/* * Copyright (C) 2021 Valve Corporation * Copyright (C) 2014 Rob Clark * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include "ir3_ra.h" #include "util/rb_tree.h" #include "util/u_math.h" #include "ir3_shader.h" /* This file implements an SSA-based register allocator. Unlike other * SSA-based allocators, it handles vector split/collect "smartly," meaning * that multiple values may share the same register interval. From the * perspective of the allocator itself, only the top-level intervals matter, * and the allocator is only concerned with allocating top-level intervals, * which may mean moving other top-level intervals around. Other intervals, * like the destination of a split instruction or the source of a collect * instruction, are "locked" to their parent interval. The details of this are * mostly handled by ir3_merge_regs and ir3_reg_ctx. * * We currently don't do any backtracking, but we do use the merge sets as a * form of affinity to try to avoid moves from phis/splits/collects. Each * merge set is what a more "classic" graph-coloring or live-range based * allocator would consider a single register, but here we use it as merely a * hint, except when multiple overlapping values are live at the same time. * Each merge set has a "preferred" register, and we try to honor that when * allocating values in the merge set. */ /* ir3_reg_ctx implementation. */ static int ir3_reg_interval_cmp(const struct rb_node *node, const void *data) { unsigned reg = *(const unsigned *)data; const struct ir3_reg_interval *interval = ir3_rb_node_to_interval_const(node); if (interval->reg->interval_start > reg) return -1; else if (interval->reg->interval_end <= reg) return 1; else return 0; } static struct ir3_reg_interval * ir3_reg_interval_search(struct rb_tree *tree, unsigned offset) { struct rb_node *node = rb_tree_search(tree, &offset, ir3_reg_interval_cmp); return node ? ir3_rb_node_to_interval(node) : NULL; } static struct ir3_reg_interval * ir3_reg_interval_search_sloppy(struct rb_tree *tree, unsigned offset) { struct rb_node *node = rb_tree_search_sloppy(tree, &offset, ir3_reg_interval_cmp); return node ? ir3_rb_node_to_interval(node) : NULL; } /* Get the interval covering the reg, or the closest to the right if it * doesn't exist. */ static struct ir3_reg_interval * ir3_reg_interval_search_right(struct rb_tree *tree, unsigned offset) { struct ir3_reg_interval *interval = ir3_reg_interval_search_sloppy(tree, offset); if (!interval) { return NULL; } else if (interval->reg->interval_end > offset) { return interval; } else { /* There is no interval covering reg, and ra_file_search_sloppy() * returned the closest range to the left, so the next interval to the * right should be the closest to the right. */ return ir3_reg_interval_next_or_null(interval); } } static int ir3_reg_interval_insert_cmp(const struct rb_node *_a, const struct rb_node *_b) { const struct ir3_reg_interval *a = ir3_rb_node_to_interval_const(_a); const struct ir3_reg_interval *b = ir3_rb_node_to_interval_const(_b); return b->reg->interval_start - a->reg->interval_start; } static void interval_insert(struct ir3_reg_ctx *ctx, struct rb_tree *tree, struct ir3_reg_interval *interval) { struct ir3_reg_interval *right = ir3_reg_interval_search_right(tree, interval->reg->interval_start); if (right && right->reg->interval_start < interval->reg->interval_end) { /* We disallow trees where different members have different half-ness. * This means that we can't treat bitcasts as copies like normal * split/collect, so something like this would require an extra copy * in mergedregs mode, and count as 4 half-units of register pressure * instead of 2: * * f16vec2 foo = unpackFloat2x16(bar) * ... = foo.x * ... = bar * * However, relaxing this rule would open a huge can of worms. What * happens when there's a vector of 16 things, and the fifth element * has been bitcasted as a half-reg? Would that element alone have to * be small enough to be used as a half-reg source? Let's keep that * can of worms firmly shut for now. */ assert((interval->reg->flags & IR3_REG_HALF) == (right->reg->flags & IR3_REG_HALF)); if (right->reg->interval_end <= interval->reg->interval_end && right->reg->interval_start >= interval->reg->interval_start) { /* Check if we're inserting something that's already inserted */ assert(interval != right); /* "right" is contained in "interval" and must become a child of * it. There may be further children too. */ for (struct ir3_reg_interval *next = ir3_reg_interval_next(right); right && right->reg->interval_start < interval->reg->interval_end; right = next, next = ir3_reg_interval_next_or_null(next)) { /* "right" must be contained in "interval." */ assert(right->reg->interval_end <= interval->reg->interval_end); assert((interval->reg->flags & IR3_REG_HALF) == (right->reg->flags & IR3_REG_HALF)); if (!right->parent) ctx->interval_delete(ctx, right); right->parent = interval; rb_tree_remove(tree, &right->node); rb_tree_insert(&interval->children, &right->node, ir3_reg_interval_insert_cmp); } } else { /* "right" must contain "interval," since intervals must form a * tree. */ assert(right->reg->interval_start <= interval->reg->interval_start); interval->parent = right; interval_insert(ctx, &right->children, interval); return; } } if (!interval->parent) ctx->interval_add(ctx, interval); rb_tree_insert(tree, &interval->node, ir3_reg_interval_insert_cmp); interval->inserted = true; } void ir3_reg_interval_insert(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *interval) { rb_tree_init(&interval->children); interval->parent = NULL; interval_insert(ctx, &ctx->intervals, interval); } /* Call after ir3_reg_interval_remove_temp() to reinsert the interval */ static void ir3_reg_interval_reinsert(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *interval) { interval->parent = NULL; interval_insert(ctx, &ctx->intervals, interval); } void ir3_reg_interval_remove(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *interval) { if (interval->parent) { rb_tree_remove(&interval->parent->children, &interval->node); } else { ctx->interval_delete(ctx, interval); rb_tree_remove(&ctx->intervals, &interval->node); } rb_tree_foreach_safe (struct ir3_reg_interval, child, &interval->children, node) { rb_tree_remove(&interval->children, &child->node); child->parent = interval->parent; if (interval->parent) { rb_tree_insert(&child->parent->children, &child->node, ir3_reg_interval_insert_cmp); } else { ctx->interval_readd(ctx, interval, child); rb_tree_insert(&ctx->intervals, &child->node, ir3_reg_interval_insert_cmp); } } interval->inserted = false; } static void _mark_free(struct ir3_reg_interval *interval) { interval->inserted = false; rb_tree_foreach (struct ir3_reg_interval, child, &interval->children, node) { _mark_free(child); } } /* Remove an interval and all its children from the tree. */ void ir3_reg_interval_remove_all(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *interval) { assert(!interval->parent); ctx->interval_delete(ctx, interval); rb_tree_remove(&ctx->intervals, &interval->node); _mark_free(interval); } /* Used when popping an interval to be shuffled around. Don't disturb children * so that it can be later reinserted. */ static void ir3_reg_interval_remove_temp(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *interval) { assert(!interval->parent); ctx->interval_delete(ctx, interval); rb_tree_remove(&ctx->intervals, &interval->node); } static void interval_dump(struct log_stream *stream, struct ir3_reg_interval *interval, unsigned indent) { for (unsigned i = 0; i < indent; i++) mesa_log_stream_printf(stream, "\t"); mesa_log_stream_printf(stream, "reg %u start %u\n", interval->reg->name, interval->reg->interval_start); rb_tree_foreach (struct ir3_reg_interval, child, &interval->children, node) { interval_dump(stream, child, indent + 1); } for (unsigned i = 0; i < indent; i++) mesa_log_stream_printf(stream, "\t"); mesa_log_stream_printf(stream, "reg %u end %u\n", interval->reg->name, interval->reg->interval_end); } void ir3_reg_interval_dump(struct log_stream *stream, struct ir3_reg_interval *interval) { interval_dump(stream, interval, 0); } /* These are the core datastructures used by the register allocator. First * ra_interval and ra_file, which are used for intra-block tracking and use * the ir3_reg_ctx infrastructure: */ struct ra_interval { struct ir3_reg_interval interval; struct rb_node physreg_node; physreg_t physreg_start, physreg_end; /* True if this is a source of the current instruction which is entirely * killed. This means we can allocate the dest over it, but we can't break * it up. */ bool is_killed; /* True if this interval cannot be moved from its position. This is only * used for precolored inputs to ensure that other inputs don't get * allocated on top of them. */ bool frozen; }; struct ra_file { struct ir3_reg_ctx reg_ctx; BITSET_DECLARE(available, RA_MAX_FILE_SIZE); BITSET_DECLARE(available_to_evict, RA_MAX_FILE_SIZE); struct rb_tree physreg_intervals; unsigned size; unsigned start; }; /* State for inter-block tracking. When we split a live range to make space * for a vector, we may need to insert fixup code when a block has multiple * predecessors that have moved the same live value to different registers. * This keeps track of state required to do that. */ struct ra_block_state { /* Map of defining ir3_register -> physreg it was allocated to at the end * of the block. */ struct hash_table *renames; /* For loops, we need to process a block before all its predecessors have * been processed. In particular, we need to pick registers for values * without knowing if all the predecessors have been renamed. This