label must be handled differently than jumps to the loop start, and it is impossible to differentiate them in this case. */ if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling failure: loop start label is gone\n"); return; } if (unroll_type == UNROLL_NAIVE && GET_CODE (last_loop_insn) == BARRIER && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn))) { /* In this case, we must copy the jump and barrier, because they will not be converted to jumps to an immediately following label. */ insert_before = NEXT_INSN (last_loop_insn); copy_end = last_loop_insn; } /* Allocate a translation table for the labels and insn numbers. They will be filled in as we copy the insns in the loop. */ max_labelno = max_label_num (); max_insnno = get_max_uid (); map = (struct inline_remap *) alloca (sizeof (struct inline_remap)); map->integrating = 0; /* Allocate the label map. */ if (max_labelno > 0) { map->label_map = (rtx *) alloca (max_labelno * sizeof (rtx)); local_label = (char *) alloca (max_labelno); bzero (local_label, max_labelno); } else map->label_map = 0; /* Search the loop and mark all local labels, i.e. the ones which have to be distinct labels when copied. For all labels which might be non-local, set their label_map entries to point to themselves. If they happen to be local their label_map entries will be overwritten before the loop body is copied. The label_map entries for local labels will be set to a different value each time the loop body is copied. */ for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn)) { if (GET_CODE (insn) == CODE_LABEL) local_label[CODE_LABEL_NUMBER (insn)] = 1; else if (GET_CODE (insn) == JUMP_INSN) { if (JUMP_LABEL (insn)) map->label_map[CODE_LABEL_NUMBER (JUMP_LABEL (insn))] = JUMP_LABEL (insn); else if (GET_CODE (PATTERN (insn)) == ADDR_VEC || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) { rtx pat = PATTERN (insn); int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC; int len = XVECLEN (pat, diff_vec_p); rtx label; for (i = 0; i < len; i++) { label = XEXP (XVECEXP (pat, diff_vec_p, i), 0); map->label_map[CODE_LABEL_NUMBER (label)] = label; } } } } /* Allocate space for the insn map. */ map->insn_map = (rtx *) alloca (max_insnno * sizeof (rtx)); /* Set this to zero, to indicate that we are doing loop unrolling, not function inlining. */ map->inline_target = 0; /* The register and constant maps depend on the number of registers present, so the final maps can't be created until after find_splittable_regs is called. However, they are needed for preconditioning, so we create temporary maps when preconditioning is performed. */ /* The preconditioning code may allocate two new pseudo registers. */ maxregnum = max_reg_num (); /* Allocate and zero out the splittable_regs and addr_combined_regs arrays. These must be zeroed here because they will be used if loop preconditioning is performed, and must be zero for that case. It is safe to do this here, since the extra registers created by the preconditioning code and find_splittable_regs will never be used to access the splittable_regs[] and addr_combined_regs[] arrays. */ splittable_regs = (rtx *) alloca (maxregnum * sizeof (rtx)); bzero (splittable_regs, maxregnum * sizeof (rtx)); splittable_regs_updates = (int *) alloca (maxregnum * sizeof (int)); bzero (splittable_regs_updates, maxregnum * sizeof (int)); addr_combined_regs = (struct induction **) alloca (maxregnum * sizeof (struct induction *)); bzero (addr_combined_regs, maxregnum * sizeof (struct induction *)); /* If this loop requires exit tests when unrolled, check to see if we can precondition the loop so as to make the exit tests unnecessary. Just like variable splitting, this is not safe if the loop is entered via a jump to the bottom. Also, can not do this if no strength reduce info, because precondition_loop_p uses this info. */ /* Must copy the loop body for preconditioning before the following find_splittable_regs call since that will emit insns which need to be after the preconditioned loop copies, but immediately before the unrolled loop copies. */ /* Also, it is not safe to split induction variables for the preconditioned copies of the loop body. If we split induction variables, then the code assumes that each induction variable can be represented as a function of its initial value and the loop iteration number. This is not true in this case, because the last preconditioned copy of the loop body could be any iteration from the first up to the `unroll_number-1'th, depending on the initial value of the iteration variable. Therefore we can not split induction variables here, because we can not calculate their value. Hence, this code must occur before find_splittable_regs is called. */ if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p) { rtx initial_value, final_value, increment; if (precondition_loop_p (&initial_value, &final_value, &increment, loop_start, loop_end)) { register rtx diff, temp; enum machine_mode mode; rtx *labels; int abs_inc, neg_inc; map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx)); map->const_equiv_map = (rtx *) alloca (maxregnum * sizeof (rtx)); map->const_age_map = (unsigned *) alloca (maxregnum * sizeof (unsigned)); map->const_equiv_map_size = maxregnum; global_const_equiv_map = map->const_equiv_map; global_const_equiv_map_size = maxregnum; init_reg_map (map, maxregnum); /* Limit loop unrolling to 4, since this will make 7 copies of the loop body. */ if (unroll_number > 4) unroll_number = 4; /* Save the absolute value of the increment, and also whether or not it is negative. */ neg_inc = 0; abs_inc = INTVAL (increment); if (abs_inc < 0) { abs_inc = - abs_inc; neg_inc = 1; } start_sequence (); /* Decide what mode to do these calculations in. Choose the larger of final_value's mode and initial_value's mode, or a full-word if both are constants. */ mode = GET_MODE (final_value); if (mode == VOIDmode) { mode = GET_MODE (initial_value); if (mode == VOIDmode) mode = word_mode; } else if (mode != GET_MODE (initial_value) && (GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (initial_value)))) mode = GET_MODE (initial_value); /* Calculate the difference between the final and initial values. Final value may be a (plus (reg x) (const_int 1)) rtx. Let the following cse pass simplify this if initial value is a constant. We must copy the final and initial values here to avoid improperly shared rtl. */ diff = expand_binop (mode, sub_optab, copy_rtx (final_value), copy_rtx (initial_value), NULL_RTX, 0, OPTAB_LIB_WIDEN); /* Now calculate (diff % (unroll * abs (increment))) by using an and instruction. */ diff = expand_binop (GET_MODE (diff), and_optab, diff, GEN_INT (unroll_number * abs_inc - 1), NULL_RTX, 0, OPTAB_LIB_WIDEN); /* Now emit a sequence of branches to jump to the proper precond loop entry point. */ labels = (rtx *) alloca (sizeof (rtx) * unroll_number); for (i = 0; i < unroll_number; i++) labels[i] = gen_label_rtx (); /* Assuming the unroll_number is 4, and the increment is 2, then for a negative increment: for a positive increment: diff = 0,1 precond 0 diff = 0,7 precond 0 diff = 2,3 precond 3 diff = 1,2 precond 1 diff = 4,5 precond 2 diff = 3,4 precond 2 diff = 6,7 precond 1 diff = 5,6 precond 3 */ /* We only need to emit (unroll_number - 1) branches here, the last case just falls through to the following code. */ /* ??? This would give better code if we emitted a tree of branches instead of the current linear list of branches. */ for (i = 0; i < unroll_number - 1; i++) { int cmp_const; /* For negative increments, must invert the constant compared against, except when comparing against zero. */ if (i == 0) cmp_const = 0; else if (neg_inc) cmp_const = unroll_number - i; else cmp_const = i; emit_cmp_insn (diff, GEN_INT (abs_inc * cmp_const), EQ, NULL_RTX, mode, 0, 0); if (i == 0) emit_jump_insn (gen_beq (labels[i])); else if (neg_inc) emit_jump_insn (gen_bge (labels[i])); else emit_jump_insn (gen_ble (labels[i])); JUMP_LABEL (get_last_insn ()) = labels[i]; LABEL_NUSES (labels[i])++; } /* If the increment is greater than one, then we need another branch, to handle other cases equivalent to 0. */ /* ??? This should be merged into the code above somehow to help simplify the code here, and reduce the number of branches emitted. For the negative increment case, the branch here could easily be merged with the `0' case branch above. For the positive increment case, it is not clear how this can be simplified. */ if (abs_inc != 1) { int cmp_const; if (neg_inc) cmp_const = abs_inc - 1; else cmp_const = abs_inc * (unroll_number - 1) + 1; emit_cmp_insn (diff, GEN_INT (cmp_const), EQ, NULL_RTX, mode, 0, 0); if (neg_inc) emit_jump_insn (gen_ble (labels[0])); else emit_jump_insn (gen_bge (labels[0])); JUMP_LABEL (get_last_insn ()) = labels[0]; LABEL_NUSES (labels[0])++; } sequence = gen_sequence (); end_sequence (); emit_insn_before (sequence, loop_start); /* Only the last copy of the loop body here needs the exit test, so set copy_end to exclude the compare/branch here, and then reset it inside the loop when get to the last copy. */ if (GET_CODE (last_loop_insn) == BARRIER) copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); else if (GET_CODE (last_loop_insn) == JUMP_INSN) { #ifdef HAVE_cc0 /* The immediately preceding insn is a compare which we do not want to copy. */ copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); #else /* The immediately preceding insn may not be a compare, so we must copy it. */ copy_end = PREV_INSN (last_loop_insn); #endif } else abort (); for (i = 1; i < unroll_number; i++) { emit_label_after (labels[unroll_number - i], PREV_INSN (loop_start)); bzero (map->insn_map, max_insnno * sizeof (rtx)); bzero (map->const_equiv_map, maxregnum * sizeof (rtx)); bzero (map->const_age_map, maxregnum * sizeof (unsigned)); map->const_age = 0; for (j = 