i965: Enable EGL_KHR_gl_texture_3D_image
[mesa.git] / src / mesa / drivers / dri / i965 / brw_cfg.cpp
index 8714b682be0b681c12814f1c9d25a822597a27f7..53b32be510a0b6f1414e23690767090976d210e6 100644 (file)
@@ -51,12 +51,9 @@ link(void *mem_ctx, bblock_t *block)
 }
 
 bblock_t::bblock_t(cfg_t *cfg) :
-   cfg(cfg), start_ip(0), end_ip(0), num(0),
-   if_block(NULL), else_block(NULL)
+   cfg(cfg), idom(NULL), start_ip(0), end_ip(0), num(0)
 {
-   start = NULL;
-   end = NULL;
-
+   instructions.make_empty();
    parents.make_empty();
    children.make_empty();
 }
@@ -119,8 +116,8 @@ bblock_t::can_combine_with(const bblock_t *that) const
    if ((const bblock_t *)this->link.next != that)
       return false;
 
-   if (ends_block(this->end) ||
-       starts_block(that->start))
+   if (ends_block(this->end()) ||
+       starts_block(that->start()))
       return false;
 
    return true;
@@ -138,21 +135,18 @@ bblock_t::combine_with(bblock_t *that)
    }
 
    this->end_ip = that->end_ip;
-   this->end = that->end;
-   this->else_block = that->else_block;
+   this->instructions.append_list(&that->instructions);
 
    this->cfg->remove_block(that);
 }
 
 void
-bblock_t::dump(backend_visitor *v)
+bblock_t::dump(backend_shader *s) const
 {
    int ip = this->start_ip;
-   for (backend_instruction *inst = (backend_instruction *)this->start;
-       inst != this->end->next;
-       inst = (backend_instruction *) inst->next) {
+   foreach_inst_in_block(backend_instruction, inst, this) {
       fprintf(stderr, "%5d: ", ip);
-      v->dump_instruction(inst);
+      s->dump_instruction(inst);
       ip++;
    }
 }
@@ -163,6 +157,7 @@ cfg_t::cfg_t(exec_list *instructions)
    block_list.make_empty();
    blocks = NULL;
    num_blocks = 0;
+   idom_dirty = true;
 
    bblock_t *cur = NULL;
    int ip = 0;
@@ -178,16 +173,16 @@ cfg_t::cfg_t(exec_list *instructions)
 
    set_next_block(&cur, entry, ip);
 
-   entry->start = (backend_instruction *) instructions->get_head();
-
-   foreach_in_list(backend_instruction, inst, instructions) {
-      cur->end = inst;
-
+   foreach_in_list_safe(backend_instruction, inst, instructions) {
       /* set_next_block wants the post-incremented ip */
       ip++;
 
+      inst->exec_node::remove();
+
       switch (inst->opcode) {
       case BRW_OPCODE_IF:
+         cur->instructions.push_tail(inst);
+
         /* Push our information onto a stack so we can recover from
          * nested ifs.
          */
@@ -202,55 +197,46 @@ cfg_t::cfg_t(exec_list *instructions)
          * instructions.
          */
         next = new_block();
-        next->start = (backend_instruction *)inst->next;
         cur_if->add_successor(mem_ctx, next);
 
         set_next_block(&cur, next, ip);
         break;
 
       case BRW_OPCODE_ELSE:
+         cur->instructions.push_tail(inst);
+
          cur_else = cur;
 
         next = new_block();
-        next->start = (backend_instruction *)inst->next;
+         assert(cur_if != NULL);
         cur_if->add_successor(mem_ctx, next);
 
         set_next_block(&cur, next, ip);
         break;
 
       case BRW_OPCODE_ENDIF: {
-         if (cur->start == inst) {
+         if (cur->instructions.is_empty()) {
             /* New block was just created; use it. */
             cur_endif = cur;
          } else {
             cur_endif = new_block();
-            cur_endif->start = inst;
 
-            cur->end = (backend_instruction *)inst->prev;
             cur->add_successor(mem_ctx, cur_endif);
 
             set_next_block(&cur, cur_endif, ip - 1);
          }
 
+         cur->instructions.push_tail(inst);
+
          if (cur_else) {
             cur_else->add_successor(mem_ctx, cur_endif);
          } else {
+            assert(cur_if != NULL);
             cur_if->add_successor(mem_ctx, cur_endif);
          }
 
-         assert(cur_if->end->opcode == BRW_OPCODE_IF);
-         assert(!cur_else || cur_else->end->opcode == BRW_OPCODE_ELSE);
-
-         cur_if->if_block = cur_if;
-         cur_if->else_block = cur_else;
-
-        if (cur_else) {
-            cur_else->if_block = cur_if;
-            cur_else->else_block = cur_else;
-         }
-
-         cur->if_block = cur_if;
-         cur->else_block = cur_else;
+         assert(cur_if->end()->opcode == BRW_OPCODE_IF);
+         assert(!cur_else || cur_else->end()->opcode == BRW_OPCODE_ELSE);
 
         /* Pop the stack so we're in the previous if/else/endif */
         cur_if = pop_stack(&if_stack);
@@ -269,25 +255,27 @@ cfg_t::cfg_t(exec_list *instructions)
          */
         cur_while = new_block();
 
-         if (cur->start == inst) {
+         if (cur->instructions.is_empty()) {
             /* New block was just created; use it. */
             cur_do = cur;
          } else {
             cur_do = new_block();
-            cur_do->start = inst;
 
