nir: Add a better out-of-SSA pass

This commit rewrites the out-of-SSA pass to not be nearly as naieve. It's based on "Revisiting Out-of-SSA Translation for Correctness, Code Quality, and Efficiency" by Boissinot et. al. It should be fairly close to state-of-the art. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2014-10-31 11:17:09 -07:00 · 2014-10-31 11:17:09 -07:00 · 943ddb9458
parent 4f44120ff5
commit 943ddb9458
1 changed files with 733 additions and 90 deletions
--- a/src/glsl/nir/nir_from_ssa.c
+++ b/src/glsl/nir/nir_from_ssa.c
@ -28,134 +28,763 @@
 #include "nir.h"

 /*
- * Implements a quick-and-dirty out-of-ssa pass.
+ * This file implements an out-of-SSA pass as described in "Revisiting
+ * Out-of-SSA Translation for Correctness, Code Quality, and Efficiency" by
+ * Boissinot et. al.
 */

 struct from_ssa_state {
   void *mem_ctx;
   void *dead_ctx;
   struct hash_table *ssa_table;
-   nir_function_impl *current_impl;
+   struct hash_table *merge_node_table;
+   nir_instr *instr;
+   nir_function_impl *impl;
 };

+/* Returns true if a dominates b */
+static bool
+ssa_def_dominates(nir_ssa_def *a, nir_ssa_def *b)
+{
+   if (a->live_index == 0) {
+      /* SSA undefs always dominate */
+      return true;
+   } else if (b->live_index < a->live_index) {
+      return false;
+   } else if (a->parent_instr->block == b->parent_instr->block) {
+      return a->live_index <= b->live_index;
+   } else {
+      nir_block *block = b->parent_instr->block;
+      while (block->imm_dom != NULL) {
+         if (block->imm_dom == a->parent_instr->block)
+            return true;
+         block = block->imm_dom;
+      }
+      return false;
+   }
+}
+
+
+/* The following data structure, which I have named merge_set is a way of
+ * representing a set registers of non-interfering registers.  This is
+ * based on the concept of a "dominence forest" presented in "Fast Copy
+ * Coalescing and Live-Range Identification" by Budimlic et. al. but the
+ * implementation concept is taken from  "Revisiting Out-of-SSA Translation
+ * for Correctness, Code Quality, and Efficiency" by Boissinot et. al..
+ *
+ * Each SSA definition is associated with a merge_node and the association
+ * is represented by a combination of a hash table and the "def" parameter
+ * in the merge_node structure.  The merge_set stores a linked list of
+ * merge_node's in dominence order of the ssa definitions.  (Since the
+ * liveness analysis pass indexes the SSA values in dominence order for us,
+ * this is an easy thing to keep up.)  It is assumed that no pair of the
+ * nodes in a given set interfere.  Merging two sets or checking for
+ * interference can be done in a single linear-time merge-sort walk of the
+ * two lists of nodes.
+ */
+struct merge_set;
+
+typedef struct {
+   struct exec_node node;
+   struct merge_set *set;
+   nir_ssa_def *def;
+} merge_node;
+
+typedef struct merge_set {
+   struct exec_list nodes;
+   unsigned size;
+   nir_register *reg;
+} merge_set;
+
+#if 0
+static void
+merge_set_dump(merge_set *set, FILE *fp)
+{
+   nir_ssa_def *dom[set->size];
+   int dom_idx = -1;
+
+   foreach_list_typed(merge_node, node, node, &set->nodes) {
+      while (dom_idx >= 0 && !ssa_def_dominates(dom[dom_idx], node->def))
+         dom_idx--;
+
+      for (int i = 0; i <= dom_idx; i++)
+         fprintf(fp, "  ");
+
+      if (node->def->name)
+         fprintf(fp, "ssa_%d /* %s */\n", node->def->index, node->def->name);
+      else
+         fprintf(fp, "ssa_%d\n", node->def->index);
+
+      dom[++dom_idx] = node->def;
+   }
+}
+#endif
+
+static merge_node *
+get_merge_node(nir_ssa_def *def, struct from_ssa_state *state)
+{
+   struct hash_entry *entry =
+      _mesa_hash_table_search(state->merge_node_table, def);
+   if (entry)
+      return entry->data;
+
+   merge_set *set = ralloc(state->dead_ctx, merge_set);
