Dec 30 patch causes large performance regression on ia64

[Bernd Schmidt]

Richard,

do you have any idea why this patch

2000-12-30  Richard Kenner  <kenner@vlsi1.ultra.nyu.edu>

	* alias.c (alias_sets_conflict_p): New function.
	(mems_in_disjoint_alias_sets_p): Use it.
	(readonly_fields_p): Moved from expr.c; check for record type.
	(objects_must_conflict_p): New function.
	* calls.c (expand_call): Use assign_temp as much as possible, use
	readonly variant if assigned once, and don't set memory attributes.
	(emit_library_call_value_1, store_one_arg): Likewise.
	* integrate.c (expand_inline_function): Likewise.
	* stmt.c (expand_asm_operands, expand_return): Likewise.
	* expr.c (copy_blkmode_from_reg, store_constructor): Likewise.
	(store_field, save_noncopied_parts, expand_expr): Likewise.
	(expand_expr_unaligned): Likewise.
	(readonly_fields_p): Moved to alias.c.
	(safe_from_p): Rework handling of SAVE_EXPR.
	MEMs ony conflict if alias sets conflict; likewise for INDIRECT_REF.
	* function.c (struct temp_slot): Delete field ALIAS_SET; add TYPE.
	(assign_stack_for_temp): Use objects_must_confict_p.
	Set all memory attributes from type, if specified.
	(mark_temp_slot): Mark TYPE field.
	* tree.h (alias_sets_conflict_p, readonly_fields_p): New decls.
	(objects_must_conflict_p): Likewise.

	* stmt.c (expand_decl): Don't use assign_stack_temp in error case.
	(add_case_node): No need to copy nodes anymore.

might cause a 20% performance drop when running the 147.vortex benchmark
from SPEC95 on ia64-linux?


Bernd