ir_gcm.c

/*
 * IR - Lightweight JIT Compilation Framework
 * (GCM - Global Code Motion and Scheduler)
 * Copyright (C) 2022 Zend by Perforce.
 * Authors: Dmitry Stogov <dmitry@php.net>
 *
 * The GCM algorithm is based on Cliff Click's publication
 * See: C. Click. "Global code motion, global value numbering" Submitted to PLDI'95.
 */

#include "ir.h"
#include "ir_private.h"

#define IR_GCM_IS_SCHEDULED_EARLY(b) (((int32_t)(b)) < 0)
#define IR_GCM_EARLY_BLOCK(b)        ((uint32_t)-((int32_t)(b)))

#define IR_GCM_SPLIT 1
#define IR_SCHEDULE_SWAP_OPS 1

static uint32_t ir_gcm_schedule_early(ir_ctx *ctx, ir_ref ref, ir_list *queue_late)
{
	ir_ref n, *p, input;
	ir_insn *insn;
	uint32_t dom_depth;
	uint32_t b, result;

	insn = &ctx->ir_base[ref];

	IR_ASSERT(insn->op != IR_PARAM && insn->op != IR_VAR);
	IR_ASSERT(insn->op != IR_PHI && insn->op != IR_PI);

	result = 1;
	dom_depth = 0;

	n = insn->inputs_count;
	for (p = insn->ops + 1; n > 0; p++, n--) {
		input = *p;
		if (input > 0) {
			b = ctx->cfg_map[input];
			if (IR_GCM_IS_SCHEDULED_EARLY(b)) {
				b = IR_GCM_EARLY_BLOCK(b);
			} else if (!b) {
				b = ir_gcm_schedule_early(ctx, input, queue_late);
			}
			if (dom_depth < ctx->cfg_blocks[b].dom_depth) {
				dom_depth = ctx->cfg_blocks[b].dom_depth;
				result = b;
			}
		}
	}

	ctx->cfg_map[ref] = IR_GCM_EARLY_BLOCK(result);
	ir_list_push_unchecked(queue_late, ref);
	return result;
}

/* Last Common Ancestor */
static uint32_t ir_gcm_find_lca(ir_ctx *ctx, uint32_t b1, uint32_t b2)
{
	uint32_t dom_depth;

	dom_depth = ctx->cfg_blocks[b2].dom_depth;
	while (ctx->cfg_blocks[b1].dom_depth > dom_depth) {
		b1 = ctx->cfg_blocks[b1].dom_parent;
	}
	dom_depth = ctx->cfg_blocks[b1].dom_depth;
	while (ctx->cfg_blocks[b2].dom_depth > dom_depth) {
		b2 = ctx->cfg_blocks[b2].dom_parent;
	}
	while (b1 != b2) {
		b1 = ctx->cfg_blocks[b1].dom_parent;
		b2 = ctx->cfg_blocks[b2].dom_parent;
	}
	return b2;
}

static uint32_t ir_gcm_select_best_block(ir_ctx *ctx, ir_ref ref, uint32_t lca)
{
	ir_block *bb = &ctx->cfg_blocks[lca];
	uint32_t loop_depth = bb->loop_depth;
	uint32_t flags, best, b;

	if (!loop_depth) {
		return lca;
	}

#if 0 /* This is not necessary anymore. Conditions may be fused with IF across BBs. */
	if (ctx->ir_base[ref].op >= IR_EQ && ctx->ir_base[ref].op <= IR_UGT) {
		ir_use_list *use_list = &ctx->use_lists[ref];

		if (use_list->count == 1) {
			ir_ref use = ctx->use_edges[use_list->refs];
			ir_insn *insn = &ctx->ir_base[use];
			if (insn->op == IR_IF || insn->op == IR_GUARD || insn->op == IR_GUARD_NOT) {
				/* Don't hoist invariant comparison */
				return lca;
			}
		}
	}
#endif

	flags = (bb->flags & IR_BB_LOOP_HEADER) ? bb->flags : ctx->cfg_blocks[bb->loop_header].flags;
	if ((flags & IR_BB_LOOP_WITH_ENTRY)
	 && !(ctx->binding && ir_binding_find(ctx, ref))) {
		/* Don't move loop invariant code across an OSR ENTRY if we can't restore it */
		return lca;
	}

	best = b = lca;
	do {
		b = bb->dom_parent;
		bb = &ctx->cfg_blocks[b];
		if (bb->loop_depth < loop_depth) {
			if (!bb->loop_depth) {
#if 1
				/* Avoid LICM if LOOP doesn't have a pre-header block */
				ir_block *loop_bb = &ctx->cfg_blocks[best];

				if (!(loop_bb->flags & IR_BB_LOOP_HEADER)) {
					loop_bb = &ctx->cfg_blocks[loop_bb->loop_header];
				}
				if (loop_bb->predecessors_count > 2) {
					int n = loop_bb->predecessors_count;
					uint32_t *p = ctx->cfg_edges + loop_bb->predecessors;

					while (n && *p != b) {
						n--; p++;
					}
					if (!n) {
						break;
					}
				}
#endif
				best = b;
				break;
			}
			flags = (bb->flags & IR_BB_LOOP_HEADER) ? bb->flags : ctx->cfg_blocks[bb->loop_header].flags;
			if ((flags & IR_BB_LOOP_WITH_ENTRY)
			 && !(ctx->binding && ir_binding_find(ctx, ref))) {
				break;
			}
			loop_depth = bb->loop_depth;
			best = b;
		}
	} while (b != ctx->cfg_map[ref]);

	return best;
}

#if IR_GCM_SPLIT
/* Partially Dead Code Elimination through splitting the node and sunking the clones
 *
 * This code is based on the Benedikt Meurer's idea first implemented in V8.
 * See: https://codereview.chromium.org/899433005
 */

typedef struct _ir_gcm_split_data {
	ir_sparse_set totally_useful;
	ir_list       worklist;
} ir_gcm_split_data;

static void _push_predecessors(ir_ctx *ctx, ir_block *bb, ir_gcm_split_data *data)
{
	uint32_t *p, i, n = bb->predecessors_count;