+   exec_list_make_empty(&set->nodes);
+   set->size = 1;
+   set->reg = NULL;
+
+   merge_node *node = ralloc(state->dead_ctx, merge_node);
+   node->set = set;
+   node->def = def;
+   exec_list_push_head(&set->nodes, &node->node);
+
+   _mesa_hash_table_insert(state->merge_node_table, def, node);
+
+   return node;
+}
+
+static bool
+merge_nodes_interfere(merge_node *a, merge_node *b)
+{
+   return nir_ssa_defs_interfere(a->def, b->def);
+}
+
+/* Merges b into a */
+static merge_set *
+merge_merge_sets(merge_set *a, merge_set *b)
+{
+   struct exec_node *an = exec_list_get_head(&a->nodes);
+   struct exec_node *bn = exec_list_get_head(&b->nodes);
+   while (!exec_node_is_tail_sentinel(bn)) {
+      merge_node *a_node = exec_node_data(merge_node, an, node);
+      merge_node *b_node = exec_node_data(merge_node, bn, node);
+
+      if (exec_node_is_tail_sentinel(an) ||
+          a_node->def->live_index > b_node->def->live_index) {
+         struct exec_node *next = bn->next;
+         exec_node_remove(bn);
+         exec_node_insert_node_before(an, bn);
+         exec_node_data(merge_node, bn, node)->set = a;
+         bn = next;
+      } else {
+         an = an->next;
+      }
+   }
+
+   a->size += b->size;
+   b->size = 0;
+
+   return a;
+}
+
+/* Checks for any interference between two merge sets
+ *
+ * This is an implementation of Algorithm 2 in "Revisiting Out-of-SSA
+ * Translation for Correctness, Code Quality, and Efficiency" by
+ * Boissinot et. al.
+ */
+static bool
+merge_sets_interfere(merge_set *a, merge_set *b)
+{
+   merge_node *dom[a->size + b->size];
+   int dom_idx = -1;
+
+   struct exec_node *an = exec_list_get_head(&a->nodes);
+   struct exec_node *bn = exec_list_get_head(&b->nodes);
+   while (!exec_node_is_tail_sentinel(an) ||
+          !exec_node_is_tail_sentinel(bn)) {
+
+      merge_node *current;
+      if (exec_node_is_tail_sentinel(an)) {
+         current = exec_node_data(merge_node, bn, node);
+         bn = bn->next;
+      } else if (exec_node_is_tail_sentinel(bn)) {
+         current = exec_node_data(merge_node, an, node);
+         an = an->next;
+      } else {
+         merge_node *a_node = exec_node_data(merge_node, an, node);
+         merge_node *b_node = exec_node_data(merge_node, bn, node);
+
+         if (a_node->def->live_index <= b_node->def->live_index) {
+            current = a_node;
+            an = an->next;
+         } else {
+            current = b_node;
+            bn = bn->next;
+         }
+      }
+
+      while (dom_idx >= 0 &&
+             !ssa_def_dominates(dom[dom_idx]->def, current->def))
+         dom_idx--;
+
+      if (dom_idx >= 0 && merge_nodes_interfere(current, dom[dom_idx]))
+         return true;
+
+      dom[++dom_idx] = current;
+   }
+
+   return false;
+}
+
+static nir_parallel_copy_instr *
+block_get_parallel_copy_at_end(nir_block *block, void *mem_ctx)
+{
+   nir_instr *last_instr = nir_block_last_instr(block);
+
+   /* First we try and find a parallel copy if it already exists.  If the
+    * last instruction is a jump, it will be right before the jump;
+    * otherwise, it will be the last instruction.
+    */
+   nir_instr *pcopy_instr;
+   if (last_instr != NULL && last_instr->type == nir_instr_type_jump)
+      pcopy_instr = nir_instr_prev(last_instr);
+   else
+      pcopy_instr = last_instr;
+
+   if (pcopy_instr != NULL &&
+       pcopy_instr->type == nir_instr_type_parallel_copy) {
+      /* A parallel copy already exists. */
+      nir_parallel_copy_instr *pcopy = nir_instr_as_parallel_copy(pcopy_instr);
+
+      /* This parallel copy may be the copy for the beginning of some
+       * block, so we need to check for that before we return it.
+       */
+      if (pcopy->at_end)
+         return pcopy;
+   }
+
+   /* At this point, we haven't found a suitable parallel copy, so we