0; j < max_labelno; j++) if (local_label[j]) map->label_map[j] = gen_label_rtx (); /* The last copy needs the compare/branch insns at the end, so reset copy_end here if the loop ends with a conditional branch. */ if (i == unroll_number - 1) { if (GET_CODE (last_loop_insn) == BARRIER) copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); else copy_end = last_loop_insn; } /* None of the copies are the `last_iteration', so just pass zero for that parameter. */ copy_loop_body (copy_start, copy_end, map, exit_label, 0, unroll_type, start_label, loop_end, loop_start, copy_end); } emit_label_after (labels[0], PREV_INSN (loop_start)); if (GET_CODE (last_loop_insn) == BARRIER) { insert_before = PREV_INSN (last_loop_insn); copy_end = PREV_INSN (insert_before); } else { #ifdef HAVE_cc0 /* The immediately preceding insn is a compare which we do not want to copy. */ insert_before = PREV_INSN (last_loop_insn); copy_end = PREV_INSN (insert_before); #else /* The immediately preceding insn may not be a compare, so we must copy it. */ insert_before = last_loop_insn; copy_end = PREV_INSN (last_loop_insn); #endif } /* Set unroll type to MODULO now. */ unroll_type = UNROLL_MODULO; loop_preconditioned = 1; } } /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll the loop unless all loops are being unrolled. */ if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops) { if (loop_dump_stream) fprintf (loop_dump_stream, "Unrolling failure: Naive unrolling not being done.\n"); return; } /* At this point, we are guaranteed to unroll the loop. */ /* For each biv and giv, determine whether it can be safely split into a different variable for each unrolled copy of the loop body. We precalculate and save this info here, since computing it is expensive. Do this before deleting any instructions from the loop, so that back_branch_in_range_p will work correctly. */ if (splitting_not_safe) temp = 0; else temp = find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before, unroll_number); /* find_splittable_regs may have created some new registers, so must reallocate the reg_map with the new larger size, and must realloc the constant maps also. */ maxregnum = max_reg_num (); map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx)); init_reg_map (map, maxregnum); /* Space is needed in some of the map for new registers, so new_maxregnum is an (over)estimate of how many registers will exist at the end. */ new_maxregnum = maxregnum + (temp * unroll_number * 2); /* Must realloc space for the constant maps, because the number of registers may have changed. */ map->const_equiv_map = (rtx *) alloca (new_maxregnum * sizeof (rtx)); map->const_age_map = (unsigned *) alloca (new_maxregnum * sizeof (unsigned)); map->const_equiv_map_size = new_maxregnum; global_const_equiv_map = map->const_equiv_map; global_const_equiv_map_size = new_maxregnum; /* Search the list of bivs and givs to find ones which need to be remapped when split, and set their reg_map entry appropriately. */ for (bl = loop_iv_list; bl; bl = bl->next) { if (REGNO (bl->biv->src_reg) != bl->regno) map->reg_map[bl->regno] = bl->biv->src_reg; #if 0 /* Currently, non-reduced/final-value givs are never split. */ for (v = bl->giv; v; v = v->next_iv) if (REGNO (v->src_reg) != bl->regno) map->reg_map[REGNO (v->dest_reg)] = v->src_reg; #endif } /* If the loop is being partially unrolled, and the iteration variables are being split, and are being renamed for the split, then must fix up the compare instruction at the end of the loop to refer to the new registers. This compare isn't copied, so the registers used in it will never be replaced if it isn't done here. */ if (unroll_type == UNROLL_MODULO) { insn = NEXT_INSN (copy_end); if (GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == SET) { #if 0 /* If non-reduced/final-value givs were split, then this would also have to remap those givs. */ #endif tem = SET_SRC (PATTERN (insn)); /* The set source is a register. */ if (GET_CODE (tem) == REG) { if (REGNO (tem) < max_reg_before_loop && reg_iv_type[REGNO (tem)] == BASIC_INDUCT) SET_SRC (PATTERN (insn)) = reg_biv_class[REGNO (tem)]->biv->src_reg; } else { /* The set source is a compare of some sort. */ tem = XEXP (SET_SRC (PATTERN (insn)), 0); if (GET_CODE (tem) == REG && REGNO (tem) < max_reg_before_loop && reg_iv_type[REGNO (tem)] == BASIC_INDUCT) XEXP (SET_SRC (PATTERN (insn)), 0) = reg_biv_class[REGNO (tem)]->biv->src_reg; tem = XEXP (SET_SRC (PATTERN (insn)), 1); if (GET_CODE (tem) == REG && REGNO (tem) < max_reg_before_loop && reg_iv_type[REGNO (tem)] == BASIC_INDUCT) XEXP (SET_SRC (PATTERN (insn)), 1) = reg_biv_class[REGNO (tem)]->biv->src_reg; } } } /* For unroll_number - 1 times, make a copy of each instruction between copy_start and copy_end, and insert these new instructions before the end of the loop. */ for (i = 0; i < unroll_number; i++) { bzero (map->insn_map, max_insnno * sizeof (rtx)); bzero (map->const_equiv_map, new_maxregnum * sizeof (rtx)); bzero (map->const_age_map, new_maxregnum * sizeof (unsigned)); map->const_age = 0; for (j = 0; j < max_labelno; j++) if (local_label[j]) map->label_map[j] = gen_label_rtx (); /* If loop starts with a branch to the test, then fix it so that it points to the test of the first unrolled copy of the loop. */ if (i == 0 && loop_start != copy_start) { insn = PREV_INSN (copy_start); pattern = PATTERN (insn); tem = map->label_map[CODE_LABEL_NUMBER (XEXP (SET_SRC (pattern), 0))]; SET_SRC (pattern) = gen_rtx (LABEL_REF, VOIDmode, tem); /* Set the jump label so that it can be used by later loop unrolling passes. */ JUMP_LABEL (insn) = tem; LABEL_NUSES (tem)++; } copy_loop_body (copy_start, copy_end, map, exit_label, i == unroll_number - 1, unroll_type, start_label, loop_end, insert_before, insert_before); } /* Before deleting any insns, emit a CODE_LABEL immediately after the last insn to be deleted. This prevents any runaway delete_insn call from more insns that it should, as it always stops at a CODE_LABEL. */ /* Delete the compare and branch at the end of the loop if completely unrolling the loop. Deleting the backward branch at the end also deletes the code label at the start of the loop. This is done at the very end to avoid problems with back_branch_in_range_p. */ if (unroll_type == UNROLL_COMPLETELY) safety_label = emit_label_after (gen_label_rtx (), last_loop_insn); else safety_label = emit_label_after (gen_label_rtx (), copy_end); /* Delete all of the original loop instructions. Don't delete the LOOP_BEG note, or the first code label in the loop. */ insn = NEXT_INSN (copy_start); while (insn != safety_label) { if (insn != start_label) insn = delete_insn (insn); else insn = NEXT_INSN (insn); } /* Can now delete the 'safety' label emitted to protect us from runaway delete_insn calls. */ if (INSN_DELETED_P (safety_label)) abort (); delete_insn (safety_label); /* If exit_label exists, emit it after the loop. Doing the emit here forces it to have a higher INSN_UID than any insn in the unrolled loop. This is needed so that mostly_true_jump in reorg.c will treat jumps to this loop end label correctly, i.e. predict that they are usually not taken. */ if (exit_label) emit_label_after (exit_label, loop_end); } /* Return true if the loop can be safely, and profitably, preconditioned so that the unrolled copies of the loop body don't need exit tests. This only works if final_value, initial_value and increment can be determined, and if increment is a constant power of 2. If increment is not a power of 2, then the preconditioning modulo operation would require a real modulo instead of a boolean AND, and this is not considered `profitable'. */ /* ??? If the loop is known to be executed very many times, or the machine has a very cheap divide instruction, then preconditioning is a win even when the increment is not a power of 2. Use RTX_COST to compute whether divide is cheap. */ static int precondition_loop_p (initial_value, final_value, increment, loop_start, loop_end) rtx *initial_value, *final_value, *increment; rtx loop_start, loop_end; { int unsigned_compare, compare_dir; if (loop_n_iterations > 0) { *initial_value = const0_rtx; *increment = const1_rtx; *final_value = GEN_INT (loop_n_iterations); if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Success, number of iterations known, %d.\n", loop_n_iterations); return 1; } if (loop_initial_value == 0) { if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Could not find initial value.\n"); return 0; } else if (loop_increment == 0) { if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Could not find increment value.\n"); return 0; } else if (GET_CODE (loop_increment) != CONST_INT) { if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Increment not a constant.\n"); return 0; } else if ((exact_log2 (INTVAL (loop_increment)) < 0) && (exact_log2 (- INTVAL (loop_increment)) < 0)) { if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Increment not a constant power of 2.\n"); return 0; } /* Unsigned_compare and compare_dir can be ignored here, since they do not matter for preconditioning. */ if (loop_final_value == 0) { if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: EQ comparison loop.\n"); return 0; } /* Must ensure that final_value is invariant, so call invariant_p to check. Before doing so, must check regno against max_reg_before_loop to make sure that the register is in the range covered by invariant_p. If it isn't, then it is most likely a biv/giv which by definition are not invariant. */ if ((GET_CODE (loop_final_value) == REG && REGNO (loop_final_value) >= max_reg_before_loop) || (GET_CODE (loop_final_value) == PLUS && REGNO (XEXP (loop_final_value, 0)) >= max_reg_before_loop) || ! invariant_p (loop_final_value)) { if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Final value not invariant.