keeps * track of the registers we chose so that when we visit the back-edge we * can move them appropriately. If all predecessors have been visited * before this block is visited then we don't need to fill this out. This * is a map from ir3_register -> physreg. */ struct hash_table *entry_regs; /* True if the block has been visited and "renames" is complete. */ bool visited; }; struct ra_parallel_copy { struct ra_interval *interval; physreg_t src; }; /* The main context: */ struct ra_ctx { /* r0.x - r47.w. On a6xx with merged-regs, hr0.x-hr47.w go into the bottom * half of this file too. */ struct ra_file full; /* hr0.x - hr63.w, only used without merged-regs. */ struct ra_file half; /* Shared regs. */ struct ra_file shared; struct ir3_liveness *live; struct ir3_block *block; const struct ir3_compiler *compiler; gl_shader_stage stage; /* Pending moves of top-level intervals that will be emitted once we're * finished: */ DECLARE_ARRAY(struct ra_parallel_copy, parallel_copies); struct ra_interval *intervals; struct ra_block_state *blocks; bool merged_regs; }; #define foreach_interval(interval, file) \ rb_tree_foreach (struct ra_interval, interval, &(file)->physreg_intervals, \ physreg_node) #define foreach_interval_rev(interval, file) \ rb_tree_foreach (struct ra_interval, interval, &(file)->physreg_intervals, \ physreg_node) #define foreach_interval_safe(interval, file) \ rb_tree_foreach_safe (struct ra_interval, interval, \ &(file)->physreg_intervals, physreg_node) #define foreach_interval_rev_safe(interval, file) \ rb_tree_foreach_rev_safe(struct ra_interval, interval, \ &(file)->physreg_intervals, physreg_node) static struct ra_interval * rb_node_to_interval(struct rb_node *node) { return rb_node_data(struct ra_interval, node, physreg_node); } static const struct ra_interval * rb_node_to_interval_const(const struct rb_node *node) { return rb_node_data(struct ra_interval, node, physreg_node); } static struct ra_interval * ra_interval_next(struct ra_interval *interval) { struct rb_node *next = rb_node_next(&interval->physreg_node); return next ? rb_node_to_interval(next) : NULL; } static struct ra_interval * ra_interval_next_or_null(struct ra_interval *interval) { return interval ? ra_interval_next(interval) : NULL; } static int ra_interval_cmp(const struct rb_node *node, const void *data) { physreg_t reg = *(const physreg_t *)data; const struct ra_interval *interval = rb_node_to_interval_const(node); if (interval->physreg_start > reg) return -1; else if (interval->physreg_end <= reg) return 1; else return 0; } static struct ra_interval * ra_interval_search_sloppy(struct rb_tree *tree, physreg_t reg) { struct rb_node *node = rb_tree_search_sloppy(tree, ®, ra_interval_cmp); return node ? rb_node_to_interval(node) : NULL; } /* Get the interval covering the reg, or the closest to the right if it * doesn't exist. */ static struct ra_interval * ra_interval_search_right(struct rb_tree *tree, physreg_t reg) { struct ra_interval *interval = ra_interval_search_sloppy(tree, reg); if (!interval) { return NULL; } else if (interval->physreg_end > reg) { return interval; } else { /* There is no interval covering reg, and ra_file_search_sloppy() * returned the closest range to the left, so the next interval to the * right should be the closest to the right. */ return ra_interval_next_or_null(interval); } } static struct ra_interval * ra_file_search_right(struct ra_file *file, physreg_t reg) { return ra_interval_search_right(&file->physreg_intervals, reg); } static int ra_interval_insert_cmp(const struct rb_node *_a, const struct rb_node *_b) { const struct ra_interval *a = rb_node_to_interval_const(_a); const struct ra_interval *b = rb_node_to_interval_const(_b); return b->physreg_start - a->physreg_start; } static struct ra_interval * ir3_reg_interval_to_ra_interval(struct ir3_reg_interval *interval) { return rb_node_data(struct ra_interval, interval, interval); } static struct ra_file * ir3_reg_ctx_to_file(struct ir3_reg_ctx *ctx) { return rb_node_data(struct ra_file, ctx, reg_ctx); } static void interval_add(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *_interval) { struct ra_interval *interval = ir3_reg_interval_to_ra_interval(_interval); struct ra_file *file = ir3_reg_ctx_to_file(ctx); /* We can assume in this case that physreg_start/physreg_end is already * initialized. */ for (physreg_t i = interval->physreg_start; i < interval->physreg_end; i++) { BITSET_CLEAR(file->available, i); BITSET_CLEAR(file->available_to_evict, i); } rb_tree_insert(&file->physreg_intervals, &interval->physreg_node, ra_interval_insert_cmp); } static void interval_delete(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *_interval) { struct ra_interval *interval = ir3_reg_interval_to_ra_interval(_interval); struct ra_file *file = ir3_reg_ctx_to_file(ctx); for (physreg_t i = interval->physreg_start; i < interval->physreg_end; i++) { BITSET_SET(file->available, i); BITSET_SET(file->available_to_evict, i); } rb_tree_remove(&file->physreg_intervals, &interval->physreg_node); } static void interval_readd(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *_parent, struct ir3_reg_interval *_child) { struct ra_interval *parent = ir3_reg_interval_to_ra_interval(_parent); struct ra_interval *child = ir3_reg_interval_to_ra_interval(_child); child->physreg_start = parent->physreg_start + (child->interval.reg->interval_start - parent->interval.reg->interval_start); child->physreg_end = child->physreg_start + (child->interval.reg->interval_end - child->interval.reg->interval_start); interval_add(ctx, _child); } static void ra_file_init(struct ra_file *file) { for (unsigned i = 0; i < file->size; i++) { BITSET_SET(file->available, i); BITSET_SET(file->available_to_evict, i); } rb_tree_init(&file->reg_ctx.intervals); rb_tree_init(&file->physreg_intervals); file->reg_ctx.interval_add = interval_add; file->reg_ctx.interval_delete = interval_delete; file->reg_ctx.interval_readd = interval_readd; } static void ra_file_insert(struct ra_file *file, struct ra_interval *interval) { assert(interval->physreg_start < interval->physreg_end); assert(interval->physreg_end <= file->size); if (interval->interval.reg->flags & IR3_REG_HALF) assert(interval->physreg_end <= RA_HALF_SIZE); ir3_reg_interval_insert(&file->reg_ctx, &interval->interval); } static void ra_file_remove(struct ra_file *file, struct ra_interval *interval) { ir3_reg_interval_remove(&file->reg_ctx, &interval->interval); } static void ra_file_mark_killed(struct ra_file *file, struct ra_interval *interval) { assert(!interval->interval.parent); for (physreg_t i = interval->physreg_start; i < interval->physreg_end; i++) { BITSET_SET(file->available, i); } interval->is_killed = true; } static void ra_file_unmark_killed(struct ra_file *file, struct ra_interval *interval) { assert(!interval->interval.parent); for (physreg_t i = interval->physreg_start; i < interval->physreg_end; i++) { BITSET_CLEAR(file->available, i); } interval->is_killed = false; } static physreg_t ra_interval_get_physreg(const struct ra_interval *interval) { unsigned child_start = interval->interval.reg->interval_start; while (interval->interval.parent) { interval = ir3_reg_interval_to_ra_interval(interval->interval.parent); } return interval->physreg_start + (child_start - interval->interval.reg->interval_start); } static unsigned ra_interval_get_num(const struct ra_interval *interval) { return ra_physreg_to_num(ra_interval_get_physreg(interval), interval->interval.reg->flags); } static void ra_interval_init(struct ra_interval *interval, struct ir3_register *reg) { ir3_reg_interval_init(&interval->interval, reg); interval->is_killed = false; interval->frozen = false; } static void ra_interval_dump(struct log_stream *stream, struct ra_interval *interval) { mesa_log_stream_printf(stream, "physreg %u ", interval->physreg_start); ir3_reg_interval_dump(stream, &interval->interval); } static void ra_file_dump(struct log_stream *stream, struct ra_file *file) { rb_tree_foreach (struct ra_interval, interval, &file->physreg_intervals, physreg_node) { ra_interval_dump(stream, interval); } unsigned start, end; mesa_log_stream_printf(stream, "available:\n"); BITSET_FOREACH_RANGE (start, end, file->available, file->size) { mesa_log_stream_printf(stream, "%u-%u ", start, end); } mesa_log_stream_printf(stream, "\n"); mesa_log_stream_printf(stream, "available to evict:\n"); BITSET_FOREACH_RANGE (start, end, file->available_to_evict, file->size) { mesa_log_stream_printf(stream, "%u-%u ", start, end); } mesa_log_stream_printf(stream, "\n"); mesa_log_stream_printf(stream, "start: %u\n", file->start); } static void ra_ctx_dump(struct ra_ctx *ctx) { struct log_stream *stream = mesa_log_streami(); mesa_log_stream_printf(stream, "full:\n"); ra_file_dump(stream, &ctx->full); mesa_log_stream_printf(stream, "half:\n"); ra_file_dump(stream, &ctx->half); mesa_log_stream_printf(stream, "shared:"); ra_file_dump(stream, &ctx->shared); mesa_log_stream_destroy(stream); } static unsigned reg_file_size(struct ra_file *file, struct ir3_register *reg) { /* Half-regs can only take up the first half of the combined regfile */ if (reg->flags & IR3_REG_HALF) return MIN2(file->size, RA_HALF_SIZE); else return file->size; } /* ra_pop_interval/ra_push_interval provide an API to shuffle around multiple * top-level intervals at once. Pop multiple intervals, then push them back in * any order. */ struct ra_removed_interval { struct ra_interval *interval; unsigned size; }; static struct ra_removed_interval ra_pop_interval(struct ra_ctx *ctx, struct ra_file *file, struct ra_interval *interval) { assert(!interval->interval.parent); /* Check if we've