Hi, the overall description how IPA optimization is structured (describing difference in between simple_ipa_opt_pass and ipa_opt_pass) is in http://arxiv.org/pdf/1010.2196v2.pdf and also in lto.texi. There is not exactly a tutrial, but generally simple_ipa_opt_pass is easier to start with since you can directly touch function bodies, while ipa_opt_pass needs you to split everything into analysis/propagation/modification stages. So you probably want to start with implementing simple_ipa_opt_pass that does what you need while keeping in mind that you may later turn it intp ipa_opt_pass.
You can look at the existing simple_ipa_opt_passes to see how they are structured. probably the easiest is pass_data_ipa_increase_alignment in tree-vectorizer.c that consist of one function walking all variables. Honza