\n"); return 0; } /* Fail for floating point values, since the caller of this function does not have code to deal with them. */ if (GET_MODE_CLASS (GET_MODE (loop_final_value)) == MODE_FLOAT || GET_MODE_CLASS (GET_MODE (loop_initial_value)) == MODE_FLOAT) { if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Floating point final or initial value.\n"); return 0; } /* Now set initial_value to be the iteration_var, since that may be a simpler expression, and is guaranteed to be correct if all of the above tests succeed. We can not use the initial_value as calculated, because it will be one too small for loops of the form "while (i-- > 0)". We can not emit code before the loop_skip_over insns to fix this problem as this will then give a number one too large for loops of the form "while (--i > 0)". Note that all loops that reach here are entered at the top, because this function is not called if the loop starts with a jump. */ /* Fail if loop_iteration_var is not live before loop_start, since we need to test its value in the preconditioning code. */ if (uid_luid[regno_first_uid[REGNO (loop_iteration_var)]] > INSN_LUID (loop_start)) { if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Iteration var not live before loop start.\n"); return 0; } *initial_value = loop_iteration_var; *increment = loop_increment; *final_value = loop_final_value; /* Success! */ if (loop_dump_stream) fprintf (loop_dump_stream, "Preconditioning: Successful.\n"); return 1; } /* All pseudo-registers must be mapped to themselves. Two hard registers must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_ REGNUM, to avoid function-inlining specific conversions of these registers. All other hard regs can not be mapped because they may be used with different modes. */ static void init_reg_map (map, maxregnum) struct inline_remap *map; int maxregnum; { int i; for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--) map->reg_map[i] = regno_reg_rtx[i]; /* Just clear the rest of the entries. */ for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--) map->reg_map[i] = 0; map->reg_map[VIRTUAL_STACK_VARS_REGNUM] = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM]; map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM] = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM]; } /* Strength-reduction will often emit code for optimized biv/givs which calculates their value in a temporary register, and then copies the result to the iv. This procedure reconstructs the pattern computing the iv; verifying that all operands are of the proper form. The return value is the amount that the giv is incremented by. */ static rtx calculate_giv_inc (pattern, src_insn, regno) rtx pattern, src_insn; int regno; { rtx increment; rtx increment_total = 0; int tries = 0; retry: /* Verify that we have an increment insn here. First check for a plus as the set source. */ if (GET_CODE (SET_SRC (pattern)) != PLUS) { /* SR sometimes computes the new giv value in a temp, then copies it to the new_reg. */ src_insn = PREV_INSN (src_insn); pattern = PATTERN (src_insn); if (GET_CODE (SET_SRC (pattern)) != PLUS) abort (); /* The last insn emitted is not needed, so delete it to avoid confusing the second cse pass. This insn sets the giv unnecessarily. */ delete_insn (get_last_insn ()); } /* Verify that we have a constant as the second operand of the plus. */ increment = XEXP (SET_SRC (pattern), 1); if (GET_CODE (increment) != CONST_INT) { /* SR sometimes puts the constant in a register, especially if it is too big to be an add immed operand. */ src_insn = PREV_INSN (src_insn); increment = SET_SRC (PATTERN (src_insn)); /* SR may have used LO_SUM to compute the constant if it is too large for a load immed operand. In this case, the constant is in operand one of the LO_SUM rtx. */ if (GET_CODE (increment) == LO_SUM) increment = XEXP (increment, 1); if (GET_CODE (increment) != CONST_INT) abort (); /* The insn loading the constant into a register is not longer needed, so delete it. */ delete_insn (get_last_insn ()); } if (increment_total) increment_total = GEN_INT (INTVAL (increment_total) + INTVAL (increment)); else increment_total = increment; /* Check that the source register is the same as the register we expected to see as the source. If not, something is seriously wrong. */ if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG || REGNO (XEXP (SET_SRC (pattern), 0)) != regno) { /* Some machines (e.g. the romp), may emit two add instructions for certain constants, so lets try looking for another add immediately before this one if we have only seen one add insn so far. */ if (tries == 0) { tries++; src_insn = PREV_INSN (src_insn); pattern = PATTERN (src_insn); delete_insn (get_last_insn ()); goto retry; } abort (); } return increment_total; } /* Copy REG_NOTES, except for insn references, because not all insn_map entries are valid yet. We do need to copy registers now though, because the reg_map entries can change during copying. */ static rtx initial_reg_note_copy (notes, map) rtx notes; struct inline_remap *map; { rtx copy; if (notes == 0) return 0; copy = rtx_alloc (GET_CODE (notes)); PUT_MODE (copy, GET_MODE (notes)); if (GET_CODE (notes) == EXPR_LIST) XEXP (copy, 0) = copy_rtx_and_substitute (XEXP (notes, 0), map); else if (GET_CODE (notes) == INSN_LIST) /* Don't substitute for these yet. */ XEXP (copy, 0) = XEXP (notes, 0); else abort (); XEXP (copy, 1) = initial_reg_note_copy (XEXP (notes, 1), map); return copy; } /* Fixup insn references in copied REG_NOTES. */ static void final_reg_note_copy (notes, map) rtx notes; struct inline_remap *map; { rtx note; for (note = notes; note; note = XEXP (note, 1)) if (GET_CODE (note) == INSN_LIST) XEXP (note, 0) = map->insn_map[INSN_UID (XEXP (note, 0))]; } /* Copy each instruction in the loop, substituting from map as appropriate. This is very similar to a loop in expand_inline_function. */ static void copy_loop_body (copy_start, copy_end, map, exit_label, last_iteration, unroll_type, start_label, loop_end, insert_before, copy_notes_from) rtx copy_start, copy_end; struct inline_remap *map; rtx exit_label; int last_iteration; enum unroll_types unroll_type; rtx start_label, loop_end, insert_before, copy_notes_from; { rtx insn, pattern; rtx tem, copy; int dest_reg_was_split, i; rtx cc0_insn = 0; rtx final_label = 0; rtx giv_inc, giv_dest_reg, giv_src_reg; /* If this isn't the last iteration, then map any references to the start_label to final_label. Final label will then be emitted immediately after the end of this loop body if it was ever used. If this is the last iteration, then map references to the start_label to itself. */ if (! last_iteration) { final_label = gen_label_rtx (); map->label_map[CODE_LABEL_NUMBER (start_label)] = final_label; } else map->label_map[CODE_LABEL_NUMBER (start_label)] = start_label; start_sequence (); insn = copy_start; do { insn = NEXT_INSN (insn); map->orig_asm_operands_vector = 0; switch (GET_CODE (insn)) { case INSN: pattern = PATTERN (insn); copy = 0; giv_inc = 0; /* Check to see if this is a giv that has been combined with some split address givs. (Combined in the sense that `combine_givs' in loop.c has put two givs in the same register.) In this case, we must search all givs based on the same biv to find the address givs. Then split the address givs. Do this before splitting the giv, since that may map the SET_DEST to a new register. */ if (GET_CODE (pattern) == SET && GET_CODE (SET_DEST (pattern)) == REG && addr_combined_regs[REGNO (SET_DEST (pattern))]) { struct iv_class *bl; struct induction *v, *tv; int regno = REGNO (SET_DEST (pattern)); v = addr_combined_regs[REGNO (SET_DEST (pattern))]; bl = reg_biv_class[REGNO (v->src_reg)]; /* Although the giv_inc amount is not needed here, we must call calculate_giv_inc here since it might try to delete the last insn emitted. If we wait until later to call it, we might accidentally delete insns generated immediately below by emit_unrolled_add. */ giv_inc = calculate_giv_inc (pattern, insn, regno); /* Now find all address giv's that were combined with this giv 'v'. */ for (tv = bl->giv; tv; tv = tv->next_iv) if (tv->giv_type == DEST_ADDR && tv->same == v) { int this_giv_inc = INTVAL (giv_inc); /* Scale this_giv_inc if the multiplicative factors of the two givs are different. */ if (tv->mult_val != v->mult_val) this_giv_inc = (this_giv_inc / INTVAL (v->mult_val) * INTVAL (tv->mult_val)); tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc); *tv->location = tv->dest_reg; if (last_iteration && unroll_type != UNROLL_COMPLETELY) { /* Must emit an insn to increment the split address giv. Add in the const_adjust field in case there was a constant eliminated from the address. */ rtx value, dest_reg; /* tv->dest_reg will be either a bare register, or else a register plus a constant. */ if (GET_CODE (tv->dest_reg) == REG) dest_reg = tv->dest_reg; else dest_reg = XEXP (tv->dest_reg, 0); /* tv->dest_reg may actually be a (PLUS (REG) (CONST)) here, so we must call plus_constant to add the const_adjust amount before calling emit_unrolled_add below. */ value = plus_constant (tv->dest_reg, tv->const_adjust); /* The constant could be too large for an add immediate, so can't directly emit an insn here. */ emit_unrolled_add (dest_reg, XEXP (value, 0), XEXP (value, 1)); /* Reset the giv to be just the register again, in case it is used after the set we have just emitted. We must subtract the const_adjust factor added in above. */ tv->dest_reg = plus_constant (dest_reg, - tv->const_adjust); *tv->location = tv->dest_reg; } } } /* If this is a setting of a splittable variable, then determine how to split the variable, create a new set based on this split, and set up the reg_map so that later uses of the variable will use the new split variable. */ dest_reg_was_split = 0; if (GET_CODE (pattern) == SET && GET_CODE (SET_DEST (pattern)) == REG && splittable_regs[REGNO (SET_DEST (pattern))]) { int regno = REGNO (SET_DEST (pattern)); dest_reg_was_split = 1; /* Compute the increment value for the giv, if it wasn't already computed above. */ if (giv_inc == 0) giv_inc = calculate_giv_inc (pattern, insn, regno); giv_dest_reg = SET_DEST (pattern); giv_src_reg = SET_DEST (pattern); if (unroll_type == UNROLL_COMPLETELY) { /* Completely unrolling the loop. Set the induction variable to a known constant value. */ /* The value in splittable_regs may be an invariant value, so we must use plus_constant here. */ splittable_regs[regno] = plus_constant (splittable_regs[regno], INTVAL (giv_inc)); if (GET_CODE (splittable_regs[regno]) == PLUS) { giv_src_reg = XEXP (splittable_regs[regno], 0); giv_inc = XEXP (splittable_regs[regno], 1); } else { /* The splittable_regs value must be a REG or a CONST_INT, so put the entire value in the giv_src_reg variable. */ giv_src_reg = splittable_regs[regno]; giv_inc = const0_rtx; } } else { /* Partially unrolling loop. Create a new pseudo register for the iteration variable, and set it to be a constant plus the original register. Except on the last iteration, when the result has to go back into the original iteration var register. */ /* Handle bivs which must be mapped to a new register when split. This happens for bivs which need their final value set before loop entry. The new register for the biv was stored in the biv's first struct induction entry by find_splittable_regs. */ if (regno < max_reg_before_loop && reg_iv_type[regno] == BASIC_INDUCT) { giv_src_reg = reg_biv_class[regno]->biv->src_reg; giv_dest_reg = giv_src_reg; } #if 0 /* If non-reduced/final-value givs were split, then this would have to remap those givs also. See find_splittable_regs. */ #endif splittable_regs[regno] = GEN_INT (INTVAL (giv_inc) + INTVAL (splittable_regs[regno])); giv_inc = splittable_regs[regno]; /* Now split the induction variable by changing the dest of this insn to a new register, and setting its reg_map entry to point to this new register. If this is the last iteration, and this is the last insn that will update the iv, then reuse the original dest, to ensure that the iv will have the proper value when the loop exits or repeats. Using splittable_regs_updates here like this is safe, because it can only be greater than one if all instructions modifying the iv are always executed in order. */ if (! last_iteration || (splittable_regs_updates[regno]-- != 1)) { tem = gen_reg_rtx (GET_MODE (giv_src_reg)); giv_dest_reg = tem; map->reg_map[regno] = tem; } else map->reg_map[regno] = giv_src_reg; } /* The constant being added could be too large for an add immediate, so can't directly emit an insn here. */ emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc); copy = get_last_insn (); pattern = PATTERN (copy); } else { pattern = copy_rtx_and_substitute (pattern, map); copy = emit_insn (pattern); } REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); #ifdef HAVE_cc0 /* If this insn is setting CC0, it may need to look at the insn that uses CC0 to see what type of insn it is. In that case, the call to recog via validate_change will fail. So don't substitute constants here. Instead, do it when we emit the following insn. For example, see the pyr.md file. That machine has signed and unsigned compares. The compare patterns must check the following branch insn to see which what kind of compare to emit. If the previous insn set CC0, substitute constants on it as well. */ if (sets_cc0_p (copy) != 0) cc0_insn = copy; else { if (cc0_insn) try_constants (cc0_insn, map); cc0_insn = 0; try_constants (copy, map); } #else try_constants (copy, map); #endif /* Make split induction variable constants `permanent' since we know there are no backward branches across iteration variable settings which would invalidate this. */ if (dest_reg_was_split) { int regno = REGNO (SET_DEST (pattern)); if (regno < map->const_equiv_map_size && map->const_age_map[regno] == map->const_age) map->const_age_map[regno] = -1; } break; case JUMP_INSN: pattern = copy_rtx_and_substitute (PATTERN (insn), map); copy = emit_jump_insn (pattern); REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); if (JUMP_LABEL (insn) == start_label && insn == copy_end && ! last_iteration) { /* This is a branch to the beginning of the loop; this is the last insn being copied; and this is not the last iteration. In this case, we want to change the original fall through case to be a branch past the end of the loop, and the original jump label case to fall_through. */ if (! invert_exp (pattern, copy) || ! redirect_exp (&pattern, map->label_map[CODE_LABEL_NUMBER (JUMP_LABEL (insn))], exit_label, copy)) abort (); } #ifdef HAVE_cc0 if (cc0_insn) try_constants (cc0_insn, map); cc0_insn = 0; #endif try_constants (copy, map); /* Set the jump label of COPY correctly to avoid problems with later passes of unroll_loop, if INSN had jump label set. */ if (JUMP_LABEL (insn)) { rtx label = 0; /* Can't use the label_map for every insn, since this may be the backward branch, and hence the label was not mapped. */ if (GET_CODE (pattern) == SET) { tem = SET_SRC (pattern); if (GET_CODE (tem) == LABEL_REF) label = XEXP (tem, 0); else if (GET_CODE (tem) == IF_THEN_ELSE) { if (XEXP (tem, 1) != pc_rtx) label = XEXP (XEXP (tem, 1), 0); else label = XEXP (XEXP (tem, 2), 0); } } if (label && GET_CODE (label) == CODE_LABEL) JUMP_LABEL (copy) = label; else { /* An unrecognizable jump insn, probably the entry jump for a switch statement. This label must have been mapped, so just use the label_map to get the new jump label. */ JUMP_LABEL (copy) = map->label_map[CODE_LABEL_NUMBER (JUMP_LABEL (insn))]; } /* If this is a non-local jump, then must increase the label use count so that the label will not be deleted when the original jump is deleted. */ LABEL_NUSES (JUMP_LABEL (copy))++; } else if (GET_CODE (PATTERN (copy)) == ADDR_VEC || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC) { rtx pat = PATTERN (copy); int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC; int len = XVECLEN (pat, diff_vec_p); int i; for (i = 0; i < len; i++) LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++; } /* If this used to be a conditional jump insn but whose branch direction is now known, we must do something special. */ if (condjump_p (insn) && !simplejump_p (insn) && map->last_pc_value) { #ifdef HAVE_cc0 /* The previous insn set cc0 for us. So delete it. */ delete_insn (PREV_INSN (copy)); #endif /* If this is now a no-op, delete it. */ if (map->last_pc_value == pc_rtx) { /* Don't let delete_insn delete the label referenced here, because we might possibly need it later for some other instruction in the loop. */ if (JUMP_LABEL (copy)) LABEL_NUSES (JUMP_LABEL (copy))++; delete_insn (copy); if (JUMP_LABEL (copy)) LABEL_NUSES (JUMP_LABEL (copy))--; copy = 0; } else /* Otherwise, this is unconditional jump so we must put a BARRIER after it. We could do some dead code elimination here, but jump.c will do it just as well. */ emit_barrier (); } break; case CALL_INSN: pattern = copy_rtx_and_substitute (PATTERN (insn), map); copy = emit_call_insn (pattern); REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); #ifdef HAVE_cc0 if (cc0_insn) try_constants (cc0_insn, map); cc0_insn = 0; #endif try_constants (copy, map); /* Be lazy and assume CALL_INSNs clobber all hard registers. */ for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) map->const_equiv_map[i] = 0; break; case CODE_LABEL: /* If this is the loop start label, then we don't need to emit a copy of this label since no one will use it. */ if (insn != start_label) { copy = emit_label (map->label_map[CODE_LABEL_NUMBER (insn)]); map->const_age++; } break; case BARRIER: copy = emit_barrier (); break; case NOTE: /* VTOP notes are valid only before the loop exit test. If placed anywhere else, loop may generate bad code. */ if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED && (NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP || (last_iteration && unroll_type != UNROLL_COMPLETELY))) copy = emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn)); else copy = 0; break; default: abort (); break; } map->insn_map[INSN_UID (insn)] = copy; } while (insn != copy_end); /* Now finish coping the REG_NOTES. */ insn = copy_start; do { insn = NEXT_INSN (insn); if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN || GET_CODE (insn) == CALL_INSN) && map->insn_map[INSN_UID (insn)]) final_reg_note_copy (REG_NOTES (map->insn_map[INSN_UID (insn)]), map); } while (insn != copy_end); /* There may be notes between copy_notes_from and loop_end. Emit a copy of each of these notes here, since there may be some important ones, such as NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last iteration, because the original notes won't be deleted. We can't use insert_before here, because when from preconditioning, insert_before points before the loop. We can't use copy_end, because there may be insns already inserted after it (which we don't want to copy) when not from preconditioning code. */ if (! last_iteration) { for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn)) { if (GET_CODE (insn) == NOTE && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED) emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn)); } } if (final_label && LABEL_NUSES (final_label) > 0) emit_label (final_label); tem = gen_sequence (); end_sequence (); emit_insn_before (tem, insert_before); } /* Emit an insn, using the expand_binop to ensure that a valid insn is emitted. This will correctly handle the case where the increment value won't fit in the immediate field of a PLUS insns. */ void emit_unrolled_add (dest_reg, src_reg, increment) rtx dest_reg, src_reg, increment; { rtx result; result = expand_binop (GET_MODE (dest_reg), add_optab, src_reg, increment, dest_reg, 