already moved this reg before */ unsigned pcopy_index; for (pcopy_index = 0; pcopy_index < ctx->parallel_copies_count; pcopy_index++) { if (ctx->parallel_copies[pcopy_index].interval == interval) break; } if (pcopy_index == ctx->parallel_copies_count) { array_insert(ctx, ctx->parallel_copies, (struct ra_parallel_copy){ .interval = interval, .src = interval->physreg_start, }); } ir3_reg_interval_remove_temp(&file->reg_ctx, &interval->interval); return (struct ra_removed_interval){ .interval = interval, .size = interval->physreg_end - interval->physreg_start, }; } static void ra_push_interval(struct ra_ctx *ctx, struct ra_file *file, const struct ra_removed_interval *removed, physreg_t dst) { struct ra_interval *interval = removed->interval; interval->physreg_start = dst; interval->physreg_end = dst + removed->size; assert(interval->physreg_end <= file->size); if (interval->interval.reg->flags & IR3_REG_HALF) assert(interval->physreg_end <= RA_HALF_SIZE); ir3_reg_interval_reinsert(&file->reg_ctx, &interval->interval); } /* Pick up the interval and place it at "dst". */ static void ra_move_interval(struct ra_ctx *ctx, struct ra_file *file, struct ra_interval *interval, physreg_t dst) { struct ra_removed_interval temp = ra_pop_interval(ctx, file, interval); ra_push_interval(ctx, file, &temp, dst); } static struct ra_file * ra_get_file(struct ra_ctx *ctx, struct ir3_register *reg) { if (reg->flags & IR3_REG_SHARED) return &ctx->shared; else if (ctx->merged_regs || !(reg->flags & IR3_REG_HALF)) return &ctx->full; else return &ctx->half; } /* Returns true if the proposed spot for "dst" or a killed source overlaps a * destination that's been allocated. */ static bool check_dst_overlap(struct ra_ctx *ctx, struct ra_file *file, struct ir3_register *dst, physreg_t start, physreg_t end) { struct ir3_instruction *instr = dst->instr; ra_foreach_dst (other_dst, instr) { /* We assume only destinations before the current one have been allocated. */ if (other_dst == dst) break; if (ra_get_file(ctx, other_dst) != file) continue; struct ra_interval *other_interval = &ctx->intervals[other_dst->name]; assert(!other_interval->interval.parent); physreg_t other_start = other_interval->physreg_start; physreg_t other_end = other_interval->physreg_end; if (other_end > start && end > other_start) return true; } return false; } /* True if the destination is "early-clobber," meaning that it cannot be * allocated over killed sources. Some destinations always require it, but it * also is implicitly true for tied destinations whose source is live-through. * If the source is killed, then we skip allocating a register for the * destination altogether so we don't need to worry about that case here. */ static bool is_early_clobber(struct ir3_register *reg) { return (reg->flags & IR3_REG_EARLY_CLOBBER) || reg->tied; } static bool get_reg_specified(struct ra_ctx *ctx, struct ra_file *file, struct ir3_register *reg, physreg_t physreg, bool is_source) { for (unsigned i = 0; i < reg_size(reg); i++) { if (!BITSET_TEST(is_early_clobber(reg) || is_source ? file->available_to_evict : file->available, physreg + i)) return false; } if (!is_source && check_dst_overlap(ctx, file, reg, physreg, physreg + reg_size(reg))) return false; return true; } /* Try to evict any registers conflicting with the proposed spot "physreg" for * "reg". That is, move them to other places so that we can allocate "physreg" * here. */ static bool try_evict_regs(struct ra_ctx *ctx, struct ra_file *file, struct ir3_register *reg, physreg_t physreg, unsigned *_eviction_count, bool is_source, bool speculative) { BITSET_DECLARE(available_to_evict, RA_MAX_FILE_SIZE); memcpy(available_to_evict, file->available_to_evict, sizeof(available_to_evict)); BITSET_DECLARE(available, RA_MAX_FILE_SIZE); memcpy(available, file->available, sizeof(available)); for (unsigned i = 0; i < reg_size(reg); i++) { BITSET_CLEAR(available_to_evict, physreg + i); BITSET_CLEAR(available, physreg + i); } unsigned eviction_count = 0; /* Iterate over each range conflicting with physreg */ for (struct ra_interval *conflicting = ra_file_search_right(file, physreg), *next = ra_interval_next_or_null(conflicting); conflicting != NULL && conflicting->physreg_start < physreg + reg_size(reg); conflicting = next, next = ra_interval_next_or_null(next)) { if (!is_early_clobber(reg) && !is_source && conflicting->is_killed) continue; if (conflicting->frozen) { assert(speculative); return false; } unsigned conflicting_file_size = reg_file_size(file, conflicting->interval.reg); unsigned avail_start, avail_end; bool evicted = false; BITSET_FOREACH_RANGE (avail_start, avail_end, available_to_evict, conflicting_file_size) { unsigned size = avail_end - avail_start; /* non-half registers must be aligned */ if (!(conflicting->interval.reg->flags & IR3_REG_HALF) && avail_start % 2 == 1) { avail_start++; size--; } unsigned conflicting_size = conflicting->physreg_end - conflicting->physreg_start; if (size >= conflicting_size && !check_dst_overlap(ctx, file, reg, avail_start, avail_start + conflicting_size)) { for (unsigned i = 0; i < conflicting->physreg_end - conflicting->physreg_start; i++) BITSET_CLEAR(available_to_evict, avail_start + i); eviction_count += conflicting->physreg_end - conflicting->physreg_start; if (!speculative) ra_move_interval(ctx, file, conflicting, avail_start); evicted = true; break; } } if (evicted) continue; /* If we couldn't evict this range, we may be able to swap it with a * killed range to acheive the same effect. */ foreach_interval (killed, file) { if (!killed->is_killed) continue; if (killed->physreg_end - killed->physreg_start != conflicting->physreg_end - conflicting->physreg_start) continue; if (killed->physreg_end > conflicting_file_size || conflicting->physreg_end > reg_file_size(file, killed->interval.reg)) continue; /* We can't swap the killed range if it partially/fully overlaps the * space we're trying to allocate or (in speculative mode) if it's * already been swapped and will overlap when we actually evict. */ bool killed_available = true; for (unsigned i = killed->physreg_start; i < killed->physreg_end; i++) { if (!BITSET_TEST(available, i)) { killed_available = false; break; } } if (!killed_available) continue; if (check_dst_overlap(ctx, file, reg, killed->physreg_start, killed->physreg_end)) continue; /* Check for alignment if one is a full reg */ if ((!(killed->interval.reg->flags & IR3_REG_HALF) || !(conflicting->interval.reg->flags & IR3_REG_HALF)) && (killed->physreg_start % 2 != 0 || conflicting->physreg_start % 2 != 0)) continue; for (unsigned i = killed->physreg_start; i < killed->physreg_end; i++) { BITSET_CLEAR(available, i); } /* Because this will generate swaps instead of moves, multiply the * cost by 2. */ eviction_count += (killed->physreg_end - killed->physreg_start) * 2; if (!speculative) { physreg_t killed_start = killed->physreg_start, conflicting_start = conflicting->physreg_start; struct ra_removed_interval killed_removed = ra_pop_interval(ctx, file, killed); struct ra_removed_interval conflicting_removed = ra_pop_interval(ctx, file, conflicting); ra_push_interval(ctx, file, &killed_removed, conflicting_start); ra_push_interval(ctx, file, &conflicting_removed, killed_start); } evicted = true; break; } if (!evicted) return false; } *_eviction_count = eviction_count; return true; } static int removed_interval_cmp(const void *_i1, const void *_i2) { const struct ra_removed_interval *i1 = _i1; const struct ra_removed_interval *i2 = _i2; /* We sort the registers as follows: * * |------------------------------------------------------------------------------------------| * | | | | | | | * | Half | Half early-clobber | Half | Full | Full early-clobber | Full | * | live-through | destination | killed | killed | destination | live-through | * | | | | | | | * |------------------------------------------------------------------------------------------| * | | * | Destination | * | | * |-----------------| * * Half-registers have to be first so that they stay in the low half of * the register file. Then half and full killed must stay together so that * there's a contiguous range where we can put the register. With this * structure we should be able to accomodate any collection of intervals * such that the total number of half components is within the half limit * and the combined components are within the full limit. */ unsigned i1_align = reg_elem_size(i1->interval->interval.reg); unsigned i2_align = reg_elem_size(i2->interval->interval.reg); if (i1_align > i2_align) return 1; if (i1_align < i2_align) return -1; if (i1_align == 1) { if (i2->interval->is_killed) return -1; if (i1->interval->is_killed) return 1; } else { if (i2->interval->is_killed) return 1; if (i1->interval->is_killed) return -1; } return 0; } static int dsts_cmp(const void *_i1, const void *_i2) { struct ir3_register *i1 = *(struct ir3_register *const *) _i1; struct ir3_register *i2 = *(struct ir3_register *const *) _i2; /* Treat tied destinations as-if they are live-through sources, and normal * destinations as killed sources. */ unsigned i1_align = reg_elem_size(i1); unsigned i2_align = reg_elem_size(i2); if (i1_align > i2_align) return 1; if (i1_align < i2_align) return -1; if (i1_align == 1) { if (!is_early_clobber(i2)) return -1; if (!is_early_clobber(i1)) return 1; } else { if (!is_early_clobber(i2)) return 1; if (!is_early_clobber(i1)) return -1; } return 0; } /* "Compress" all the live intervals so that there is enough space for the * destination register. As there can be gaps when a more-aligned interval * follows a less-aligned interval, this also sorts them to remove such * "padding", which may be required when space is very tight. This isn't * amazing, but should be used only as a last resort in case the register file * is almost full and badly fragmented. * * Return the physreg to use. */ static physreg_t compress_regs_left(struct ra_ctx *ctx, struct ra_file *file, struct ir3_register *reg) { unsigned align = reg_elem_size(reg); DECLARE_ARRAY(struct ra_removed_interval, intervals); intervals_count = intervals_sz = 0; intervals = NULL; DECLARE_ARRAY(struct ir3_register *, dsts); dsts_count = dsts_sz = 0; dsts = NULL; array_insert(ctx, dsts, reg); bool dst_inserted[reg->instr->dsts_count]; unsigned dst_size = reg->tied ? 