	IR_ASSERT(n > 0);
	p = ctx->cfg_edges + bb->predecessors;
	do {
		i = *p;
		if (!ir_sparse_set_in(&data->totally_useful, i)) {
			ir_list_push(&data->worklist, i);
		}
		p++;
		n--;
	} while (n > 0);
}

static bool _check_successors(ir_ctx *ctx, ir_block *bb, ir_gcm_split_data *data)
{
	uint32_t *p, i, n = bb->successors_count;

	if (n <= 1) {
		IR_ASSERT(ir_sparse_set_in(&data->totally_useful, ctx->cfg_edges[bb->successors]));
		return 1;
	}

	p = ctx->cfg_edges + bb->successors;
	do {
		i = *p;
		if (!ir_sparse_set_in(&data->totally_useful, i)) {
			return 0;
		}
		p++;
		n--;
	} while (n > 0);

	return 1;
}

static bool ir_split_partially_dead_node(ir_ctx *ctx, ir_ref ref, uint32_t b)
{
	ir_use_list *use_list;
	ir_insn *insn;
	ir_ref n, *p, use;
	uint32_t i;
	ir_gcm_split_data *data = ctx->data;

	IR_ASSERT(b > 0 && b <= ctx->cfg_blocks_count);

	/* 1. Find a set of blocks where the node is TOTALLY_USEFUL (not PARTIALLY_DEAD)
	 * 1.1. Collect the blocks where the node is really USED.
	 */
	ir_sparse_set_clear(&data->totally_useful);

	use_list = &ctx->use_lists[ref];
	n = use_list->count;
	for (p = &ctx->use_edges[use_list->refs]; n > 0; p++, n--) {
		use = *p;
		insn = &ctx->ir_base[use];
		if (insn->op == IR_PHI) {
			ir_ref *p = insn->ops + 2; /* PHI data inputs */
			ir_ref *q = ctx->ir_base[insn->op1].ops + 1; /* MERGE inputs */
			ir_ref n = insn->inputs_count - 1;

			for (;n > 0; p++, q++, n--) {
				if (*p == ref) {
					i = ctx->cfg_map[*q];
					IR_ASSERT(i > 0 && i <= ctx->cfg_blocks_count);
					if (!ir_sparse_set_in(&data->totally_useful, i)) {
						if (i == b) return 0; /* node is totally-useful in the scheduled block */
						ir_sparse_set_add(&data->totally_useful, i);
					}
				}
			}
		} else {
			i = ctx->cfg_map[use];
			if (!i) {
				continue;
			}
			IR_ASSERT(i > 0 && i <= ctx->cfg_blocks_count);
			if (!ir_sparse_set_in(&data->totally_useful, i)) {
				if (i == b) return 0; /* node is totally-useful in the scheduled block */
				ir_sparse_set_add(&data->totally_useful, i);
			}
		}
	}

#ifdef IR_DEBUG
	if (ctx->flags & IR_DEBUG_GCM_SPLIT) {
		bool first = 1;
		fprintf(stderr, "*** Split partially dead node d_%d scheduled to BB%d\n", ref, b);
		IR_SPARSE_SET_FOREACH(&data->totally_useful, i) {
			if (first) {
				fprintf(stderr, "\td_%d is USED in [BB%d", ref, i);
				first = 0;
			} else {
				fprintf(stderr, ", BB%d", i);
			}
		} IR_SPARSE_SET_FOREACH_END();
		fprintf(stderr, "]\n");
	}
#endif

	/* 1.2. Iteratively check the predecessors of already found TOTALLY_USEFUL blocks and
	 *      add them into TOTALLY_USEFUL set if all of their sucessors are already there.
	 */
	IR_SPARSE_SET_FOREACH(&data->totally_useful, i) {
		_push_predecessors(ctx, &ctx->cfg_blocks[i], data);
	} IR_SPARSE_SET_FOREACH_END();

	while (ir_list_len(&data->worklist)) {
		i = ir_list_pop(&data->worklist);
		if (!ir_sparse_set_in(&data->totally_useful, i)) {
			ir_block *bb = &ctx->cfg_blocks[i];

			if (_check_successors(ctx, bb, data)) {
				if (i == b) {
					/* node is TOTALLY_USEFUL in the scheduled block */
					ir_list_clear(&data->worklist);
					return 0;
				}
				ir_sparse_set_add(&data->totally_useful, i);
				_push_predecessors(ctx, bb, data);
			}
		}
	}

	IR_ASSERT(!ir_sparse_set_in(&data->totally_useful, b));

#ifdef IR_DEBUG
	if (ctx->flags & IR_DEBUG_GCM_SPLIT) {
		bool first = 1;
		IR_SPARSE_SET_FOREACH(&data->totally_useful, i) {
			if (first) {
				fprintf(stderr, "\td_%d is TOTALLY_USEFUL in [BB%d", ref, i);
				first = 0;
			} else {
				fprintf(stderr, ", BB%d", i);
			}
		} IR_SPARSE_SET_FOREACH_END();
		fprintf(stderr, "]\n");
	}
#endif

	/* 2. Split the USEs into partitions */
	use_list = &ctx->use_lists[ref];
	ir_hashtab hash;
	uint32_t j, clone, clones_count = 0, uses_count = 0;
	struct {
		ir_ref   ref;
		uint32_t block;
		uint32_t use_count;
		uint32_t use;
	} *clones = ir_mem_malloc(sizeof(*clones) * use_list->count);
	struct {
		ir_ref   ref;
		uint32_t block;
		uint32_t next;
	} *uses = ir_mem_malloc(sizeof(*uses) * use_list->count);

	ir_hashtab_init(&hash, use_list->count);
	n = use_list->count;
	for (p = &ctx->use_edges[use_list->refs]; n > 0; p++, n--) {
		use = *p;
		insn = &ctx->ir_base[use];
		if (insn->op == IR_PHI) {
			ir_ref *p = insn->ops + 2; /* PHI data inputs */
			ir_ref *q = ctx->ir_base[insn->op1].ops + 1; /* MERGE inputs */
			ir_ref n = insn->inputs_count - 1;

