Safety analysis is becoming more and more important in a wide class of systems. In the automotive field, the recent ISO26262 foresees safety analysis to be performed at different levels: system, software and hardware. The assessment of architecture with respect to safety is typically better understood at system and HW levels, while an equivalent analysis at SW level has not such an established background. In literature, approaches exist to handle specific activities related to the safety assessment of software, but they are typically not so well integrated within a more general assessment and certification process. Recent safety standards put more and more emphasis on software-level safety analysis, therefore calling for a precise methodology for the assessment of software architectures. While ISO26262 requirements prescribe safety analysis of the software architecture, clear guidelines on how it should be performed are not provided, thus leaving an important gap for its industrial adoption. In this paper we provide our view on how such analysis should be performed, through the identification of well defined and repeatable activities, thus providing our contribution to a timely problem of great relevance in the automotive domain.