+    * have to create one.
+    */
+   nir_parallel_copy_instr *pcopy = nir_parallel_copy_instr_create(mem_ctx);
+   pcopy->at_end = true;
+
+   if (last_instr && last_instr->type == nir_instr_type_jump) {
+      nir_instr_insert_before(last_instr, &pcopy->instr);
+   } else {
+      nir_instr_insert_after_block(block, &pcopy->instr);
+   }
+
+   return pcopy;
+}
+
+static bool
+isolate_phi_nodes_block(nir_block *block, void *void_state)
+{
+   struct from_ssa_state *state = void_state;
+
+   nir_instr *last_phi_instr = NULL;
+   nir_foreach_instr(block, instr) {
+      /* Phi nodes only ever come at the start of a block */
+      if (instr->type != nir_instr_type_phi)
+         break;
+
+      last_phi_instr = instr;
+   }
+
+   /* If we don't have any phi's, then there's nothing for us to do. */
+   if (last_phi_instr == NULL)
+      return true;
+
+   /* If we have phi nodes, we need to create a parallel copy at the
+    * start of this block but after the phi nodes.
+    */
+   nir_parallel_copy_instr *block_pcopy =
+      nir_parallel_copy_instr_create(state->dead_ctx);
+   nir_instr_insert_after(last_phi_instr, &block_pcopy->instr);
+
+   nir_foreach_instr(block, instr) {
+      /* Phi nodes only ever come at the start of a block */
+      if (instr->type != nir_instr_type_phi)
+         break;
+
+      nir_phi_instr *phi = nir_instr_as_phi(instr);
+      assert(phi->dest.is_ssa);
+      foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
+         nir_parallel_copy_instr *pcopy =
+            block_get_parallel_copy_at_end(src->pred, state->dead_ctx);
+
+         nir_parallel_copy_copy *copy = ralloc(state->dead_ctx,
+                                               nir_parallel_copy_copy);
+         exec_list_push_tail(&pcopy->copies, &copy->node);
+
+         copy->src = nir_src_copy(src->src, state->dead_ctx);
+         _mesa_set_add(src->src.ssa->uses,
+                       _mesa_hash_pointer(&pcopy->instr), &pcopy->instr);
+
+         copy->dest.is_ssa = true;
+         nir_ssa_def_init(state->impl, &pcopy->instr, &copy->dest.ssa,
+                          phi->dest.ssa.num_components, src->src.ssa->name);
+
+         struct set_entry *entry = _mesa_set_search(src->src.ssa->uses,
+                                                    _mesa_hash_pointer(instr),
+                                                    instr);
+         if (entry)
+            /* It is possible that a phi node can use the same source twice
+             * but for different basic blocks.  If that happens, entry will
+             * be NULL because we already deleted it.  This is safe
+             * because, by the time the loop is done, we will have deleted
+             * all of the sources of the phi from their respective use sets
+             * and moved them to the parallel copy definitions.
+             */
+            _mesa_set_remove(src->src.ssa->uses, entry);
+
+         src->src.ssa = &copy->dest.ssa;
+         _mesa_set_add(copy->dest.ssa.uses, _mesa_hash_pointer(instr), instr);
+      }
+
+      nir_parallel_copy_copy *copy = ralloc(state->dead_ctx,
+                                            nir_parallel_copy_copy);
+      exec_list_push_tail(&block_pcopy->copies, &copy->node);
+
+      copy->dest.is_ssa = true;
+      nir_ssa_def_init(state->impl, &block_pcopy->instr, &copy->dest.ssa,
+                       phi->dest.ssa.num_components, phi->dest.ssa.name);
+
+      nir_src copy_dest_src = {
+         .ssa = &copy->dest.ssa,
+         .is_ssa = true,
+      };
+      nir_ssa_def_rewrite_uses(&phi->dest.ssa, copy_dest_src, state->mem_ctx);
+
+      copy->src.is_ssa = true;
+      copy->src.ssa = &phi->dest.ssa;
+      _mesa_set_add(phi->dest.ssa.uses,
+                    _mesa_hash_pointer(&block_pcopy->instr),