			/* PHIs must be processed once */
			if (ir_hashtab_find(&hash, -use) != (ir_ref)IR_INVALID_VAL) {
				continue;
			}
			ir_hashtab_add(&hash, -use, IR_NULL);
			for (;n > 0; p++, q++, n--) {
				if (*p == ref) {
					j = i = ctx->cfg_map[*q];
					while (ir_sparse_set_in(&data->totally_useful, ctx->cfg_blocks[j].idom)) {
						j = ctx->cfg_blocks[j].idom;
					}
					clone = ir_hashtab_find(&hash, j);
					if (clone == IR_INVALID_VAL) {
						clone = clones_count++;
						ir_hashtab_add(&hash, j, clone);
						clones[clone].block = j;
						clones[clone].use_count = 0;
						clones[clone].use = (uint32_t)-1;
					}
					uses[uses_count].ref = use;
					uses[uses_count].block = i;
					uses[uses_count].next = clones[clone].use;
					clones[clone].use_count++;
					clones[clone].use = uses_count++;
				}
			}
		} else {
			j = i = ctx->cfg_map[use];
			IR_ASSERT(i > 0);
			while (ir_sparse_set_in(&data->totally_useful, ctx->cfg_blocks[j].idom)) {
				j = ctx->cfg_blocks[j].idom;
			}
			clone = ir_hashtab_find(&hash, j);
			if (clone == IR_INVALID_VAL) {
				clone = clones_count++;
				ir_hashtab_add(&hash, j, clone);
				clones[clone].block = j;
				clones[clone].use_count = 0;
				clones[clone].use = -1;
			}
			uses[uses_count].ref = use;
			uses[uses_count].block = i;
			uses[uses_count].next = clones[clone].use;
			clones[clone].use_count++;
			clones[clone].use = uses_count++;
		}
	}

#ifdef IR_DEBUG
	if (ctx->flags & IR_DEBUG_GCM_SPLIT) {
		for (i = 0; i < clones_count; i++) {
			uint32_t u = clones[i].use;

			fprintf(stderr, "\tCLONE #%d in BB%d USES(%d)=[d_%d/BB%d",
				i, clones[i].block, clones[i].use_count, uses[u].ref, uses[u].block);
			u = uses[u].next;
			while (u != (uint32_t)-1) {
				fprintf(stderr, ", d_%d/BB%d", uses[u].ref, uses[u].block);
				u = uses[u].next;
			}
			fprintf(stderr, "]\n");
		}
	}
#endif

	/* Create Clones */
	insn = &ctx->ir_base[ref];
	clones[0].ref = ref;
	for (i = 1; i < clones_count; i++) {
		clones[i].ref = clone = ir_emit(ctx, insn->optx, insn->op1, insn->op2, insn->op3);
		insn = &ctx->ir_base[ref];
		if (insn->op1 > 0) ir_use_list_add(ctx, insn->op1, clone);
		if (insn->op2 > 0) ir_use_list_add(ctx, insn->op2, clone);
		if (insn->op3 > 0) ir_use_list_add(ctx, insn->op3, clone);
	}

	/* Reconstruct IR: Update DEF->USE lists, CFG mapping and etc */
	ctx->use_lists = ir_mem_realloc(ctx->use_lists, ctx->insns_count * sizeof(ir_use_list));
	ctx->cfg_map = ir_mem_realloc(ctx->cfg_map, ctx->insns_count * sizeof(uint32_t));
	n = ctx->use_lists[ref].refs;
	for (i = 0; i < clones_count; i++) {
		clone = clones[i].ref;
		if (clones[i].use_count == 1) {
			/* TOTALLY_USEFUL block may be a head of a diamond above the real usage.
			 * Sink it down to the real usage block.
			 * Clones with few uses we be sunk into the LCA block.
			 */
			clones[i].block = uses[clones[i].use].block;
		}
		ctx->cfg_map[clone] = clones[i].block;
		ctx->use_lists[clone].count = clones[i].use_count;
		ctx->use_lists[clone].refs = n;

		uint32_t u = clones[i].use;
		while (u != (uint32_t)-1) {
			use = uses[u].ref;
			ctx->use_edges[n++] = use;
			u = uses[u].next;
			if (i > 0) {
				/* replace inputs */
				ir_insn *insn = &ctx->ir_base[use];
				ir_ref k, l = insn->inputs_count;

				if (insn->op == IR_PHI) {
					for (k = 1; k <= l; k++) {
						if (ir_insn_op(insn, k) == ref) {
							j = ctx->cfg_map[ir_insn_op(&ctx->ir_base[insn->op1], k - 1)];
							if (j != clones[i].block) {
								uint32_t dom_depth = ctx->cfg_blocks[clones[i].block].dom_depth;
								while (ctx->cfg_blocks[j].dom_depth > dom_depth) {
									j = ctx->cfg_blocks[j].dom_parent;
								}
								if (j != clones[i].block) {
									continue;
								}
							}
							ir_insn_set_op(insn, k, clone);
							break;
						}
					}
				} else {
					for (k = 1; k <= l; k++) {
						if (ir_insn_op(insn, k) == ref) {
							ir_insn_set_op(insn, k, clone);
							break;
						}
					}
				}
			}
		}
	}

	ir_mem_free(uses);
	ir_mem_free(clones);
	ir_hashtab_free(&hash);

#ifdef IR_DEBUG
	if (ctx->flags & IR_DEBUG_GCM_SPLIT) {
		ir_check(ctx);
	}
#endif

	return 1;
}
#endif

#ifdef IR_DEBUG
static bool ir_gcm_dominates(ir_ctx *ctx, uint32_t b1, uint32_t b2)
{
	uint32_t b1_depth = ctx->cfg_blocks[b1].dom_depth;
	const ir_block *bb2 = &ctx->cfg_blocks[b2];

	while (bb2->dom_depth > b1_depth) {
		b2 = bb2->dom_parent;
		bb2 = &ctx->cfg_blocks[b2];
	}
	return b1 == b2;
}
#endif

static void ir_gcm_schedule_late(ir_ctx *ctx, ir_ref ref, uint32_t b)
{
	ir_ref n, use;
	uint32_t lca = 0;