0 : reg_size(reg); unsigned ec_dst_size = is_early_clobber(reg) ? reg_size(reg) : 0; unsigned half_dst_size = 0, ec_half_dst_size = 0; if (align == 1) { half_dst_size = dst_size; ec_half_dst_size = ec_dst_size; } unsigned removed_size = 0, removed_half_size = 0; unsigned removed_killed_size = 0, removed_killed_half_size = 0; unsigned file_size = align == 1 ? MIN2(file->size, RA_HALF_SIZE) : file->size; physreg_t start_reg = 0; foreach_interval_rev_safe (interval, file) { /* We'll check if we can compact the intervals starting here. */ physreg_t candidate_start = interval->physreg_end; /* Check if there are any other destinations we need to compact. */ ra_foreach_dst_n (other_dst, n, reg->instr) { if (other_dst == reg) break; if (ra_get_file(ctx, other_dst) != file) continue; if (dst_inserted[n]) continue; struct ra_interval *other_interval = &ctx->intervals[other_dst->name]; /* if the destination partially overlaps this interval, we need to * extend candidate_start to the end. */ if (other_interval->physreg_start < candidate_start) { candidate_start = MAX2(candidate_start, other_interval->physreg_end); continue; } dst_inserted[n] = true; /* dst intervals with a tied killed source are considered attached to * that source. Don't actually insert them. This means we have to * update them below if their tied source moves. */ if (other_dst->tied) { struct ra_interval *tied_interval = &ctx->intervals[other_dst->tied->def->name]; if (tied_interval->is_killed) continue; } d("popping destination %u physreg %u\n", other_interval->interval.reg->name, other_interval->physreg_start); array_insert(ctx, dsts, other_dst); unsigned interval_size = reg_size(other_dst); if (is_early_clobber(other_dst)) { ec_dst_size += interval_size; if (other_interval->interval.reg->flags & IR3_REG_HALF) ec_half_dst_size += interval_size; } else { dst_size += interval_size; if (other_interval->interval.reg->flags & IR3_REG_HALF) half_dst_size += interval_size; } } /* Check if we can sort the intervals *after* this one and have enough * space leftover to accomodate all intervals, keeping in mind that killed * sources overlap non-tied destinations. Also check that we have enough * space leftover for half-registers, if we're inserting a half-register * (otherwise we only shift any half-registers down so they should be * safe). */ if (candidate_start + removed_size + ec_dst_size + MAX2(removed_killed_size, dst_size) <= file->size && (align != 1 || candidate_start + removed_half_size + ec_half_dst_size + MAX2(removed_killed_half_size, half_dst_size) <= file_size)) { start_reg = candidate_start; break; } /* We assume that all frozen intervals are at the start and that we * can avoid popping them. */ assert(!interval->frozen); /* Killed sources are different because they go at the end and can * overlap the register we're trying to add. */ unsigned interval_size = interval->physreg_end - interval->physreg_start; if (interval->is_killed) { removed_killed_size += interval_size; if (interval->interval.reg->flags & IR3_REG_HALF) removed_killed_half_size += interval_size; } else { removed_size += interval_size; if (interval->interval.reg->flags & IR3_REG_HALF) removed_half_size += interval_size; } /* Now that we've done the accounting, pop this off */ d("popping interval %u physreg %u%s\n", interval->interval.reg->name, interval->physreg_start, interval->is_killed ? ", killed" : ""); array_insert(ctx, intervals, ra_pop_interval(ctx, file, interval)); } /* TODO: In addition to skipping registers at the beginning that are * well-packed, we should try to skip registers at the end. */ qsort(intervals, intervals_count, sizeof(*intervals), removed_interval_cmp); qsort(dsts, dsts_count, sizeof(*dsts), dsts_cmp); physreg_t live_reg = start_reg; physreg_t dst_reg = (physreg_t)~0; physreg_t ret_reg = (physreg_t)~0; unsigned dst_index = 0; unsigned live_index = 0; /* We have two lists of intervals to process, live intervals and destination * intervals. Process them in the order of the disgram in insert_cmp(). */ while (live_index < intervals_count || dst_index < dsts_count) { bool process_dst; if (live_index == intervals_count) { process_dst = true; } else if (dst_index == dsts_count) { process_dst = false; } else { struct ir3_register *dst = dsts[dst_index]; struct ra_interval *live_interval = intervals[live_index].interval; bool live_half = live_interval->interval.reg->flags & IR3_REG_HALF; bool live_killed = live_interval->is_killed; bool dst_half = dst->flags & IR3_REG_HALF; bool dst_early_clobber = is_early_clobber(dst); if (live_half && !live_killed) { /* far-left of diagram. */ process_dst = false; } else if (dst_half && dst_early_clobber) { /* mid-left of diagram. */ process_dst = true; } else if (!dst_early_clobber) { /* bottom of disagram. */ process_dst = true; } else if (live_killed) { /* middle of diagram. */ process_dst = false; } else if (!dst_half && dst_early_clobber) { /* mid-right of diagram. */ process_dst = true; } else { /* far right of diagram. */ assert(!live_killed && !live_half); process_dst = false; } } struct ir3_register *cur_reg = process_dst ? dsts[dst_index] : intervals[live_index].interval->interval.reg; physreg_t physreg; if (process_dst && !is_early_clobber(cur_reg)) { if (dst_reg == (physreg_t)~0) dst_reg = live_reg; physreg = dst_reg; } else { physreg = live_reg; struct ra_interval *live_interval = intervals[live_index].interval; bool live_killed = live_interval->is_killed; /* If this is live-through and we've processed the destinations, we * need to make sure we take into account any overlapping destinations. */ if (!live_killed && dst_reg != (physreg_t)~0) physreg = MAX2(physreg, dst_reg); } if (!(cur_reg->flags & IR3_REG_HALF)) physreg = ALIGN(physreg, 2); d("pushing reg %u physreg %u\n", cur_reg->name, physreg); unsigned interval_size = reg_size(cur_reg); if (physreg + interval_size > reg_file_size(file, cur_reg)) { d("ran out of room for interval %u!\n", cur_reg->name); unreachable("reg pressure calculation was wrong!"); return 0; } if (process_dst) { if (cur_reg == reg) { ret_reg = physreg; } else { struct ra_interval *interval = &ctx->intervals[cur_reg->name]; interval->physreg_start = physreg; interval->physreg_end = physreg + interval_size; } dst_index++; } else { ra_push_interval(ctx, file, &intervals[live_index], physreg); live_index++; } physreg += interval_size; if (process_dst && !is_early_clobber(cur_reg)) { dst_reg = physreg; } else { live_reg = physreg; } } /* If we shuffled around a tied source that is killed, we may have to update * its corresponding destination since we didn't insert it above. */ ra_foreach_dst (dst, reg->instr) { if (dst == reg) break; struct ir3_register *tied = dst->tied; if (!tied) continue; struct ra_interval *tied_interval = &ctx->intervals[tied->def->name]; if (!tied_interval->is_killed) continue; struct ra_interval *dst_interval = &ctx->intervals[dst->name]; unsigned dst_size = reg_size(dst); dst_interval->physreg_start = ra_interval_get_physreg(tied_interval); dst_interval->physreg_end = dst_interval->physreg_start + dst_size; } return ret_reg; } static void update_affinity(struct ra_file *file, struct ir3_register *reg, physreg_t physreg) { if (!reg->merge_set || reg->merge_set->preferred_reg != (physreg_t)~0) return; if (physreg < reg->merge_set_offset) return; if ((physreg - reg->merge_set_offset + reg->merge_set->size) > file->size) return; reg->merge_set->preferred_reg = physreg - reg->merge_set_offset; } /* Try to find free space for a register without shuffling anything. This uses * a round-robin algorithm to reduce false dependencies. */ static physreg_t find_best_gap(struct ra_ctx *ctx, struct ra_file *file, struct ir3_register *dst, unsigned file_size, unsigned size, unsigned align) { /* This can happen if we create a very large merge set. Just bail out in that * case. */ if (size > file_size) return (physreg_t) ~0; BITSET_WORD *available = is_early_clobber(dst) ? file->available_to_evict : file->available; unsigned start = ALIGN(file->start, align) % (file_size - size + align); unsigned candidate = start; do { bool is_available = true; for (unsigned i = 0; i < size; i++) { if (!BITSET_TEST(available, candidate + i)) { is_available = false; break; } } if (is_available) { is_available = !check_dst_overlap(ctx, file, dst, candidate, candidate + size); } if (is_available) { file->start = (candidate + size) % file_size; return candidate; } candidate += align; if (candidate + size > file_size) candidate = 0; } while (candidate != start); return (physreg_t)~0; } /* This is the main entrypoint for picking a register. Pick a free register * for "reg", shuffling around sources if necessary. In the normal case where * "is_source" is false, this register can overlap with killed sources * (intervals with "is_killed == true"). If "is_source" is true, then * is_killed is ignored and the register returned must not overlap with killed * sources. This must be used for tied registers, because we're actually * allocating the destination and the tied