[RFC] RTL prologue/epilogue for SPARC

[Eric Botcazou <ebotcazou at libertysurf dot fr>]

Hi,

This is a first attempt at implementing RTL prologue/epilogue on SPARC, at 
least partially.  I wrote it a few months ago, after David had proposed the 
idea.  It has passsed bootstrap/regtest on sparc-sun-solaris2.8 
and sparc64-sun-solaris2.9 on mainline today.

I immediately stumbled upon the following difficulty: by the time RTL 
prologue/epilogue are normally expanded, we don't know yet if the function 
is truly leaf (uses only leaf regs) because, for example, regrename has not 
yet been run.  My solution was to delay the expansion until after regname, 
which is just before sched2.  I also needed to fix two minor problems in 
reorg.c which relate to the interaction of return and delay slots.  Nothing 
controversial I think.

The patch makes it possible to get rid of a fair amount of duplicated code in 
the back-end between sibcalls and returns; in particular, all the 'return 
peepholes' were essentially duplicated by the sibcall code, so I removed 
them.

I didn't run any benchmark but I diffed the assembly emitted for the 
gcc.c-torture/compile testsuite, both for UltraSPARC and UltraSPARC III: the 
sibcalls were correctly emitted, the return delay slots filled, scheduling 
was improved in the prologue for leaf functions.  The only problem I found 
is that, in a few cases, delay slot in insns right before the return were 
not correctly filled.  Tree-SSA was approaching so I temporarily gave up.

David recently expressed interest in the patch so I'm resurrecting it.


2004-06-14  Eric Botcazou  <ebotcazou@libertysurf.fr>

	RTL prologue/epilogue for SPARC
	* config/sparc/sparc-protos.h (sparc_emitting_epilogue): Delete.
	(sparc_skip_caller_unimp): Likewise.
	(load_pic_register): Likewise.
	(leaf_return_peephole_ok): Likewise.
	(compute_frame_size): Rename into sparc_compute_frame_size.
	(sparc_expand_prologue): New prototype.
	(sparc_expand_epilogue): Likewise.
	(output_return): Likewise.
	(eligible_for_epilogue_delay): Rename into eligible_for_return_delay.
	* config/sparc/sparc.h (INITIAL_ELIMINATION_OFFSET): Adjust
	call to compute_frame_size.  Move comment up.
	(DELAY_SLOTS_FOR_EPILOGUE): Delete.
	(ELIGIBLE_FOR_EPILOGUE_DELAY): Likewise.
	(EPILOGUE_USES): Return true for %g1 if the function uses EH return.
	* config/sparc/sparc.md (UNSPECV_SAVEW): New constant.
	(type attribute): Add 'return' and 'savew'.
	(eligible_for_return_delay): New attribute.
	(return): New delay_slot.
	(sibcall_epilogue): Call sparc_expand_epilogue.
	(prologue): Call sparc_expand_prologue.  Move up.
	(save_register_window): New expander.
	(save_register_windowsi): New pattern.
	(save_register_windowdi): New pattern.
	(epilogue): New expander.
	(return_internal): New pattern.
	(Return peepholes): Delete.
	* config/sparc/sparc.c  (SIBCALL_SLOT_EMPTY_P): New macro.
	(sparc_emitting_epilogue): Delete.
	(sparc_skip_caller_unimp): Likewise.
	(sparc_sr_alias_set): New global variable.
	(frame_base_name): Delete.
	(frame_base_reg): New global variable.
	(sparc_override_options): Get new alias set for save/restore.
	(leaf_return_peephole_ok): Delete.
	(eligible_for_epilogue_delay): Rename into eligible_for_return_delay.
	Factor out code into eligible_for_restore_insn_delay.
	(eligible_for_restore_insn_delay): New function extraced from above.
	Use IN_UNCOND_BRANCH_DELAY_TRUE instead of IN_BRANCH_DELAY_TRUE.
	(eligible_for_sibcall_delay): Use SIBCALL_SLOT_EMPTY_P.
	Factor out code into eligible_for_restore_insn_delay.
	(load_pic_register): Make static.  Remove check.
	(save_regs): Delete.
	(restore_regs): Likewise.
	(compute_frame_size): Rename into sparc_compute_frame_size.
	Rename leaf_function into leaf_function_p.
	(build_big_number): Delete.
	(save_or_restore_regs): New function.
	(emit_save_regs): Likewise.
	(emit_restore_regs): Likewise.
	(emit_stack_pointer_increment ): Likewise.
	(emit_stack_pointer_decrement): Likewise.
	(sparc_expand_prologue): Likewise.
	(sparc_function_prologue): Rename into sparc_asm_function_prologue.
	Remove all code to emit instructions.
	(sparc_expand_epilogue): New function.
	(sparc_function_epilogue): Rename into sparc_asm_function_epilogue.
	Remove all code to emit instructions.
	(output_restore): New function.
	(output_return): New function.
	(output_sibcall): Factor out code into output_restore.
	(print_operand): Adjust for frame_base_reg.
	* target.h (struct gcc_target): New field 'late_rtl_prologue_epilogue'.
	* target-def.h (TARGET_LATE_RTL_PROLOGUE_EPILOGUE): New define.
	(TARGET_INITIALIZER): Add it.
	* passes.c (rest_of_compilation): Set current_function_uses_only_leaf_regs
	before sched2.  If target has  'late_rtl_prologue_epilogue', emit RTL
	prologue/epilogue right before sched2.
	* reorg.c (return_insn_p): New predicate.
	(find_end_label): Use it.
	(relax_delay_slots): Do not thread an unconditional jump that points
	to the end return label.


-- 
Eric Botcazou

Attachment: sparc_rtl_prologue-6.diff.gz
Description: GNU Zip compressed data