0, OPTAB_LIB_WIDEN); if (dest_reg != result) emit_move_insn (dest_reg, result); } /* Searches the insns between INSN and LOOP_END. Returns 1 if there is a backward branch in that range that branches to somewhere between LOOP_START and INSN. Returns 0 otherwise. */ /* ??? This is quadratic algorithm. Could be rewritten to be linear. In practice, this is not a problem, because this function is seldom called, and uses a negligible amount of CPU time on average. */ static int back_branch_in_range_p (insn, loop_start, loop_end) rtx insn; rtx loop_start, loop_end; { rtx p, q, target_insn; /* Stop before we get to the backward branch at the end of the loop. */ loop_end = prev_nonnote_insn (loop_end); if (GET_CODE (loop_end) == BARRIER) loop_end = PREV_INSN (loop_end); /* Check in case insn has been deleted, search forward for first non deleted insn following it. */ while (INSN_DELETED_P (insn)) insn = NEXT_INSN (insn); /* Check for the case where insn is the last insn in the loop. */ if (insn == loop_end) return 0; for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p)) { if (GET_CODE (p) == JUMP_INSN) { target_insn = JUMP_LABEL (p); /* Search from loop_start to insn, to see if one of them is the target_insn. We can't use INSN_LUID comparisons here, since insn may not have an LUID entry. */ for (q = loop_start; q != insn; q = NEXT_INSN (q)) if (q == target_insn) return 1; } } return 0; } /* Try to generate the simplest rtx for the expression (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial value of giv's. */ static rtx fold_rtx_mult_add (mult1, mult2, add1, mode) rtx mult1, mult2, add1; enum machine_mode mode; { rtx temp, mult_res; rtx result; /* The modes must all be the same. This should always be true. For now, check to make sure. */ if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode) || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode) || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode)) abort (); /* Ensure that if at least one of mult1/mult2 are constant, then mult2 will be a constant. */ if (GET_CODE (mult1) == CONST_INT) { temp = mult2; mult2 = mult1; mult1 = temp; } mult_res = simplify_binary_operation (MULT, mode, mult1, mult2); if (! mult_res) mult_res = gen_rtx (MULT, mode, mult1, mult2); /* Again, put the constant second. */ if (GET_CODE (add1) == CONST_INT) { temp = add1; add1 = mult_res; mult_res = temp; } result = simplify_binary_operation (PLUS, mode, add1, mult_res); if (! result) result = gen_rtx (PLUS, mode, add1, mult_res); return result; } /* Searches the list of induction struct's for the biv BL, to try to calculate the total increment value for one iteration of the loop as a constant. Returns the increment value as an rtx, simplified as much as possible, if it can be calculated. Otherwise, returns 0. */ rtx biv_total_increment (bl, loop_start, loop_end) struct iv_class *bl; rtx loop_start, loop_end; { struct induction *v; rtx result; /* For increment, must check every instruction that sets it. Each instruction must be executed only once each time through the loop. To verify this, we check that the the insn is always executed, and that there are no backward branches after the insn that branch to before it. Also, the insn must have a mult_val of one (to make sure it really is an increment). */ result = const0_rtx; for (v = bl->biv; v; v = v->next_iv) { if (v->always_computable && v->mult_val == const1_rtx && ! back_branch_in_range_p (v->insn, loop_start, loop_end)) result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode); else return 0; } return result; } /* Determine the initial value of the iteration variable, and the amount that it is incremented each loop. Use the tables constructed by the strength reduction pass to calculate these values. Initial_value and/or increment are set to zero if their values could not be calculated. */ static void iteration_info (iteration_var, initial_value, increment, loop_start, loop_end) rtx iteration_var, *initial_value, *increment; rtx loop_start, loop_end; { struct iv_class *bl; struct induction *v, *b; /* Clear the result values, in case no answer can be found. */ *initial_value = 0; *increment = 0; /* The iteration variable can be either a giv or a biv. Check to see which it is, and compute the variable's initial value, and increment value if possible. */ /* If this is a new register, can't handle it since we don't have any reg_iv_type entry for it. */ if (REGNO (iteration_var) >= max_reg_before_loop) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: No reg_iv_type entry for iteration var.\n"); return; } /* Reject iteration variables larger than the host long size, since they could result in a number of iterations greater than the range of our `unsigned long' variable loop_n_iterations. */ else if (GET_MODE_BITSIZE (GET_MODE (iteration_var)) > HOST_BITS_PER_LONG) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Iteration var rejected because mode larger than host long.\n"); return; } else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Iteration var not an integer.\n"); return; } else if (reg_iv_type[REGNO (iteration_var)] == BASIC_INDUCT) { /* Grab initial value, only useful if it is a constant. */ bl = reg_biv_class[REGNO (iteration_var)]; *initial_value = bl->initial_value; *increment = biv_total_increment (bl, loop_start, loop_end); } else if (reg_iv_type[REGNO (iteration_var)] == GENERAL_INDUCT) { #if 1 /* ??? The code below does not work because the incorrect number of iterations is calculated when the biv is incremented after the giv is set (which is the usual case). This can probably be accounted for by biasing the initial_value by subtracting the amount of the increment that occurs between the giv set and the giv test. However, a giv as an iterator is very rare, so it does not seem worthwhile to handle this. */ /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */ if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Giv iterators are not handled.\n"); return; #else /* Initial value is mult_val times the biv's initial value plus add_val. Only useful if it is a constant. */ v = reg_iv_info[REGNO (iteration_var)]; bl = reg_biv_class[REGNO (v->src_reg)]; *initial_value = fold_rtx_mult_add (v->mult_val, bl->initial_value, v->add_val, v->mode); /* Increment value is mult_val times the increment value of the biv. */ *increment = biv_total_increment (bl, loop_start, loop_end); if (*increment) *increment = fold_rtx_mult_add (v->mult_val, *increment, const0_rtx, v->mode); #endif } else { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Not basic or general induction var.\n"); return; } } /* Calculate the approximate final value of the iteration variable which has an loop exit test with code COMPARISON_CODE and comparison value of COMPARISON_VALUE. Also returns an indication of whether the comparison was signed or unsigned, and the direction of the comparison. This info is needed to calculate the number of loop iterations. */ static rtx approx_final_value (comparison_code, comparison_value, unsigned_p, compare_dir) enum rtx_code comparison_code; rtx comparison_value; int *unsigned_p; int *compare_dir; { /* Calculate the final value of the induction variable. The exact final value depends on the branch operator, and increment sign. This is only an approximate value. It will be wrong if the iteration variable is not incremented by one each time through the loop, and approx final value - start value % increment != 0. */ *unsigned_p = 0; switch (comparison_code) { case LEU: *unsigned_p = 1; case LE: *compare_dir = 1; return plus_constant (comparison_value, 1); case GEU: *unsigned_p = 1; case GE: *compare_dir = -1; return plus_constant (comparison_value, -1); case EQ: /* Can not calculate a final value for this case. */ *compare_dir = 0; return 0; case LTU: *unsigned_p = 1; case LT: *compare_dir = 1; return comparison_value; break; case GTU: *unsigned_p = 1; case GT: *compare_dir = -1; return comparison_value; case NE: *compare_dir = 0; return comparison_value; default: abort (); } } /* For each biv and giv, determine whether it can be safely split into a different variable for each unrolled copy of the loop body. If it is safe to split, then indicate that by saving some useful info in the splittable_regs array. If the loop is being completely unrolled, then splittable_regs will hold the current value of the induction variable while the loop is unrolled. It must be set to the initial value of the induction variable here. Otherwise, splittable_regs will hold the difference between the current value of the induction variable and the value the induction variable had at the top of the loop. It must be set to the value 0 here. Returns the total number of instructions that set registers that are splittable. */ /* ?? If the loop is only unrolled twice, then most of the restrictions to constant values are unnecessary, since we can easily calculate increment values in this case even if nothing is constant. The increment value should not involve a multiply however. */ /* ?? Even if the biv/giv increment values aren't constant, it may still be beneficial to split the variable if the loop is only unrolled a few times, since multiplies by small integers (1,2,3,4) are very cheap. */ static int find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before, unroll_number) enum unroll_types unroll_type; rtx loop_start, loop_end; rtx end_insert_before; int unroll_number; { struct iv_class *bl; struct induction *v; rtx increment, tem; rtx biv_final_value; int biv_splittable; int result = 0; for (bl = loop_iv_list; bl; bl = bl->next) { /* Biv_total_increment must return a constant value, otherwise we can not calculate the split values. */ increment = biv_total_increment (bl, loop_start, loop_end); if (! increment || GET_CODE (increment) != CONST_INT) continue; /* The