sibcall, sibcall_pop, sibcall_value and sibcall_pop_value named patterns not documented

["Dave Korn" <dave dot korn at artimi dot com>]

  At least as far as I have been able to find, there's no mention of these
anywhere in any version of the internals manual.

  I think they should at least be mentioned and the similarities/differences
to ordinary call/call_value/call_pop/call_value_pop explained, because they
actually turn out to be vital, but the dependency isn't immediately obvious:
if you implement epilogue and sibcall_epilogue (as schedulable RTL generated
by expanders, rather than as insn patterns with code-generating functions
which just opencode the epilogue at assembly-generation time), and you haven't
implemented sibcall/sibcall_value patterns, the compiler falls back to the
ordinary call/call_value pattern and then you get that abort in
propagate_one_insn when it finds itself trying to delete an epilogue insn.  I
initially thought it would be sufficient to just test for SIBLING_CALL_P
(insn) in the code generation for the call/call_XXX patterns, but it turns out
to be subtly not so.

  This may be only apply on architectures that have a link register for all I
know.  It seems at first glance to have been because of using an ordinary call
pattern (which clobbers LR) to match a sibcall (which does not, in fact, use
LR at all); it was the reload of the LR from the stack prior to the sibcall
jump that got deleted, but then again it was only a simple testcase from the
testsuite and there weren't any other items on the stack frame in that
instance, so I can't be sure.

  Since I have only deduced the existence of these patterns from stumbling
across HAVE_sibcall etc. in calls.c, I won't immediately volunteer to write
the doco patch.  Shall I file a bugzilla, or is there some documentation about
the sibcall patterns somewhere that I missed?


    cheers,
      DaveK
-- 
Can't think of a witty .sigline today....