add garbage collector for 64 bit windows and linux

6 files changed, 8271 insertions, 0 deletions

diff --git a/acompact.asm b/acompact.asm
new file mode 100644
index 0000000..5db500e
--- /dev/null
+++ b/acompact.asm
@@ -0,0 +1,1156 @@
+
+; mark used nodes and pointers in argument parts and link backward pointers
+
+	mov	rax,qword ptr heap_size_65
+	shl	rax,6
+	mov	qword ptr heap_size_64_65,rax 
+
+	mov	rax,qword ptr caf_list
+
+	test	qword ptr flags,4096
+	jne	pmarkr
+
+	test	rax,rax 
+	je	end_mark_cafs
+
+mark_cafs_lp:
+	push	(-8)[rax]
+
+	lea	rsi,8[rax]
+	mov	rax,qword ptr [rax]
+	lea	rcx,[rsi+rax*8]
+
+	mov	qword ptr end_vector,rcx 
+
+	call	rmark_stack_nodes
+	
+	pop	rax
+	test	rax,rax 
+	jne	mark_cafs_lp
+
+end_mark_cafs:
+	mov	rsi,qword ptr stack_p
+
+	mov	rcx,qword ptr stack_top
+	mov	qword ptr end_vector,rcx 
+
+	call	rmark_stack_nodes
+
+	call	add_mark_compact_garbage_collect_time
+	
+	jmp	compact_heap
+
+pmarkr:
+	test	rax,rax 
+	je	end_rmarkp_cafs
+
+rmarkp_cafs_lp:
+	push	(-8)[rax]
+
+	lea	rsi,8[rax]
+	mov	rax,qword ptr [rax]
+	lea	rcx,[rsi+rax*8]
+
+	mov	qword ptr end_vector,rcx 
+
+	call	rmarkp_stack_nodes
+	
+	pop	rax
+	test	rax,rax 
+	jne	rmarkp_cafs_lp
+
+end_rmarkp_cafs:
+	mov	rsi,qword ptr stack_p
+
+	mov	rcx,qword ptr stack_top
+	mov	qword ptr end_vector,rcx 
+
+	call	rmarkp_stack_nodes
+
+	call	add_mark_compact_garbage_collect_time
+	
+	jmp	compact_heap
+
+	include	acompact_rmark.asm
+
+	include	acompact_rmark_prefetch.asm
+
+; compact the heap
+
+compact_heap:
+
+	mov	rcx,offset finalizer_list
+	mov	rdx,offset free_finalizer_list
+
+	mov	rbp,qword ptr [rcx]
+determine_free_finalizers_after_compact1:
+	lea	r9,__Nil-8
+	cmp	rbp,r9
+	je	end_finalizers_after_compact1
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rbp 
+	mov	rbx,rax 
+	and	rax,31*9
+	shr	rbx,8
+	mov	esi,dword ptr (bit_set_table2)[rax]
+	test	esi,dword ptr [rdi+rbx*4]
+	je	finalizer_not_used_after_compact1
+
+	mov	rax,qword ptr [rbp]
+	mov	rsi,rbp 
+	jmp	finalizer_find_descriptor
+
+finalizer_find_descriptor_lp:
+	and	rax,-4
+	mov	rsi,rax 
+	mov	rax,qword ptr [rax]
+finalizer_find_descriptor:
+	test	rax,1
+	jne	finalizer_find_descriptor_lp
+
+	mov	qword ptr [rsi],offset e____system__kFinalizerGCTemp+2
+
+	cmp	rbp,rcx 
+	ja	finalizer_no_reverse
+
+	mov	rax,qword ptr [rbp]
+	lea	rsi,1[rcx]
+	mov	qword ptr [rbp],rsi 
+	mov	qword ptr [rcx],rax 
+
+finalizer_no_reverse:
+	lea	rcx,8[rbp]
+	mov	rbp,qword ptr 8[rbp]
+	jmp	determine_free_finalizers_after_compact1
+
+finalizer_not_used_after_compact1:
+	mov	qword ptr [rbp],offset e____system__kFinalizerGCTemp+2
+
+	mov	qword ptr [rdx],rbp 
+	lea	rdx,8[rbp]
+
+	mov	rbp ,qword ptr 8[rbp]
+	mov	qword ptr [rcx],rbp 
+
+	jmp	determine_free_finalizers_after_compact1
+
+end_finalizers_after_compact1:
+	mov	qword ptr [rdx],rbp 
+
+	mov	rcx,qword ptr finalizer_list
+	lea	r9,__Nil-8
+	cmp	rcx,r9
+	je	finalizer_list_empty
+	test	rcx,3
+	jne	finalizer_list_already_reversed
+	mov	rax ,qword ptr [rcx]
+	mov	qword ptr [rcx],offset finalizer_list+1
+	mov	qword ptr finalizer_list,rax 
+finalizer_list_already_reversed:
+finalizer_list_empty:
+
+	mov	rbp ,offset free_finalizer_list
+	lea	r9,__Nil-8
+	cmp	qword ptr [rbp],r9
+	je	free_finalizer_list_empty
+
+	mov	qword ptr end_vector,offset free_finalizer_list+8
+
+	test	qword ptr flags,4096
+	je	no_pmarkr
+	call	rmarkp_stack_nodes
+	jmp	free_finalizer_list_empty
+no_pmarkr:
+	call	rmark_stack_nodes
+
+free_finalizer_list_empty:
+
+	mov	rax,qword ptr heap_size_65
+	mov	rbx,rax 
+	shl	rbx,6
+
+	add	rbx,qword ptr heap_p3
+
+	mov	qword ptr end_heap_p3,rbx 
+
+	add	rax,3
+	shr	rax,2
+	mov	r12,rax
+	
+	mov	r8,qword ptr heap_vector
+
+	lea	rbx,4[r8]
+	neg	rbx 
+	mov	qword ptr neg_heap_vector_plus_4,rbx 
+
+	mov	rdi,qword ptr heap_p3
+	xor	rsi,rsi 
+	jmp	skip_zeros
+
+; %rax ,%rcx ,%rbp : free
+find_non_zero_long:
+skip_zeros:
+	sub	r12,1
+	jc	end_move
+	mov	esi,dword ptr [r8]
+	add	r8,4
+	test	rsi,rsi 
+	je	skip_zeros
+; %rbp : free
+end_skip_zeros:
+	mov	rbp,qword ptr neg_heap_vector_plus_4
+
+	add	rbp,r8
+
+	shl	rbp,6
+	add	rbp,qword ptr heap_p3
+
+bsf_and_copy_nodes:
+	movzx	rax,sil
+	test	rax,rax
+	jne	found_bit1
+	movzx	rcx,si
+	shr	rcx,8
+	jne	found_bit2
+	mov	rax,rsi 
+	and	rax,0ff0000h
+	jne	found_bit3
+	mov	rcx,rsi
+	shr	rcx,24
+	movzx	rcx,byte ptr first_one_bit_table[rcx*1]
+	add	rcx,24
+	jmp	copy_nodes
+
+found_bit3:
+	shr	rax,16
+	movzx	rcx,byte ptr first_one_bit_table[rax*1]
+	add	rcx,16
+	jmp	copy_nodes
+
+found_bit2:
+	movzx	rcx,byte ptr first_one_bit_table[rcx*1]
+	add	rcx,8
+	jmp	copy_nodes
+
+found_bit1:
+	movzx	rcx,byte ptr first_one_bit_table[rax*1]
+
+copy_nodes:
+	mov	rax,qword ptr [rbp+rcx*8]
+	shr	esi,1
+	lea	rbp,8[rbp+rcx*8]
+	shr	esi,cl
+	mov	rcx,rbp
+
+	dec	rax
+
+	test	rax,2
+	je	begin_update_list_2
+
+move_argument_part:
+	mov	rbx,qword ptr (-18)[rax]
+	sub	rax,2
+
+	test	rbx,1
+	je	end_list_2
+find_descriptor_2:
+	and	rbx,-4
+	mov	rbx,qword ptr [rbx]
+	test	rbx,1
+	jne	find_descriptor_2
+
+end_list_2:
+	mov	rdx,rbx 
+	movzx	rbx,word ptr (-2)[rbx]
+	cmp	rbx,256
+	jb	no_record_arguments
+
+	movzx	rdx,word ptr (-2+2)[rdx]
+	sub	rdx,2
+	jae	copy_record_arguments_aa
+
+	sub	rbx,256+3
+
+copy_record_arguments_all_b:
+	push	rbx 
+	mov	rbx,qword ptr heap_vector
+
+update_up_list_1r:
+	mov	rdx,rax 
+	add	rax,qword ptr neg_heap_p3
+
+	push	rcx 
+
+	mov	rcx,rax 
+
+	shr	rax,8
+	and	rcx,31*8
+
+	mov	ecx,dword ptr bit_set_table2[rcx*1]
+	mov	eax,dword ptr [rbx+rax*4]
+
+	and	rax,rcx 
+
+	pop	rcx 
+	je	copy_argument_part_1r
+
+	mov	rax,qword ptr [rdx]
+	mov	qword ptr [rdx],rdi 
+	sub	rax,3
+	jmp	update_up_list_1r
+
+copy_argument_part_1r:
+	mov	rax,qword ptr [rdx]
+	mov	qword ptr [rdx],rdi 
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	shr	rax,3
+
+	mov	rbx,rax
+	and	rbx,31
+	cmp	rbx,1
+	jae	bit_in_this_word
+
+	dec	r12
+	mov	esi,dword ptr [r8]
+	add	r8,4
+
+	mov	rbp,qword ptr neg_heap_vector_plus_4
+	add	rbp,r8
+	shl	rbp,6
+	add	rbp,qword ptr heap_p3
+
+bit_in_this_word:
+	shr	esi,1
+	add	rbp,8
+
+	pop	rbx 
+
+copy_b_record_argument_part_arguments:
+	mov	rax,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+	sub	rbx,1
+	jnc	copy_b_record_argument_part_arguments
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+copy_record_arguments_aa:
+	sub	rbx,256+2
+	sub	rbx,rdx 
+	
+	push	rbx 
+	push	rdx 
+
+update_up_list_2r:
+	mov	rdx,rax 
+	mov	rax,qword ptr [rdx]
+	mov	rbx,3
+	and	rbx,rax 
+	sub	rbx,3
+	jne	copy_argument_part_2r
+
+	mov	qword ptr [rdx],rdi 
+	sub	rax,3
+	jmp	update_up_list_2r
+
+copy_argument_part_2r:
+	mov	qword ptr [rdx],rdi 
+	cmp	rax,rcx 
+	jb	copy_record_argument_2
+
+	cmp	rax,qword ptr end_heap_p3
+	jae	copy_record_argument_2
+
+	mov	rdx,rax 
+	mov	rax,qword ptr [rdx]
+	lea	rbx,1[rdi]
+	mov	qword ptr [rdx],rbx 
+copy_record_argument_2:
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+
+	pop	rbx 
+	sub	rbx,1
+	jc	no_pointers_in_record
+
+copy_record_pointers:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jb	copy_record_pointers_2
+
+	cmp	rdx,qword ptr end_heap_p3
+	jae	copy_record_pointers_2
+
+	mov	rax,qword ptr [rdx]
+	inc	rdi 
+	mov	qword ptr [rdx],rdi 
+	dec	rdi 
+	mov	rdx,rax 
+copy_record_pointers_2:
+	mov	qword ptr [rdi],rdx 
+	add	rdi,8
+	sub	rbx,1
+	jnc	copy_record_pointers
+
+no_pointers_in_record:
+	pop	rbx 
+	
+	sub	rbx,1
+	jc	no_non_pointers_in_record
+
+copy_non_pointers_in_record:
+	mov	rax,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+	sub	rbx,1
+	jnc	copy_non_pointers_in_record
+
+no_non_pointers_in_record:
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+no_record_arguments:
+	sub	rbx,3
+update_up_list_2:
+	mov	rdx,rax
+	mov	rax,qword ptr [rax]
+	inc	rax
+	mov	qword ptr [rdx],rdi
+	test	al,3
+	jne	copy_argument_part_2
+
+	sub	rax,4
+	jmp	update_up_list_2
+
+copy_argument_part_2:
+	dec	rax 