	IR_ASSERT(ctx->ir_base[ref].op != IR_PARAM && ctx->ir_base[ref].op != IR_VAR);
	IR_ASSERT(ctx->ir_base[ref].op != IR_PHI && ctx->ir_base[ref].op != IR_PI);

	IR_ASSERT(IR_GCM_IS_SCHEDULED_EARLY(b));
	b = IR_GCM_EARLY_BLOCK(b);
	ctx->cfg_map[ref] = b;

	for (n = 0; n < ctx->use_lists[ref].count; n++) {
		use = ctx->use_edges[ctx->use_lists[ref].refs + n];
		b = ctx->cfg_map[use];
		if (IR_GCM_IS_SCHEDULED_EARLY(b)) {
			ir_gcm_schedule_late(ctx, use, b);
			b = ctx->cfg_map[use];
			IR_ASSERT(b != 0);
		} else if (!b) {
			continue;
		} else if (ctx->ir_base[use].op == IR_PHI) {
			ir_insn *insn = &ctx->ir_base[use];
			ir_ref *p = insn->ops + 2; /* PHI data inputs */
			ir_ref *q = ctx->ir_base[insn->op1].ops + 1; /* MERGE inputs */
			ir_ref n = insn->inputs_count - 1;

			for (;n > 0; p++, q++, n--) {
				if (*p == ref) {
					b = ctx->cfg_map[*q];
					lca = !lca ? b : ir_gcm_find_lca(ctx, lca, b);
				}
			}
			continue;
		}
		lca = !lca ? b : ir_gcm_find_lca(ctx, lca, b);
	}

	IR_ASSERT(lca != 0 && "No Common Ancestor");
	IR_ASSERT(ir_gcm_dominates(ctx, ctx->cfg_map[ref], lca) && "Early placement doesn't dominate the late");

#if IR_GCM_SPLIT
	if (ctx->use_lists[ref].count > 1
	 && ir_split_partially_dead_node(ctx, ref, lca)) {
		return;
	}
#endif

	if (lca != ctx->cfg_map[ref]) {
		b = ir_gcm_select_best_block(ctx, ref, lca);

		ctx->cfg_map[ref] = b;

		/* OVERFLOW is a projection of ADD/SUB/MUL_OV and must be scheduled into the same block */
		if (ctx->ir_base[ref].op >= IR_ADD_OV && ctx->ir_base[ref].op <= IR_MUL_OV) {
			ir_use_list *use_list = &ctx->use_lists[ref];
			ir_ref n, *p, use;

			for (n = use_list->count, p = &ctx->use_edges[use_list->refs]; n < 0; p++, n--) {
				use = *p;
				if (ctx->ir_base[use].op == IR_OVERFLOW) {
					ctx->cfg_map[use] = b;
					break;
				}
			}
		}
	}
}

int ir_gcm(ir_ctx *ctx)
{
	ir_ref k, n, *p, ref;
	ir_block *bb;
	ir_list queue_early;
	ir_list queue_late;
	uint32_t *_blocks, b;
	ir_insn *insn, *use_insn;
	ir_use_list *use_list;

	IR_ASSERT(ctx->cfg_map);
	_blocks = ctx->cfg_map;

	ir_list_init(&queue_early, ctx->insns_count);

	if (ctx->cfg_blocks_count == 1) {
		ref = ctx->cfg_blocks[1].end;
		do {
			insn = &ctx->ir_base[ref];
			_blocks[ref] = 1; /* pin to block */
			if (insn->inputs_count > 1) {
				/* insn has input data edges */
				ir_list_push_unchecked(&queue_early, ref);
			}
			ref = insn->op1; /* control predecessor */
		} while (ref != 1); /* IR_START */
		_blocks[1] = 1; /* pin to block */

		use_list = &ctx->use_lists[1];
		n = use_list->count;
		for (p = &ctx->use_edges[use_list->refs]; n > 0; n--, p++) {
			ref = *p;
			use_insn = &ctx->ir_base[ref];
			if (use_insn->op == IR_PARAM || use_insn->op == IR_VAR) {
				ctx->cfg_blocks[1].flags |= (use_insn->op == IR_PARAM) ? IR_BB_HAS_PARAM : IR_BB_HAS_VAR;
				_blocks[ref] = 1; /* pin to block */
			}
		}

		/* Place all live nodes to the first block */
		while (ir_list_len(&queue_early)) {
			ref = ir_list_pop(&queue_early);
			insn = &ctx->ir_base[ref];
			n = insn->inputs_count;
			for (p = insn->ops + 1; n > 0; p++, n--) {
				ref = *p;
				if (ref > 0 && _blocks[ref] == 0) {
					_blocks[ref] = 1;
					ir_list_push_unchecked(&queue_early, ref);
				}
			}
		}

		ir_list_free(&queue_early);

		return 1;
	}

	ir_list_init(&queue_late, ctx->insns_count);

	/* pin and collect control and control depended (PARAM, VAR, PHI, PI) instructions */
	b = ctx->cfg_blocks_count;
	for (bb = ctx->cfg_blocks + b; b > 0; bb--, b--) {
		IR_ASSERT(!(bb->flags & IR_BB_UNREACHABLE));
		ref = bb->end;

		/* process the last instruction of the block */
		insn = &ctx->ir_base[ref];
		_blocks[ref] = b; /* pin to block */
		if (insn->inputs_count > 1) {
			/* insn has input data edges */
			ir_list_push_unchecked(&queue_early, ref);
		}
		ref = insn->op1; /* control predecessor */

		while (ref != bb->start) {
			insn = &ctx->ir_base[ref];
			_blocks[ref] = b; /* pin to block */
			if (insn->inputs_count > 1) {
				/* insn has input data edges */
				ir_list_push_unchecked(&queue_early, ref);
			}
			if (insn->type != IR_VOID) {
				IR_ASSERT(ir_op_flags[insn->op] & IR_OP_FLAG_MEM);
			}
			ref = insn->op1; /* control predecessor */
		}

