nvc0/ir: implement better placement of texture barriers

Put them before first uses instead of right after the texturing
instruction and cull unnecessary barriers.

7 files changed, 58 insertions, 6 deletions

diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir.h b/src/gallium/drivers/nv50/codegen/nv50_ir.h
index c299cab3f52..da9042066ad 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir.h
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir.h
@@ -135,6 +135,8 @@ enum operation
    OP_LAST
 };
 
+// various instruction-specific modifier definitions Instruction::subOp
+// MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 #define NV50_IR_SUBOP_MUL_HIGH     1
 #define NV50_IR_SUBOP_EMIT_RESTART 1
 #define NV50_IR_SUBOP_LDC_IL       1
@@ -143,6 +145,7 @@ enum operation
 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 #define NV50_IR_SUBOP_EMU_PRERET   1
 #define NV50_IR_SUBOP_TEXBAR(n)    n
+#define NV50_IR_SUBOP_MOV_FINAL    1
 
 enum DataType
 {
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_graph.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir_graph.cpp
index 90147b6bfb8..f1bff973636 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_graph.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_graph.cpp
@@ -21,6 +21,9 @@
  */
 
 #include "nv50_ir_graph.h"
+#include <limits>
+#include <list>
+#include "nv50_ir.h"
 
 namespace nv50_ir {
 
@@ -228,8 +231,7 @@ public:
    virtual bool end() const { return pos >= count; }
    virtual void next() { if (pos < count) ++pos; }
    virtual void *get() const { return nodes[pos]; }
-
-   void reset() { pos = 0; }
+   virtual void reset() { pos = 0; }
 
 protected:
    Graph::Node **nodes;
@@ -274,6 +276,7 @@ public:
    virtual void *get() const { return nodes[pos]; }
    virtual bool end() const { return pos >= count; }
    virtual void next() { if (pos < count) ++pos; }
+   virtual void reset() { pos = 0; }
 
 private:
    void search(Graph::Node *node, const int sequence)
@@ -389,4 +392,43 @@ void Graph::classifyDFS(Node *curr, int& seq)
    curr->tag = 0;
 }
 
+// @dist is indexed by Node::tag, returns -1 if no path found
+int
+Graph::findLightestPathWeight(Node *a, Node *b, const std::vector<int> &weight)
+{
+   std::vector<int> path(weight.size(), std::numeric_limits<int>::max());
+   std::list<Node *> nodeList;
+   const int seq = nextSequence();
+
+   path[a->tag] = 0;
+   for (Node *c = a; c && c != b;) {
+      const int p = path[c->tag] + weight[c->tag];
+      for (EdgeIterator ei = c->outgoing(); !ei.end(); ei.next()) {
+         Node *t = ei.getNode();
+         if (t->getSequence() < seq) {
+            if (path[t->tag] == std::numeric_limits<int>::max())
+               nodeList.push_front(t);
+            if (p < path[t->tag])
+               path[t->tag] = p;
+         }
+      }
+      c->visit(seq);
+      Node *next = NULL;
+      for (std::list<Node *>::iterator n = nodeList.begin();
+           n != nodeList.end(); ++n) {
+         if (!next || path[(*n)->tag] < path[next->tag])
+            next = *n;
+         if ((*n) == c) {
+            // erase visited
+            n = nodeList.erase(n);
+            --n;
+         }
+      }
+      c = next;
+   }
+   if (path[b->tag] == std::numeric_limits<int>::max())
+      return -1;
+   return path[b->tag];
+}
+
 } // namespace nv50_ir
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_graph.h b/src/gallium/drivers/nv50/codegen/nv50_ir_graph.h
index cfa73c3dd1a..9ef317f943c 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_graph.h
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_graph.h
@@ -24,6 +24,7 @@
 #define __NV50_IR_GRAPH_H__
 
 #include "nv50_ir_util.h"
+#include <vector>
 
 namespace nv50_ir {
 
@@ -165,6 +166,9 @@ public:
 
    void classifyEdges();
 
+   // @weights: indexed by Node::tag
+   int findLightestPathWeight(Node *, Node *, const std::vector<int>& weights);
+
 private:
    void classifyDFS(Node *, int&);
 
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_lowering_nv50.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir_lowering_nv50.cpp
index 30d8acee3bc..27373b4cc47 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_lowering_nv50.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_lowering_nv50.cpp
@@ -998,6 +998,7 @@ NV50LoweringPreSSA::handleEXPORT(Instruction *i)
          int id = i->getSrc(0)->reg.data.offset / 4; // in 32 bit reg units
 
          i->op = OP_MOV;
+         i->subOp = NV50_IR_SUBOP_MOV_FINAL;
          i->src(0).set(i->src(1));
          i->setSrc(1, NULL);
          i->setDef(0, new_LValue(func, FILE_GPR));
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_peephole.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir_peephole.cpp
index 7d0ebc69710..5bc3a450779 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_peephole.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_peephole.cpp
@@ -33,7 +33,7 @@ namespace nv50_ir {
 bool
 Instruction::isNop() const
 {
-   if (op == OP_CONSTRAINT || op == OP_PHI)
+   if (op == OP_PHI || op == OP_SPLIT || op == OP_MERGE || op == OP_CONSTRAINT)
       return true;
    if (terminator || join) // XXX: should terminator imply flow ?
       return false;
@@ -1855,7 +1855,10 @@ FlatteningPass::visit(BasicBlock *bb)
    Instruction *insn = bb->getExit();
    if (insn && insn->op == OP_JOIN && !insn->getPredicate()) {
       insn = insn->prev;
-      if (insn && !insn->getPredicate() && !insn->asFlow() && !insn->isNop()) {
+      if (insn && !insn->getPredicate() &&
+          !insn->asFlow() &&
+          insn->op != OP_TEXBAR &&
+          !insn->isNop()) {
          insn->join = 1;
          bb->remove(bb->getExit());
          return true;
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_ra.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir_ra.cpp
index 5d2aa587a31..77edaa6067a 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_ra.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_ra.cpp
@@ -1615,7 +1615,6 @@ GCRA::resolveSplitsAndMerges()
          v->join = v;
          reg += v->reg.size;
       }
-      delete_Instruction(prog, split);
    }
    splits.clear();
 
@@ -1630,7 +1629,6 @@ GCRA::resolveSplitsAndMerges()
          v->join = v;
          reg += v->reg.size;
       }
-      delete_Instruction(prog, merge);
    }
    merges.clear();
 }
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_util.h b/src/gallium/drivers/nv50/codegen/nv50_ir_util.h
index f348f1811b8..a0020789c11 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_util.h
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_util.h
@@ -91,6 +91,7 @@ public:
    virtual void next() = 0;
    virtual void *get() const = 0;
    virtual bool end() const = 0; // if true, get will return 0
+   virtual void reset() { assert(0); } // only for graph iterators
 };
 
 typedef std::auto_ptr<Iterator> IteratorRef;