+                    &block_pcopy->instr);
+   }
+
+   return true;
+}
+
+static bool
+coalesce_phi_nodes_block(nir_block *block, void *void_state)
+{
+   struct from_ssa_state *state = void_state;
+
+   nir_foreach_instr(block, instr) {
+      /* Phi nodes only ever come at the start of a block */
+      if (instr->type != nir_instr_type_phi)
+         break;
+
+      nir_phi_instr *phi = nir_instr_as_phi(instr);
+
+      assert(phi->dest.is_ssa);
+      merge_node *dest_node = get_merge_node(&phi->dest.ssa, state);
+
+      foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
+         assert(src->src.is_ssa);
+         merge_node *src_node = get_merge_node(src->src.ssa, state);
+         if (src_node->set != dest_node->set)
+            merge_merge_sets(dest_node->set, src_node->set);
+      }
+   }
+
+   return true;
+}
+
+static void
+agressive_coalesce_parallel_copy(nir_parallel_copy_instr *pcopy,
+                                 struct from_ssa_state *state)
+{
+   foreach_list_typed_safe(nir_parallel_copy_copy, copy, node, &pcopy->copies) {
+      if (!copy->src.is_ssa)
+         continue;
+
+      /* Don't try and coalesce these */
+      if (copy->dest.ssa.num_components != copy->src.ssa->num_components)
+         continue;
+
+      merge_node *src_node = get_merge_node(copy->src.ssa, state);
+      merge_node *dest_node = get_merge_node(&copy->dest.ssa, state);
+
+      if (src_node->set == dest_node->set)
+         continue;
+
+      if (!merge_sets_interfere(src_node->set, dest_node->set))
+         merge_merge_sets(src_node->set, dest_node->set);
+   }
+}
+
+static bool
+agressive_coalesce_block(nir_block *block, void *void_state)
+{
+   struct from_ssa_state *state = void_state;
+
+   nir_foreach_instr(block, instr) {
+      /* Phi nodes only ever come at the start of a block */
+      if (instr->type != nir_instr_type_phi) {
+         if (instr->type != nir_instr_type_parallel_copy)
+            break; /* The parallel copy must be right after the phis */
+
+         nir_parallel_copy_instr *pcopy = nir_instr_as_parallel_copy(instr);
+
+         agressive_coalesce_parallel_copy(pcopy, state);
+
+         if (pcopy->at_end)
+            return true;
+
+         break;
+      }
+   }
+
+   nir_instr *last_instr = nir_block_last_instr(block);
+   if (last_instr && last_instr->type == nir_instr_type_parallel_copy) {
+      nir_parallel_copy_instr *pcopy = nir_instr_as_parallel_copy(last_instr);
+      if (pcopy->at_end)
+         agressive_coalesce_parallel_copy(pcopy, state);
+   }
+
+   return true;
+}
+
+static nir_register *
+get_register_for_ssa_def(nir_ssa_def *def, struct from_ssa_state *state)
+{
+   struct hash_entry *entry =
+      _mesa_hash_table_search(state->merge_node_table, def);
+   if (entry) {
+      merge_node *node = (merge_node *)entry->data;
+
+      /* If it doesn't have a register yet, create one.  Note that all of
+       * the things in the merge set should be the same so it doesn't
+       * matter which node's definition we use.
+       */
+      if (node->set->reg == NULL) {
+         node->set->reg = nir_local_reg_create(state->impl);
+         node->set->reg->name = def->name;
+         node->set->reg->num_components = def->num_components;
+         node->set->reg->num_array_elems = 0;
+      }
+
+      return node->set->reg;
+   }
+
+   entry = _mesa_hash_table_search(state->ssa_table, def);
+   if (entry) {
+      return (nir_register *)entry->data;
+   } else {
+      nir_register *reg = nir_local_reg_create(state->impl);
+      reg->name = def->name;
+      reg->num_components = def->num_components;
+      reg->num_array_elems = 0;
+
+      _mesa_hash_table_insert(state->ssa_table, def, reg);
+      return reg;
+   }
+}
+
 static bool
 rewrite_ssa_src(nir_src *src, void *void_state)
 {
   struct from_ssa_state *state = void_state;