+	cmp	rax,rcx 
+	jc	copy_arguments_1
+
+	cmp	rax,qword ptr end_heap_p3
+	jnc	copy_arguments_1
+
+	mov	rdx,rax
+	mov	rax,qword ptr [rax]
+	inc	rdi 
+	mov	qword ptr [rdx],rdi 
+	dec	rdi 
+copy_arguments_1:
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+
+copy_argument_part_arguments:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	copy_arguments_2
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	copy_arguments_2
+
+	mov	rax,qword ptr [rdx]
+	inc	rdi 
+	mov	qword ptr [rdx],rdi 
+	dec	rdi
+	mov	rdx,rax 
+copy_arguments_2:
+	mov	qword ptr [rdi],rdx 
+	add	rdi,8
+	sub	rbx,1
+	jnc	copy_argument_part_arguments
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+update_list_2_:
+	dec	rax 
+	mov	qword ptr [rdx],rdi 
+begin_update_list_2:
+	mov	rdx,rax 
+	mov	rax,qword ptr [rax]
+update_list__2:
+	test	rax,1
+	jz	end_update_list_2
+	test	rax,2
+	jz	update_list_2_
+	lea	rdx,(-3)[rax]
+	mov	rax,qword ptr (-3)[rax]
+	jmp	update_list__2
+
+end_update_list_2:
+	mov	qword ptr [rdx],rdi 
+
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+
+	test	al,2
+	je	move_lazy_node
+
+	movzx	rbx,word ptr (-2)[rax]
+	test	rbx,rbx 
+	je	move_hnf_0
+
+	cmp	rbx,256
+	jae	move_record
+
+	sub	rbx,2
+	jc	move_hnf_1
+	je	move_hnf_2
+
+move_hnf_3:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	move_hnf_3_1
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	move_hnf_3_1
+
+	lea	rax,1[rdi]
+	mov	rbx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx 
+move_hnf_3_1:
+	mov	qword ptr [rdi],rdx 
+	
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	move_hnf_3_2
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	move_hnf_3_2
+
+	lea	rax,(8+2+1)[rdi]
+	mov	rbx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx
+move_hnf_3_2:
+	mov	qword ptr 8[rdi],rdx 
+	add	rdi,16
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_hnf_2:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	move_hnf_2_1
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	move_hnf_2_1
+
+	lea	rax,1[rdi]
+	mov	rbx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx 
+move_hnf_2_1:
+	mov	qword ptr [rdi],rdx 
+	
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	move_hnf_2_2
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	move_hnf_2_2
+
+	lea	rax,(8+1)[rdi]
+	mov	rbx ,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx
+move_hnf_2_2:
+	mov	qword ptr 8[rdi],rdx 
+	add	rdi,16
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_hnf_1:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	move_hnf_1_
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	move_hnf_1_
+
+	lea	rax,1[rdi]
+	mov	rbx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx
+move_hnf_1_:
+	mov	qword ptr [rdi],rdx 
+	add	rdi,8
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_record:
+	sub	rbx,258
+	jb	move_record_1
+	je	move_record_2
+
+move_record_3:
+	movzx	rbx,word ptr (-2+2)[rax]
+	sub	rbx,1
+	ja	move_hnf_3
+
+	mov	rdx,qword ptr [rcx]
+	lea	rcx,8[rcx]
+	jb	move_record_3_1b
+
+move_record_3_1a:
+	cmp	rdx,rcx 
+	jb	move_record_3_1b
+
+	cmp	rdx,qword ptr end_heap_p3
+	jae	move_record_3_1b
+
+	lea	rax,1[rdi]
+	mov	rbx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx 
+move_record_3_1b:
+	mov	qword ptr [rdi],rdx 
+	add	rdi,8
+
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jb	move_record_3_2
+
+	cmp	rdx,qword ptr end_heap_p3
+	jae	move_record_3_2
+
+	mov	rax,qword ptr neg_heap_p3
+
+	push	rbp 
+
+	add	rax,rdx 
+
+	mov	rbx,qword ptr heap_vector
+	add	rax,8
+	mov	rbp,rax 
+	and	rbp,31*8
+	shr	rax,8
+	mov	ebp,dword ptr bit_set_table2[rbp]
+	test	ebp,dword ptr [rbx+rax*4]
+	je	not_linked_record_argument_part_3_b
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdi 
+
+	mov	rbp,rax 
+	and	rbp,31*8
+	shr	rax,8
+	mov	ebp,dword ptr bit_set_table2[rbp]
+	or	dword ptr [rbx+rax*4],ebp
+	pop	rbp
+
+	jmp	linked_record_argument_part_3_b
+
+not_linked_record_argument_part_3_b:
+	or	dword ptr [rbx+rax*4],ebp 
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdi 
+
+	mov	rbp,rax 
+	and	rbp,31*8
+	shr	rax,8
+	mov	ebp,dword ptr bit_clear_table2[rbp]
+	and	dword ptr [rbx+rax*4],ebp 
+	pop	rbp 
+
+linked_record_argument_part_3_b:
+	mov	rbx,qword ptr [rdx]
+	lea	rax,(2+1)[rdi]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx 
+move_record_3_2:
+	mov	qword ptr [rdi],rdx 
+	add	rdi,8
+
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,rcx
+	shr	rbx,3
+	dec	rbx
+	and	rbx,31
+	cmp	rbx,2
+	jb	bit_in_next_word
+
+	shr	esi,2
+	add	rbp,16
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+bit_in_next_word:
+	dec	r12
+	mov	esi,dword ptr [r8]
+	add	r8,4
+
+	and	esi,dword ptr bit_clear_table[rbx*4]
+
+	test	rsi,rsi
+	je	skip_zeros
+	jmp	end_skip_zeros
+
+move_record_2:
+	cmp	word ptr (-2+2)[rax],1
+	ja	move_hnf_2
+	jb	move_real_or_file
+
+move_record_2_ab:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jb	move_record_2_1
+
+	cmp	rdx,qword ptr end_heap_p3
+	jae	move_record_2_1
+
+	lea	rax,1[rdi]
+	mov	rbx ,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx
+move_record_2_1:
+	mov	qword ptr [rdi],rdx 
+	mov	rbx,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr 8[rdi],rbx 
+	add	rdi,16
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_record_1:
+	movzx	rbx,word ptr (-2+2)[rax]
+	test	rbx,rbx 
+	jne	move_hnf_1
+	jmp	move_int_bool_or_char
+
+move_real_or_file:
+	mov	rax ,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+move_int_bool_or_char:
+	mov	rax,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+copy_normal_hnf_0:
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_hnf_0:
+	cmp	rax,offset dINT+2
+	jb	move_real_file_string_or_array
+	cmp	rax,offset CHAR+2
+	jbe	move_int_bool_or_char
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_real_file_string_or_array:
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+	ja	move_real_or_file
+	jne	move_array
+
+	mov	rax,qword ptr [rcx]
+	add	rax,7
+	shr	rax,3
+
+cp_s_arg_lp3:
+	mov	rbx,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rbx 
+	add	rdi,8
+	sub	rax,1
+	jnc	cp_s_arg_lp3
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_array:
+	test	rsi,rsi
+	push	rcx
+	jne	bsf_and_end_array_bit
+
+skip_zeros_a:
+	sub	r12,1
+	mov	esi,dword ptr [r8]
+	add	r8,4
+	test	rsi,rsi
+	je	skip_zeros_a
+
+	mov	rbp,qword ptr neg_heap_vector_plus_4
+	add	rbp,r8
+
+	shl	rbp,6
+
+	add	rbp,qword ptr heap_p3
+
+bsf_and_end_array_bit:
+	mov	rax,rsi
+	mov	rdx,rsi
+	and	rax,0ffh
+	jne	a_found_bit1
+	and	rdx,0ff00h
+	jne	a_found_bit2
+	mov	rax,rsi
+	mov	rdx,rsi
+	and	rax,0ff0000h
+	jne	a_found_bit3
+	shr	rdx,24
+	movzx	rcx,byte ptr first_one_bit_table[rdx*1]
+	add	rcx,24
+	jmp	end_array_bit
+a_found_bit3:
+	shr	rax,16
+	movzx	rcx,byte ptr first_one_bit_table[rax*1]
+	add	rcx,16
+	jmp	end_array_bit
+a_found_bit2:
+	shr	rdx,8
+	movzx	rcx,byte ptr first_one_bit_table[rdx*1]
+	add	rcx,8
+	jmp	end_array_bit
+a_found_bit1:
+	movzx	rcx,byte ptr first_one_bit_table[rax*1]
+
+end_array_bit:
+	lea	rbx,[rbp+rcx*8]
+	shr	esi,1
+	lea	rbp,8[rbp+rcx*8]
+	shr	esi,cl
+	pop	rcx
+
+	cmp	rcx,rbx
+	jne	move_a_array
+
+move_b_array:
+	mov	rdx,qword ptr [rcx]
+	mov	qword ptr [rdi],rdx
+	mov	rbx,qword ptr 8[rcx]
+	add	rcx,8
+
+	movzx	rax,word ptr (-2)[rbx]
+	add	rdi,8
+	test	rax,rax 
+	je	move_strict_basic_array
+
+	sub	rax,256
+	imul	rdx,rax 
+	mov	rax,rdx
+	jmp	cp_s_arg_lp3
+
+move_strict_basic_array:
+	mov	rax,rdx 
+	cmp	rbx,offset BOOL+2
+	jne	cp_s_arg_lp3
+
+move_bool_array:
+	add	rax,7
+	shr	rax,3
+	jmp	cp_s_arg_lp3
+
+move_a_array:
+	mov	rdx,rbx
+	sub	rbx,rcx
+	shr	rbx,3
+
+	push	rsi
+	sub	rbx,1
+	jb	end_array
+	mov	rsi,qword ptr [rcx]
+
+	mov	rax,qword ptr (-8)[rdx]
+	mov	qword ptr (-8)[rdx],rsi
+
+	mov	qword ptr [rdi],rax 
+
+	mov	rax,qword ptr [rdx]
+
+	mov	rsi,qword ptr 8[rcx]
+	add	rcx,16
+
+	mov	qword ptr [rdx],rsi 
+
+	mov	qword ptr 8[rdi],rax 
