nir: calculate dominance information
authorConnor Abbott <connor.abbott@intel.com>
Fri, 18 Jul 2014 23:13:11 +0000 (16:13 -0700)
committerJason Ekstrand <jason.ekstrand@intel.com>
Thu, 15 Jan 2015 15:18:58 +0000 (07:18 -0800)
src/glsl/Makefile.sources
src/glsl/nir/nir.c
src/glsl/nir/nir.h
src/glsl/nir/nir_dominance.c [new file with mode: 0644]

index 6bc7cd7cd387e1cd9f68e3e7bd75e4356977df25..c736d6a56ec7e0c072883362da15f609e6f768f8 100644 (file)
@@ -16,6 +16,7 @@ LIBGLCPP_GENERATED_FILES = \
 NIR_FILES = \
        $(GLSL_SRCDIR)/nir/nir.c \
        $(GLSL_SRCDIR)/nir/nir.h \
+       $(GLSL_SRCDIR)/nir/nir_dominance.c \
        $(GLSL_SRCDIR)/nir/nir_intrinsics.c \
        $(GLSL_SRCDIR)/nir/nir_intrinsics.h \
        $(GLSL_SRCDIR)/nir/nir_lower_atomics.c \
index be293fb029c2bb33749366505da7b5119690b393..c3b0ee4e12018da0cf8f029ba0966af2b3649b30 100644 (file)
@@ -248,6 +248,7 @@ nir_function_impl_create(nir_function_overload *overload)
    impl->reg_alloc = 0;
    impl->ssa_alloc = 0;
    impl->block_index_dirty = true;
+   impl->dominance_dirty = true;
 
    /* create start & end blocks */
    nir_block *start_block = nir_block_create(mem_ctx);
@@ -274,6 +275,8 @@ nir_block_create(void *mem_ctx)
 
    block->successors[0] = block->successors[1] = NULL;
    block->predecessors = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
+   block->imm_dom = NULL;
+   block->dom_frontier = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
 
    exec_list_make_empty(&block->instr_list);
 
@@ -817,6 +820,9 @@ handle_jump(nir_block *block)
 
    unlink_block_successors(block);
 
+   nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
+   impl->dominance_dirty = true;
+
    if (jump_instr->type == nir_jump_break ||
        jump_instr->type == nir_jump_continue) {
       nir_loop *loop = nearest_loop(&block->cf_node);
@@ -841,7 +847,6 @@ handle_jump(nir_block *block)
       }
    } else {
       assert(jump_instr->type == nir_jump_return);
-      nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
       link_blocks(block, impl->end_block, NULL);
    }
 }
@@ -912,6 +917,9 @@ handle_remove_jump(nir_block *block, nir_jump_type type)
          block_add_pred(next_block, last_block);
       }
    }
+
+   nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
+   impl->dominance_dirty = true;
 }
 
 /**
@@ -1019,6 +1027,7 @@ nir_cf_node_insert_after(nir_cf_node *node, nir_cf_node *after)
 
    nir_function_impl *impl = nir_cf_node_get_function(node);
    impl->block_index_dirty = true;
+   impl->dominance_dirty = true;
 }
 
 void
@@ -1061,6 +1070,7 @@ nir_cf_node_insert_before(nir_cf_node *node, nir_cf_node *before)
 
    nir_function_impl *impl = nir_cf_node_get_function(node);
    impl->block_index_dirty = true;
+   impl->dominance_dirty = true;
 }
 
 void
index 7d55948ab8635566244c16c9334fea898b12aaa3..97687cda8f06d75f446075bbc98d6683d0b20600 100644 (file)
@@ -979,6 +979,18 @@ typedef struct nir_block {
    struct nir_block *successors[2];
 
    struct set *predecessors;
+
+   /*
+    * this node's immediate dominator in the dominance tree - set to NULL for
+    * the start block.
+    */
+   struct nir_block *imm_dom;
+
+   /* This node's children in the dominance tree */
+   unsigned num_dom_children;
+   struct nir_block **dom_children;
+
+   struct set *dom_frontier;
 } nir_block;
 
 #define nir_block_first_instr(block) \
@@ -1052,6 +1064,7 @@ typedef struct {
    unsigned num_blocks;
 
    bool block_index_dirty;
+   bool dominance_dirty;
 } nir_function_impl;
 
