Re: tail call optimizations

[Richard Henderson <rth at cygnus dot com>]

Here's a revised patch that works on x86, as well as
cleaning up some of the BLOCK note handling now that
I understand it a bit more.


r~

	* Makefile.in (OBJS): Add sibcall.o.
	(sibcall.o): New.
	* sibcall.c: New file.
	* calls.c (FUNCTION_OK_FOR_SIBCALL): Provide default.
	(ECF_IS_CONST, ECF_NOTHROW, ECF_SIBCALL): New.
	(emit_call_1): Replace `is_const' and `nothrow' with `ecf_flags'.
	Emit sibcall patterns when requested.  Update all callers.
	(expand_call): Generate CALL_PLACEHOLDER insns when tail call
	elimination seems feasable.
	* final.c (leaf_function_p): Sibling calls don't discount being
	a leaf function.
	* flow.c (HAVE_sibcall_epilogue): Provide default.
	(find_basic_blocks_1): Sibling calls don't throw.
	(make_edges): Make edge from sibling call to EXIT.
	(propagate_block): Don't remove sibcall_epilogue insns.
	* function.c (prologue, epilogue): Turn into varrays.  Update all uses.
	(sibcall_epilogue): New.
	(fixup_var_refs): Scan CALL_PLACEHOLDER sub-sequences.
	(identify_blocks_1): Likewise.  Break out from ...
	(identify_blocks): ... here.
	(reorder_blocks_1): Scan CALL_PLACEHOLDER.  Break out from ...
	(reorder_blocks): ... here.
	(init_function_for_compilation): Zap prologue/epilogue as varrays.
	(record_insns): Extend a varray instead of mallocing new memory.
	(contains): Read a varray not array of ints.
	(sibcall_epilogue_contains): New.
	(thread_prologue_and_epilogue_insns): Emit and record
	sibcall_epilogue patterns.
	(init_function_once): Allocate prologue/epilogue varrays.
	* genflags.c (gen_insn): Treat sibcall patterns as calls.
	* integrate.c (save_parm_insns): Recurse on CALL_PLACEHOLDER patterns.
	Broken out from ...
	(save_for_inline_nocopy): ... here.
	(copy_insn_list): Recurse on CALL_PLACEHOLDER patterns. 
	Broken out from ...
	(expand_inline_function): ... here.
	(copy_rtx_and_substitute): Handle NOTE_INSN_DELETED_LABEL.
	(subst_constants): Handle 'n' formats.
	* jump.c (jump_optimize_minimal): New.
	(jump_optimize_1): New arg `minimal'; update callers.  Elide most
	optimizations if it's set.
	* rtl.c (copy_rtx): Do copy jump & call for insns.
	* rtl.h (struct rtx_def): Document use of jump and call for insns.
	(SIBLING_CALL_P): New.
	(sibcall_use_t): New.
	* toplev.c (rest_of_compilation): Do init_EXPR_INSN_LIST_cache earlier.
	Invoke optimize_sibling_and_tail_recursive_calls.
	* tree.c (lang_safe_for_unsave): New.
	(safe_for_unsave): New.
	* tree.h (lang_safe_for_unsave, safe_for_unsave): Declare.

	* alpha.h (FUNCTION_OK_FOR_SIBCALL): New.
	* alpha.md (sibcall, sibcall_value, sibcall_epilogue): New.
	(*sibcall_osf_1, *sibcall_value_osf_1): New.

	* i386.c (call_insn_operand): Always allow SYMBOL_REF,
	care for HALF_PIC_P.
	(expander_call_insn_operand): Remove.
	(ix86_expand_epilogue): New arg `emit_return' to control return insn.
	* i386.h (PREDICATE_CODES): Update.
	* i386.md (all call expanders): Remove predicates, remove special	
	handling for half-pic.
	(*call_1, *call_value_1): Handle SIBLING_CALL_P insns.
	(*call_pop_pic2, *call_pic2, *call_value_pop_2, *call_value_2): Remove.
	(sibcall, sibcall_value, sibcall_epilogue): New.

cp/
	* optimize.c (copy_body_r): Don't abort on UNSAVE_EXPR.
	* tree.c (init_tree): Set lang_safe_for_unsave.
	(cp_safe_for_unsave): New.

d-sibcall-2.bz2