Created attachment 304 [details]
callstack1

Here is the code snippet (part of the function body):


void solve(const VectorXd& initial_alpha) {
  typedef Eigen::SparseMatrix<double, Eigen::RowMajor> SparseMatrixType;
  typedef Eigen::Triplet<double> DoubleTriplet; 

  SparseMatrixType laplacian;
  fill_laplacian(laplacian); 
/* laplacian's size is 2450x2450, it is symmetric, positive definite */

  SparseMatrixType D(laplacian.rows(), laplacian.cols());
  vector<DoubleTriplet> D_triplets;
  const double lambda = 0.1;

  for(int i = 0; i < laplacian.rows(); ++i) {
    double value = /* calculated value */;
    D_triplets.push_back(DoubleTriplet(i, i, value));
  }

  D.setFromTriplets(D_triplets.begin(), D_triplets.end());

  VectorXd b = ;

  ConjugateGradient<SparseMatrixType, Upper> cg;
  cg.setTolerance(1e-15);
  cg.setMaxIterations(500);

  cg.compute(laplacian + lambda * D); 
  VectorXd alpha = cg.solve(lambda * D * initial_alpha); // <- crash
}

I've observed 2 types of crashes.

1. When both cg.compute() and cg.solve() are called with expressions, visual studio gives Access violation reading location <some address>.

Crash is in the file Eigen/src/SparseCore/SparseMatrix.h, line 950, method
SparseMatrix<Scalar,_Options,_Index>::InnerIterator::index()

m_id = -544149818

Call stack is in the attached file callstack1.txt 

2. If I replace an argument of solve() with pre-calculated variable:

    VectorXd b = lambda * D * initial_alpha;
    cg.compute(laplacian + lambda * D); 
    VectorXd alpha = cg.solve(lambda * D * initial_alpha); // <- crash

then I've got failed Eigen assertion:

Assertion failed: 
(i>=0) && 
(
 ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && i<xpr.rows()) ||
 ((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && i<xpr.cols())
), file c:\eigen\eigen\src/Core/Block.h, line 280

here i = 1007402692;
xpr.rows() return 2450, xpr.cols() returns 1.
Call stack is in the attached file callstack2.txt


3. If I replace an argument of cg.compute() with precalculated variable, everything works fine:

  SparseMatrixType A = laplacian + lambda * D;
  cg.compute(A);
  alpha = cg.solve(lambda * D * initial_alpha);

I can attach text files with concrete matrices, if needed.