   if (src->is_ssa) {
-      struct hash_entry *entry =
-         _mesa_hash_table_search(state->ssa_table, src->ssa);
-      assert(entry);
+      /* We don't need to remove it from the uses set because that is going
+       * away.  We just need to add it to the one for the register. */
+      nir_register *reg = get_register_for_ssa_def(src->ssa, state);
      memset(src, 0, sizeof *src);
-      src->reg.reg = (nir_register *)entry->data;
+      src->reg.reg = reg;
+
+      _mesa_set_add(reg->uses, _mesa_hash_pointer(state->instr), state->instr);
   }

   return true;
 }

-static nir_register *
-reg_create_from_def(nir_ssa_def *def, struct from_ssa_state *state)
-{
-   nir_register *reg = nir_local_reg_create(state->current_impl);
-   reg->name = def->name;
-   reg->num_components = def->num_components;
-   reg->num_array_elems = 0;
-
-   /* Might as well steal the use-def information from SSA */
-   _mesa_set_destroy(reg->uses, NULL);
-   reg->uses = def->uses;
-   _mesa_set_destroy(reg->if_uses, NULL);
-   reg->if_uses = def->if_uses;
-   _mesa_set_add(reg->defs, _mesa_hash_pointer(def->parent_instr),
-                 def->parent_instr);
-
-   /* Add the new register to the table and rewrite the destination */
-   _mesa_hash_table_insert(state->ssa_table, def, reg);
-
-   return reg;
-}
-
 static bool
 rewrite_ssa_dest(nir_dest *dest, void *void_state)
 {
   struct from_ssa_state *state = void_state;

   if (dest->is_ssa) {
-      nir_register *reg = reg_create_from_def(&dest->ssa, state);
+      _mesa_set_destroy(dest->ssa.uses, NULL);
+      _mesa_set_destroy(dest->ssa.if_uses, NULL);
+
+      nir_register *reg = get_register_for_ssa_def(&dest->ssa, state);
      memset(dest, 0, sizeof *dest);
      dest->reg.reg = reg;
+
+      _mesa_set_add(reg->defs, _mesa_hash_pointer(state->instr), state->instr);
   }

   return true;
 }

+/* Resolves ssa definitions to registers.  While we're at it, we also
+ * remove phi nodes and ssa_undef instructions
+ */
 static bool
-convert_from_ssa_block(nir_block *block, void *void_state)
+resolve_registers_block(nir_block *block, void *void_state)
 {
   struct from_ssa_state *state = void_state;

   nir_foreach_instr_safe(block, instr) {
-      if (instr->type == nir_instr_type_ssa_undef) {
-         nir_ssa_undef_instr *undef = nir_instr_as_ssa_undef(instr);
-         reg_create_from_def(&undef->def, state);
-         exec_node_remove(&instr->node);
-         ralloc_steal(state->dead_ctx, instr);
-      } else {
-         nir_foreach_src(instr, rewrite_ssa_src, state);
-         nir_foreach_dest(instr, rewrite_ssa_dest, state);
-      }
-   }
-
-   nir_if *following_if = nir_block_following_if(block);
-   if (following_if)
-      rewrite_ssa_src(&following_if->condition, state);
-
-   return true;
-}
-
-static bool
-remove_phi_nodes(nir_block *block, void *void_state)
-{
-   struct from_ssa_state *state = void_state;
-
-   nir_foreach_instr_safe(block, instr) {
-      /* Phi nodes only ever come at the start of a block */
-      if (instr->type != nir_instr_type_phi)
-         break;
-
+      state->instr = instr;
+      nir_foreach_src(instr, rewrite_ssa_src, state);
      nir_foreach_dest(instr, rewrite_ssa_dest, state);