source at the same time. */ static physreg_t get_reg(struct ra_ctx *ctx, struct ra_file *file, struct ir3_register *reg) { unsigned file_size = reg_file_size(file, reg); if (reg->merge_set && reg->merge_set->preferred_reg != (physreg_t)~0) { physreg_t preferred_reg = reg->merge_set->preferred_reg + reg->merge_set_offset; if (preferred_reg < file_size && preferred_reg % reg_elem_size(reg) == 0 && get_reg_specified(ctx, file, reg, preferred_reg, false)) return preferred_reg; } /* If this register is a subset of a merge set which we have not picked a * register for, first try to allocate enough space for the entire merge * set. */ unsigned size = reg_size(reg); if (reg->merge_set && reg->merge_set->preferred_reg == (physreg_t)~0 && size < reg->merge_set->size) { physreg_t best_reg = find_best_gap(ctx, file, reg, file_size, reg->merge_set->size, reg->merge_set->alignment); if (best_reg != (physreg_t)~0u) { best_reg += reg->merge_set_offset; return best_reg; } } /* For ALU and SFU instructions, if the src reg is avail to pick, use it. * Because this doesn't introduce unnecessary dependencies, and it * potentially avoids needing (ss) syncs for write after read hazards for * SFU instructions: */ if (is_sfu(reg->instr) || is_alu(reg->instr)) { for (unsigned i = 0; i < reg->instr->srcs_count; i++) { struct ir3_register *src = reg->instr->srcs[i]; if (!ra_reg_is_src(src)) continue; if (ra_get_file(ctx, src) == file && reg_size(src) >= size) { struct ra_interval *src_interval = &ctx->intervals[src->def->name]; physreg_t src_physreg = ra_interval_get_physreg(src_interval); if (src_physreg % reg_elem_size(reg) == 0 && src_physreg + size <= file_size && get_reg_specified(ctx, file, reg, src_physreg, false)) return src_physreg; } } } physreg_t best_reg = find_best_gap(ctx, file, reg, file_size, size, reg_elem_size(reg)); if (best_reg != (physreg_t)~0u) { return best_reg; } /* Ok, we couldn't find anything that fits. Here is where we have to start * moving things around to make stuff fit. First try solely evicting * registers in the way. */ unsigned best_eviction_count = ~0; for (physreg_t i = 0; i + size <= file_size; i += reg_elem_size(reg)) { unsigned eviction_count; if (try_evict_regs(ctx, file, reg, i, &eviction_count, false, true)) { if (eviction_count < best_eviction_count) { best_eviction_count = eviction_count; best_reg = i; } } } if (best_eviction_count != ~0) { ASSERTED bool result = try_evict_regs( ctx, file, reg, best_reg, &best_eviction_count, false, false); assert(result); return best_reg; } /* Use the dumb fallback only if try_evict_regs() fails. */ return compress_regs_left(ctx, file, reg); } static void assign_reg(struct ir3_instruction *instr, struct ir3_register *reg, unsigned num) { if (reg->flags & IR3_REG_ARRAY) { reg->array.base = num; if (reg->flags & IR3_REG_RELATIV) reg->array.offset += num; else reg->num = num + reg->array.offset; } else { reg->num = num; } } static void mark_src_killed(struct ra_ctx *ctx, struct ir3_register *src) { struct ra_interval *interval = &ctx->intervals[src->def->name]; if (!(src->flags & IR3_REG_FIRST_KILL) || interval->is_killed || interval->interval.parent || !rb_tree_is_empty(&interval->interval.children)) return; ra_file_mark_killed(ra_get_file(ctx, src), interval); } static void insert_dst(struct ra_ctx *ctx, struct ir3_register *dst) { struct ra_file *file = ra_get_file(ctx, dst); struct ra_interval *interval = &ctx->intervals[dst->name]; d("insert dst %u physreg %u", dst->name, ra_interval_get_physreg(interval)); if (!(dst->flags & IR3_REG_UNUSED)) ra_file_insert(file, interval); assign_reg(dst->instr, dst, ra_interval_get_num(interval)); } static void allocate_dst_fixed(struct ra_ctx *ctx, struct ir3_register *dst, physreg_t physreg) { struct ra_file *file = ra_get_file(ctx, dst); struct ra_interval *interval = &ctx->intervals[dst->name]; update_affinity(file, dst, physreg); ra_interval_init(interval, dst); interval->physreg_start = physreg; interval->physreg_end = physreg + reg_size(dst); } /* If a tied destination interferes with its source register, we have to insert * a copy beforehand to copy the source to the destination. Because we are using * the parallel_copies array and not creating a separate copy, this copy will * happen in parallel with any shuffling around of the tied source, so we have * to copy the source *as it exists before it is shuffled around*. We do this by * inserting the copy early, before any other copies are inserted. We don't * actually know the destination of the copy, but that's ok because the * dst_interval will be filled out later. */ static void insert_tied_dst_copy(struct ra_ctx *ctx, struct ir3_register *dst) { struct ir3_register *tied = dst->tied; if (!tied) return; struct ra_interval *tied_interval = &ctx->intervals[tied->def->name]; struct ra_interval *dst_interval = &ctx->intervals[dst->name]; if (tied_interval->is_killed) return; physreg_t tied_physreg = ra_interval_get_physreg(tied_interval); array_insert(ctx, ctx->parallel_copies, (struct ra_parallel_copy){ .interval = dst_interval, .src = tied_physreg, }); } static void allocate_dst(struct ra_ctx *ctx, struct ir3_register *dst) { struct ra_file *file = ra_get_file(ctx, dst); struct ir3_register *tied = dst->tied; if (tied) { struct ra_interval *tied_interval = &ctx->intervals[tied->def->name]; if (tied_interval->is_killed) { /* The easy case: the source is killed, so we can just reuse it * for the destination. */ allocate_dst_fixed(ctx, dst, ra_interval_get_physreg(tied_interval)); return; } } /* All the hard work is done by get_reg here. */ physreg_t physreg = get_reg(ctx, file, dst); allocate_dst_fixed(ctx, dst, physreg); } static void assign_src(struct ra_ctx *ctx, struct ir3_instruction *instr, struct ir3_register *src) { struct ra_interval *interval = &ctx->intervals[src->def->name]; struct ra_file *file = ra_get_file(ctx, src); struct ir3_register *tied = src->tied; physreg_t physreg; if (tied) { struct ra_interval *tied_interval = &ctx->intervals[tied->name]; physreg = ra_interval_get_physreg(tied_interval); } else { physreg = ra_interval_get_physreg(interval); } assign_reg(instr, src, ra_physreg_to_num(physreg, src->flags)); if (src->flags & IR3_REG_FIRST_KILL) ra_file_remove(file, interval); } /* Insert a parallel copy instruction before the instruction with the parallel * copy entries we've built up. */ static void insert_parallel_copy_instr(struct ra_ctx *ctx, struct ir3_instruction *instr) { if (ctx->parallel_copies_count == 0) return; struct ir3_instruction *pcopy = ir3_instr_create(instr->block, OPC_META_PARALLEL_COPY, ctx->parallel_copies_count, ctx->parallel_copies_count); for (unsigned i = 0; i < ctx->parallel_copies_count; i++) { struct ra_parallel_copy *entry = &ctx->parallel_copies[i]; struct ir3_register *reg = ir3_dst_create(pcopy, INVALID_REG, entry->interval->interval.reg->flags & (IR3_REG_HALF | IR3_REG_ARRAY)); reg->size = entry->interval->interval.reg->size; reg->wrmask = entry->interval->interval.reg->wrmask; assign_reg(pcopy, reg, ra_interval_get_num(entry->interval)); } for (unsigned i = 0; i < ctx->parallel_copies_count; i++) { struct ra_parallel_copy *entry = &ctx->parallel_copies[i]; struct ir3_register *reg = ir3_src_create(pcopy, INVALID_REG, entry->interval->interval.reg->flags & (IR3_REG_HALF | IR3_REG_ARRAY)); reg->size = entry->interval->interval.reg->size; reg->wrmask = entry->interval->interval.reg->wrmask; assign_reg(pcopy, reg, ra_physreg_to_num(entry->src, reg->flags)); } list_del(&pcopy->node); list_addtail(&pcopy->node, &instr->node); ctx->parallel_copies_count = 0; } static void handle_normal_instr(struct ra_ctx *ctx, struct ir3_instruction *instr) { /* First, mark sources as going-to-be-killed while allocating the dest. */ ra_foreach_src (src, instr) { mark_src_killed(ctx, src); } /* Pre-insert tied dst copies. */ ra_foreach_dst (dst, instr) { insert_tied_dst_copy(ctx, dst); } /* Allocate the destination. */ ra_foreach_dst (dst, instr) { allocate_dst(ctx, dst); } /* Now handle sources. Go backward so that in case there are multiple * sources with the same def and that def is killed we only remove it at * the end. */ ra_foreach_src_rev (src, instr) { assign_src(ctx, instr, src); } /* Now finally insert the destination into the map. */ ra_foreach_dst (dst, instr) { insert_dst(ctx, dst); } insert_parallel_copy_instr(ctx, instr); } static void handle_split(struct ra_ctx *ctx, struct ir3_instruction *instr) { struct ir3_register *dst = instr->dsts[0]; struct ir3_register *src = instr->srcs[0]; if (dst->merge_set == NULL || src->def->merge_set != dst->merge_set) { handle_normal_instr(ctx, instr); return; } struct ra_interval *src_interval = &ctx->intervals[src->def->name]; physreg_t physreg = ra_interval_get_physreg(src_interval); assign_src(ctx, instr, src); allocate_dst_fixed( ctx, dst, physreg - src->def->merge_set_offset + dst->merge_set_offset); insert_dst(ctx, dst); } static void handle_collect(struct ra_ctx *ctx, struct ir3_instruction *instr) { struct ir3_merge_set *dst_set = instr->dsts[0]->merge_set; unsigned dst_offset = instr->dsts[0]->merge_set_offset; if (!dst_set || dst_set->regs_count == 1) { handle_normal_instr(ctx, instr); return; } /* We need to check if any of the sources are contained in an interval * that is at least as large as the vector. In this case, we should put * the vector inside that larger interval. (There should be one * unambiguous place to put it, because values sharing the same merge set * should be allocated together.) This can happen in a case like: * * ssa_1 (wrmask=0xf) = ... * ssa_2 = split ssa_1 off:0 * ssa_3 = split ssa_1 off:1 * ssa_4 (wrmask=0x3) = collect (kill)ssa_2, (kill)ssa_3 * ... = (kill)ssa_1 * ... = (kill)ssa_4 * * ssa_4 will be coalesced with ssa_1 and needs to be allocated inside it. */ physreg_t dst_fixed = (physreg_t)~0u; ra_foreach_src (src, instr) { if (src->flags & IR3_REG_FIRST_KILL) { mark_src_killed(ctx, src); } struct ra_interval *interval = &ctx->intervals[src->def->name]; if (src->def->merge_set != dst_set || interval->is_killed) continue; while (interval->interval.parent != NULL) { interval = ir3_reg_interval_to_ra_interval(interval->interval.parent); } if (reg_size(interval->interval.reg) >= reg_size(instr->dsts[0])) { dst_fixed = interval->physreg_start - interval->interval.reg->merge_set_offset + dst_offset; } else { /* For sources whose root interval is smaller than the * destination (i.e. the normal case), we will shuffle them * around after allocating the destination. Mark them killed so * that the destination can be allocated over them, even if they * aren't actually killed. */ ra_file_mark_killed(ra_get_file(ctx, src), interval); } } if (dst_fixed != (physreg_t)~0u) allocate_dst_fixed(ctx, instr->dsts[0], dst_fixed); else allocate_dst(ctx, instr->dsts[0]); /* Remove the temporary is_killed we added */ ra_foreach_src (src, instr) { struct ra_interval *interval = &ctx->intervals[src->def->name]; while (interval->interval.parent != NULL) { interval = ir3_reg_interval_to_ra_interval(interval->interval.parent); } /* Filter out cases where it actually should be killed */ if (interval != &ctx->intervals[src->def->name] || !(src->flags & IR3_REG_KILL)) { ra_file_unmark_killed(ra_get_file(ctx, src), interval); } } ra_foreach_src_rev (src, instr) { assign_src(ctx, instr, src); } /* We need to do this before insert_dst(), so that children of the * destination which got marked as killed and then shuffled around to make * space for the destination have the correct pcopy destination that * matches what we assign the source of the collect to in assign_src(). * * TODO: In this case we'll wind up copying the value in the pcopy and * then again in the collect. We could avoid one of those by updating the * pcopy destination to match up with the final location of the source * after the collect and making the collect a no-op. However this doesn't * seem to happen often. */ insert_parallel_copy_instr(ctx, instr); /* Note: insert_dst will automatically shuffle around any intervals that * are a child of the collect by making them children of the collect. */ insert_dst(ctx, instr->dsts[0]); } /* Parallel copies before RA should only be at the end of the block, for * phi's. For these we only need to fill in the sources, and then we fill in * the destinations in the successor block. */ static void handle_pcopy(struct ra_ctx *ctx, struct ir3_instruction *instr) { ra_foreach_src_rev (src, instr) { assign_src(ctx, instr, src); } } /* Some inputs may need to be precolored. We need to handle those first, so * that other non-precolored inputs don't accidentally get allocated over * them. Inputs are the very first thing in the shader, so it shouldn't be a * problem to allocate them to a specific physreg. */ static void handle_precolored_input(struct ra_ctx *ctx, struct ir3_instruction *instr) { if (instr->dsts[0]->num == INVALID_REG) return; struct ra_interval *interval = &ctx->intervals[instr->dsts[0]->name]; physreg_t physreg = ra_reg_get_physreg(instr->dsts[0]); allocate_dst_fixed(ctx, instr->dsts[0], physreg); insert_dst(ctx, instr->dsts[0]); interval->frozen = true; } static void handle_input(struct ra_ctx *ctx, struct ir3_instruction *instr) { if (instr->dsts[0]->num != INVALID_REG) return; allocate_dst(ctx, instr->dsts[0]); struct ra_file *file = ra_get_file(ctx, instr->dsts[0]); struct ra_interval *interval = &ctx->intervals[instr->dsts[0]->name]; ra_file_insert(file, interval); } static void assign_input(struct ra_ctx *ctx, struct ir3_instruction *instr) { struct ra_interval *interval = &ctx->intervals[instr->dsts[0]->name]; struct ra_file *file = ra_get_file(ctx, instr->dsts[0]); if (instr->dsts[0]->num == INVALID_REG) { assign_reg(instr, instr->dsts[0], ra_interval_get_num(interval)); } else { interval->frozen = false; } if (instr->dsts[0]->flags & IR3_REG_UNUSED) ra_file_remove(file, interval); ra_foreach_src_rev (src, instr) assign_src(ctx, instr, src); } /* chmask is a bit weird, because it has pre-colored sources due to the need * to pass some registers to the next stage. Fortunately there are only at * most two, and there should be no other live values by the time we get to * this instruction, so we only have to do the minimum and don't need any * fancy fallbacks. * * TODO: Add more complete handling of precolored sources, e.g. for function * argument handling. We'd need a way to mark sources as fixed so that they * don't get moved around when placing other sources in the fallback case, and * a duplication of much of the logic in get_reg(). This also opens another * can of worms, e.g. what if the precolored source is a split of a vector * which is still live -- this breaks our assumption that splits don't incur * any "extra" register requirements and we'd have to break it out of the * parent ra_interval. */ static void handle_precolored_source(struct ra_ctx *ctx, struct ir3_register *src) { struct ra_file *file = ra_get_file(ctx, src); struct ra_interval *interval = &ctx->intervals[src->def->name]; physreg_t physreg = ra_reg_get_physreg(src); if (ra_interval_get_num(interval) == src->num) return; /* Try evicting stuff in our way if it isn't free. This won't move * anything unless it overlaps with our precolored physreg, so we don't * have to worry about evicting other precolored sources. */ if (!get_reg_specified(ctx, file, src, physreg, true)) { unsigned eviction_count; if (!try_evict_regs(ctx, file, src, physreg, &eviction_count, true, false)) { unreachable("failed to evict for precolored source!"); return; } } ra_move_interval(ctx, file, interval, physreg); } static void handle_chmask(struct ra_ctx *ctx, struct ir3_instruction *instr) { /* Note: we purposely don't mark sources as killed, so that we can reuse * some of the get_reg() machinery as-if the source is a destination. * Marking it as killed would make e.g. get_reg_specified() wouldn't work * correctly. */ ra_foreach_src (src, instr) { assert(src->num != INVALID_REG); handle_precolored_source(ctx, src); } ra_foreach_src (src, instr) { struct ra_file *file = ra_get_file(ctx, src); struct ra_interval *interval = &ctx->intervals[src->def->name]; if (src->flags & IR3_REG_FIRST_KILL) ra_file_remove(file, interval); } insert_parallel_copy_instr(ctx, instr); } static physreg_t read_register(struct ra_ctx *ctx, struct ir3_block *block, struct ir3_register *def) { struct ra_block_state *state = &ctx->blocks[block->index]; if (state->renames) { struct hash_entry *entry = _mesa_hash_table_search(state->renames, def); if (entry) { return (physreg_t)(uintptr_t)entry->data; } } return ra_reg_get_physreg(def); } static void handle_live_in(struct ra_ctx *ctx, struct ir3_register *def) { physreg_t physreg = ~0; for (unsigned i = 0; i < ctx->block->predecessors_count; i++) { struct ir3_block *pred = ctx->block->predecessors[i]; struct ra_block_state *pred_state = &ctx->blocks[pred->index]; if (!pred_state->visited) continue; physreg = read_register(ctx, pred, def); break; } assert(physreg != (physreg_t)~0); struct ra_interval *interval = &ctx->intervals[def->name]; struct ra_file *file = ra_get_file(ctx, def); ra_interval_init(interval, def); interval->physreg_start = physreg; interval->physreg_end = physreg + reg_size(def); ra_file_insert(file, interval); } static void handle_live_out(struct ra_ctx *ctx, struct ir3_register *def) { /* Skip parallelcopy's which in the original program are only used as phi * arguments. Even though phi arguments are live out, they are only * assigned when the phi is. */ if (def->instr->opc == OPC_META_PARALLEL_COPY) return; struct ra_block_state *state = &ctx->blocks[ctx->block->index]; struct ra_interval *interval = &ctx->intervals[def->name]; physreg_t physreg = ra_interval_get_physreg(interval); if (physreg != ra_reg_get_physreg(def)) { if (!state->renames) state->renames = _mesa_pointer_hash_table_create(ctx); _mesa_hash_table_insert(state->renames, def, (void *)(uintptr_t)physreg); } } static void handle_phi(struct ra_ctx *ctx, struct ir3_register *def) { struct ra_file *file = ra_get_file(ctx, def); struct ra_interval *interval = &ctx->intervals[def->name]; /* phis are always scalar, so they should already be the smallest possible * size. However they may be coalesced with other live-in values/phi * nodes, so check for that here. */ struct ir3_reg_interval *parent_ir3 = ir3_reg_interval_search(&file->reg_ctx.intervals, def->interval_start); physreg_t physreg; if (parent_ir3) { struct ra_interval *parent = ir3_reg_interval_to_ra_interval(parent_ir3); physreg = ra_interval_get_physreg(parent) + (def->interval_start - parent_ir3->reg->interval_start); } else { physreg = get_reg(ctx, file, def); } allocate_dst_fixed(ctx, def, physreg); ra_file_insert(file, interval); } static void assign_phi(struct ra_ctx *ctx, struct ir3_instruction *phi) { struct ra_file *file = ra_get_file(ctx, phi->dsts[0]); struct ra_interval *interval = &ctx->intervals[phi->dsts[0]->name]; assert(!interval->interval.parent); unsigned num = ra_interval_get_num(interval); assign_reg(phi, phi->dsts[0], num); /* Assign the parallelcopy sources of this phi */ for (unsigned i = 0; i < phi->srcs_count; i++) { if (phi->srcs[i]->def) { assign_reg(phi, phi->srcs[i], num); assign_reg(phi, phi->srcs[i]->def, num); } } if (phi->dsts[0]->flags & IR3_REG_UNUSED) ra_file_remove(file, interval); } /* When we