loop must be unrolled completely, or else have a known number of iterations and only one exit, or else the biv must be dead outside the loop, or else the final value must be known. Otherwise, it is unsafe to split the biv since it may not have the proper value on loop exit. */ /* loop_number_exit_labels is non-zero if the loop has an exit other than a fall through at the end. */ biv_splittable = 1; biv_final_value = 0; if (unroll_type != UNROLL_COMPLETELY && (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]] || unroll_type == UNROLL_NAIVE) && (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)) || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set))) && ! (biv_final_value = final_biv_value (bl, loop_start, loop_end))) biv_splittable = 0; /* If any of the insns setting the BIV don't do so with a simple PLUS, we don't know how to split it. */ for (v = bl->biv; biv_splittable && v; v = v->next_iv) if ((tem = single_set (v->insn)) == 0 || GET_CODE (SET_DEST (tem)) != REG || REGNO (SET_DEST (tem)) != bl->regno || GET_CODE (SET_SRC (tem)) != PLUS) biv_splittable = 0; /* If final value is non-zero, then must emit an instruction which sets the value of the biv to the proper value. This is done after handling all of the givs, since some of them may need to use the biv's value in their initialization code. */ /* This biv is splittable. If completely unrolling the loop, save the biv's initial value. Otherwise, save the constant zero. */ if (biv_splittable == 1) { if (unroll_type == UNROLL_COMPLETELY) { /* If the initial value of the biv is itself (i.e. it is too complicated for strength_reduce to compute), or is a hard register, then we must create a new pseudo reg to hold the initial value of the biv. */ if (GET_CODE (bl->initial_value) == REG && (REGNO (bl->initial_value) == bl->regno || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER)) { rtx tem = gen_reg_rtx (bl->biv->mode); emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), loop_start); if (loop_dump_stream) fprintf (loop_dump_stream, "Biv %d initial value remapped to %d.\n", bl->regno, REGNO (tem)); splittable_regs[bl->regno] = tem; } else splittable_regs[bl->regno] = bl->initial_value; } else splittable_regs[bl->regno] = const0_rtx; /* Save the number of instructions that modify the biv, so that we can treat the last one specially. */ splittable_regs_updates[bl->regno] = bl->biv_count; result += bl->biv_count; if (loop_dump_stream) fprintf (loop_dump_stream, "Biv %d safe to split.\n", bl->regno); } /* Check every giv that depends on this biv to see whether it is splittable also. Even if the biv isn't splittable, givs which depend on it may be splittable if the biv is live outside the loop, and the givs aren't. */ result += find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment, unroll_number); /* If final value is non-zero, then must emit an instruction which sets the value of the biv to the proper value. This is done after handling all of the givs, since some of them may need to use the biv's value in their initialization code. */ if (biv_final_value) { /* If the loop has multiple exits, emit the insns 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_labels[uid_loop_num[INSN_UID (loop_start)]]) emit_insn_before (gen_move_insn (bl->biv->src_reg, biv_final_value), end_insert_before); else { /* Create a new register to hold the value of the biv, and then set the biv to its final value before the loop start. The biv is set to its final value before loop start to ensure that this insn will always be executed, no matter how the loop exits. */ rtx tem = gen_reg_rtx (bl->biv->mode); emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), loop_start); emit_insn_before (gen_move_insn (bl->biv->src_reg, biv_final_value), loop_start); if (loop_dump_stream) fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n", REGNO (bl->biv->src_reg), REGNO (tem)); /* Set up the mapping from the original biv register to the new register. */ bl->biv->src_reg = tem; } } } return result; } /* For every giv based on the biv BL, check to determine whether it is splittable. This is a subroutine to find_splittable_regs (). Return the number of instructions that set splittable registers. */ static int find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment, unroll_number) struct iv_class *bl; enum unroll_types unroll_type; rtx loop_start, loop_end; rtx increment; int unroll_number; { struct induction *v; rtx final_value; rtx tem; int result = 0; for (v = bl->giv; v; v = v->next_iv) { rtx giv_inc, value; /* Only split the giv if it has already been reduced, or if the loop is being completely unrolled. */ if (unroll_type != UNROLL_COMPLETELY && v->ignore) continue; /* The giv can be split if the insn that sets the giv is executed once and only once on every iteration of the loop. */ /* An address giv can always be split. v->insn is just a use not a set, and hence it does not matter whether it is always executed. All that matters is that all the biv increments are always executed, and we won't reach here if they aren't. */ if (v->giv_type != DEST_ADDR && (! v->always_computable || back_branch_in_range_p (v->insn, loop_start, loop_end))) continue; /* The giv increment value must be a constant. */ giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx, v->mode); if (! giv_inc || GET_CODE (giv_inc) != CONST_INT) continue; /* The loop must be unrolled completely, or else have a known number of iterations and only one exit, or else the giv must be dead outside the loop, or else the final value of the giv must be known. Otherwise, it is not safe to split the giv since it may not have the proper value on loop exit. */ /* The used outside loop test will fail for DEST_ADDR givs. They are never used outside the loop anyways, so it is always safe to split a DEST_ADDR giv. */ final_value = 0; if (unroll_type != UNROLL_COMPLETELY && (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]] || unroll_type == UNROLL_NAIVE) && v->giv_type != DEST_ADDR && ((regno_first_uid[REGNO (v->dest_reg)] != INSN_UID (v->insn) /* Check for the case where the pseudo is set by a shift/add sequence, in which case the first insn setting the pseudo is the first insn of the shift/add sequence. */ && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX)) || (regno_first_uid[REGNO (v->dest_reg)] != INSN_UID (XEXP (tem, 0))))) /* Line above always fails if INSN was moved by loop opt. */ || (uid_luid[regno_last_uid[REGNO (v->dest_reg)]] >= INSN_LUID (loop_end))) && ! (final_value = v->final_value)) continue; #if 0 /* Currently, non-reduced/final-value givs are never split. */ /* Should emit insns after the loop if possible, as the biv final value code below does. */ /* If the final value is non-zero, and the giv has not been reduced, then must emit an instruction to set the final value. */ if (final_value && !v->new_reg) { /* Create a new register to hold the value of the giv, and then set the giv to its final value before the loop start. The giv is set to its final value before loop start to ensure that this insn will always be executed, no matter how we exit. */ tem = gen_reg_rtx (v->mode); emit_insn_before (gen_move_insn (tem, v->dest_reg), loop_start); emit_insn_before (gen_move_insn (v->dest_reg, final_value), loop_start); if (loop_dump_stream) fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n", REGNO (v->dest_reg), REGNO (tem)); v->src_reg = tem; } #endif /* This giv is splittable. If completely unrolling the loop, save the giv's initial value. Otherwise, save the constant zero for it. */ if (unroll_type == UNROLL_COMPLETELY) { /* It is not safe to use bl->initial_value here, because it may not be invariant. It is safe to use the initial value stored in the splittable_regs array if it is set. In rare cases, it won't be set, so then we do exactly the same thing as find_splittable_regs does to get a safe value. */ rtx biv_initial_value; if (splittable_regs[bl->regno]) biv_initial_value = splittable_regs[bl->regno]; else if (GET_CODE (bl->initial_value) != REG || (REGNO (bl->initial_value) != bl->regno && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER)) biv_initial_value = bl->initial_value; else { rtx tem = gen_reg_rtx (bl->biv->mode); emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), loop_start); biv_initial_value = tem; } value = fold_rtx_mult_add (v->mult_val, biv_initial_value, v->add_val, v->mode); } else value = const0_rtx; if (v->new_reg) { /* If a giv was combined with another giv, then we can only split this giv if the giv it was combined with was reduced. This is because the value of v->new_reg is meaningless in this case. */ if (v->same && ! v->same->new_reg) { if (loop_dump_stream) fprintf (loop_dump_stream, "giv combined with unreduced giv not split.