+	add	rdi,16
+
+	test	rax,rax 
+	je	st_move_array_lp
+
+	movzx	rsi,word ptr (-2+2)[rax]
+	movzx	rax,word ptr (-2)[rax]
+	sub	rax,256
+	cmp	rax,rsi 
+	je	st_move_array_lp
+
+move_array_ab:
+	push	rcx 
+
+	mov	rdx,qword ptr (-16)[rdi]
+	mov	rbx,rsi 
+	imul	rdx,rax 
+	shl	rdx,3
+
+	sub	rax,rbx 
+	add	rdx,rcx 
+	call	reorder
+
+	pop	rcx 
+	sub	rbx,1
+	sub	rax,1
+
+	push	rbx 
+	push	rax 
+	push	(-16)[rdi]
+	jmp	st_move_array_lp_ab
+
+move_array_ab_lp1:
+	mov	rax,qword ptr 16[rsp]
+move_array_ab_a_elements:
+	mov	rbx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rbx,rcx 
+	jb	move_array_element_ab
+
+	cmp	rbx,qword ptr end_heap_p3
+	jnc	move_array_element_ab
+
+	mov	rdx,rbx 
+	mov	rbx,qword ptr [rdx]
+	inc	rdi
+	mov	qword ptr [rdx],rdi 
+	dec	rdi
+move_array_element_ab:
+	mov	qword ptr [rdi],rbx 
+	add	rdi,8
+	sub	rax,1
+	jnc	move_array_ab_a_elements
+
+	mov	rax,qword ptr 8[rsp]
+move_array_ab_b_elements:
+	mov	rbx,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rbx 
+	add	rdi,8
+	sub	rax,1
+	jnc	move_array_ab_b_elements
+
+st_move_array_lp_ab:
+	sub	qword ptr [rsp],1
+	jnc	move_array_ab_lp1
+
+	add	rsp,24
+	jmp	end_array	
+
+move_array_lp1:
+	mov	rax,qword ptr [rcx]
+	add	rcx,8
+	add	rdi,8
+	cmp	rax,rcx 
+	jb	move_array_element
+
+	cmp	rax,qword ptr end_heap_p3
+	jnc	move_array_element
+
+	mov	rsi,qword ptr [rax]
+	mov	rdx,rax 
+	mov	qword ptr (-8)[rdi],rsi 
+	lea	rax,(-8+1)[rdi]
+	mov	qword ptr [rdx],rax 
+
+	sub	rbx,1
+	jnc	move_array_lp1
+
+	jmp	end_array
+
+move_array_element:
+	mov	qword ptr (-8)[rdi],rax 
+st_move_array_lp:
+	sub	rbx,1
+	jnc	move_array_lp1
+
+end_array:
+	pop	rsi
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_lazy_node:
+	mov	rdx,rax 
+	movsxd	rbx,dword ptr (-4)[rdx]
+	test	rbx,rbx 
+	je	move_lazy_node_0
+
+	sub	rbx,1
+	jle	move_lazy_node_1
+
+	cmp	rbx,256
+	jge	move_closure_with_unboxed_arguments
+
+move_lazy_node_arguments:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	move_lazy_node_arguments_
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	move_lazy_node_arguments_
+
+	mov	rax,qword ptr [rdx]
+	mov	qword ptr [rdi],rax 
+	lea	rax,1[rdi]
+	add	rdi,8
+	mov	qword ptr [rdx],rax 
+	sub	rbx,1
+	jnc	move_lazy_node_arguments
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_lazy_node_arguments_:
+	mov	qword ptr [rdi],rdx 
+	add	rdi,8
+	sub	rbx,1
+	jnc	move_lazy_node_arguments
+	
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_lazy_node_1:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	move_lazy_node_1_
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	move_lazy_node_1_
+
+	lea	rax,1[rdi]
+	mov	rbx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+	mov	rdx,rbx 
+move_lazy_node_1_:
+	mov	qword ptr [rdi],rdx 
+	add	rdi,16
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_lazy_node_0:
+	add	rdi,16
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_closure_with_unboxed_arguments:
+	je	move_closure_with_unboxed_arguments_1
+	add	rbx,1
+	mov	rax,rbx 
+	and	rbx,255
+	shr	rax,8
+	sub	rbx,rax 
+	je	move_non_pointers_of_closure
+
+	push	rax 
+
+move_closure_with_unboxed_arguments_lp:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jc	move_closure_with_unboxed_arguments_
+
+	cmp	rdx,qword ptr end_heap_p3
+	jnc	move_closure_with_unboxed_arguments_
+
+	mov	rax,qword ptr [rdx]
+	mov	qword ptr [rdi],rax 
+	lea	rax,1[rdi]
+	add	rdi,8
+	mov	qword ptr [rdx],rax 
+	sub	rbx,1
+	jne	move_closure_with_unboxed_arguments_lp
+
+	pop	rax 
+	jmp	move_non_pointers_of_closure
+
+move_closure_with_unboxed_arguments_:
+	mov	qword ptr [rdi],rdx 
+	add	rdi,8
+	sub	rbx,1
+	jne	move_closure_with_unboxed_arguments_lp
+
+	pop	rax 
+
+move_non_pointers_of_closure:
+	mov	rbx,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rbx 
+	add	rdi,8
+	sub	rax,1
+	jne	move_non_pointers_of_closure
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_closure_with_unboxed_arguments_1:
+	mov	rax,qword ptr [rcx]
+	mov	qword ptr [rdi],rax 
+	add	rdi,16
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+end_move:
+
+	mov	rcx,qword ptr finalizer_list
+
+restore_finalizer_descriptors:
+	lea	r9,__Nil-8
+	cmp	rcx,r9
+	je	end_restore_finalizer_descriptors
+
+	mov	qword ptr [rcx],offset e____system__kFinalizer+2
+	mov	rcx,qword ptr 8[rcx]
+	jmp	restore_finalizer_descriptors
+
+end_restore_finalizer_descriptors:
+
diff --git a/acompact_rmark.asm b/acompact_rmark.asm
new file mode 100644
index 0000000..d173b59
--- /dev/null
+++ b/acompact_rmark.asm
@@ -0,0 +1,956 @@
+
+rmark_stack_nodes1:
+	mov	rbx,qword ptr [rcx]
+	lea	rax,1[rsi]
+	mov	qword ptr [rsi],rbx
+	mov	qword ptr [rcx],rax
+
+rmark_next_stack_node:
+	add	rsi,8
+	cmp	rsi,qword ptr end_vector
+	je	end_rmark_nodes
+
+rmark_stack_nodes:
+	mov	rcx,qword ptr [rsi]
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmark_next_stack_node
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	mov	ebp,dword ptr [rdi+rbx*4]
+	test	rbp,rax 
+	jne	rmark_stack_nodes1
+
+	or	rbp,rax
+	mov	dword ptr [rdi+rbx*4],ebp 
+
+	mov	rax,qword ptr [rcx]
+	call	rmark_stack_node
+
+	add	rsi,8
+	cmp	rsi,qword ptr end_vector
+	jne	rmark_stack_nodes
+	ret
+
+rmark_stack_node:
+	sub	rsp,16
+	mov	qword ptr [rsi],rax 
+	lea	rbp,1[rsi]
+	mov	qword ptr 8[rsp],rsi 
+	mov	rbx,-1
+	mov	qword ptr [rsp],0
+	mov	qword ptr [rcx],rbp 
+	jmp	rmark_no_reverse
+
+rmark_node_d1:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmark_next_node
+
+	jmp	rmark_node_
+
+rmark_hnf_2:
+	lea	rbx,8[rcx]
+	mov	rax,qword ptr 8[rcx]
+	sub	rsp,16
+
+	mov	rsi,rcx 
+	mov	rcx,qword ptr [rcx]
+
+	mov	qword ptr 8[rsp],rbx 
+	mov	qword ptr [rsp],rax 
+
+rmark_node:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmark_next_node
+
+	mov	rbx,rsi 
+
+rmark_node_:
+	mov	rdx,rax 
+	and	rax,31*8
+	shr	rdx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	mov	ebp,dword ptr [rdi+rdx*4]
+	test	rbp,rax 
+	jne	rmark_reverse_and_mark_next_node
+	
+	or	rbp,rax 
+	mov	dword ptr [rdi+rdx*4],ebp 
+
+	mov	rax,qword ptr [rcx]
+rmark_arguments:
+	cmp	rcx,rbx 
+	ja	rmark_no_reverse
+
+	lea	rbp,1[rsi]
+	mov	qword ptr [rsi],rax 
+	mov	qword ptr [rcx],rbp 
+
+rmark_no_reverse:
+	test	al,2
+	je	rmark_lazy_node
+
+	movzx	rbp,word ptr (-2)[rax]
+	test	rbp,rbp 
+	je	rmark_hnf_0
+
+	add	rcx,8
+
+	cmp	rbp,256
+	jae	rmark_record
+
+	sub	rbp,2
+	je	rmark_hnf_2
+	jc	rmark_hnf_1
+
+rmark_hnf_3:
+	mov	rdx,qword ptr 8[rcx]
+rmark_hnf_3_:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdx 
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,[rdi+rbx*4]
+	jne	rmark_shared_argument_part
+
+	or	dword ptr [rdi+rbx*4],eax 
+
+rmark_no_shared_argument_part:
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rcx 
+	lea	rsi,8[rcx]
+	mov	rcx,qword ptr [rcx]
+	lea	rdx,[rdx+rbp*8]
+	mov	qword ptr [rsp],rcx 
+
+rmark_push_hnf_args:
+	mov	rbx,qword ptr [rdx]
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rdx 
+	sub	rdx,8
+	mov	qword ptr [rsp],rbx 
+
+	sub	rbp,1
+	jg	rmark_push_hnf_args
+
+	mov	rcx,qword ptr [rdx]
+
+	cmp	rdx,rsi 
+	ja	rmark_no_reverse_argument_pointer
+
+	lea	rbp,3[rsi]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rdx],rbp 
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmark_next_node
+
+	mov	rbx,rdx 
+	jmp	rmark_node_
+
+rmark_no_reverse_argument_pointer:
+	mov	rsi,rdx 
+	jmp	rmark_node
+
+rmark_shared_argument_part:
+	cmp	rdx,rcx 
+	ja	rmark_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	lea	rax,(8+2+1)[rcx]
+	mov	qword ptr [rdx],rax 
+	mov	qword ptr 8[rcx],rbx 
+	jmp	rmark_hnf_1
+
+rmark_record:
+	sub	rbp,258
+	je	rmark_record_2
+	jb	rmark_record_1
+
+rmark_record_3:
+	movzx	rbp,word ptr (-2+2)[rax]
+	mov	rdx,qword ptr (16-8)[rcx]
+	sub	rbp,1
+	jb	rmark_record_3_bb
+	je	rmark_record_3_ab
+	sub	rbp,1
+	je	rmark_record_3_aab
+	jmp	rmark_hnf_3_
+
+rmark_record_3_bb:
+	sub	rcx,8
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdx 