-            cur->end = (backend_instruction *)inst->prev;
             cur->add_successor(mem_ctx, cur_do);
 
             set_next_block(&cur, cur_do, ip - 1);
          }
+
+         cur->instructions.push_tail(inst);
         break;
 
       case BRW_OPCODE_CONTINUE:
+         cur->instructions.push_tail(inst);
+
+         assert(cur_do != NULL);
         cur->add_successor(mem_ctx, cur_do);
 
         next = new_block();
-        next->start = (backend_instruction *)inst->next;
         if (inst->predicate)
            cur->add_successor(mem_ctx, next);
 
@@ -295,10 +283,12 @@ cfg_t::cfg_t(exec_list *instructions)
         break;
 
       case BRW_OPCODE_BREAK:
+         cur->instructions.push_tail(inst);
+
+         assert(cur_while != NULL);
         cur->add_successor(mem_ctx, cur_while);
 
         next = new_block();
-        next->start = (backend_instruction *)inst->next;
         if (inst->predicate)
            cur->add_successor(mem_ctx, next);
 
@@ -306,9 +296,14 @@ cfg_t::cfg_t(exec_list *instructions)
         break;
 
       case BRW_OPCODE_WHILE:
-        cur_while->start = (backend_instruction *)inst->next;
+         cur->instructions.push_tail(inst);
 
+         assert(cur_do != NULL && cur_while != NULL);
         cur->add_successor(mem_ctx, cur_do);
+
+         if (inst->predicate)
+            cur->add_successor(mem_ctx, cur_while);
+
         set_next_block(&cur, cur_while, ip);
 
         /* Pop the stack so we're in the previous loop */
@@ -317,13 +312,12 @@ cfg_t::cfg_t(exec_list *instructions)
         break;
 
       default:
+         cur->instructions.push_tail(inst);
         break;
       }
    }
 
-   assert(cur->end);
-
-   cur->end_ip = ip;
+   cur->end_ip = ip - 1;
 
    make_block_array();
 }
@@ -383,6 +377,7 @@ cfg_t::remove_block(bblock_t *block)
 
    this->blocks[this->num_blocks - 1]->num = this->num_blocks - 2;
    this->num_blocks--;
+   idom_dirty = true;
 }
 
 bblock_t *
@@ -397,7 +392,6 @@ void
 cfg_t::set_next_block(bblock_t **cur, bblock_t *block, int ip)
 {
    if (*cur) {
-      assert((*cur)->end->next == block->start);
       (*cur)->end_ip = ip - 1;
    }
 
@@ -420,16 +414,24 @@ cfg_t::make_block_array()
 }
 
 void
-cfg_t::dump(backend_visitor *v)
+cfg_t::dump(backend_shader *s)
 {
+   if (idom_dirty)
+      calculate_idom();
+
    foreach_block (block, this) {
-      fprintf(stderr, "START B%d", block->num);
+      if (block->idom)
+         fprintf(stderr, "START B%d IDOM(B%d)", block->num, block->idom->num);
+      else
+         fprintf(stderr, "START B%d IDOM(none)", block->num);
+
       foreach_list_typed(bblock_link, link, link, &block->parents) {
          fprintf(stderr, " <-B%d",
                  link->block->num);
       }
       fprintf(stderr, "\n");
-      block->dump(v);
+      if (s != NULL)
+         block->dump(s);
       fprintf(stderr, "END B%d", block->num);
       foreach_list_typed(bblock_link, link, link, &block->children) {
          fprintf(stderr, " ->B%d",
@@ -438,3 +440,91 @@ cfg_t::dump(backend_visitor *v)
       fprintf(stderr, "\n");
    }
 }
+
+/* Calculates the immediate dominator of each block, according to "A Simple,
+ * Fast Dominance Algorithm" by Keith D. Cooper, Timothy J. Harvey, and Ken
+ * Kennedy.
+ *
+ * The authors claim that for control flow graphs of sizes normally encountered
+ * (less than 1000 nodes) that this algorithm is significantly faster than
+ * others like Lengauer-Tarjan.
+ */
+void
+cfg_t::calculate_idom()
+{
+   foreach_block(block, this) {
+      block->idom = NULL;
+   }
+   blocks[0]->idom = blocks[0];
+
+   bool changed;
+   do {
+      changed = false;
+
+      foreach_block(block, this) {
+         if (block->num == 0)
+            continue;
+
+         bblock_t *new_idom = NULL;
+         foreach_list_typed(bblock_link, parent, link, &block->parents) {
+            if (parent->block->idom) {
+               if (new_idom == NULL) {
+                  new_idom = parent->block;
+               } else if (parent->block->idom != NULL) {
+                  new_idom = intersect(parent->block, new_idom);
+               }
+            }
+         }
+
+         if (block->idom != new_idom) {
+            block->idom = new_idom;
+            changed = true;
+         }
+      }
+   } while (changed);
+
+   idom_dirty = false;
+}
+
+bblock_t *
+cfg_t::intersect(bblock_t *b1, bblock_t *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->num != b2->num) {
+      while (b1->num > b2->num)
+         b1 = b1->idom;
+      while (b2->num > b1->num)
+         b2 = b2->idom;
+   }
+   assert(b1);
+   return b1;
+}
+
+void
+cfg_t::dump_cfg()
+{
+   printf("digraph CFG {\n");
+   for (int b = 0; b < num_blocks; b++) {
+      bblock_t *block = this->blocks[b];
+
+      foreach_list_typed_safe (bblock_link, child, link, &block->children) {
+         printf("\t%d -> %d\n", b, child->block->num);
+      }
+   }
+   printf("}\n");
+}
+
+void
+cfg_t::dump_domtree()
+{
+   printf("digraph DominanceTree {\n");
+   foreach_block(block, this) {
+      if (block->idom) {
+         printf("\t%d -> %d\n", block->idom->num, block->num);
+      }
+   }
+   printf("}\n");
+}