-      nir_phi_instr *phi = nir_instr_as_phi(instr);
-      foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
-         assert(src->src.is_ssa);
-         struct hash_entry *entry =
-            _mesa_hash_table_search(state->ssa_table, src->src.ssa);
-         nir_alu_instr *mov = nir_alu_instr_create(state->mem_ctx, nir_op_imov);
-         mov->dest.dest = nir_dest_copy(phi->dest, state->mem_ctx);
-         if (entry) {
-            nir_register *reg = (nir_register *)entry->data;
-            mov->src[0].src.reg.reg = reg;
-            mov->dest.write_mask = (1 << reg->num_components) - 1;
-         } else {
-            mov->src[0].src = nir_src_copy(src->src, state->mem_ctx);
-            mov->dest.write_mask = (1 << src->src.ssa->num_components) - 1;
-         }
+      if (instr->type == nir_instr_type_ssa_undef ||
+          instr->type == nir_instr_type_phi) {
+         nir_instr_remove(instr);
+         ralloc_steal(state->dead_ctx, instr);
+         continue;
+      }
+   }
+   state->instr = NULL;

-         nir_instr *block_end = nir_block_last_instr(src->pred);
-         if (block_end && block_end->type == nir_instr_type_jump) {
-            /* If the last instruction in the block is a jump, we want to
-             * place the moves after the jump.  Otherwise, we want to place
-             * them at the very end.
-             */
-            exec_node_insert_node_before(&block_end->node, &mov->instr.node);
-         } else {
-            exec_list_push_tail(&src->pred->instr_list, &mov->instr.node);
-         }
+   nir_if *following_if = nir_block_following_if(block);
+   if (following_if && following_if->condition.is_ssa) {
+      nir_register *reg = get_register_for_ssa_def(following_if->condition.ssa,
+                                                   state);
+      memset(&following_if->condition, 0, sizeof following_if->condition);
+      following_if->condition.reg.reg = reg;
+
+      _mesa_set_add(reg->if_uses, _mesa_hash_pointer(following_if),
+                    following_if);
+   }
+
+   return true;
+}
+
+static void
+emit_copy(nir_parallel_copy_instr *pcopy, nir_src src, nir_src dest_src,
+          void *mem_ctx)
+{
+   assert(!dest_src.is_ssa &&
+          dest_src.reg.indirect == NULL &&
+          dest_src.reg.base_offset == 0);
+   nir_dest dest = {
+      .reg.reg = dest_src.reg.reg,
+      .reg.indirect = NULL,
+      .reg.base_offset = 0,
+      .is_ssa = false,
+   };
+
+   if (src.is_ssa)
+      assert(src.ssa->num_components >= dest.reg.reg->num_components);
+   else
+      assert(src.reg.reg->num_components >= dest.reg.reg->num_components);
+
+   nir_alu_instr *mov = nir_alu_instr_create(mem_ctx, nir_op_imov);
+   mov->src[0].src = nir_src_copy(src, mem_ctx);
+   mov->dest.dest = nir_dest_copy(dest, mem_ctx);
+   mov->dest.write_mask = (1 << dest.reg.reg->num_components) - 1;
+
+   nir_instr_insert_before(&pcopy->instr, &mov->instr);
+}
+
+/* Resolves a single parallel copy operation into a sequence of mov's
+ *
+ * This is based on Algorithm 1 from "Revisiting Out-of-SSA Translation for
+ * Correctness, Code Quality, and Efficiency" by Boissinot et. al..
+ * However, I never got the algorithm to work as written, so this version
+ * is slightly modified.
+ *
+ * The algorithm works by playing this little shell game with the values.
+ * We start by recording where every source value is and which source value
+ * each destination value should recieve.  We then grab any copy whose
+ * destination is "empty", i.e. not used as a source, and do the following:
+ *  - Find where its source value currently lives
+ *  - Emit the move instruction
+ *  - Set the location of the source value to the destination
+ *  - Mark the location containing the source value
+ *  - Mark the destination as no longer needing to be copied
+ *
+ * When we run out of "empty" destinations, we have a cycle and so we
+ * create a temporary register, copy to that register, and mark the value
+ * we copied as living in that temporary.  Now, the cycle is broken, so we
+ * can continue with the above steps.
+ */
+static void
+resolve_parallel_copy(nir_parallel_copy_instr *pcopy,
+                      struct from_ssa_state *state)
+{
+   unsigned num_copies = 0;
+   foreach_list_typed_safe(nir_parallel_copy_copy, copy, node, &pcopy->copies) {
+      /* Sources may be SSA */
+      if (!copy->src.is_ssa && copy->src.reg.reg == copy->dest.reg.reg)
+         continue;
+
+      /* Set both indices equal to UINT_MAX to mark them as not indexed yet. */
+      num_copies++;
+   }
+
+   if (num_copies == 0) {
+      /* Hooray, we don't need any copies! */
+      nir_instr_remove(&pcopy->instr);
+      return;
+   }
+
+   /* The register/source corresponding to the given index */
+   nir_src values[num_copies * 2];
+   memset(values, 0, sizeof values);
+
+   /* The current location of a given piece of data */
+   int loc[num_copies * 2];
+
+   /* The piece of data that the given piece of data is to be copied from */
+   int pred[num_copies * 2];
+
+   /* Initialize loc and pred.  We will use -1 for "null" */
+   memset(loc, -1, sizeof loc);
+   memset(pred, -1, sizeof pred);
+
+   /* The destinations we have yet to properly fill */
+   int to_do[num_copies * 2];
+   int to_do_idx = -1;
+
+   /* Now we set everything up:
+    *  - All values get assigned a temporary index
+    *  - Current locations are set from sources
+    *  - Predicessors are recorded from sources and destinations
+    */
+   int num_vals = 0;
+   foreach_list_typed(nir_parallel_copy_copy, copy, node, &pcopy->copies) {
+      /* Sources may be SSA */
+      if (!copy->src.is_ssa && copy->src.reg.reg == copy->dest.reg.reg)
+         continue;
+
+      int src_idx = -1;
+      for (int i = 0; i < num_vals; ++i) {
+         if (nir_srcs_equal(values[i], copy->src))
+            src_idx = i;
+      }
+      if (src_idx < 0) {
+         src_idx = num_vals++;
+         values[src_idx] = copy->src;
      }