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax 
+	
+	cmp	rdx,rcx 
+	ja	rmark_next_node
+
+	add	eax,eax 
+	jne	rmark_bit_in_same_word1
+	inc	rbp
+	mov	rax,1
+rmark_bit_in_same_word1:
+	test	eax,dword ptr [rdi+rbp*4]
+	je	rmark_not_yet_linked_bb
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	add	rax,16
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax 
+
+	mov	rbp,qword ptr [rdx]
+	lea	rax,(16+2+1)[rcx]
+	mov	qword ptr 16[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmark_next_node
+
+rmark_not_yet_linked_bb:
+	or	dword ptr [rdi+rbp*4],eax 
+	mov	rbp,qword ptr [rdx]
+	lea	rax,(16+2+1)[rcx]
+	mov	qword ptr 16[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmark_next_node
+
+rmark_record_3_ab:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdx 
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax 
+
+	cmp	rdx,rcx 
+	ja	rmark_hnf_1
+
+	add	eax,eax 
+	jne	rmark_bit_in_same_word2
+	inc	rbp
+	mov	rax,1
+rmark_bit_in_same_word2:
+	test	eax,dword ptr [rdi+rbp*4]
+	je	rmark_not_yet_linked_ab
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	add	rax,8
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax 
+
+	mov	rbp,qword ptr [rdx]
+	lea	rax,(8+2+1)[rcx]
+	mov	qword ptr 8[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmark_hnf_1
+
+rmark_not_yet_linked_ab:
+	or	dword ptr [rdi+rbp*4],eax 
+	mov	rbp,qword ptr [rdx]
+	lea	rax,(8+2+1)[rcx]
+	mov	qword ptr 8[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmark_hnf_1
+
+rmark_record_3_aab:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdx 
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rbp*4]
+	jne	rmark_shared_argument_part
+	or	dword ptr [rdi+rbp*4],eax 
+
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rcx 
+	lea	rsi,8[rcx]
+	mov	rcx,qword ptr [rcx]
+	mov	qword ptr [rsp],rcx 
+
+	mov	rcx,qword ptr [rdx]
+
+	cmp	rdx,rsi 
+	ja	rmark_no_reverse_argument_pointer
+
+	lea	rbp,3[rsi]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rdx],rbp 
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmark_next_node
+
+	mov	rbx,rdx 
+	jmp	rmark_node_
+
+rmark_record_2:
+	cmp	word ptr (-2+2)[rax],1
+	ja	rmark_hnf_2
+	je	rmark_hnf_1
+	jmp	rmark_next_node
+
+rmark_record_1:
+	cmp	word ptr (-2+2)[rax],0
+	jne	rmark_hnf_1
+	jmp	rmark_next_node
+
+rmark_lazy_node_1:
+; selectors:
+	jne	rmark_selector_node_1
+
+rmark_hnf_1:
+	mov	rsi,rcx 
+	mov	rcx,qword ptr [rcx]
+	jmp	rmark_node
+
+; selectors
+rmark_indirection_node:
+	mov	rdx,qword ptr neg_heap_p3
+	sub	rcx,8
+	add	rdx,rcx 
+
+	mov	rbp,rdx 
+	and	rbp,31*8
+	shr	rdx,8
+	mov	ebp,dword ptr (bit_clear_table2)[rbp]
+	and	dword ptr [rdi+rdx*4],ebp 
+
+	mov	rdx,rcx
+	cmp	rcx,rbx 
+	mov	rcx,qword ptr 8[rcx]
+	mov	qword ptr [rsi],rcx 
+	ja	rmark_node_d1
+	mov	qword ptr [rdx],rax 
+	jmp	rmark_node_d1
+
+rmark_selector_node_1:
+	add	rbp,3
+	je	rmark_indirection_node
+
+	mov	rdx,qword ptr [rcx]
+	mov	qword ptr pointer_compare_address,rbx 
+
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,rdx 
+	shr	rbx,3
+
+	add	rbp,1
+	jle	rmark_record_selector_node_1
+
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmark_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmark_hnf_1
+
+	cmp	word ptr (-2)[rbx],2
+	jbe	rmark_small_tuple_or_record
+
+rmark_large_tuple_or_record:
+	mov	d2,qword ptr 16[rdx]
+
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,d2
+	shr	rbx,3
+
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmark_hnf_1
+
+ ifdef NEW_DESCRIPTORS
+	mov	rbx,qword ptr neg_heap_p3
+	lea	rbx,(-8)[rcx+rbx]
+
+	mov	eax,dword ptr (-8)[rax]
+
+	mov	d3,rbx 
+	and	d3,31*8
+	shr	rbx,8
+	mov	d3d,dword ptr (bit_clear_table2)[d3]
+	and	dword ptr [rdi+rbx*4],d3d 
+
+	movzx	eax,word ptr 4[rax]
+	mov	rbx,qword ptr pointer_compare_address
+
+	mov	qword ptr (-8)[rcx],offset __indirection
+
+	cmp	rax,16
+	jl	rmark_tuple_or_record_selector_node_2
+
+	mov	rdx,rcx
+	je	rmark_tuple_selector_node_2
+
+	mov	rcx,qword ptr (-24)[d2+rax]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rdx],rcx 
+	jmp	rmark_node_d1
+
+rmark_tuple_selector_node_2:
+	mov	rcx,qword ptr [d2]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rdx],rcx 
+	jmp	rmark_node_d1
+ else
+rmark_small_tuple_or_record:
+	mov	rbx,qword ptr neg_heap_p3
+	lea	rbx,(-8)[rcx+rbx]
+
+	push	rcx
+
+	mov	rcx,rbx 
+	and	rcx,31*8
+	shr	rbx,8
+	mov	ecx,dword ptr (bit_clear_table2)[rcx]
+	and	dword ptr [rdi+rbx*4],ecx 
+
+	mov	eax,(-8)[rax]
+
+	mov	rcx,rdx 
+	push	rsi
+	mov	eax,4[rax]
+	call	near ptr rax
+	pop	rsi
+	pop	rdx
+
+	mov	qword ptr [rsi],rcx 
+
+	mov	rbx,qword ptr pointer_compare_address
+
+	mov	qword ptr (-8)[rdx],offset __indirection
+	mov	qword ptr [rdx],rcx 
+	jmp	rmark_node_d1
+ endif
+
+rmark_record_selector_node_1:
+	je	rmark_strict_record_selector_node_1
+
+	mov	rbp,rbx
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmark_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmark_hnf_1
+
+	cmp	word ptr (-2)[rbx],258
+	jbe	rmark_small_tuple_or_record
+
+ ifdef NEW_DESCRIPTORS
+	mov	d2,qword ptr 16[rdx]
+
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,d2
+	shr	rbx,3
+
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmark_hnf_1
+
+rmark_small_tuple_or_record:
+	mov	rbx,qword ptr neg_heap_p3
+	lea	rbx,(-8)[rcx+rbx]
+
+	mov	eax,(-8)[rax]
+
+	mov	d3,rbx 
+	and	d3,31*8
+	shr	rbx,8
+	mov	d3d,dword ptr (bit_clear_table2)[d3]
+	and	dword ptr [rdi+rbx*4],d3d 
+
+	movzx	eax,word ptr 4[rax]
+	mov	rbx,qword ptr pointer_compare_address
+
+	mov	qword ptr (-8)[rcx],offset __indirection
+
+	cmp	rax,16
+	jle	rmark_tuple_or_record_selector_node_2
+	mov	rdx,d2
+	sub	rax,24
+rmark_tuple_or_record_selector_node_2:
+	mov	rbp,rcx
+	mov	rcx,qword ptr [rdx+rax]
+	mov	qword ptr [rsi],rcx
+	mov	qword ptr [rbp],rcx
+	mov	rdx,rbp
+	jmp	rmark_node_d1
+ else
+	jmp	rmark_large_tuple_or_record
+ endif
+
+rmark_strict_record_selector_node_1:
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmark_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmark_hnf_1
+
+	cmp	word ptr (-2)[rbx],258
+	jbe	rmark_select_from_small_record
+
+	mov	d2,qword ptr 16[rdx]
+
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,d2
+	mov	rbp,rbx 
+
+	shr	rbx,8
+	and	rbp,31*8
+	mov	ebp,dword ptr (bit_set_table2)[rbp]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp
+	jne	rmark_hnf_1
+
+rmark_select_from_small_record:
+	mov	ebx,(-8)[rax]
+	sub	rcx,8
+
+	cmp	rcx,qword ptr pointer_compare_address
+	ja	rmark_selector_pointer_not_reversed
+
+ ifdef NEW_DESCRIPTORS
+	movzx	eax,word ptr 4[rbx]
+	cmp	rax,16
+	jle	rmark_strict_record_selector_node_2
+	mov	rax,qword ptr (-24)[d2+rax]
+	jmp	rmark_strict_record_selector_node_3
+rmark_strict_record_selector_node_2:
+	mov	rax,qword ptr [rdx+rax]
+rmark_strict_record_selector_node_3:
+	mov	qword ptr 8[rcx],rax
+
+	movzx	eax,word ptr 6[rbx]
+	test	rax,rax
+	je	rmark_strict_record_selector_node_5
+	cmp	rax,16
+	jle	rmark_strict_record_selector_node_4
+	mov	rdx,d2
+	sub	rax,24
+rmark_strict_record_selector_node_4:
+	mov	rax,qword ptr [rdx+rax]
+	mov	qword ptr 16[rcx],rax
+rmark_strict_record_selector_node_5:
+
+	mov	rax,qword ptr (-8)[rbx]
+ else
+	mov	qword ptr [rcx],rax 
+	mov	qword ptr [rsi],rcx 
+	
+	push	rsi
+	mov	ebx,4[rbx]
+	call	near ptr rbx
+	pop	rsi 
+
+	mov	rax,qword ptr [rcx]
+ endif
+	add	rsi,1
+	mov	qword ptr [rcx],rsi 
+	mov	qword ptr (-1)[rsi],rax 
+	jmp	rmark_next_node
+
+rmark_selector_pointer_not_reversed:
+ ifdef NEW_DESCRIPTORS
+	movzx	eax,word ptr 4[rbx]
+	cmp	rax,16
+	jle	rmark_strict_record_selector_node_6
+	mov	rax,qword ptr (-24)[d2+rax]
+	jmp	rmark_strict_record_selector_node_7
+rmark_strict_record_selector_node_6:
+	mov	rax,qword ptr [rdx+rax]
+rmark_strict_record_selector_node_7:
+	mov	qword ptr 8[rcx],rax
+
+	movzx	eax,word ptr 6[rbx]
+	test	rax,rax
+	je	rmark_strict_record_selector_node_9