split a live range, we sometimes need to emit fixup code at the end * of a block. For example, something like: * * a = ... * if (...) { * ... * a' = a * b = ... // a evicted to make room for b * ... * } * ... = a * * When we insert the copy to a' in insert_parallel_copy_instr(), this forces * to insert another copy "a = a'" at the end of the if. Normally this would * also entail adding a phi node, but since we're about to go out of SSA * anyway we just insert an extra move. Note, however, that "b" might be used * in a phi node at the end of the if and share registers with "a", so we * have to be careful to extend any preexisting parallelcopy instruction * instead of creating our own in order to guarantee that they properly get * swapped. */ static void insert_liveout_copy(struct ir3_block *block, physreg_t dst, physreg_t src, struct ir3_register *reg) { struct ir3_instruction *old_pcopy = NULL; if (!list_is_empty(&block->instr_list)) { struct ir3_instruction *last = list_entry(block->instr_list.prev, struct ir3_instruction, node); if (last->opc == OPC_META_PARALLEL_COPY) old_pcopy = last; } unsigned old_pcopy_srcs = old_pcopy ? old_pcopy->srcs_count : 0; struct ir3_instruction *pcopy = ir3_instr_create( block, OPC_META_PARALLEL_COPY, old_pcopy_srcs + 1, old_pcopy_srcs + 1); for (unsigned i = 0; i < old_pcopy_srcs; i++) { old_pcopy->dsts[i]->instr = pcopy; pcopy->dsts[pcopy->dsts_count++] = old_pcopy->dsts[i]; } unsigned flags = reg->flags & (IR3_REG_HALF | IR3_REG_ARRAY); struct ir3_register *dst_reg = ir3_dst_create(pcopy, INVALID_REG, flags); dst_reg->wrmask = reg->wrmask; dst_reg->size = reg->size; assign_reg(pcopy, dst_reg, ra_physreg_to_num(dst, reg->flags)); for (unsigned i = 0; i < old_pcopy_srcs; i++) { pcopy->srcs[pcopy->srcs_count++] = old_pcopy->srcs[i]; } struct ir3_register *src_reg = ir3_src_create(pcopy, INVALID_REG, flags); src_reg->wrmask = reg->wrmask; src_reg->size = reg->size; assign_reg(pcopy, src_reg, ra_physreg_to_num(src, reg->flags)); if (old_pcopy) list_del(&old_pcopy->node); } static void insert_live_in_move(struct ra_ctx *ctx, struct ra_interval *interval) { physreg_t physreg = ra_interval_get_physreg(interval); bool shared = interval->interval.reg->flags & IR3_REG_SHARED; struct ir3_block **predecessors = shared ? ctx->block->physical_predecessors : ctx->block->predecessors; unsigned predecessors_count = shared ? ctx->block->physical_predecessors_count : ctx->block->predecessors_count; for (unsigned i = 0; i < predecessors_count; i++) { struct ir3_block *pred = predecessors[i]; struct ra_block_state *pred_state = &ctx->blocks[pred->index]; if (!pred_state->visited) continue; physreg_t pred_reg = read_register(ctx, pred, interval->interval.reg); if (pred_reg != physreg) { insert_liveout_copy(pred, physreg, pred_reg, interval->interval.reg); /* This is a bit tricky, but when visiting the destination of a * physical-only edge, we have two predecessors (the if and the * header block) and both have multiple successors. We pick the * register for all live-ins from the normal edge, which should * guarantee that there's no need for shuffling things around in * the normal predecessor as long as there are no phi nodes, but * we still may need to insert fixup code in the physical * predecessor (i.e. the last block of the if) and that has * another successor (the block after the if) so we need to update * the renames state for when we process the other successor. This * crucially depends on the other successor getting processed * after this. * * For normal (non-physical) edges we disallow critical edges so * that hacks like this aren't necessary. */ if (!pred_state->renames) pred_state->renames = _mesa_pointer_hash_table_create(ctx); _mesa_hash_table_insert(pred_state->renames, interval->interval.reg, (void *)(uintptr_t)physreg); } } } static void insert_file_live_in_moves(struct ra_ctx *ctx, struct ra_file *file) { BITSET_WORD *live_in = ctx->live->live_in[ctx->block->index]; rb_tree_foreach (struct ra_interval, interval, &file->physreg_intervals, physreg_node) { /* Skip phi nodes. This needs to happen after phi nodes are allocated, * because we may have to move live-ins around to make space for phi * nodes, but we shouldn't be handling phi nodes here. */ if (BITSET_TEST(live_in, interval->interval.reg->name)) insert_live_in_move(ctx, interval); } } static void insert_entry_regs(struct ra_block_state *state, struct ra_file *file) { rb_tree_foreach (struct ra_interval, interval, &file->physreg_intervals, physreg_node) { _mesa_hash_table_insert(state->entry_regs, interval->interval.reg, (void *)(uintptr_t)interval->physreg_start); } } static void insert_live_in_moves(struct ra_ctx *ctx) { insert_file_live_in_moves(ctx, &ctx->full); insert_file_live_in_moves(ctx, &ctx->half); insert_file_live_in_moves(ctx, &ctx->shared); /* If not all predecessors are visited, insert live-in regs so that * insert_live_out_moves() will work. */ bool all_preds_visited = true; for (unsigned i = 0; i < ctx->block->predecessors_count; i++) { if (!ctx->blocks[ctx->block->predecessors[i]->index].visited) { all_preds_visited = false; break; } } if (!all_preds_visited) { struct ra_block_state *state = &ctx->blocks[ctx->block->index]; state->entry_regs = _mesa_pointer_hash_table_create(ctx); insert_entry_regs(state, &ctx->full); insert_entry_regs(state, &ctx->half); insert_entry_regs(state, &ctx->shared); } } static void insert_live_out_move(struct ra_ctx *ctx, struct ra_interval *interval) { for (unsigned i = 0; i < 2; i++) { if (!ctx->block->successors[i]) continue; struct ir3_block *succ = ctx->block->successors[i]; struct ra_block_state *succ_state = &ctx->blocks[succ->index]; if (!succ_state->visited) continue; struct hash_entry *entry = _mesa_hash_table_search( succ_state->entry_regs, interval->interval.reg); if (!entry) continue; physreg_t new_reg = (physreg_t)(uintptr_t)entry->data; if (new_reg != interval->physreg_start) { insert_liveout_copy(ctx->block, new_reg, interval->physreg_start, interval->interval.reg); } } } static void insert_file_live_out_moves(struct ra_ctx *ctx, struct ra_file *file) { rb_tree_foreach (struct ra_interval, interval, &file->physreg_intervals, physreg_node) { insert_live_out_move(ctx, interval); } } static void insert_live_out_moves(struct ra_ctx *ctx) { insert_file_live_out_moves(ctx, &ctx->full); insert_file_live_out_moves(ctx, &ctx->half); insert_file_live_out_moves(ctx, &ctx->shared); } static void handle_block(struct ra_ctx *ctx, struct ir3_block *block) { ctx->block = block; /* Reset the register files from the last block */ ra_file_init(&ctx->full); ra_file_init(&ctx->half); ra_file_init(&ctx->shared); /* Handle live-ins, phis, and input meta-instructions. These all appear * live at the beginning of the block, and interfere with each other * therefore need to be allocated "in parallel". This means that we * have to allocate all of them, inserting them into the file, and then * delay updating the IR until all of them are allocated. * * Handle precolored inputs first, because we need to make sure that other * inputs don't overwrite them. We shouldn't have both live-ins/phi nodes * and inputs at the same time, because the first block doesn't have * predecessors. Therefore handle_live_in doesn't have to worry about * them. */ foreach_instr (instr, &block->instr_list) { if (instr->opc == OPC_META_INPUT) handle_precolored_input(ctx, instr); else break; } unsigned name; BITSET_FOREACH_SET (name, ctx->live->live_in[block->index], ctx->live->definitions_count) { struct ir3_register *reg = ctx->live->definitions[name]; handle_live_in(ctx, reg); } foreach_instr (instr, &block->instr_list) { if (instr->opc == OPC_META_PHI) handle_phi(ctx, instr->dsts[0]); else if (instr->opc == OPC_META_INPUT || instr->opc == OPC_META_TEX_PREFETCH) handle_input(ctx, instr); else break; } /* After this point, every live-in/phi/input has an interval assigned to * it. We delay actually assigning values until everything has been * allocated, so we can simply ignore any parallel copy entries created * when shuffling them around. */ ctx->parallel_copies_count = 0; insert_live_in_moves(ctx); if (RA_DEBUG) { d("after live-in block %u:\n", block->index); ra_ctx_dump(ctx); } /* Now we're done with processing live-ins, and can handle the body of the * block. */ foreach_instr (instr, &block->instr_list) { di(instr, "processing"); if (instr->opc == OPC_META_PHI) assign_phi(ctx, instr); else if (instr->opc == OPC_META_INPUT || instr->opc == OPC_META_TEX_PREFETCH) assign_input(ctx, instr); else if (instr->opc == OPC_META_SPLIT) handle_split(ctx, instr); else if (instr->opc == OPC_META_COLLECT) handle_collect(ctx, instr); else if (instr->opc == OPC_META_PARALLEL_COPY) handle_pcopy(ctx, instr); else if (instr->opc == OPC_CHMASK) handle_chmask(ctx, instr); else handle_normal_instr(ctx, instr); if (RA_DEBUG) ra_ctx_dump(ctx); } insert_live_out_moves(ctx); BITSET_FOREACH_SET (name, ctx->live->live_out[block->index], ctx->live->definitions_count) { struct ir3_register *reg = ctx->live->definitions[name]; handle_live_out(ctx, reg); } ctx->blocks[block->index].visited = true; } static unsigned calc_target_full_pressure(struct ir3_shader_variant *v, unsigned pressure) { /* Registers are allocated in units of vec4, so switch from units of * half-regs to vec4. */ unsigned reg_count = DIV_ROUND_UP(pressure, 2 * 4); bool double_threadsize = ir3_should_double_threadsize(v, reg_count); unsigned target = reg_count; unsigned reg_independent_max_waves = ir3_get_reg_independent_max_waves(v, double_threadsize); unsigned reg_dependent_max_waves = ir3_get_reg_dependent_max_waves( v->compiler, reg_count, double_threadsize); unsigned target_waves = MIN2(reg_independent_max_waves, reg_dependent_max_waves); while (target <= RA_FULL_SIZE / (2 * 4) && ir3_should_double_threadsize(v, target) == double_threadsize && ir3_get_reg_dependent_max_waves(v->compiler, target, double_threadsize) >= target_waves) target++; return (target - 1) * 2 * 4; } static void add_pressure(struct ir3_pressure *pressure, struct ir3_register *reg, bool merged_regs) { unsigned size = reg_size(reg); if (reg->flags & IR3_REG_HALF) pressure->half += size; if (!