\n"); continue; } /* If the giv is an address destination, it could be something other than a simple register, these have to be treated differently. */ else if (v->giv_type == DEST_REG) { /* If value is not a constant, register, or register plus constant, then compute its value into a register before loop start. This prevents illegal rtx sharing, and should generate better code. We can use bl->initial_value here instead of splittable_regs[bl->regno] because this code is going before the loop start. */ if (unroll_type == UNROLL_COMPLETELY && GET_CODE (value) != CONST_INT && GET_CODE (value) != REG && (GET_CODE (value) != PLUS || GET_CODE (XEXP (value, 0)) != REG || GET_CODE (XEXP (value, 1)) != CONST_INT)) { rtx tem = gen_reg_rtx (v->mode); emit_iv_add_mult (bl->initial_value, v->mult_val, v->add_val, tem, loop_start); value = tem; } splittable_regs[REGNO (v->new_reg)] = value; } else { /* Splitting address givs is useful since it will often allow us to eliminate some increment insns for the base giv as unnecessary. */ /* If the addr giv is combined with a dest_reg giv, then all references to that dest reg will be remapped, which is NOT what we want for split addr regs. We always create a new register for the split addr giv, just to be safe. */ /* ??? If there are multiple address givs which have been combined with the same dest_reg giv, then we may only need one new register for them. Pulling out constants below will catch some of the common cases of this. Currently, I leave the work of simplifying multiple address givs to the following cse pass. */ v->const_adjust = 0; if (unroll_type != UNROLL_COMPLETELY) { /* If not completely unrolling the loop, then create a new register to hold the split value of the DEST_ADDR giv. Emit insn to initialize its value before loop start. */ tem = gen_reg_rtx (v->mode); /* If the address giv has a constant in its new_reg value, then this constant can be pulled out and put in value, instead of being part of the initialization code. */ if (GET_CODE (v->new_reg) == PLUS && GET_CODE (XEXP (v->new_reg, 1)) == CONST_INT) { v->dest_reg = plus_constant (tem, INTVAL (XEXP (v->new_reg,1))); /* Only succeed if this will give valid addresses. Try to validate both the first and the last address resulting from loop unrolling, if one fails, then can't do const elim here. */ if (memory_address_p (v->mem_mode, v->dest_reg) && memory_address_p (v->mem_mode, plus_constant (v->dest_reg, INTVAL (giv_inc) * (unroll_number - 1)))) { /* Save the negative of the eliminated const, so that we can calculate the dest_reg's increment value later. */ v->const_adjust = - INTVAL (XEXP (v->new_reg, 1)); v->new_reg = XEXP (v->new_reg, 0); if (loop_dump_stream) fprintf (loop_dump_stream, "Eliminating constant from giv %d\n", REGNO (tem)); } else v->dest_reg = tem; } else v->dest_reg = tem; /* If the address hasn't been checked for validity yet, do so now, and fail completely if either the first or the last unrolled copy of the address is not a valid address. */ if (v->dest_reg == tem && (! memory_address_p (v->mem_mode, v->dest_reg) || ! memory_address_p (v->mem_mode, plus_constant (v->dest_reg, INTVAL (giv_inc) * (unroll_number -1))))) { if (loop_dump_stream) fprintf (loop_dump_stream, "Illegal address for giv at insn %d\n", INSN_UID (v->insn)); continue; } /* To initialize the new register, just move the value of new_reg into it. This is not guaranteed to give a valid instruction on machines with complex addressing modes. If we can't recognize it, then delete it and emit insns to calculate the value from scratch. */ emit_insn_before (gen_rtx (SET, VOIDmode, tem, copy_rtx (v->new_reg)), loop_start); if (recog_memoized (PREV_INSN (loop_start)) < 0) { rtx sequence, ret; /* We can't use bl->initial_value to compute the initial value, because the loop may have been preconditioned. We must calculate it from NEW_REG. Try using force_operand instead of emit_iv_add_mult. */ delete_insn (PREV_INSN (loop_start)); start_sequence (); ret = force_operand (v->new_reg, tem); if (ret != tem) emit_move_insn (tem, ret); sequence = gen_sequence (); end_sequence (); emit_insn_before (sequence, loop_start); if (loop_dump_stream) fprintf (loop_dump_stream, "Illegal init insn, rewritten.\n"); } } else { v->dest_reg = value; /* Check the resulting address for validity, and fail if the resulting address would be illegal. */ if (! memory_address_p (v->mem_mode, v->dest_reg) || ! memory_address_p (v->mem_mode, plus_constant (v->dest_reg, INTVAL (giv_inc) * (unroll_number -1)))) { if (loop_dump_stream) fprintf (loop_dump_stream, "Illegal address for giv at insn %d\n", INSN_UID (v->insn)); continue; } } /* Store the value of dest_reg into the insn. This sharing will not be a problem as this insn will always be copied later. */ *v->location = v->dest_reg; /* If this address giv is combined with a dest reg giv, then save the base giv's induction pointer so that we will be able to handle this address giv properly. The base giv itself does not have to be splittable. */ if (v->same && v->same->giv_type == DEST_REG) addr_combined_regs[REGNO (v->same->new_reg)] = v->same; if (GET_CODE (v->new_reg) == REG) { /* This giv maybe hasn't been combined with any others. Make sure that it's giv is marked as splittable here. */ splittable_regs[REGNO (v->new_reg)] = value; /* Make it appear to depend upon itself, so that the giv will be properly split in the main loop above. */ if (! v->same) { v->same = v; addr_combined_regs[REGNO (v->new_reg)] = v; } } if (loop_dump_stream) fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n"); } } else { #if 0 /* Currently, unreduced giv's can't be split. This is not too much of a problem since unreduced giv's are not live across loop iterations anyways. When unrolling a loop completely though, it makes sense to reduce&split givs when possible, as this will result in simpler instructions, and will not require that a reg be live across loop iterations. */ splittable_regs[REGNO (v->dest_reg)] = value; fprintf (stderr, "Giv %d at insn %d not reduced\n", REGNO (v->dest_reg), INSN_UID (v->insn)); #else continue; #endif } /* Givs are only updated once by definition. Mark it so if this is a splittable register. Don't need to do anything for address givs where this may not be a register. */ if (GET_CODE (v->new_reg) == REG) splittable_regs_updates[REGNO (v->new_reg)] = 1; result++; if (loop_dump_stream) { int regnum; if (GET_CODE (v->dest_reg) == CONST_INT) regnum = -1; else if (GET_CODE (v->dest_reg) != REG) regnum = REGNO (XEXP (v->dest_reg, 0)); else regnum = REGNO (v->dest_reg); fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n", regnum, INSN_UID (v->insn)); } } return result; } /* Try to prove that the register is dead after the loop exits. Trace every loop exit looking for an insn that will always be executed, which sets the register to some value, and appears before the first use of the register is found. If successful, then return 1, otherwise return 0. */ /* ?? Could be made more intelligent in the handling of jumps, so that it can search past if statements and other similar structures. */ static int reg_dead_after_loop (reg, loop_start, loop_end) rtx reg, loop_start, loop_end; { rtx insn, label; enum rtx_code code; int jump_count = 0; /* HACK: Must also search the loop fall through exit, create a label_ref here which points to the loop_end, and append the loop_number_exit_labels list to it. */ label = gen_rtx (LABEL_REF, VOIDmode, loop_end); LABEL_NEXTREF (label) = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]; for ( ; label; label = LABEL_NEXTREF (label)) { /* Succeed if find an insn which sets the biv or if reach end of function. Fail if find an insn that uses the biv, or if come to a conditional jump. */ insn = NEXT_INSN (XEXP (label, 0)); while (insn) { code = GET_CODE (insn); if (GET_RTX_CLASS (code) == 'i') { rtx set; if (reg_referenced_p (reg, PATTERN (insn))) return 0; set = single_set (insn); if (set && rtx_equal_p (SET_DEST (set), reg)) break; } if (code == JUMP_INSN) { if (GET_CODE (PATTERN (insn)) == RETURN) break; else if (! simplejump_p (insn) /* Prevent infinite loop following infinite loops. */ || jump_count++ > 20) return 0; else insn = JUMP_LABEL (insn); } insn = NEXT_INSN (insn); } } /* Success, the register is dead on all loop exits. */ return 1; } /* Try to calculate the final value of the biv, the value it will have at the end of the loop. If we can do it, return that value. */ rtx final_biv_value (bl, loop_start, loop_end) struct iv_class *bl; rtx loop_start, loop_end; { rtx increment, tem; /* ??? This only works for MODE_INT biv's. Reject all others for now. */ if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT) return 0; /* The final value for reversed bivs must be calculated differently than for ordinary bivs. In this case, there is already an insn after the loop which sets this biv's final value (if necessary), and there are no other loop exits, so we can return any value. */ if (bl->reversed) { if (loop_dump_stream) fprintf (loop_dump_stream, "Final biv value for %d, reversed biv.\n", bl->regno); return const0_rtx; } /* Try to calculate the final value as initial value + (number of iterations * increment). For this to work, increment must be invariant, the only exit from the loop must be the fall through at the bottom (otherwise it may not have its final value when the loop exits), and the initial value of the biv must be invariant. */ if (loop_n_iterations != 0 && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]] && invariant_p (bl->initial_value)) { increment = biv_total_increment (bl, loop_start, loop_end); if (increment && invariant_p (increment)) { /* Can calculate the loop exit value, emit insns after loop end to calculate this value into a temporary register in case it is needed later. */ tem = gen_reg_rtx (bl->biv->mode); /* Make sure loop_end is not the last insn. */ if (NEXT_INSN (loop_end) == 0) emit_note_after (NOTE_INSN_DELETED, loop_end); emit_iv_add_mult (increment, GEN_INT (loop_n_iterations), bl->initial_value, tem, NEXT_INSN (loop_end)); if (loop_dump_stream) fprintf (loop_dump_stream, "Final biv value for %d, calculated.\n", bl->regno); return tem; } } /* Check to see if the biv is dead at all loop exits. */ if (reg_dead_after_loop (bl->biv->src_reg, loop_start, loop_end)) { if (loop_dump_stream) fprintf (loop_dump_stream, "Final biv value for %d, biv dead after loop exit.