+	cmp	rax,16
+	jle	rmark_strict_record_selector_node_8
+	mov	rdx,d2
+	sub	rax,24
+rmark_strict_record_selector_node_8:
+	mov	rax,qword ptr [rdx+rax]
+	mov	qword ptr 16[rcx],rax
+rmark_strict_record_selector_node_9:
+
+	mov	rax,qword ptr (-8)[rbx]
+	mov	qword ptr [rcx],rax
+ else
+	mov	ebx,4[rbx]
+	call	near ptr rbx
+ endif
+	jmp	rmark_next_node
+
+rmark_reverse_and_mark_next_node:
+	cmp	rcx,rbx 
+	ja	rmark_next_node
+
+	mov	rax,qword ptr [rcx]
+	mov	qword ptr [rsi],rax 
+	add	rsi,1
+	mov	qword ptr [rcx],rsi 
+
+; %rbp ,%rbx : free
+
+rmark_next_node:
+	mov	rcx,qword ptr [rsp]
+	mov	rsi,qword ptr 8[rsp]
+	add	rsp,16
+
+	cmp	rcx,1
+	ja	rmark_node
+
+end_rmark_nodes:
+	ret
+
+rmark_lazy_node:
+	movsxd	rbp,dword ptr (-4)[rax]
+	test	rbp,rbp
+	je	rmark_next_node
+
+	add	rcx,8
+
+	sub	rbp,1
+	jle	rmark_lazy_node_1
+
+	cmp	rbp,255
+	jge	rmark_closure_with_unboxed_arguments
+
+rmark_closure_with_unboxed_arguments_:
+	lea	rcx,[rcx+rbp*8]
+
+rmark_push_lazy_args:
+	mov	rbx,qword ptr [rcx]
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rcx 
+	sub	rcx,8
+	mov	qword ptr [rsp],rbx 
+	sub	rbp,1
+	jg	rmark_push_lazy_args
+
+	mov	rsi,rcx 
+	mov	rcx,qword ptr [rcx]
+	jmp	rmark_node
+
+rmark_closure_with_unboxed_arguments:
+; (a_size+b_size)+(b_size<<8)
+;	addl	$1,%rbp 
+	mov	rax,rbp 
+	and	rbp,255
+	shr	rax,8
+	sub	rbp,rax 
+;	subl	$1,%rbp 
+	jg	rmark_closure_with_unboxed_arguments_
+	je	rmark_hnf_1
+	jmp	rmark_next_node
+
+rmark_hnf_0:
+	cmp	rax,offset dINT+2
+	je	rmark_int_3
+
+	cmp	rax,offset CHAR+2
+ 	je	rmark_char_3
+
+	jb	rmark_no_normal_hnf_0
+
+	mov	rbp,qword ptr neg_heap_p3
+	add	rbp,rcx 
+
+	mov	rdx,rbp 
+	and	rdx,31*8
+	shr	rbp,8
+	mov	edx,dword ptr (bit_clear_table2)[rdx]
+	and	dword ptr [rdi+rbp*4],edx 
+
+ ifdef NEW_DESCRIPTORS
+	lea	rdx,((-8)-2)[rax]
+ else
+	lea	rdx,((-12)-2)[rax]
+ endif
+	mov	qword ptr [rsi],rdx 
+	cmp	rcx,rbx 
+	ja	rmark_next_node
+	mov	qword ptr [rcx],rax 
+	jmp	rmark_next_node
+
+rmark_int_3:
+	mov	rbp,qword ptr 8[rcx]
+	cmp	rbp,33
+	jnc	rmark_next_node
+
+	shl	rbp,4
+	lea	rdx,(small_integers)[rbp]
+	mov	rbp,qword ptr neg_heap_p3
+	mov	qword ptr [rsi],rdx 
+	add	rbp,rcx 
+
+	mov	rdx,rbp 
+	and	rdx,31*8
+	shr	rbp,8
+	mov	edx,dword ptr (bit_clear_table2)[rdx]
+	and	dword ptr [rdi+rbp*4],edx 
+
+	cmp	rcx,rbx
+	ja	rmark_next_node
+	mov	qword ptr [rcx],rax 
+	jmp	rmark_next_node
+
+rmark_char_3:
+	movzx	rdx,byte ptr 8[rcx]
+	mov	rbp,qword ptr neg_heap_p3
+
+	shl	rdx,4
+	add	rbp,rcx 
+	add	rdx,offset static_characters
+	mov	qword ptr [rsi],rdx 
+
+	mov	rdx,rbp 
+	and	rdx,31*8
+	shr	rbp,8
+	mov	edx,dword ptr (bit_clear_table2)[rdx]
+	and	dword ptr [rdi+rbp*4],edx
+
+	cmp	rcx,rbx
+	ja	rmark_next_node
+	mov	qword ptr [rcx],rax 
+	jmp	rmark_next_node
+
+rmark_no_normal_hnf_0:
+	lea	r9,__ARRAY__+2
+	cmp	rax,r9
+	jne	rmark_next_node
+
+	mov	rax,qword ptr 16[rcx]
+	test	rax,rax 
+	je	rmark_lazy_array
+
+	movzx	rdx,word ptr (-2+2)[rax]
+	test	rdx,rdx
+	je	rmark_b_array
+
+	movzx	rax,word ptr (-2)[rax]
+	test	rax,rax 
+	je	rmark_b_array
+
+	sub	rax,256
+	cmp	rdx,rax 
+	mov	rbx,rdx
+	je	rmark_a_record_array
+
+rmark_ab_record_array:
+	mov	rdx,qword ptr 8[rcx]
+	add	rcx,16
+	push	rcx
+
+	imul	rdx,rax 
+	shl	rdx,3
+
+	sub	rax,rbx 
+	add	rcx,8
+	add	rdx,rcx 
+	call	reorder
+	
+	pop	rcx 
+	mov	rax,rbx
+	imul	rax,qword ptr (-8)[rcx]
+	jmp	rmark_lr_array
+
+rmark_b_array:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	add	rax,8
+	mov	rbp,rax
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax
+
+	jmp	rmark_next_node
+
+rmark_a_record_array:
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+	cmp	rbx,2
+	jb	rmark_lr_array
+
+	imul	rax,rbx 
+	jmp	rmark_lr_array
+
+rmark_lazy_array:
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+
+rmark_lr_array:
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,rcx 
+	shr	rbx,3
+	add	rbx,rax 
+
+	mov	rdx,rbx 
+	and	rbx,31
+	shr	rdx,5
+	mov	ebx,dword ptr (bit_set_table)[rbx*4]
+	or	dword ptr [rdi+rdx*4],ebx 
+
+	cmp	rax,1
+	jbe	rmark_array_length_0_1
+
+	mov	rdx,rcx 
+	lea	rcx,[rcx+rax*8]
+
+	mov	rax,qword ptr [rcx]
+
+	mov	rbx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+
+	mov	qword ptr [rcx],rbx 
+
+	mov	rax,qword ptr (-8)[rcx]
+	sub	rcx,8
+
+	mov	rbx,qword ptr (-8)[rdx]
+
+	sub	rdx,8
+	mov	qword ptr [rcx],rbx 
+
+	mov	qword ptr [rdx],rax 
+
+	push	rcx
+	mov	rsi,rdx 
+	jmp	rmark_array_nodes
+
+rmark_array_nodes1:
+	cmp	rcx,rsi 
+	ja	rmark_next_array_node
+
+	mov	rbx,qword ptr [rcx]
+	lea	rax,1[rsi]
+	mov	qword ptr [rsi],rbx 
+	mov	qword ptr [rcx],rax 
+
+rmark_next_array_node:
+	add	rsi,8
+	cmp	rsi,qword ptr [rsp]
+	je	end_rmark_array_node
+
+rmark_array_nodes:
+	mov	rcx,qword ptr [rsi]
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmark_next_array_node
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	mov	ebp,dword ptr [rdi+rbx*4]
+	test	rbp,rax 
+	jne	rmark_array_nodes1
+	
+	or	rbp,rax 
+	mov	dword ptr [rdi+rbx*4],ebp 
+
+	mov	rax,qword ptr [rcx]
+	call	rmark_array_node
+
+	add	rsi,8
+	cmp	rsi,qword ptr [rsp]
+	jne	rmark_array_nodes
+
+end_rmark_array_node:
+	add	rsp,8
+	jmp	rmark_next_node
+
+rmark_array_node:
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rsi 
+	mov	rbx,rsi
+	mov	qword ptr [rsp],1
+	jmp	rmark_arguments
+
+rmark_array_length_0_1:
+	lea	rcx,-16[rcx]
+	jb	rmark_next_node
+
+	mov	rbx,qword ptr 24[rcx]
+	mov	rbp,qword ptr 16[rcx]
+	mov	qword ptr 24[rcx],rbp
+	mov	rbp,qword ptr 8[rcx] 
+	mov	qword ptr 16[rcx],rbp
+	mov	qword ptr 8[rcx],rbx
+	add	rcx,8
+	jmp	rmark_hnf_1
+
+_TEXT	ends
+	_DATA segment
+pointer_compare_address:
+	dq	0
+_DATA	ends
+	_TEXT segment
diff --git a/acompact_rmark_prefetch.asm b/acompact_rmark_prefetch.asm
new file mode 100644
index 0000000..376a119
--- /dev/null
+++ b/acompact_rmark_prefetch.asm
@@ -0,0 +1,1178 @@
+
+_TEXT	ends
+	_DATA segment
+rmarkp_n_queue_items_16:
+	dq	0
+rmarkp_queue_first:
+	dq	0
+rmarkp_queue:
+	dq	0,0,0,0,0,0,0,0
+	dq	0,0,0,0,0,0,0,0
+	dq	0,0,0,0,0,0,0,0
+	dq	0,0,0,0,0,0,0,0
+_DATA	ends
+	_TEXT segment
+
+rmarkp_stack_nodes1:
+	mov	rbx,qword ptr [rcx]
+	lea	rax,1[rsi]
+	mov	qword ptr [rsi],rbx
+	mov	qword ptr [rcx],rax
+
+rmarkp_next_stack_node:
+	add	rsi,8
+	cmp	rsi,qword ptr end_vector
+	je	end_rmarkp_nodes
+
+rmarkp_stack_nodes:
+	mov	rcx,qword ptr [rsi]
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_next_stack_node
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	mov	ebp,dword ptr [rdi+rbx*4]
+	test	rbp,rax 
+	jne	rmarkp_stack_nodes1
+
+	or	rbp,rax
+	mov	dword ptr [rdi+rbx*4],ebp 
+
+	mov	rax,qword ptr [rcx]
+	call	rmarkp_stack_node
+
+	add	rsi,8
+	cmp	rsi,qword ptr end_vector
+	jne	rmarkp_stack_nodes
+	ret
+
+rmarkp_stack_node:
+	sub	rsp,16
+	mov	qword ptr [rsi],rax 
+	lea	rbp,1[rsi]
+	mov	qword ptr 8[rsp],rsi 
+	mov	rbx,-1
+	mov	qword ptr [rsp],0
+	mov	qword ptr [rcx],rbp 
+	jmp	rmarkp_no_reverse
+
+rmarkp_node_d1:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_next_node
+
+	jmp	rmarkp_node_
+
+rmarkp_hnf_2:
+	lea	rbx,8[rcx]
+	mov	rax,qword ptr 8[rcx]
+	sub	rsp,16
+
+	mov	rsi,rcx 
+	mov	rcx,qword ptr [rcx]
+
+	mov	qword ptr 8[rsp],rbx 
+	mov	qword ptr [rsp],rax 
+
+rmarkp_node:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_next_node
+
+	mov	rbx,rsi 
+
+rmarkp_node_:
+
+
+
+	mov	rdx,rax 
+	and	rax,31*8
+	shr	rdx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rdx*4]
+	jne	rmarkp_reverse_and_mark_next_node
+
+	mov	rbp,qword ptr rmarkp_queue_first
+	mov	rdx,qword ptr rmarkp_n_queue_items_16
+
+	prefetch	[rcx]
+	mov	qword ptr rmarkp_queue[rbp],rcx
+	mov	qword ptr rmarkp_queue+8[rbp],rsi
+	mov	qword ptr rmarkp_queue+16[rbp],rbx
+	lea	rbx,[rbp+rdx]
+	add	rbp,32
+
+	and	rbp,7*32
+	and	rbx,7*32
+
+	mov	qword ptr rmarkp_queue_first,rbp
+
+	cmp	rdx,-(4*32)
+	je	rmarkp_last_item_in_queue
+
+rmarkp_add_items:
+	mov	rcx,[rsp]
+	cmp	rcx,1
+	jbe	rmarkp_add_stacked_item