-      exec_node_remove(&instr->node);
-      ralloc_steal(state->dead_ctx, instr);
+      nir_src dest_src = {
+         .reg.reg = copy->dest.reg.reg,
+         .reg.indirect = NULL,
+         .reg.base_offset = 0,
+         .is_ssa = false,
+      };
+
+      int dest_idx = -1;
+      for (int i = 0; i < num_vals; ++i) {
+         if (nir_srcs_equal(values[i], dest_src)) {
+            /* Each destination of a parallel copy instruction should be
+             * unique.  A destination may get used as a source, so we still
+             * have to walk the list.  However, the predecessor should not,
+             * at this point, be set yet, so we should have -1 here.
+             */
+            assert(pred[i] == -1);
+            dest_idx = i;
+         }
+      }
+      if (dest_idx < 0) {
+         dest_idx = num_vals++;
+         values[dest_idx] = dest_src;
+      }
+
+      loc[src_idx] = src_idx;
+      pred[dest_idx] = src_idx;
+
+      to_do[++to_do_idx] = dest_idx;
+   }
+
+   /* Currently empty destinations we can go ahead and fill */
+   int ready[num_copies * 2];
+   int ready_idx = -1;
+
+   /* Mark the ones that are ready for copying.  We know an index is a
+    * destination if it has a predecessor and it's ready for copying if
+    * it's not marked as containing data.
+    */
+   for (int i = 0; i < num_vals; i++) {
+      if (pred[i] != -1 && loc[i] == -1)
+         ready[++ready_idx] = i;
+   }
+
+   while (to_do_idx >= 0) {
+      while (ready_idx >= 0) {
+         int b = ready[ready_idx--];
+         int a = pred[b];
+         emit_copy(pcopy, values[loc[a]], values[b], state->mem_ctx);
+
+         /* If any other copies want a they can find it at b */
+         loc[a] = b;
+
+         /* b has been filled, mark it as not needing to be copied */
+         pred[b] = -1;
+
+         /* If a needs to be filled, it's ready for copying now */
+         if (pred[a] != -1)
+            ready[++ready_idx] = a;
+      }
+      int b = to_do[to_do_idx--];
+      if (pred[b] == -1)
+         continue;
+
+      /* If we got here, then we don't have any more trivial copies that we
+       * can do.  We have to break a cycle, so we create a new temporary
+       * register for that purpose.  Normally, if going out of SSA after
+       * register allocation, you would want to avoid creating temporary
+       * registers.  However, we are going out of SSA before register
+       * allocation, so we would rather not create extra register
+       * dependencies for the backend to deal with.  If it wants, the
+       * backend can coalesce the (possibly multiple) temporaries.
+       */
+      assert(num_vals < num_copies * 2);
+      nir_register *reg = nir_local_reg_create(state->impl);
+      reg->name = "copy_temp";
+      reg->num_array_elems = 0;
+      if (values[b].is_ssa)
+         reg->num_components = values[b].ssa->num_components;
+      else
+         reg->num_components = values[b].reg.reg->num_components;
+      values[num_vals].is_ssa = false;
+      values[num_vals].reg.reg = reg;
+
+      emit_copy(pcopy, values[b], values[num_vals], state->mem_ctx);
+      loc[b] = num_vals;
+      ready[++ready_idx] = b;
+      num_vals++;
+   }
+
+   nir_instr_remove(&pcopy->instr);
+}
+
+/* Resolves the parallel copies in a block.  Each block can have at most
+ * two:  One at the beginning, right after all the phi noces, and one at
+ * the end (or right before the final jump if it exists).
+ */
+static bool
+resolve_parallel_copies_block(nir_block *block, void *void_state)
+{
+   struct from_ssa_state *state = void_state;
+
+   /* At this point, we have removed all of the phi nodes.  If a parallel
+    * copy existed right after the phi nodes in this block, it is now the
+    * first instruction.
+    */
+   nir_instr *first_instr = nir_block_first_instr(block);
+   if (first_instr == NULL)
+      return true; /* Empty, nothing to do. */
+
+   if (first_instr->type == nir_instr_type_parallel_copy) {
+      nir_parallel_copy_instr *pcopy = nir_instr_as_parallel_copy(first_instr);
+
+      resolve_parallel_copy(pcopy, state);
+   }
+
+   nir_instr *last_instr = nir_block_last_instr(block);
+   if (last_instr == NULL)
+      return true; /* Now empty, nothing to do. */
+
+   /* If the last instruction is a jump, the parallel copy will be before
+    * the jump.
+    */
+   if (last_instr->type == nir_instr_type_jump)
+      last_instr = nir_instr_prev(last_instr);
+
+   if (last_instr && last_instr->type == nir_instr_type_parallel_copy) {
+      nir_parallel_copy_instr *pcopy = nir_instr_as_parallel_copy(last_instr);
+      if (pcopy->at_end)
+         resolve_parallel_copy(pcopy, state);
   }