		/* process the first instruction of the block */
		_blocks[ref] = b; /* pin to block */

		use_list = &ctx->use_lists[ref];
		n = use_list->count;
		if (n > 1) {
			for (p = &ctx->use_edges[use_list->refs]; n > 0; n--, p++) {
				ref = *p;
				use_insn = &ctx->ir_base[ref];
				if (use_insn->op == IR_PHI || use_insn->op == IR_PI) {
					bb->flags |= (use_insn->op == IR_PHI) ? IR_BB_HAS_PHI : IR_BB_HAS_PI;
					if (EXPECTED(ctx->use_lists[ref].count != 0)) {
						_blocks[ref] = b; /* pin to block */
						ir_list_push_unchecked(&queue_early, ref);
					}
				} else if (use_insn->op == IR_PARAM) {
					bb->flags |= IR_BB_HAS_PARAM;
					_blocks[ref] = b; /* pin to block */
				} else if (use_insn->op == IR_VAR) {
					bb->flags |= IR_BB_HAS_VAR;
					_blocks[ref] = b; /* pin to block */
				}
			}
		}
	}

	n = ir_list_len(&queue_early);
	while (n > 0) {
		n--;
		ref = ir_list_at(&queue_early, n);
		insn = &ctx->ir_base[ref];
		k = insn->inputs_count - 1;
		for (p = insn->ops + 2; k > 0; p++, k--) {
			ref = *p;
			if (ref > 0 && _blocks[ref] == 0) {
				ir_gcm_schedule_early(ctx, ref, &queue_late);
			}
		}
	}

#ifdef IR_DEBUG
	if (ctx->flags & IR_DEBUG_GCM) {
		fprintf(stderr, "GCM Schedule Early\n");
		for (n = 1; n < ctx->insns_count; n++) {
			fprintf(stderr, "%d -> %d\n", n, ctx->cfg_map[n]);
		}
	}
#endif

#if IR_GCM_SPLIT
	ir_gcm_split_data data;

	ir_sparse_set_init(&data.totally_useful, ctx->cfg_blocks_count + 1);
	ir_list_init(&data.worklist, ctx->cfg_blocks_count + 1);
	ctx->data = &data;
#endif

	n = ir_list_len(&queue_late);
	while (n > 0) {
		n--;
		ref = ir_list_at(&queue_late, n);
		b = ctx->cfg_map[ref];
		if (IR_GCM_IS_SCHEDULED_EARLY(b)) {
			ir_gcm_schedule_late(ctx, ref, b);
		}
	}

#if IR_GCM_SPLIT
	ir_list_free(&data.worklist);
	ir_sparse_set_free(&data.totally_useful);
	ctx->data = NULL;
#endif

	ir_list_free(&queue_early);
	ir_list_free(&queue_late);

#ifdef IR_DEBUG
	if (ctx->flags & IR_DEBUG_GCM) {
		fprintf(stderr, "GCM Schedule Late\n");
		for (n = 1; n < ctx->insns_count; n++) {
			fprintf(stderr, "%d -> %d\n", n, ctx->cfg_map[n]);
		}
	}
#endif

	return 1;
}

static void ir_xlat_binding(ir_ctx *ctx, ir_ref *_xlat)
{
	uint32_t n1, n2, pos;
	ir_ref key;
	ir_hashtab_bucket *b1, *b2;
	ir_hashtab *binding = ctx->binding;
	uint32_t hash_size = (uint32_t)(-(int32_t)binding->mask);

	memset((char*)binding->data - (hash_size * sizeof(uint32_t)), -1, hash_size * sizeof(uint32_t));
	n1 = binding->count;
	n2 = 0;
	pos = 0;
	b1 = binding->data;
	b2 = binding->data;
	while (n1 > 0) {
		key = b1->key;
		IR_ASSERT(key < ctx->insns_count);
		if (_xlat[key]) {
			key = _xlat[key];
			b2->key = key;
			if (b1->val > 0) {
				IR_ASSERT(_xlat[b1->val]);
				b2->val = _xlat[b1->val];
			} else {
				b2->val = b1->val;
			}
			key |= binding->mask;
			b2->next = ((uint32_t*)binding->data)[key];
			((uint32_t*)binding->data)[key] = pos;
			pos += sizeof(ir_hashtab_bucket);
			b2++;
			n2++;
		}
		b1++;
		n1--;
	}
	binding->count = n2;
}

IR_ALWAYS_INLINE ir_ref ir_count_constant(ir_ref *_xlat, ir_ref ref)
{
	if (!_xlat[ref]) {
		_xlat[ref] = ref; /* this is only a "used constant" marker */
		return 1;
	}
	return 0;
}

int ir_schedule(ir_ctx *ctx)
{
	ir_ctx new_ctx;
	ir_ref i, j, k, n, *p, *q, ref, new_ref, prev_ref, insns_count, consts_count, use_edges_count;
	ir_ref *_xlat;
	ir_ref *edges;
	ir_ref prev_b_end;
	uint32_t b, prev_b;
	uint32_t *_blocks = ctx->cfg_map;
	ir_ref *_next = ir_mem_malloc(ctx->insns_count * sizeof(ir_ref));
	ir_ref *_prev = ir_mem_malloc(ctx->insns_count * sizeof(ir_ref));
	ir_ref _move_down = 0;
	ir_block *bb;
	ir_insn *insn, *new_insn;
	ir_use_list *lists, *use_list, *new_list;