+
+	mov	rsi,8[rsp]
+	add	rsp,16
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_add_items
+
+	mov	rdx,rax
+	and	rax,31*8
+	shr	rdx,8
+	mov	eax,dword ptr bit_set_table2[rax]
+	mov	ebp,dword ptr [rdi+rdx*4]
+	test	rbp,rax
+	je	rmarkp_add_item
+
+	cmp	rcx,rsi
+	ja	rmarkp_add_items
+
+	mov	rax,[rcx]
+	mov	[rsi],rax
+	add	rsi,1
+	mov	[rcx],rsi
+	jmp	rmarkp_add_items
+
+rmarkp_add_stacked_item:
+	je	rmarkp_last_item_in_queue
+rmarkp_add_items2:
+	mov	rsi,8[rsp]
+	add	rsi,8
+	cmp	rsi,qword ptr end_vector
+	je	rmarkp_last_item_in_queue
+
+	mov	rcx,[rsi]
+	mov	8[rsp],rsi
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_add_items2
+
+	mov	rdx,rax
+	and	rax,31*8
+	shr	rdx,8
+	mov	eax,dword ptr bit_set_table2[rax]
+	mov	ebp,dword ptr [rdi+rdx*4]
+	test	rbp,rax
+	je	rmarkp_add_item2
+
+	mov	rax,[rcx]
+	mov	[rsi],rax
+	add	rsi,1
+	mov	[rcx],rsi
+	jmp	rmarkp_add_items2
+
+rmarkp_add_item2:
+	prefetch	[rcx]
+	mov	rbp,qword ptr rmarkp_queue_first
+	mov	rdx,qword ptr rmarkp_n_queue_items_16
+
+	mov	qword ptr rmarkp_queue[rbp],rcx
+	mov	qword ptr rmarkp_queue+8[rbp],rsi
+	mov	qword ptr rmarkp_queue+16[rbp],-1
+	add	rbp,32
+	and	rbp,7*32
+
+	sub	rdx,32
+
+	mov	qword ptr rmarkp_queue_first,rbp
+	mov	qword ptr rmarkp_n_queue_items_16,rdx
+
+	cmp	rdx,-(4*32)
+	jne	rmarkp_add_items2
+	jmp	rmarkp_last_item_in_queue
+
+rmarkp_add_items3:
+	mov	rsi,8[rsp]
+	add	rsi,8
+	cmp	rsi,24[rsp]
+	je	rmarkp_last_item_in_queue
+
+	mov	rcx,[rsi]
+	mov	8[rsp],rsi
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_add_items3
+
+	mov	rdx,rax
+	and	rax,31*8
+	shr	rdx,8
+	mov	eax,dword ptr bit_set_table2[rax]
+	mov	ebp,[rdi+rdx*4]
+	test	rbp,rax
+	je	rmarkp_add_item3
+
+	cmp	rcx,rsi
+	ja	rmarkp_add_items3
+
+	mov	rax,[rcx]
+	mov	[rsi],rax
+	add	rsi,1
+	mov	[rcx],rsi
+	jmp	rmarkp_add_items3
+
+rmarkp_add_item3:
+	prefetch	[rcx]
+	mov	rbp,qword ptr rmarkp_queue_first
+	mov	rdx,qword ptr rmarkp_n_queue_items_16
+
+	mov	qword ptr rmarkp_queue[rbp],rcx
+	mov	qword ptr rmarkp_queue+8[rbp],rsi
+	mov	qword ptr rmarkp_queue+16[rbp],rsi
+	add	rbp,32
+	and	rbp,7*32
+
+	sub	rdx,32
+
+	mov	qword ptr rmarkp_queue_first,rbp
+	mov	qword ptr rmarkp_n_queue_items_16,rdx
+
+	cmp	rdx,-(4*32)
+	jne	rmarkp_add_items3
+	jmp	rmarkp_last_item_in_queue
+
+rmarkp_add_item:
+	prefetch	[rcx]
+	mov	rbp,qword ptr rmarkp_queue_first
+	mov	rdx,qword ptr rmarkp_n_queue_items_16
+
+	mov	qword ptr rmarkp_queue[rbp],rcx
+	mov	qword ptr rmarkp_queue+8[rbp],rsi
+	mov	qword ptr rmarkp_queue+16[rbp],rsi
+	add	rbp,32
+	and	rbp,7*32
+
+	sub	rdx,32
+
+	mov	qword ptr rmarkp_queue_first,rbp
+	mov	qword ptr rmarkp_n_queue_items_16,rdx
+
+	cmp	rdx,-(4*32)
+	jne	rmarkp_add_items
+
+rmarkp_last_item_in_queue:
+	mov	rcx,qword ptr rmarkp_queue[rbx]
+
+	mov	rax,qword ptr neg_heap_p3
+
+	mov	rsi,qword ptr rmarkp_queue+8[rbx]
+	mov	rbx,qword ptr rmarkp_queue+16[rbx]
+
+	add	rax,rcx
+
+rmarkp_node_no_prefetch:
+
+
+
+	mov	rdx,rax 
+	and	rax,31*8
+	shr	rdx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	mov	ebp,dword ptr [rdi+rdx*4]
+	test	rbp,rax 
+	jne	rmarkp_reverse_and_mark_next_node
+	
+	or	rbp,rax 
+	mov	dword ptr [rdi+rdx*4],ebp 
+
+	mov	rax,qword ptr [rcx]
+rmarkp_arguments:
+	cmp	rcx,rbx 
+	ja	rmarkp_no_reverse
+
+	lea	rbp,1[rsi]
+	mov	qword ptr [rsi],rax 
+	mov	qword ptr [rcx],rbp 
+
+rmarkp_no_reverse:
+	test	al,2
+	je	rmarkp_lazy_node
+
+	movzx	rbp,word ptr (-2)[rax]
+	test	rbp,rbp 
+	je	rmarkp_hnf_0
+
+	add	rcx,8
+
+	cmp	rbp,256
+	jae	rmarkp_record
+
+	sub	rbp,2
+	je	rmarkp_hnf_2
+	jc	rmarkp_hnf_1
+
+rmarkp_hnf_3:
+	mov	rdx,qword ptr 8[rcx]
+rmarkp_hnf_3_:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdx 
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,[rdi+rbx*4]
+	jne	rmarkp_shared_argument_part
+
+	or	dword ptr [rdi+rbx*4],eax 
+
+rmarkp_no_shared_argument_part:
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rcx 
+	lea	rsi,8[rcx]
+	mov	rcx,qword ptr [rcx]
+	lea	rdx,[rdx+rbp*8]
+	mov	qword ptr [rsp],rcx 
+
+rmarkp_push_hnf_args:
+	mov	rbx,qword ptr [rdx]
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rdx 
+	sub	rdx,8
+	mov	qword ptr [rsp],rbx 
+
+	sub	rbp,1
+	jg	rmarkp_push_hnf_args
+
+	mov	rcx,qword ptr [rdx]
+
+	cmp	rdx,rsi 
+	ja	rmarkp_no_reverse_argument_pointer
+
+	lea	rbp,3[rsi]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rdx],rbp 
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_next_node
+
+	mov	rbx,rdx 
+	jmp	rmarkp_node_
+
+rmarkp_no_reverse_argument_pointer:
+	mov	rsi,rdx 
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_next_node
+	mov	rbx,rsi
+	jmp	rmarkp_node_no_prefetch
+
+rmarkp_shared_argument_part:
+	cmp	rdx,rcx 
+	ja	rmarkp_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	lea	rax,(8+2+1)[rcx]
+	mov	qword ptr [rdx],rax 
+	mov	qword ptr 8[rcx],rbx 
+	jmp	rmarkp_hnf_1
+
+rmarkp_record:
+	sub	rbp,258
+	je	rmarkp_record_2
+	jb	rmarkp_record_1
+
+rmarkp_record_3:
+	movzx	rbp,word ptr (-2+2)[rax]
+	mov	rdx,qword ptr (16-8)[rcx]
+	sub	rbp,1
+	jb	rmarkp_record_3_bb
+	je	rmarkp_record_3_ab
+	sub	rbp,1
+	je	rmarkp_record_3_aab
+	jmp	rmarkp_hnf_3_
+
+rmarkp_record_3_bb:
+	sub	rcx,8
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdx 
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax 
+	
+	cmp	rdx,rcx 
+	ja	rmarkp_next_node
+
+	add	eax,eax 
+	jne	rmarkp_bit_in_same_word1
+	inc	rbp
+	mov	rax,1
+rmarkp_bit_in_same_word1:
+	test	eax,dword ptr [rdi+rbp*4]
+	je	rmarkp_not_yet_linked_bb
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	add	rax,16
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax 
+
+	mov	rbp,qword ptr [rdx]
+	lea	rax,(16+2+1)[rcx]
+	mov	qword ptr 16[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmarkp_next_node
+
+rmarkp_not_yet_linked_bb:
+	or	dword ptr [rdi+rbp*4],eax 
+	mov	rbp,qword ptr [rdx]
+	lea	rax,(16+2+1)[rcx]
+	mov	qword ptr 16[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmarkp_next_node
+
+rmarkp_record_3_ab:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdx 
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax 
+
+	cmp	rdx,rcx 
+	ja	rmarkp_hnf_1
+
+	add	eax,eax 
+	jne	rmarkp_bit_in_same_word2
+	inc	rbp
+	mov	rax,1
+rmarkp_bit_in_same_word2:
+	test	eax,dword ptr [rdi+rbp*4]
+	je	rmarkp_not_yet_linked_ab
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	add	rax,8
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax 
+
+	mov	rbp,qword ptr [rdx]
+	lea	rax,(8+2+1)[rcx]
+	mov	qword ptr 8[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmarkp_hnf_1
+
+rmarkp_not_yet_linked_ab:
+	or	dword ptr [rdi+rbp*4],eax 
+	mov	rbp,qword ptr [rdx]
+	lea	rax,(8+2+1)[rcx]
+	mov	qword ptr 8[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmarkp_hnf_1
+
+rmarkp_record_3_aab:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rdx 
+
+	mov	rbp,rax 
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rbp*4]
+	jne	rmarkp_shared_argument_part
+	or	dword ptr [rdi+rbp*4],eax 
+
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rcx 
+	lea	rsi,8[rcx]
+	mov	rcx,qword ptr [rcx]
+	mov	qword ptr [rsp],rcx 
+
+	mov	rcx,qword ptr [rdx]
+
+	cmp	rdx,rsi 
+	ja	rmarkp_no_reverse_argument_pointer
+
+	lea	rbp,3[rsi]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rdx],rbp 
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_next_node
+
+	mov	rbx,rdx 
+	jmp	rmarkp_node_
+
+rmarkp_record_2:
+	cmp	word ptr (-2+2)[rax],1
+	ja	rmarkp_hnf_2
+	je	rmarkp_hnf_1
+	jmp	rmarkp_next_node
+
+rmarkp_record_1:
+	cmp	word ptr (-2+2)[rax],0
+	jne	rmarkp_hnf_1
+	jmp	rmarkp_next_node
+
+rmarkp_lazy_node_1:
+; selectors:
+	jne	rmarkp_selector_node_1
+
+rmarkp_hnf_1:
+	mov	rsi,rcx 
+	mov	rcx,qword ptr [rcx]
+	jmp	rmarkp_node
+
+; selectors
+rmarkp_indirection_node:
+	mov	rdx,qword ptr neg_heap_p3
+	sub	rcx,8
+	add	rdx,rcx 
+
+	mov	rbp,rdx 
+	and	rbp,31*8
+	shr	rdx,8
+	mov	ebp,dword ptr (bit_clear_table2)[rbp]