   return true;
@ -168,16 +797,30 @@ nir_convert_from_ssa_impl(nir_function_impl *impl)

   state.mem_ctx = ralloc_parent(impl);
   state.dead_ctx = ralloc_context(NULL);
-   state.current_impl = impl;
+   state.impl = impl;
+   state.merge_node_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+                                                    _mesa_key_pointer_equal);
+
+   nir_foreach_block(impl, isolate_phi_nodes_block, &state);
+
+   nir_metadata_dirty(impl, nir_metadata_block_index |
+                            nir_metadata_dominance);
+   nir_metadata_require(impl, nir_metadata_live_variables |
+                              nir_metadata_dominance);
+
+   nir_foreach_block(impl, coalesce_phi_nodes_block, &state);
+   nir_foreach_block(impl, agressive_coalesce_block, &state);
+
   state.ssa_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
                                             _mesa_key_pointer_equal);
+   nir_foreach_block(impl, resolve_registers_block, &state);

-   nir_foreach_block(impl, remove_phi_nodes, &state);
-   nir_foreach_block(impl, convert_from_ssa_block, &state);
+   nir_foreach_block(impl, resolve_parallel_copies_block, &state);

-   /* Clean up dead instructions and the hash table */
-   ralloc_free(state.dead_ctx);
+   /* Clean up dead instructions and the hash tables */
   _mesa_hash_table_destroy(state.ssa_table, NULL);
+   _mesa_hash_table_destroy(state.merge_node_table, NULL);
+   ralloc_free(state.dead_ctx);
 }

 void