http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54602
--- Comment #2 from Oleg Endo <olegendo at gcc dot gnu.org> 2012-10-09 22:57:40 UTC --- (In reply to comment #1) > The problem is the "*movsi_pop" insn. It prevents delay-slot stuffing of pop > insns on everything < SH3. This was probably added to avoid pop insns in > interrupt handler functions, which return with 'rte' instead of 'rts', see PR > 53689. > Adding "&& current_function_interrupt" to the condition of the "*movsi_pop" > insn fixes the problem, while maintaining correct code for interrupt handler > functions. > > There is also the following in sh.md: > > (define_delay > (eq_attr "type" "return") > [(and (eq_attr "in_delay_slot" "yes") > (ior (and (eq_attr "interrupt_function" "no") > (eq_attr "type" "!pload,prset")) > (and (eq_attr "interrupt_function" "yes") > (ior > (not (match_test "TARGET_SH3")) > (eq_attr "hit_stack" "no") > (eq_attr "banked" "no"))))) (nil) (nil)]) > > which should actually prevent normal movsi pop insn from slipping into the > delay slot. I've tapped the "hit_stack" test and it is not invoked for > interrupt functions for some reason .... ... because this define_delay applies for return insns, not for mem loads. D'uh. :T
