kc3-lang/angle/src/compiler/translator/BreakVariableAliasingInInnerLoops.cpp

Branch : - branch - main - tag -

• Show log

  Commit

• Author : Olli Etuaho
  Date : 2017-07-05 17:02:55
  Hash : 3ec75686
  Message : Collect AST transform utilities to a separate file Collect static functions that are used to create nodes in AST transformations into a single file. BUG=angleproject:1490 TEST=angle_unittests Change-Id: I6f87422988fa088f2f4b48986e378a2909705cb7

• src/compiler/translator/BreakVariableAliasingInInnerLoops.cpp

• //
  // Copyright (c) 2016 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.
  //
  
  // BreakVariableAliasingInInnerLoops.h: To optimize simple assignments, the HLSL compiler frontend
  //      may record a variable as aliasing another. Sometimes the alias information gets garbled
  //      so we work around this issue by breaking the aliasing chain in inner loops.
  
  #include "BreakVariableAliasingInInnerLoops.h"
  
  #include "compiler/translator/IntermNode_util.h"
  #include "compiler/translator/IntermTraverse.h"
  
  // A HLSL compiler developer gave us more details on the root cause and the workaround needed:
  //     The root problem is that if the HLSL compiler is applying aliasing information even on
  //     incomplete simulations (in this case, a single pass). The bug is triggered by an assignment
  //     that comes from a series of assignments, possibly with swizzled or ternary operators with
  //     known conditionals, where the source is before the loop.
  //     So, a workaround is to add a +0 term to variables the first time they are assigned to in
  //     an inner loop (if they are declared in an outside scope, otherwise there is no need).
  //     This will break the aliasing chain.
  
  // For simplicity here we add a +0 to any assignment that is in at least two nested loops. Because
  // the bug only shows up with swizzles, and ternary assignment, whole array or whole structure
  // assignment don't need a workaround.
  
  namespace sh
  {
  
  namespace
  {
  
  class AliasingBreaker : public TIntermTraverser
  {
    public:
      AliasingBreaker() : TIntermTraverser(true, false, true) {}
  
    protected:
      bool visitBinary(Visit visit, TIntermBinary *binary)
      {
          if (visit != PreVisit)
          {
              return false;
          }
  
          if (mLoopLevel < 2 || !binary->isAssignment())
          {
              return true;
          }
  
          TIntermTyped *B = binary->getRight();
          TType type      = B->getType();
  
          if (!type.isScalar() && !type.isVector() && !type.isMatrix())
          {
              return true;
          }
  
          if (type.isArray() || IsSampler(type.getBasicType()))
          {
              return true;
          }
  
          // We have a scalar / vector / matrix assignment with loop depth 2.
          // Transform it from
          //    A = B
          // to
          //    A = (B + typeof<B>(0));
  
          TIntermBinary *bPlusZero = new TIntermBinary(EOpAdd, B, CreateZeroNode(type));
          bPlusZero->setLine(B->getLine());
  
          binary->replaceChildNode(B, bPlusZero);
  
          return true;
      }
  
      bool visitLoop(Visit visit, TIntermLoop *loop)
      {
          if (visit == PreVisit)
          {
              mLoopLevel++;
          }
          else
          {
              ASSERT(mLoopLevel > 0);
              mLoopLevel--;
          }
  
          return true;
      }
  
    private:
      int mLoopLevel = 0;
  };
  
  }  // anonymous namespace
  
  void BreakVariableAliasingInInnerLoops(TIntermNode *root)
  {
      AliasingBreaker breaker;
      root->traverse(&breaker);
  }
  
  }  // namespace sh