\n", bl->regno); return const0_rtx; } return 0; } /* Try to calculate the final value of the giv, the value it will have at the end of the loop. If we can do it, return that value. */ rtx final_giv_value (v, loop_start, loop_end) struct induction *v; rtx loop_start, loop_end; { struct iv_class *bl; rtx insn; rtx increment, tem; enum rtx_code code; rtx insert_before, seq; bl = reg_biv_class[REGNO (v->src_reg)]; /* The final value for givs which depend on reversed bivs must be calculated differently than for ordinary givs. In this case, there is already an insn after the loop which sets this giv's final value (if necessary), and there are no other loop exits, so we can return any value. */ if (bl->reversed) { if (loop_dump_stream) fprintf (loop_dump_stream, "Final giv value for %d, depends on reversed biv\n", REGNO (v->dest_reg)); return const0_rtx; } /* Try to calculate the final value as a function of the biv it depends upon. The only exit from the loop must be the fall through at the bottom (otherwise it may not have its final value when the loop exits). */ /* ??? Can calculate the final giv value by subtracting off the extra biv increments times the giv's mult_val. The loop must have only one exit for this to work, but the loop iterations does not need to be known. */ if (loop_n_iterations != 0 && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]) { /* ?? It is tempting to use the biv's value here since these insns will be put after the loop, and hence the biv will have its final value then. However, this fails if the biv is subsequently eliminated. Perhaps determine whether biv's are eliminable before trying to determine whether giv's are replaceable so that we can use the biv value here if it is not eliminable. */ increment = biv_total_increment (bl, loop_start, loop_end); if (increment && invariant_p (increment)) { /* Can calculate the loop exit value of its biv as (loop_n_iterations * increment) + initial_value */ /* The loop exit value of the giv is then (final_biv_value - extra increments) * mult_val + add_val. The extra increments are any increments to the biv which occur in the loop after the giv's value is calculated. We must search from the insn that sets the giv to the end of the loop to calculate this value. */ insert_before = NEXT_INSN (loop_end); /* Put the final biv value in tem. */ tem = gen_reg_rtx (bl->biv->mode); emit_iv_add_mult (increment, GEN_INT (loop_n_iterations), bl->initial_value, tem, insert_before); /* Subtract off extra increments as we find them. */ for (insn = NEXT_INSN (v->insn); insn != loop_end; insn = NEXT_INSN (insn)) { struct induction *biv; for (biv = bl->biv; biv; biv = biv->next_iv) if (biv->insn == insn) { start_sequence (); tem = expand_binop (GET_MODE (tem), sub_optab, tem, biv->add_val, NULL_RTX, 0, OPTAB_LIB_WIDEN); seq = gen_sequence (); end_sequence (); emit_insn_before (seq, insert_before); } } /* Now calculate the giv's final value. */ emit_iv_add_mult (tem, v->mult_val, v->add_val, tem, insert_before); if (loop_dump_stream) fprintf (loop_dump_stream, "Final giv value for %d, calc from biv's value.\n", REGNO (v->dest_reg)); return tem; } } /* Replaceable giv's should never reach here. */ if (v->replaceable) abort (); /* Check to see if the biv is dead at all loop exits. */ if (reg_dead_after_loop (v->dest_reg, loop_start, loop_end)) { if (loop_dump_stream) fprintf (loop_dump_stream, "Final giv value for %d, giv dead after loop exit.\n", REGNO (v->dest_reg)); return const0_rtx; } return 0; } /* Calculate the number of loop iterations. Returns the exact number of loop iterations if it can be calculated, otherwise returns zero. */ unsigned HOST_WIDE_INT loop_iterations (loop_start, loop_end) rtx loop_start, loop_end; { rtx comparison, comparison_value; rtx iteration_var, initial_value, increment, final_value; enum rtx_code comparison_code; HOST_WIDE_INT i; int increment_dir; int unsigned_compare, compare_dir, final_larger; unsigned long tempu; rtx last_loop_insn; /* First find the iteration variable. If the last insn is a conditional branch, and the insn before tests a register value, make that the iteration variable. */ loop_initial_value = 0; loop_increment = 0; loop_final_value = 0; loop_iteration_var = 0; last_loop_insn = prev_nonnote_insn (loop_end); comparison = get_condition_for_loop (last_loop_insn); if (comparison == 0) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: No final conditional branch found.\n"); return 0; } /* ??? Get_condition may switch position of induction variable and invariant register when it canonicalizes the comparison. */ comparison_code = GET_CODE (comparison); iteration_var = XEXP (comparison, 0); comparison_value = XEXP (comparison, 1); if (GET_CODE (iteration_var) != REG) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Comparison not against register.\n"); return 0; } /* Loop iterations is always called before any new registers are created now, so this should never occur. */ if (REGNO (iteration_var) >= max_reg_before_loop) abort (); iteration_info (iteration_var, &initial_value, &increment, loop_start, loop_end); if (initial_value == 0) /* iteration_info already printed a message. */ return 0; if (increment == 0) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Increment value can't be calculated.\n"); return 0; } if (GET_CODE (increment) != CONST_INT) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Increment value not constant.\n"); return 0; } if (GET_CODE (initial_value) != CONST_INT) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Initial value not constant.\n"); return 0; } /* If the comparison value is an invariant register, then try to find its value from the insns before the start of the loop. */ if (GET_CODE (comparison_value) == REG && invariant_p (comparison_value)) { rtx insn, set; for (insn = PREV_INSN (loop_start); insn ; insn = PREV_INSN (insn)) { if (GET_CODE (insn) == CODE_LABEL) break; else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' && reg_set_p (comparison_value, insn)) { /* We found the last insn before the loop that sets the register. If it sets the entire register, and has a REG_EQUAL note, then use the value of the REG_EQUAL note. */ if ((set = single_set (insn)) && (SET_DEST (set) == comparison_value)) { rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX); if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST) comparison_value = XEXP (note, 0); } break; } } } final_value = approx_final_value (comparison_code, comparison_value, &unsigned_compare, &compare_dir); /* Save the calculated values describing this loop's bounds, in case precondition_loop_p will need them later. These values can not be recalculated inside precondition_loop_p because strength reduction optimizations may obscure the loop's structure. */ loop_iteration_var = iteration_var; loop_initial_value = initial_value; loop_increment = increment; loop_final_value = final_value; if (final_value == 0) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: EQ comparison loop.\n"); return 0; } else if (GET_CODE (final_value) != CONST_INT) { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Final value not constant.\n"); return 0; } /* ?? Final value and initial value do not have to be constants. Only their difference has to be constant. When the iteration variable is an array address, the final value and initial value might both be addresses with the same base but different constant offsets. Final value must be invariant for this to work. To do this, need some way to find the values of registers which are invariant. */ /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */ if (unsigned_compare) final_larger = ((unsigned HOST_WIDE_INT) INTVAL (final_value) > (unsigned HOST_WIDE_INT) INTVAL (initial_value)) - ((unsigned HOST_WIDE_INT) INTVAL (final_value) < (unsigned HOST_WIDE_INT) INTVAL (initial_value)); else final_larger = (INTVAL (final_value) > INTVAL (initial_value)) - (INTVAL (final_value) < INTVAL (initial_value)); if (INTVAL (increment) > 0) increment_dir = 1; else if (INTVAL (increment) == 0) increment_dir = 0; else increment_dir = -1; /* There are 27 different cases: compare_dir = -1, 0, 1; final_larger = -1, 0, 1; increment_dir = -1, 0, 1. There are 4 normal cases, 4 reverse cases (where the iteration variable will overflow before the loop exits), 4 infinite loop cases, and 15 immediate exit (0 or 1 iteration depending on loop type) cases. Only try to optimize the normal cases. */ /* (compare_dir/final_larger/increment_dir) Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1) Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1) Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0) Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */ /* ?? If the meaning of reverse loops (where the iteration variable will overflow before the loop exits) is undefined, then could eliminate all of these special checks, and just always assume the loops are normal/immediate/infinite. Note that this means the sign of increment_dir does not have to be known. Also, since it does not really hurt if immediate exit loops or infinite loops are optimized, then that case could be ignored also, and hence all loops can be optimized. According to ANSI Spec, the reverse loop case result is undefined, because the action on overflow is undefined. See also the special test for NE loops below. */ if (final_larger == increment_dir && final_larger != 0 && (final_larger == compare_dir || compare_dir == 0)) /* Normal case. */ ; else { if (loop_dump_stream) fprintf (loop_dump_stream, "Loop unrolling: Not normal loop.\n"); return 0; } /* Calculate the number of iterations, final_value is only an approximation, so correct for that. Note that tempu and loop_n_iterations are unsigned, because they can be as large as 2^n - 1. */ i = INTVAL (increment); if (i > 0) tempu = INTVAL (final_value) - INTVAL (initial_value); else if (i < 0) { tempu = INTVAL (initial_value) - INTVAL (final_value); i = -i; } else abort (); /* For NE tests, make sure that the iteration variable won't miss the final value. If tempu mod i is not zero, then the iteration variable will overflow before the loop exits, and we can not calculate the number of iterations. */ if (compare_dir == 0 && (tempu % i) != 0) return 0; return tempu / i + ((tempu % i) != 0); }