On Nov 4, 2011, at 2:07 PM, Alan Modra wrote: > Lots of bugs here. Most of them in TARGET_SPE_ABI code, but some also > for other ABIs.
... Another bug we're running into here is an unwarranted move of the sp restore prior to register fetches despite an attempt at preventing that with a stack_tie instruction (VxWorks target). http://gcc.gnu.org/ml/gcc/2011-03/msg00123.html The failure can still be exposed on mainline with a minor adjustment to the C testcase quoted in the msg. I have been looking into this issue again recently and I'm convinced that the the current stack_tie insn just is not quite adequate. I have a local patch which fixes the case at hand but I'm still unclear about a couple of aspects. There are lots of subtle inter-section dependencies and redundancies in emit_epilogue, which has grown pretty difficult to understand IMHO. I can see two tracks to improve things in this area: - Concentrate on the sp-move problem at hand and get to an agreement on the proper correction for it. I'll be happy to send my current patch with comments and questions for that. - Start working on reducing emit_epilogue's complexity, e.g. by first extracting portions of it out into separate functions, then see what concepts we can identify and materialize to simplify the sequence organization as a whole. I did an extraction attempt as an exercise to better understand the various sections and dependencies. It did help me and I'd be happy to provide patches gather feedback etc if you believe it might be useful to others. Olivier