	/* Create a double-linked list of nodes ordered by BB, respecting BB->start and BB->end */
	IR_ASSERT(_blocks[1] == 1);
	prev_b = 1;
	prev_b_end = ctx->cfg_blocks[1].end;
	_prev[1] = 0;
	_prev[prev_b_end] = 0;
	for (i = 2, j = 1; i < ctx->insns_count; i++) {
		b = _blocks[i];
		IR_ASSERT((int32_t)b >= 0);
		if (b == prev_b && i <= prev_b_end) {
			/* add to the end of the list */
			_next[j] = i;
			_prev[i] = j;
			j = i;
		} else if (b > prev_b) {
			bb = &ctx->cfg_blocks[b];
			if (i == bb->start) {
				IR_ASSERT(bb->end > bb->start);
				prev_b = b;
				prev_b_end = bb->end;
				_prev[bb->end] = 0;
				/* add to the end of the list */
				_next[j] = i;
				_prev[i] = j;
				j = i;
			} else {
				IR_ASSERT(i != bb->end);
				/* move down late (see the following loop) */
				_next[i] = _move_down;
				_move_down = i;
			}
		} else if (b) {
			bb = &ctx->cfg_blocks[b];
			IR_ASSERT(i != bb->start);
			if (_prev[bb->end]) {
				/* move up, insert before the end of the already scheduled BB */
				k = bb->end;
			} else {
				/* move up, insert at the end of the block */
				k = ctx->cfg_blocks[b + 1].start;
			}
			/* insert before "k" */
			_prev[i] = _prev[k];
			_next[i] = k;
			_next[_prev[k]] = i;
			_prev[k] = i;
		}
	}
	_next[j] = 0;

	while (_move_down) {
		i = _move_down;
		_move_down = _next[i];
		b = _blocks[i];
		bb = &ctx->cfg_blocks[b];
		k = _next[bb->start];

		if (bb->flags & (IR_BB_HAS_PHI|IR_BB_HAS_PI|IR_BB_HAS_PARAM|IR_BB_HAS_VAR)) {
			/* insert after the start of the block and all PARAM, VAR, PI, PHI */
			insn = &ctx->ir_base[k];
			while (insn->op == IR_PHI || insn->op == IR_PARAM || insn->op == IR_VAR || insn->op == IR_PI) {
				k = _next[k];
				insn = &ctx->ir_base[k];
			}
		}

		/* insert before "k" */
		_prev[i] = _prev[k];
		_next[i] = k;
		_next[_prev[k]] = i;
		_prev[k] = i;
	}

#ifdef IR_DEBUG
	if (ctx->flags & IR_DEBUG_SCHEDULE) {
		fprintf(stderr, "Before Schedule\n");
		for (i = 1; i != 0; i = _next[i]) {
			fprintf(stderr, "%d -> %d\n", i, _blocks[i]);
		}
	}
#endif

	_xlat = ir_mem_calloc((ctx->consts_count + ctx->insns_count), sizeof(ir_ref));
	_xlat += ctx->consts_count;
	_xlat[IR_TRUE] = IR_TRUE;
	_xlat[IR_FALSE] = IR_FALSE;
	_xlat[IR_NULL] = IR_NULL;
	_xlat[IR_UNUSED] = IR_UNUSED;
	insns_count = 1;
	consts_count = -(IR_TRUE - 1);

	/* Topological sort according dependencies inside each basic block */
	for (b = 1, bb = ctx->cfg_blocks + 1; b <= ctx->cfg_blocks_count; b++, bb++) {
		IR_ASSERT(!(bb->flags & IR_BB_UNREACHABLE));
		/* Schedule BB start */
		i = bb->start;
		_xlat[i] = bb->start = insns_count;
		insn = &ctx->ir_base[i];
		if (insn->op == IR_CASE_VAL) {
			IR_ASSERT(insn->op2 < IR_TRUE);
			consts_count += ir_count_constant(_xlat, insn->op2);
		}
		n = insn->inputs_count;
		insns_count += ir_insn_inputs_to_len(n);
		i = _next[i];
		insn = &ctx->ir_base[i];
		if (bb->flags & (IR_BB_HAS_PHI|IR_BB_HAS_PI|IR_BB_HAS_PARAM|IR_BB_HAS_VAR)) {
			/* Schedule PARAM, VAR, PI */
			while (insn->op == IR_PARAM || insn->op == IR_VAR || insn->op == IR_PI) {
				_xlat[i] = insns_count;
				insns_count += 1;
				i = _next[i];
				insn = &ctx->ir_base[i];
			}
			/* Schedule PHIs */
			while (insn->op == IR_PHI) {
				ir_ref j, *p, input;

				_xlat[i] = insns_count;
				/* Reuse "n" from MERGE and skip first input */
				insns_count += ir_insn_inputs_to_len(n + 1);
				for (j = n, p = insn->ops + 2; j > 0; p++, j--) {
					input = *p;
					if (input < IR_TRUE) {
						consts_count += ir_count_constant(_xlat, input);
					}
				}
				i = _next[i];
				insn = &ctx->ir_base[i];
			}
		}
		if (bb->successors_count > 1) {
			ir_ref input, j = bb->end;
			ir_insn *end = &ctx->ir_base[j];

			if (end->op == IR_IF) {
				/* Move condition closer to IF */
				input = end->op2;
				if (input > 0 && _blocks[input] == b && !_xlat[input] && _prev[j] != input) {
					if (input == i) {
						i = _next[i];
						insn = &ctx->ir_base[i];
					}
					/* remove "input" */
					_prev[_next[input]] = _prev[input];
					_next[_prev[input]] = _next[input];
					/* insert before "j" */
					_prev[input] = _prev[j];
					_next[input] = j;
					_next[_prev[j]] = input;
					_prev[j] = input;
				}
			}
		}
		while (i != bb->end) {
			ir_ref n, j, *p, input;

restart:
			n = insn->inputs_count;
			for (j = n, p = insn->ops + 1; j > 0; p++, j--) {
				input = *p;
				if (!_xlat[input]) {
					/* input is not scheduled yet */
					if (input > 0) {
						if (_blocks[input] == b) {
							/* "input" should be before "i" to satisfy dependency */
#ifdef IR_DEBUG
							if (ctx->flags & IR_DEBUG_SCHEDULE) {
								fprintf(stderr, "Wrong dependency %d:%d -> %d\n", b, input, i);
							}
#endif
							/* remove "input" */
							_prev[_next[input]] = _prev[input];
							_next[_prev[input]] = _next[input];
							/* insert before "i" */
							_prev[input] = _prev[i];
							_next[input] = i;
							_next[_prev[i]] = input;
							_prev[i] = input;
							/* restart from "input" */
							i = input;
							insn = &ctx->ir_base[i];
							goto restart;
						}
					} else if (input < IR_TRUE) {
						consts_count += ir_count_constant(_xlat, input);
					}
				}
			}
			_xlat[i] = insns_count;
			insns_count += ir_insn_inputs_to_len(n);
			i = _next[i];
			insn = &ctx->ir_base[i];
		}
		/* Schedule BB end */
		_xlat[i] = bb->end = insns_count;
		insns_count++;
		if (IR_INPUT_EDGES_COUNT(ir_op_flags[insn->op]) == 2) {
			if (insn->op2 < IR_TRUE) {
				consts_count += ir_count_constant(_xlat, insn->op2);
			}
		}
	}