+	and	dword ptr [rdi+rdx*4],ebp 
+
+	mov	rdx,rcx
+	cmp	rcx,rbx 
+	mov	rcx,qword ptr 8[rcx]
+	mov	qword ptr [rsi],rcx 
+	ja	rmarkp_node_d1
+	mov	qword ptr [rdx],rax 
+	jmp	rmarkp_node_d1
+
+rmarkp_selector_node_1:
+	add	rbp,3
+	je	rmarkp_indirection_node
+
+	mov	rdx,qword ptr [rcx]
+	mov	qword ptr pointer_compare_address,rbx 
+
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,rdx 
+	shr	rbx,3
+
+	add	rbp,1
+	jle	rmarkp_record_selector_node_1
+
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmarkp_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmarkp_hnf_1
+
+	cmp	word ptr (-2)[rbx],2
+	jbe	rmarkp_small_tuple_or_record
+
+rmarkp_large_tuple_or_record:
+	mov	d2,qword ptr 16[rdx]
+
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,d2
+	shr	rbx,3
+
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmarkp_hnf_1
+
+ ifdef NEW_DESCRIPTORS
+	mov	rbx,qword ptr neg_heap_p3
+	lea	rbx,(-8)[rcx+rbx]
+
+	mov	eax,(-8)[rax]
+
+	mov	d3,rbx 
+	and	d3,31*8
+	shr	rbx,8
+	mov	d3d,dword ptr (bit_clear_table2)[d3]
+	and	dword ptr [rdi+rbx*4],d3d 
+
+	movzx	eax,word ptr 4[rax]
+	mov	rbx,qword ptr pointer_compare_address
+
+	mov	qword ptr (-8)[rcx],offset __indirection
+
+	cmp	rax,16
+	jl	rmarkp_tuple_or_record_selector_node_2
+
+	mov	rdx,rcx
+	je	rmarkp_tuple_selector_node_2
+
+	mov	rcx,qword ptr (-24)[d2+rax]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rdx],rcx 
+	jmp	rmarkp_node_d1
+
+rmarkp_tuple_selector_node_2:
+	mov	rcx,qword ptr [d2]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rdx],rcx 
+	jmp	rmarkp_node_d1
+ else
+rmarkp_small_tuple_or_record:
+	mov	rbx,qword ptr neg_heap_p3
+	lea	rbx,(-8)[rcx+rbx]
+
+	push	rcx
+
+	mov	rcx,rbx 
+	and	rcx,31*8
+	shr	rbx,8
+	mov	ecx,dword ptr (bit_clear_table2)[rcx]
+	and	dword ptr [rdi+rbx*4],ecx 
+
+	mov	eax,(-8)[rax]
+
+	mov	rcx,rdx 
+	push	rsi
+	mov	eax,4[rax]
+	call	near ptr rax
+	pop	rsi
+	pop	rdx
+
+	mov	qword ptr [rsi],rcx 
+
+	mov	rbx,qword ptr pointer_compare_address
+
+	mov	qword ptr (-8)[rdx],offset __indirection
+	mov	qword ptr [rdx],rcx 
+	jmp	rmarkp_node_d1
+ endif
+
+rmarkp_record_selector_node_1:
+	je	rmarkp_strict_record_selector_node_1
+
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmarkp_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmarkp_hnf_1
+
+	cmp	word ptr (-2)[rbx],258
+	jbe	rmarkp_small_tuple_or_record
+
+ ifdef NEW_DESCRIPTORS
+	mov	d2,qword ptr 16[rdx]
+
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,d2
+	shr	rbx,3
+
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmarkp_hnf_1
+
+rmarkp_small_tuple_or_record:
+	mov	rbx,qword ptr neg_heap_p3
+	lea	rbx,(-8)[rcx+rbx]
+
+	mov	eax,(-8)[rax]
+
+	mov	d3,rbx 
+	and	d3,31*8
+	shr	rbx,8
+	mov	d3d,dword ptr (bit_clear_table2)[d3]
+	and	dword ptr [rdi+rbx*4],d3d 
+
+	movzx	eax,word ptr 4[rax]
+	mov	rbx,qword ptr pointer_compare_address
+
+	mov	qword ptr (-8)[rcx],offset __indirection
+
+	cmp	rax,16
+	jle	rmarkp_tuple_or_record_selector_node_2
+	mov	rdx,d2
+	sub	rax,24
+rmarkp_tuple_or_record_selector_node_2:
+	mov	rbp,rcx
+	mov	rcx,qword ptr [rdx+rax]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr [rbp],rcx
+	mov	rdx,rbp
+	jmp	rmarkp_node_d1
+ else
+	jmp	rmarkp_large_tuple_or_record
+ endif
+
+rmarkp_strict_record_selector_node_1:
+	mov	rbp,rbx 
+	shr	rbx,5
+	and	rbp,31
+	mov	ebp,dword ptr (bit_set_table)[rbp*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp 
+	jne	rmarkp_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmarkp_hnf_1
+
+	cmp	word ptr (-2)[rbx],258
+	jbe	rmarkp_select_from_small_record
+
+	mov	d2,qword ptr 16[rdx]
+
+	mov	d2,qword ptr neg_heap_p3
+	add	rbx,d2
+	mov	rbp,rbx 
+
+	shr	rbx,8
+	and	rbp,31*8
+	mov	ebp,dword ptr (bit_set_table2)[rbp]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rbp
+	jne	rmarkp_hnf_1
+
+rmarkp_select_from_small_record:
+	mov	ebx,(-8)[rax]
+	sub	rcx,8
+
+	cmp	rcx,qword ptr pointer_compare_address
+	ja	rmarkp_selector_pointer_not_reversed
+
+ ifdef NEW_DESCRIPTORS
+	movzx	eax,word ptr 4[rbx]
+	cmp	rax,16
+	jle	rmarkp_strict_record_selector_node_2
+	mov	rax,qword ptr (-24)[d2+rax]
+	jmp	rmarkp_strict_record_selector_node_3
+rmarkp_strict_record_selector_node_2:
+	mov	rax,qword ptr [rdx+rax]
+rmarkp_strict_record_selector_node_3:
+	mov	qword ptr 8[rcx],rax
+
+	movzx	eax,word ptr 6[rbx]
+	test	rax,rax
+	je	rmarkp_strict_record_selector_node_5
+	cmp	rax,16
+	jle	rmarkp_strict_record_selector_node_4
+	mov	rdx,d2
+	sub	rax,24
+rmarkp_strict_record_selector_node_4:
+	mov	rax,qword ptr [rdx+rax]
+	mov	qword ptr 16[rcx],rax
+rmarkp_strict_record_selector_node_5:
+
+	mov	rax,qword ptr (-8)[rbx]
+ else
+	mov	qword ptr [rcx],rax 
+	mov	qword ptr [rsi],rcx 
+	
+	push	rsi
+	mov	ebx,4[rbx]
+	call	near ptr rbx
+	pop	rsi 
+
+	mov	rax,qword ptr [rcx]
+ endif
+	add	rsi,1
+	mov	qword ptr [rcx],rsi 
+	mov	qword ptr (-1)[rsi],rax 
+	jmp	rmarkp_next_node
+
+rmarkp_selector_pointer_not_reversed:
+ ifdef NEW_DESCRIPTORS
+	movzx	eax,word ptr 4[rbx]
+	cmp	rax,16
+	jle	rmarkp_strict_record_selector_node_6
+	mov	rax,qword ptr (-24)[d2+rax]
+	jmp	rmarkp_strict_record_selector_node_7
+rmarkp_strict_record_selector_node_6:
+	mov	rax,qword ptr [rdx+rax]
+rmarkp_strict_record_selector_node_7:
+	mov	qword ptr 8[rcx],rax
+
+	movzx	eax,word ptr 6[rbx]
+	test	rax,rax
+	je	rmarkp_strict_record_selector_node_9
+	cmp	rax,16
+	jle	rmarkp_strict_record_selector_node_8
+	mov	rdx,d2
+	sub	rax,24
+rmarkp_strict_record_selector_node_8:
+	mov	rax,qword ptr [rdx+rax]
+	mov	qword ptr 16[rcx],rax
+rmarkp_strict_record_selector_node_9:
+
+	mov	rax,qword ptr (-8)[rbx]
+	mov	qword ptr [rcx],rax
+ else
+ 	mov	ebx,4[rbx]
+	call	near ptr rbx
+ endif
+	jmp	rmarkp_next_node
+
+rmarkp_reverse_and_mark_next_node:
+	cmp	rcx,rbx 
+	ja	rmarkp_next_node
+
+	mov	rax,qword ptr [rcx]
+	mov	qword ptr [rsi],rax 
+	add	rsi,1
+	mov	qword ptr [rcx],rsi 
+
+; %rbp ,%rbx : free
+
+rmarkp_next_node:
+	mov	rcx,qword ptr [rsp]
+	mov	rsi,qword ptr 8[rsp]
+	add	rsp,16
+
+	cmp	rcx,1
+	ja	rmarkp_node
+
+rmarkp_next_node_:
+	mov	rdx,qword ptr rmarkp_n_queue_items_16
+	test	rdx,rdx
+	je	end_rmarkp_nodes
+
+	sub	rsp,16
+
+	mov	rbp,qword ptr rmarkp_queue_first
+
+	lea	rbx,[rbp+rdx]
+	add	rdx,32
+
+	and	rbx,7*32
+
+	mov	qword ptr rmarkp_n_queue_items_16,rdx
+	jmp	rmarkp_last_item_in_queue
+
+end_rmarkp_nodes:
+	ret
+
+rmarkp_lazy_node:
+	movsxd	rbp,dword ptr (-4)[rax]
+	test	rbp,rbp
+	je	rmarkp_next_node
+
+	add	rcx,8
+
+	sub	rbp,1
+	jle	rmarkp_lazy_node_1
+
+	cmp	rbp,255
+	jge	rmarkp_closure_with_unboxed_arguments
+
+rmarkp_closure_with_unboxed_arguments_:
+	lea	rcx,[rcx+rbp*8]
+
+rmarkp_push_lazy_args:
+	mov	rbx,qword ptr [rcx]
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rcx 
+	sub	rcx,8
+	mov	qword ptr [rsp],rbx 
+	sub	rbp,1
+	jg	rmarkp_push_lazy_args
+
+	mov	rsi,rcx 
+	mov	rcx,qword ptr [rcx]
+	jmp	rmarkp_node
+
+rmarkp_closure_with_unboxed_arguments:
+; (a_size+b_size)+(b_size<<8)
+;	addl	$1,%rbp 
+	mov	rax,rbp 
+	and	rbp,255
+	shr	rax,8
+	sub	rbp,rax 
+;	subl	$1,%rbp 
+	jg	rmarkp_closure_with_unboxed_arguments_
+	je	rmarkp_hnf_1
+	jmp	rmarkp_next_node
+
+rmarkp_hnf_0:
+	cmp	rax,offset dINT+2
+	je	rmarkp_int_3
+
+	cmp	rax,offset CHAR+2
+ 	je	rmarkp_char_3
+
+	jb	rmarkp_no_normal_hnf_0
+
+	mov	rbp,qword ptr neg_heap_p3
+	add	rbp,rcx 
+
+	mov	rdx,rbp 
+	and	rdx,31*8
+	shr	rbp,8
+	mov	edx,dword ptr (bit_clear_table2)[rdx]
+	and	dword ptr [rdi+rbp*4],edx 
+
+ ifdef NEW_DESCRIPTORS
+	lea	rdx,((-8)-2)[rax]
+ else
+	lea	rdx,((-12)-2)[rax]
+ endif
+	mov	qword ptr [rsi],rdx 
+	cmp	rcx,rbx 
+	ja	rmarkp_next_node
+	mov	qword ptr [rcx],rax 
+	jmp	rmarkp_next_node
+
+rmarkp_int_3:
+	mov	rbp,qword ptr 8[rcx]
+	cmp	rbp,33
+	jnc	rmarkp_next_node
+
+	shl	rbp,4
+	lea	rdx,(small_integers)[rbp]
+	mov	rbp,qword ptr neg_heap_p3
+	mov	qword ptr [rsi],rdx 
+	add	rbp,rcx 
+
+	mov	rdx,rbp 
+	and	rdx,31*8
+	shr	rbp,8