(reg->flags & IR3_REG_HALF) || merged_regs) pressure->full += size; } static void dummy_interval_add(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *interval) { } static void dummy_interval_delete(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *interval) { } static void dummy_interval_readd(struct ir3_reg_ctx *ctx, struct ir3_reg_interval *parent, struct ir3_reg_interval *child) { } /* Calculate the minimum possible limit on register pressure so that spilling * still succeeds. Used to implement IR3_SHADER_DEBUG=spillall. */ static void calc_min_limit_pressure(struct ir3_shader_variant *v, struct ir3_liveness *live, struct ir3_pressure *limit) { struct ir3_block *start = ir3_start_block(v->ir); struct ir3_reg_ctx *ctx = ralloc(NULL, struct ir3_reg_ctx); struct ir3_reg_interval *intervals = rzalloc_array(ctx, struct ir3_reg_interval, live->definitions_count); ctx->interval_add = dummy_interval_add; ctx->interval_delete = dummy_interval_delete; ctx->interval_readd = dummy_interval_readd; limit->full = limit->half = 0; struct ir3_pressure cur_pressure = {0}; foreach_instr (input, &start->instr_list) { if (input->opc != OPC_META_INPUT && input->opc != OPC_META_TEX_PREFETCH) break; add_pressure(&cur_pressure, input->dsts[0], v->mergedregs); } limit->full = MAX2(limit->full, cur_pressure.full); limit->half = MAX2(limit->half, cur_pressure.half); foreach_instr (input, &start->instr_list) { if (input->opc != OPC_META_INPUT && input->opc != OPC_META_TEX_PREFETCH) break; /* pre-colored inputs may have holes, which increases the pressure. */ struct ir3_register *dst = input->dsts[0]; if (dst->num != INVALID_REG) { unsigned physreg = ra_reg_get_physreg(dst) + reg_size(dst); if (dst->flags & IR3_REG_HALF) limit->half = MAX2(limit->half, physreg); if (!(dst->flags & IR3_REG_HALF) || v->mergedregs) limit->full = MAX2(limit->full, physreg); } } foreach_block (block, &v->ir->block_list) { rb_tree_init(&ctx->intervals); unsigned name; BITSET_FOREACH_SET (name, live->live_in[block->index], live->definitions_count) { struct ir3_register *reg = live->definitions[name]; ir3_reg_interval_init(&intervals[reg->name], reg); ir3_reg_interval_insert(ctx, &intervals[reg->name]); } foreach_instr (instr, &block->instr_list) { ra_foreach_dst (dst, instr) { ir3_reg_interval_init(&intervals[dst->name], dst); } /* phis and parallel copies can be deleted via spilling */ if (instr->opc == OPC_META_PHI) { ir3_reg_interval_insert(ctx, &intervals[instr->dsts[0]->name]); continue; } if (instr->opc == OPC_META_PARALLEL_COPY) continue; cur_pressure = (struct ir3_pressure) {0}; ra_foreach_dst (dst, instr) { if (dst->tied && !(dst->tied->flags & IR3_REG_KILL)) add_pressure(&cur_pressure, dst, v->mergedregs); } ra_foreach_src_rev (src, instr) { /* We currently don't support spilling the parent of a source when * making space for sources, so we have to keep track of the * intervals and figure out the root of the tree to figure out how * much space we need. * * TODO: We should probably support this in the spiller. */ struct ir3_reg_interval *interval = &intervals[src->def->name]; while (interval->parent) interval = interval->parent; add_pressure(&cur_pressure, interval->reg, v->mergedregs); if (src->flags & IR3_REG_FIRST_KILL) ir3_reg_interval_remove(ctx, &intervals[src->def->name]); } limit->full = MAX2(limit->full, cur_pressure.full); limit->half = MAX2(limit->half, cur_pressure.half); cur_pressure = (struct ir3_pressure) {0}; ra_foreach_dst (dst, instr) { ir3_reg_interval_init(&intervals[dst->name], dst); ir3_reg_interval_insert(ctx, &intervals[dst->name]); add_pressure(&cur_pressure, dst, v->mergedregs); } limit->full = MAX2(limit->full, cur_pressure.full); limit->half = MAX2(limit->half, cur_pressure.half); } } /* Account for the base register, which needs to be available everywhere. */ limit->full += 2; ralloc_free(ctx); } /* * If barriers are used, it must be possible for all waves in the workgroup * to execute concurrently. Thus we may have to reduce the registers limit. */ static void calc_limit_pressure_for_cs_with_barrier(struct ir3_shader_variant *v, struct ir3_pressure *limit_pressure) { const struct ir3_compiler *compiler = v->compiler; unsigned threads_per_wg; if (v->local_size_variable) { /* We have to expect the worst case. */ threads_per_wg = compiler->max_variable_workgroup_size; } else { threads_per_wg = v->local_size[0] * v->local_size[1] * v->local_size[2]; } /* The register file is grouped into reg_size_vec4 number of parts. * Each part has enough registers to add a single vec4 register to * each thread of a single-sized wave-pair. With double threadsize * each wave-pair would consume two parts of the register file to get * a single vec4 for a thread. The more active wave-pairs the less * parts each could get. */ bool double_threadsize = ir3_should_double_threadsize(v, 0); unsigned waves_per_wg = DIV_ROUND_UP( threads_per_wg, compiler->threadsize_base * (double_threadsize ? 2 : 1) * compiler->wave_granularity); uint32_t vec4_regs_per_thread = compiler->reg_size_vec4 / (waves_per_wg * (double_threadsize ? 2 : 1)); assert(vec4_regs_per_thread > 0); uint32_t half_regs_per_thread = vec4_regs_per_thread * 4 * 2; if (limit_pressure->full > half_regs_per_thread) { if (v->mergedregs) { limit_pressure->full = half_regs_per_thread; } else { /* TODO: Handle !mergedregs case, probably we would have to do this * after the first register pressure pass. */ } } } int ir3_ra(struct ir3_shader_variant *v) { ir3_calc_dominance(v->ir); ir3_create_parallel_copies(v->ir); struct ra_ctx *ctx = rzalloc(NULL, struct ra_ctx); ctx->merged_regs = v->mergedregs; ctx->compiler = v->compiler; ctx->stage = v->type; struct ir3_liveness *live = ir3_calc_liveness(ctx, v->ir); ir3_debug_print(v->ir, "AFTER: create_parallel_copies"); ir3_merge_regs(live, v->ir); struct ir3_pressure max_pressure; ir3_calc_pressure(v, live, &max_pressure); d("max pressure:"); d("\tfull: %u", max_pressure.full); d("\thalf: %u", max_pressure.half); d("\tshared: %u", max_pressure.shared); struct ir3_pressure limit_pressure; limit_pressure.full = RA_FULL_SIZE; limit_pressure.half = RA_HALF_SIZE; limit_pressure.shared = RA_SHARED_SIZE; if (gl_shader_stage_is_compute(v->type) && v->has_barrier) { calc_limit_pressure_for_cs_with_barrier(v, &limit_pressure); } /* If the user forces a doubled threadsize, we may have to lower the limit * because on some gens the register file is not big enough to hold a * double-size wave with all 48 registers in use. */ if (v->real_wavesize == IR3_DOUBLE_ONLY) { limit_pressure.full = MAX2(limit_pressure.full, ctx->compiler->reg_size_vec4 / 2 * 16); } /* If requested, lower the limit so that spilling happens more often. */ if (ir3_shader_debug & IR3_DBG_SPILLALL) calc_min_limit_pressure(v, live, &limit_pressure); if (max_pressure.shared > limit_pressure.shared) { /* TODO shared reg -> normal reg spilling */ d("shared max pressure exceeded!"); goto fail; } bool spilled = false; if (max_pressure.full > limit_pressure.full || max_pressure.half > limit_pressure.half) { if (!v->compiler->has_pvtmem) { d("max pressure exceeded!"); goto fail; } d("max pressure exceeded, spilling!"); IR3_PASS(v->ir, ir3_spill, v, &live, &limit_pressure); ir3_calc_pressure(v, live, &max_pressure); assert(max_pressure.full <= limit_pressure.full && max_pressure.half <= limit_pressure.half); spilled = true; } ctx->live = live; ctx->intervals = rzalloc_array(ctx, struct ra_interval, live->definitions_count); ctx->blocks = rzalloc_array(ctx, struct ra_block_state, live->block_count); ctx->full.size = calc_target_full_pressure(v, max_pressure.full); d("full size: %u", ctx->full.size); if (!v->mergedregs) ctx->half.size = RA_HALF_SIZE; ctx->shared.size = RA_SHARED_SIZE; ctx->full.start = ctx->half.start = ctx->shared.start = 0; foreach_block (block, &v->ir->block_list) handle_block(ctx, block); ir3_ra_validate(v, ctx->full.size, ctx->half.size, live->block_count); /* Strip array-ness and SSA-ness at the end, because various helpers still * need to work even on definitions that have already been assigned. For * example, we need to preserve array-ness so that array live-ins have the * right size. */ foreach_block (block, &v->ir->block_list) { foreach_instr (instr, &block->instr_list) { for (unsigned i = 0; i < instr->dsts_count; i++) { instr->dsts[i]->flags &= ~IR3_REG_SSA; /* Parallel copies of array registers copy the whole register, and * we need some way to let the parallel copy code know that this was * an array whose size is determined by reg->size. So keep the array * flag on those. spill/reload also need to work on the entire * array. */ if (!is_meta(instr) && instr->opc != OPC_RELOAD_MACRO) instr->dsts[i]->flags &= ~IR3_REG_ARRAY; } for (unsigned i = 0; i < instr->srcs_count; i++) { instr->srcs[i]->flags &= ~IR3_REG_SSA; if (!is_meta(instr) && instr->opc != OPC_SPILL_MACRO) instr->srcs[i]->flags &= ~IR3_REG_ARRAY; } } } ir3_debug_print(v->ir, "AFTER: register allocation"); if (spilled) { IR3_PASS(v->ir, ir3_lower_spill); } ir3_lower_copies(v); ir3_debug_print(v->ir, "AFTER: ir3_lower_copies"); ralloc_free(ctx); return 0; fail: ralloc_free(ctx); return -1; }