PreconditionChebyshev: reduce number of writes by one

[peterrum]

At the moment, we are modifying temp_vector1 during pre and vmult:

dealii/include/deal.II/lac/precondition.h

Lines 3017 to 3019 in 978b425

	std::memset(temp_vector1.begin() + start_range,
	0,
	sizeof(Number) * (end_range - start_range));

dealii/include/deal.II/lac/precondition.h

Lines 3010 to 3012 in 978b425

	matrix.vmult(
	temp_vector1,
	solution,

and solution_old during post:

dealii/include/deal.II/lac/precondition.h

Lines 2822 to 2824 in 978b425

	solution_old[i] =
	factor1_plus_1 * solution[i] - factor1 * solution_old[i] +
	factor2 * matrix_diagonal_inverse[i] * (rhs[i] - tmp_vector[i]);

Modifying temp_vector1 during post and appropriate swappings allow to reduce the write operations by one.

[kronbichler]

This is a good improvement, thank you. I do not understand how this could be overlooked previously.

[kronbichler]

I think this comment is wrong. It is not the read that gets avoid (as we read solution_old below), but we can avoid writing temp_vector back to memory by replacing its volatile content (only needed in the subsequent formulas) by over-writing it with the useful content that previously went into solution_old.

[peterrum]

I have updated the comment.

[kronbichler]

/rebuild