#ifdef IR_DEBUG
	if (ctx->flags & IR_DEBUG_SCHEDULE) {
		fprintf(stderr, "After Schedule\n");
		for (i = 1; i != 0; i = _next[i]) {
			fprintf(stderr, "%d -> %d\n", i, _blocks[i]);
		}
	}
#endif

#if 1
	/* Check if scheduling didn't make any modifications */
	if (consts_count == ctx->consts_count && insns_count == ctx->insns_count) {
		bool changed = 0;

		for (i = 1; i != 0; i = _next[i]) {
			if (_xlat[i] != i) {
				changed = 1;
				break;
			}
		}
		if (!changed) {
			_xlat -= ctx->consts_count;
			ir_mem_free(_xlat);
			ir_mem_free(_next);

			ctx->prev_ref = _prev;
			ctx->flags2 |= IR_LINEAR;
			ir_truncate(ctx);

			return 1;
		}
	}
#endif

	ir_mem_free(_prev);

	ir_init(&new_ctx, ctx->flags, consts_count, insns_count);
	new_ctx.insns_count = insns_count;
	new_ctx.flags2 = ctx->flags2;
	new_ctx.ret_type = ctx->ret_type;
	new_ctx.mflags = ctx->mflags;
	new_ctx.spill_base = ctx->spill_base;
	new_ctx.fixed_stack_red_zone = ctx->fixed_stack_red_zone;
	new_ctx.fixed_stack_frame_size = ctx->fixed_stack_frame_size;
	new_ctx.fixed_call_stack_size = ctx->fixed_call_stack_size;
	new_ctx.fixed_regset = ctx->fixed_regset;
	new_ctx.fixed_save_regset = ctx->fixed_save_regset;
	new_ctx.entries_count = ctx->entries_count;
#if defined(IR_TARGET_AARCH64)
	new_ctx.deoptimization_exits = ctx->deoptimization_exits;
	new_ctx.get_exit_addr = ctx->get_exit_addr;
	new_ctx.get_veneer = ctx->get_veneer;
	new_ctx.set_veneer = ctx->set_veneer;
#endif
	new_ctx.loader = ctx->loader;

	/* Copy constants */
	if (consts_count == ctx->consts_count) {
		new_ctx.consts_count = consts_count;
		ref = 1 - consts_count;
		insn = &ctx->ir_base[ref];
		new_insn = &new_ctx.ir_base[ref];

		memcpy(new_insn, insn, sizeof(ir_insn) * (IR_TRUE - ref));
		if (ctx->strtab.data) {
			while (ref != IR_TRUE) {
				if (new_insn->op == IR_FUNC_ADDR) {
					if (new_insn->proto) {
						size_t len;
						const char *proto = ir_get_strl(ctx, new_insn->proto, &len);
						new_insn->proto = ir_strl(&new_ctx, proto, len);
					}
				} else if (new_insn->op == IR_FUNC) {
					new_insn->val.u64 = ir_str(&new_ctx, ir_get_str(ctx, new_insn->val.name));
					if (new_insn->proto) {
						size_t len;
						const char *proto = ir_get_strl(ctx, new_insn->proto, &len);
						new_insn->proto = ir_strl(&new_ctx, proto, len);
					}
				} else if (new_insn->op == IR_SYM || new_insn->op == IR_STR) {
					new_insn->val.u64 = ir_str(&new_ctx, ir_get_str(ctx, new_insn->val.name));
				}
				new_insn++;
				ref++;
			}
		}
	} else {
		new_ref = -new_ctx.consts_count;
		new_insn = &new_ctx.ir_base[new_ref];
		for (ref = IR_TRUE - 1, insn = &ctx->ir_base[ref]; ref > -ctx->consts_count; insn--, ref--) {
			if (!_xlat[ref]) {
				continue;
			}
			new_insn->optx = insn->optx;
			new_insn->prev_const = 0;
			if (insn->op == IR_FUNC_ADDR) {
				new_insn->val.u64 = insn->val.u64;
				if (insn->proto) {
					size_t len;
					const char *proto = ir_get_strl(ctx, insn->proto, &len);
					new_insn->proto = ir_strl(&new_ctx, proto, len);
				} else {
					new_insn->proto = 0;
				}
			} else if (insn->op == IR_FUNC) {
				new_insn->val.u64 = ir_str(&new_ctx, ir_get_str(ctx, insn->val.name));
				if (insn->proto) {
					size_t len;
					const char *proto = ir_get_strl(ctx, insn->proto, &len);
					new_insn->proto = ir_strl(&new_ctx, proto, len);
				} else {
					new_insn->proto = 0;
				}
			} else if (insn->op == IR_SYM || insn->op == IR_STR) {
				new_insn->val.u64 = ir_str(&new_ctx, ir_get_str(ctx, insn->val.name));
			} else {
				new_insn->val.u64 = insn->val.u64;
			}
			_xlat[ref] = new_ref;
			new_ref--;
			new_insn--;
		}
		new_ctx.consts_count = -new_ref;
	}

	new_ctx.cfg_map = ir_mem_calloc(ctx->insns_count, sizeof(uint32_t));
	new_ctx.prev_ref = _prev = ir_mem_malloc(insns_count * sizeof(ir_ref));
	new_ctx.use_lists = lists = ir_mem_malloc(insns_count * sizeof(ir_use_list));
	new_ctx.use_edges = edges = ir_mem_malloc(ctx->use_edges_count * sizeof(ir_ref));

