Work around pow() issue in NVIDIA 331.x drivers

include/GLSLANG/ShaderLang.h

@@ -48,7 +48,7 @@ typedef unsigned int GLenum;
 
 // Version number for shader translation API.
 // It is incremented every time the API changes.
-#define ANGLE_SH_VERSION 136
+#define ANGLE_SH_VERSION 137
 
 typedef enum {
   SH_GLES2_SPEC = 0x8B40,
@@ -188,6 +188,10 @@ typedef enum {
   // compilation process. Pruning coupled with SH_LIMIT_CALL_STACK_DEPTH
   // helps avoid bad shaders causing stack overflows.
   SH_DONT_PRUNE_UNUSED_FUNCTIONS = 0x100000,
+
+  // This flag works around a bug in NVIDIA 331 series drivers related
+  // to pow(x, y) where y is a constant vector.
+  SH_REMOVE_POW_WITH_CONSTANT_EXPONENT = 0x200000,
 } ShCompileOptions;
 
 // Defines alternate strategies for implementing array index clamping.

src/compiler.gypi

@@ -84,6 +84,8 @@
             'compiler/translator/QualifierAlive.h',
             'compiler/translator/RegenerateStructNames.cpp',
             'compiler/translator/RegenerateStructNames.h',
+            'compiler/translator/RemovePow.cpp',
+            'compiler/translator/RemovePow.h',
             'compiler/translator/RenameFunction.h',
             'compiler/translator/ScalarizeVecAndMatConstructorArgs.cpp',
             'compiler/translator/ScalarizeVecAndMatConstructorArgs.h',

src/compiler/translator/Compiler.cpp

@@ -13,6 +13,7 @@
 #include "compiler/translator/ParseContext.h"
 #include "compiler/translator/PruneEmptyDeclarations.h"
 #include "compiler/translator/RegenerateStructNames.h"
+#include "compiler/translator/RemovePow.h"
 #include "compiler/translator/RenameFunction.h"
 #include "compiler/translator/ScalarizeVecAndMatConstructorArgs.h"
 #include "compiler/translator/UnfoldShortCircuitAST.h"
@@ -312,6 +313,11 @@ TIntermNode *TCompiler::compileTreeImpl(const char* const shaderStrings[],
             unfoldShortCircuit.updateTree();
         }
 
+        if (success && (compileOptions & SH_REMOVE_POW_WITH_CONSTANT_EXPONENT))
+        {
+            RemovePow(root);
+        }
+
         if (success && (compileOptions & SH_VARIABLES))
         {
             collectVariables(root);

src/compiler/translator/RemovePow.cpp

+//
+// Copyright (c) 2002-2015 The ANGLE Project Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+// RemovePow is an AST traverser to convert pow(x, y) built-in calls where y is a
+// constant to exp2(y * log2(x)). This works around an issue in NVIDIA 311 series
+// OpenGL drivers.
+//
+
+#include "compiler/translator/RemovePow.h"
+
+#include "compiler/translator/InfoSink.h"
+#include "compiler/translator/IntermNode.h"
+
+namespace
+{
+
+bool IsProblematicPow(TIntermTyped *node)
+{
+    TIntermAggregate *agg = node->getAsAggregate();
+    if (agg != nullptr && agg->getOp() == EOpPow)
+    {
+        ASSERT(agg->getSequence()->size() == 2);
+        return agg->getSequence()->at(1)->getAsConstantUnion() != nullptr;
+    }
+    return false;
+}
+
+// Traverser that converts all pow operations simultaneously.
+class RemovePowTraverser : public TIntermTraverser
+{
+  public:
+    RemovePowTraverser();
+
+    bool visitAggregate(Visit visit, TIntermAggregate *node) override;
+
+    void nextIteration() { mNeedAnotherIteration = false; }
+    bool needAnotherIteration() const { return mNeedAnotherIteration; }
+
+  protected:
+    bool mNeedAnotherIteration;
+};
+
+RemovePowTraverser::RemovePowTraverser()
+    : TIntermTraverser(true, false, false),
+      mNeedAnotherIteration(false)
+{
+}
+
+bool RemovePowTraverser::visitAggregate(Visit visit, TIntermAggregate *node)
+{
+    if (IsProblematicPow(node))
+    {
+        TInfoSink nullSink;
+
+        TIntermTyped *x = node->getSequence()->at(0)->getAsTyped();
+        TIntermTyped *y = node->getSequence()->at(1)->getAsTyped();
+
+        TIntermUnary *log = new TIntermUnary(EOpLog2);
+        log->setOperand(x);
+        log->setLine(node->getLine());
+        log->setType(x->getType());
+
+        TIntermBinary *mul = new TIntermBinary(EOpMul);
+        mul->setLeft(y);
+        mul->setRight(log);
+        mul->setLine(node->getLine());
+        bool valid = mul->promote(nullSink);
+        UNUSED_ASSERTION_VARIABLE(valid);
+        ASSERT(valid);
+
+        TIntermUnary *exp = new TIntermUnary(EOpExp2);
+        exp->setOperand(mul);
+        exp->setLine(node->getLine());
+        exp->setType(node->getType());
+
+        NodeUpdateEntry replacePow(getParentNode(), node, exp, false);
+        mReplacements.push_back(replacePow);
+
+        // If the x parameter also needs to be replaced, we need to do that in another traversal,
+        // since it's parent node will change in a way that's not handled correctly by updateTree().
+        if (IsProblematicPow(x))
+        {
+            mNeedAnotherIteration = true;
+            return false;
+        }
+    }
+    return true;
+}
+
+} // namespace
+
+void RemovePow(TIntermNode *root)
+{
+    RemovePowTraverser traverser;
+    // Iterate as necessary, and reset the traverser between iterations.
+    do
+    {
+        traverser.nextIteration();
+        root->traverse(&traverser);
+        traverser.updateTree();
+    }
+    while (traverser.needAnotherIteration());
+}

src/compiler/translator/RemovePow.h

+//
+// Copyright (c) 2002-2015 The ANGLE Project Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+// RemovePow is an AST traverser to convert pow(x, y) built-in calls where y is a
+// constant to exp2(y * log2(x)). This works around an issue in NVIDIA 311 series
+// OpenGL drivers.
+//
+
+#ifndef COMPILER_TRANSLATOR_REMOVEPOW_H_
+#define COMPILER_TRANSLATOR_REMOVEPOW_H_
+
+class TIntermNode;
+
+void RemovePow(TIntermNode *root);
+
+#endif   // COMPILER_TRANSLATOR_REMOVEPOW_H_