 #define nir_cf_node_next(_node) \
@@ -1249,6 +1262,18 @@ void nir_print_shader(nir_shader *shader, FILE *fp);
 
 void nir_validate_shader(nir_shader *shader);
 
+void nir_calc_dominance_impl(nir_function_impl *impl);
+void nir_calc_dominance(nir_shader *shader);
+
+void nir_dump_dom_tree_impl(nir_function_impl *impl, FILE *fp);
+void nir_dump_dom_tree(nir_shader *shader, FILE *fp);
+
+void nir_dump_dom_frontier_impl(nir_function_impl *impl, FILE *fp);
+void nir_dump_dom_frontier(nir_shader *shader, FILE *fp);
+
+void nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp);
+void nir_dump_cfg(nir_shader *shader, FILE *fp);
+
 void nir_lower_variables_scalar(nir_shader *shader, bool lower_globals,
                                 bool lower_io, bool add_names,
                                 bool native_integers);
diff --git a/src/glsl/nir/nir_dominance.c b/src/glsl/nir/nir_dominance.c
new file mode 100644 (file)
index 0000000..988a3fc
--- /dev/null
@@ -0,0 +1,300 @@
+/*
+ * Copyright © 2014 Intel Corporation
+ *
+ * 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.
+ *
+ * Authors:
+ *    Connor Abbott (cwabbott0@gmail.com)
+ *
+ */
+
+#include "nir.h"
+
+/*
+ * Implements the algorithms for computing the dominance tree and the
+ * dominance frontier from "A Simple, Fast Dominance Algorithm" by Cooper,
+ * Harvey, and Kennedy.
+ */
+
+typedef struct {
+   nir_function_impl *impl;
+   bool progress;
+} dom_state;
+
+static bool
+init_block_cb(nir_block *block, void *_state)
+{
+   dom_state *state = (dom_state *) _state;
+   if (block == state->impl->start_block)
+      block->imm_dom = block;
+   else
+      block->imm_dom = NULL;
+   block->num_dom_children = 0;
+
+   struct set_entry *entry;
+   set_foreach(block->dom_frontier, entry) {
+      _mesa_set_remove(block->dom_frontier, entry);
+   }
+
+   return true;
+}
+
+static nir_block *
+intersect(nir_block *b1, nir_block *b2)
+{
+   while (b1 != b2) {
+      /*
+       * Note, the comparisons here are the opposite of what the paper says
+       * because we index blocks from beginning -> end (i.e. reverse
+       * post-order) instead of post-order like they assume.
+       */
+      while (b1->index > b2->index)
+         b1 = b1->imm_dom;
+      while (b2->index > b1->index)
+         b2 = b2->imm_dom;
+   }
+
+   return b1;
+}
+
+static bool
+calc_dominance_cb(nir_block *block, void *_state)
+{
+   dom_state *state = (dom_state *) _state;
+   if (block == state->impl->start_block)
+      return true;
+
+   nir_block *new_idom = NULL;
+   struct set_entry *entry;
+   set_foreach(block->predecessors, entry) {
+      nir_block *pred = (nir_block *) entry->key;
+
+      if (pred->imm_dom) {
+         if (new_idom)
+            new_idom = intersect(pred, new_idom);
+         else
+            new_idom = pred;
+      }
+   }
+
+   assert(new_idom);
+   if (block->imm_dom != new_idom) {
+      block->imm_dom = new_idom;
+      state->progress = true;
+   }
+
+   return true;
+}
+
+static bool
+calc_dom_frontier_cb(nir_block *block, void *state)
+{
+   (void) state;
+
+   if (block->predecessors->entries > 1) {
+      struct set_entry *entry;
+      set_foreach(block->predecessors, entry) {
+         nir_block *runner = (nir_block *) entry->key;
+         while (runner != block->imm_dom) {
+            _mesa_set_add(runner->dom_frontier, _mesa_hash_pointer(block),
+                          block);
+            runner = runner->imm_dom;
+         }
+      }
+   }
+
+   return true;
+}
+
+/*
+ * Compute each node's children in the dominance tree from the immediate
+ * dominator information. We do this in three stages:
+ *
+ * 1. Calculate the number of children each node has
+ * 2. Allocate arrays, setting the number of children to 0 again
+ * 3. For each node, add itself to its parent's list of children, using