	/* Copy instructions, use lists and use edges */
	prev_ref = 0;
	use_edges_count = 0;
	for (i = 1; i != 0; i = _next[i]) {
		new_ref = _xlat[i];
		new_ctx.cfg_map[new_ref] = _blocks[i];
		_prev[new_ref] = prev_ref;
		prev_ref = new_ref;

		use_list = &ctx->use_lists[i];
		n = use_list->count;
		k = 0;
		if (n == 1) {
			ref = ctx->use_edges[use_list->refs];
			if (_xlat[ref]) {
				*edges = _xlat[ref];
				edges++;
				k = 1;
			}
		} else {
			p = &ctx->use_edges[use_list->refs];
			while (n--) {
				ref = *p;
				if (_xlat[ref]) {
					*edges = _xlat[ref];
					edges++;
					k++;
				}
				p++;
			}
		}
		new_list = &lists[new_ref];
		new_list->refs = use_edges_count;
		use_edges_count += k;
		new_list->count = k;

		insn = &ctx->ir_base[i];
		new_insn = &new_ctx.ir_base[new_ref];

		new_insn->optx = insn->optx;
		n = new_insn->inputs_count;
		switch (n) {
			case 0:
				new_insn->op1 = insn->op1;
				new_insn->op2 = insn->op2;
				new_insn->op3 = insn->op3;
				break;
			case 1:
				new_insn->op1 = _xlat[insn->op1];
				if (new_insn->op == IR_PARAM || insn->op == IR_VAR) {
					new_insn->op2 = ir_str(&new_ctx, ir_get_str(ctx, insn->op2));
				} else if (new_insn->op == IR_PROTO) {
					size_t len;
					const char *proto = ir_get_strl(ctx, insn->op2, &len);
					new_insn->op2 = ir_strl(&new_ctx, proto, len);
				} else {
					new_insn->op2 = insn->op2;
				}
				new_insn->op3 = insn->op3;
				break;
			case 2:
				new_insn->op1 = _xlat[insn->op1];
				new_insn->op2 = _xlat[insn->op2];
				new_insn->op3 = insn->op3;
#if IR_SCHEDULE_SWAP_OPS
				/* Swap operands according to folding rules */
				if (new_insn->op1 < new_insn->op2) {
					switch (new_insn->op) {
						case IR_EQ:
						case IR_NE:
						case IR_ADD:
						case IR_MUL:
						case IR_ADD_OV:
						case IR_MUL_OV:
						case IR_OR:
						case IR_AND:
						case IR_XOR:
						case IR_MIN:
						case IR_MAX:
							SWAP_REFS(new_insn->op1, new_insn->op2);
							break;
						case IR_LT:
						case IR_GE:
						case IR_LE:
						case IR_GT:
						case IR_ULT:
						case IR_UGE:
						case IR_ULE:
						case IR_UGT:
							SWAP_REFS(new_insn->op1, new_insn->op2);
							new_insn->op ^= 3; /* [U]LT <-> [U]GT, [U]LE <-> [U]GE */
							break;
					}
				}
#endif
				break;
			case 3:
				new_insn->op1 = _xlat[insn->op1];
				new_insn->op2 = _xlat[insn->op2];
				new_insn->op3 = _xlat[insn->op3];
				break;
			default:
				for (j = n, p = insn->ops + 1, q = new_insn->ops + 1; j > 0; p++, q++, j--) {
					*q = _xlat[*p];
				}
				break;
		}
	}

	/* Update list of terminators (IR_OPND_CONTROL_REF) */
	insn = &new_ctx.ir_base[1];
	ref = insn->op1;
	if (ref) {
		insn->op1 = ref = _xlat[ref];
		while (1) {
			insn = &new_ctx.ir_base[ref];
			ref = insn->op3;
			if (!ref) {
				break;
			}
			insn->op3 = ref = _xlat[ref];
		}
	}

	IR_ASSERT(ctx->use_edges_count >= use_edges_count);
	new_ctx.use_edges_count = use_edges_count;
	new_ctx.use_edges = ir_mem_realloc(new_ctx.use_edges, use_edges_count * sizeof(ir_ref));

	if (ctx->binding) {
		ir_xlat_binding(ctx, _xlat);
		new_ctx.binding = ctx->binding;
		ctx->binding = NULL;
	}

	_xlat -= ctx->consts_count;
	ir_mem_free(_xlat);

	new_ctx.cfg_blocks_count = ctx->cfg_blocks_count;
	new_ctx.cfg_edges_count = ctx->cfg_edges_count;
	new_ctx.cfg_blocks = ctx->cfg_blocks;
	new_ctx.cfg_edges = ctx->cfg_edges;
	ctx->cfg_blocks = NULL;
	ctx->cfg_edges = NULL;

	ir_free(ctx);
	IR_ASSERT(new_ctx.consts_count == new_ctx.consts_limit);
	IR_ASSERT(new_ctx.insns_count == new_ctx.insns_limit);
	memcpy(ctx, &new_ctx, sizeof(ir_ctx));
	ctx->flags2 |= IR_LINEAR;

	ir_mem_free(_next);

	return 1;
}

void ir_build_prev_refs(ir_ctx *ctx)
{
	uint32_t b;
	ir_block *bb;
	ir_ref i, n, prev;
	ir_insn *insn;

	ctx->prev_ref = ir_mem_malloc(ctx->insns_count * sizeof(ir_ref));
	prev = 0;
	for (b = 1, bb = ctx->cfg_blocks + b; b <= ctx->cfg_blocks_count; b++, bb++) {
		IR_ASSERT(!(bb->flags & IR_BB_UNREACHABLE));
		for (i = bb->start, insn = ctx->ir_base + i; i < bb->end;) {
			ctx->prev_ref[i] = prev;
			n = ir_insn_len(insn);
			prev = i;
			i += n;
			insn += n;
		}
		ctx->prev_ref[i] = prev;
	}
}