+ *    num_dom_children as an index - at the end of this step, num_dom_children
+ *    for each node will be the same as it was at the end of step #1.
+ */
+
+static bool
+block_count_children(nir_block *block, void *state)
+{
+   (void) state;
+
+   if (block->imm_dom)
+      block->imm_dom->num_dom_children++;
+
+   return true;
+}
+
+static bool
+block_alloc_children(nir_block *block, void *state)
+{
+   void *mem_ctx = state;
+
+   block->dom_children = ralloc_array(mem_ctx, nir_block *,
+                                      block->num_dom_children);
+   block->num_dom_children = 0;
+
+   return true;
+}
+
+static bool
+block_add_child(nir_block *block, void *state)
+{
+   (void) state;
+
+   if (block->imm_dom)
+      block->imm_dom->dom_children[block->imm_dom->num_dom_children++] = block;
+
+   return true;
+}
+
+static void
+calc_dom_children(nir_function_impl* impl)
+{
+   void *mem_ctx = ralloc_parent(impl);
+
+   nir_foreach_block(impl, block_count_children, NULL);
+   nir_foreach_block(impl, block_alloc_children, mem_ctx);
+   nir_foreach_block(impl, block_add_child, NULL);
+}
+
+void
+nir_calc_dominance_impl(nir_function_impl *impl)
+{
+   if (!impl->dominance_dirty)
+      return;
+
+   nir_index_blocks(impl);
+
+   dom_state state;
+   state.impl = impl;
+   state.progress = true;
+
+   nir_foreach_block(impl, init_block_cb, &state);
+
+   while (state.progress) {
+      state.progress = false;
+      nir_foreach_block(impl, calc_dominance_cb, &state);
+   }
+
+   nir_foreach_block(impl, calc_dom_frontier_cb, &state);
+
+   impl->start_block->imm_dom = NULL;
+
+   calc_dom_children(impl);
+
+   impl->dominance_dirty = false;
+}
+
+void
+nir_calc_dominance(nir_shader *shader)
+{
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl)
+         nir_calc_dominance_impl(overload->impl);
+   }
+}
+
+static bool
+dump_block_dom(nir_block *block, void *state)
+{
+   FILE *fp = (FILE *) state;
+   if (block->imm_dom)
+      fprintf(fp, "\t%u -> %u\n", block->imm_dom->index, block->index);
+   return true;
+}
+
+void
+nir_dump_dom_tree_impl(nir_function_impl *impl, FILE *fp)
+{
+   fprintf(fp, "digraph doms_%s {\n", impl->overload->function->name);
+   nir_foreach_block(impl, dump_block_dom, fp);
+   fprintf(fp, "}\n\n");
+}
+
+void
+nir_dump_dom_tree(nir_shader *shader, FILE *fp)
+{
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl)
+         nir_dump_dom_tree_impl(overload->impl, fp);
+   }
+}
+
+static bool
+dump_block_dom_frontier(nir_block *block, void *state)
+{
+   FILE *fp = (FILE *) state;
+
+   fprintf(fp, "DF(%u) = {", block->index);
+   struct set_entry *entry;
+   set_foreach(block->dom_frontier, entry) {
+      nir_block *df = (nir_block *) entry->key;
+      fprintf(fp, "%u, ", df->index);
+   }
+   fprintf(fp, "}\n");
+   return true;
+}
+
+void
+nir_dump_dom_frontier_impl(nir_function_impl *impl, FILE *fp)
+{
+   nir_foreach_block(impl, dump_block_dom_frontier, fp);
+}
+
+void
+nir_dump_dom_frontier(nir_shader *shader, FILE *fp)
+{
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl)
+         nir_dump_dom_frontier_impl(overload->impl, fp);
+   }
+}
+
+static bool
+dump_block_succs(nir_block *block, void *state)
+{
+   FILE *fp = (FILE *) state;
+   if (block->successors[0])
+      fprintf(fp, "\t%u -> %u\n", block->index, block->successors[0]->index);
+   if (block->successors[1])
+      fprintf(fp, "\t%u -> %u\n", block->index, block->successors[1]->index);
+   return true;
+}
+
+void
+nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp)
+{
+   fprintf(fp, "digraph cfg_%s {\n", impl->overload->function->name);
+   nir_foreach_block(impl, dump_block_succs, fp);
+   fprintf(fp, "}\n\n");
+}
+
+void
+nir_dump_cfg(nir_shader *shader, FILE *fp)
+{
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl)
+         nir_dump_cfg_impl(overload->impl, fp);
+   }
+}