+	mov	edx,dword ptr (bit_clear_table2)[rdx]
+	and	dword ptr [rdi+rbp*4],edx 
+
+	cmp	rcx,rbx
+	ja	rmarkp_next_node
+	mov	qword ptr [rcx],rax 
+	jmp	rmarkp_next_node
+
+rmarkp_char_3:
+	movzx	rdx,byte ptr 8[rcx]
+	mov	rbp,qword ptr neg_heap_p3
+
+	shl	rdx,4
+	add	rbp,rcx 
+	add	rdx,offset static_characters
+	mov	qword ptr [rsi],rdx 
+
+	mov	rdx,rbp 
+	and	rdx,31*8
+	shr	rbp,8
+	mov	edx,dword ptr (bit_clear_table2)[rdx]
+	and	dword ptr [rdi+rbp*4],edx
+
+	cmp	rcx,rbx
+	ja	rmarkp_next_node
+	mov	qword ptr [rcx],rax 
+	jmp	rmarkp_next_node
+
+rmarkp_no_normal_hnf_0:
+	lea	r9,__ARRAY__+2
+	cmp	rax,r9
+	jne	rmarkp_next_node
+
+	mov	rax,qword ptr 16[rcx]
+	test	rax,rax 
+	je	rmarkp_lazy_array
+
+	movzx	rdx,word ptr (-2+2)[rax]
+	test	rdx,rdx
+	je	rmarkp_b_array
+
+	movzx	rax,word ptr (-2)[rax]
+	test	rax,rax 
+	je	rmarkp_b_array
+
+	sub	rax,256
+	cmp	rdx,rax 
+	mov	rbx,rdx
+	je	rmarkp_a_record_array
+
+rmarkp_ab_record_array:
+	mov	rdx,qword ptr 8[rcx]
+	add	rcx,16
+	push	rcx
+
+	imul	rdx,rax 
+	shl	rdx,3
+
+	sub	rax,rbx 
+	add	rcx,8
+	add	rdx,rcx 
+	call	reorder
+	
+	pop	rcx 
+	mov	rax,rbx
+	imul	rax,qword ptr (-8)[rcx]
+	jmp	rmarkp_lr_array
+
+rmarkp_b_array:
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx
+	add	rax,8
+	mov	rbp,rax
+	and	rax,31*8
+	shr	rbp,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	or	dword ptr [rdi+rbp*4],eax
+
+	jmp	rmarkp_next_node
+
+rmarkp_a_record_array:
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+	cmp	rbx,2
+	jb	rmarkp_lr_array
+
+	imul	rax,rbx 
+	jmp	rmarkp_lr_array
+
+rmarkp_lazy_array:
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+
+rmarkp_lr_array:
+	mov	rbx,qword ptr neg_heap_p3
+	add	rbx,rcx 
+	shr	rbx,3
+	add	rbx,rax 
+
+	mov	rdx,rbx 
+	and	rbx,31
+	shr	rdx,5
+	mov	ebx,dword ptr (bit_set_table)[rbx*4]
+	or	dword ptr [rdi+rdx*4],ebx 
+
+	cmp	rax,1
+	jbe	rmarkp_array_length_0_1
+
+	mov	rdx,rcx 
+	lea	rcx,[rcx+rax*8]
+
+	mov	rax,qword ptr [rcx]
+
+	mov	rbx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax 
+
+	mov	qword ptr [rcx],rbx 
+
+	mov	rax,qword ptr (-8)[rcx]
+	sub	rcx,8
+
+	mov	rbx,qword ptr (-8)[rdx]
+
+	sub	rdx,8
+	mov	qword ptr [rcx],rbx 
+
+	mov	qword ptr [rdx],rax 
+
+	push	rcx
+	mov	rsi,rdx 
+	jmp	rmarkp_array_nodes
+
+rmarkp_array_nodes1:
+	cmp	rcx,rsi 
+	ja	rmarkp_next_array_node
+
+	mov	rbx,qword ptr [rcx]
+	lea	rax,1[rsi]
+	mov	qword ptr [rsi],rbx 
+	mov	qword ptr [rcx],rax 
+
+rmarkp_next_array_node:
+	add	rsi,8
+	cmp	rsi,qword ptr [rsp]
+	je	end_rmarkp_array_node
+
+rmarkp_array_nodes:
+	mov	rcx,qword ptr [rsi]
+
+	mov	rax,qword ptr neg_heap_p3
+	add	rax,rcx 
+
+	cmp	rax,qword ptr heap_size_64_65
+	jnc	rmarkp_next_array_node
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	mov	ebp,dword ptr [rdi+rbx*4]
+	test	rbp,rax 
+	jne	rmarkp_array_nodes1
+	
+	or	rbp,rax 
+	mov	dword ptr [rdi+rbx*4],ebp 
+
+	mov	rax,qword ptr [rcx]
+	call	rmarkp_array_node
+
+	add	rsi,8
+	cmp	rsi,qword ptr [rsp]
+	jne	rmarkp_array_nodes
+
+end_rmarkp_array_node:
+	add	rsp,8
+	jmp	rmarkp_next_node
+
+rmarkp_array_node:
+	sub	rsp,16
+	mov	qword ptr 8[rsp],rsi 
+	mov	rbx,rsi
+	mov	qword ptr [rsp],1
+	jmp	rmarkp_arguments
+
+rmarkp_array_length_0_1:
+	lea	rcx,-16[rcx]
+	jb	rmarkp_next_node
+
+	mov	rbx,qword ptr 24[rcx]
+	mov	rbp,qword ptr 16[rcx]
+	mov	qword ptr 24[rcx],rbp
+	mov	rbp,qword ptr 8[rcx] 
+	mov	qword ptr 16[rcx],rbp
+	mov	qword ptr 8[rcx],rbx
+	add	rcx,8
+	jmp	rmarkp_hnf_1
diff --git a/acopy.asm b/acopy.asm
new file mode 100644
index 0000000..fadf47e
--- /dev/null
+++ b/acopy.asm
@@ -0,0 +1,1306 @@
+
+COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP = 1
+
+	push	rsi
+
+	mov	rdi,heap_p2
+
+	mov	rax,heap_size_257
+	shl	rax,7
+	mov	semi_space_size,rax 
+	lea	rsi,[rdi+rax]
+
+	mov	qword ptr (heap2_begin_and_end+8),rsi 
+
+	mov	rax,qword ptr caf_list
+	test	rax,rax 
+	je	end_copy_cafs
+
+copy_cafs_lp:
+	push	(-8)[rax]
+	
+	lea	rbp,8[rax]
+	mov	rbx,qword ptr [rax]
+	sub	rbx,1
+	call	copy_lp2
+	
+	pop	rax 
+	test	rax,rax 
+	jne	copy_cafs_lp
+
+end_copy_cafs:
+	mov	rbx,qword ptr [rsp]
+	mov	rbp,stack_p
+	sub	rbx,rbp 
+	shr	rbx,3
+
+	sub	rbx,1
+	jb	end_copy0
+	call	copy_lp2
+end_copy0:
+	mov	rbp,heap_p2
+
+	jmp	copy_lp1
+;
+;	Copy all referenced nodes to the other semi space
+;
+
+in_hnf_1_2:
+	dec	rbx 
+copy_lp2_lp1:
+	call	copy_lp2
+copy_lp1:
+	cmp	rbp,rdi 
+	jae	end_copy1
+
+	mov	rax,[rbp]
+	add	rbp,8
+	test	al,2
+	je	not_in_hnf_1
+in_hnf_1:
+	movzx	rbx,word ptr (-2)[rax]
+
+	test	rbx,rbx 
+	je	copy_array_21
+
+	cmp	rbx,2
+	jbe	in_hnf_1_2
+
+	cmp	rbx,256
+	jae	copy_record_21
+	
+	mov	rax,8[rbp]
+
+	test	al,1
+	jne	node_without_arguments_part
+
+	push	rbx 
+	xor	rbx,rbx 
+	
+	call	copy_lp2
+
+	pop	rbx 
+	add	rbp,8
+
+	sub	rbx,2
+	jmp	copy_lp2_lp1
+
+node_without_arguments_part:
+	dec	rax 
+	xor	rbx,rbx 
+
+	mov	8[rbp],rax 
+	call	copy_lp2
+	
+	add	rbp,8
+	jmp	copy_lp1
+
+copy_record_21:
+	sub	rbx,258
+	ja	copy_record_arguments_3
+
+	movzx	rbx,word ptr (-2+2)[rax]
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+	jb	in_hnf_1_2
+
+	sub	rbx,1
+	ja	copy_lp2_lp1
+	jmp	copy_node_arity1
+ else
+	jb	copy_record_arguments_1
+
+	sub	rbx,1
+	ja	copy_lp2_lp1
+	je	copy_node_arity1
+	add	rbp,16
+	jmp	copy_lp1
+
+copy_record_arguments_1:
+	dec	rbx
+	jmp	copy_lp2_lp1
+	je	copy_lp2_lp1
+	add	rbp,8
+	jmp	copy_lp1
+ endif
+
+copy_record_arguments_3:
+	test	byte ptr 8[rbp],1
+	jne	record_node_without_arguments_part
+
+	movzx	rdx,word ptr (-2+2)[rax]
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+	sub	rdx,1
+ else
+	test	rdx,rdx 
+	je	copy_record_arguments_3b
+	sub	rdx,1
+	je	copy_record_arguments_3abb
+ endif
+
+	lea	rcx,(3*8)[rbp+rbx*8]
+	push	rcx 
+	push	rdx 
+
+	sub	rbx,rbx 
+	call	copy_lp2
+	
+	add	rbp,8
+	pop	rbx 
+	dec	rbx 
+	call	copy_lp2
+
+	pop	rbp 
+	jmp	copy_lp1
+
+ ife COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+copy_record_arguments_3abb:
+	push	rbx 
+	sub	rbx,rbx 
+	
+	call	copy_lp2
+	
+	pop	rbx 
+	
+	lea	rbp,(2*8)[rbp+rbx*8]
+	jmp	copy_lp1
+
+copy_record_arguments_3b:
+	lea	rbp,(3*8)[rbp+rbx*8]
+	jmp	copy_lp1
+ endif
+
+record_node_without_arguments_part:
+	and	qword ptr 8[rbp],-2
+
+ ife COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+	cmp	word ptr (-2+2)[rax],0
+	je	record_node_without_arguments_part_3b
+ endif
+
+	sub	rbx,rbx 
+	call	copy_lp2
+
+	add	rbp,8
+	jmp	copy_lp1
+
+ ife COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+record_node_without_arguments_part_3b:
+	add	rbp,16
+	jmp	copy_lp1
+ endif
+
+not_in_hnf_1:
+	movsxd	rbx,dword ptr (-4)[rax]
+	cmp	rbx,257
+	jge	copy_unboxed_closure_arguments
+	sub	rbx,1
+	jg	copy_lp2_lp1
+
+copy_node_arity1:
+	xor	rbx,rbx 
+	call	copy_lp2
+
+	add	rbp,8
+	jmp	copy_lp1
+
+copy_unboxed_closure_arguments:
+	je	copy_unboxed_closure_arguments1
+
+	xor	rax,rax
+	mov	al,bh 
+	and	rbx,255
+	sub	rbx,rax 
+
+	sub	rbx,1
+	jl	copy_unboxed_closure_arguments_without_pointers
+	
+	push	rax 
+	call	copy_lp2
+	pop	rax 
+
+copy_unboxed_closure_arguments_without_pointers:
+	lea	rbp,[rbp+rax*8]
+	jmp	copy_lp1
+
+copy_unboxed_closure_arguments1:
+	add	rbp,16
+	jmp	copy_lp1
+
+copy_array_21:
+	mov	rbx,qword ptr 8[rbp]
+	add	rbp,16
+	test	rbx,rbx 
+	je	copy_array_21_a
+
+	movzx	rax,word ptr (-2)[rbx]
+	movzx	rbx,word ptr (-2+2)[rbx]
+	sub	rax,256
+	test	rbx,rbx 
+	je	copy_array_21_b
+
+	cmp	rbx,rax 
+	je	copy_array_21_r_a
+
+copy_array_21_ab:
+	cmp	qword ptr (-16)[rbp],0
+	je	copy_lp1
+
+	sub	rax,rbx 
+	shl	rax,3
+	sub	rbx,1
+
+	push	rbx 
+	push	rax 
+	mov	rbx,qword ptr (-16)[rbp]
+	sub	rbx,1
+	push	rbx 
+
+copy_array_21_lp_ab:
+	mov	rbx,qword ptr 16[rsp]
+	call	copy_lp2
+
+	add	rbp,qword ptr 8[rsp]
+	sub	qword ptr [rsp],1
+	jnc	copy_array_21_lp_ab
+	
+	add	rsp,24
+	jmp	copy_lp1
+
+copy_array_21_b:
+	mov	rbx,qword ptr (-16)[rbp]
+	imul	rbx,rax 
+	lea	rbp,[rbp+rbx*8]
+	jmp	copy_lp1
+
+copy_array_21_r_a:
+	mov	rbx,qword ptr (-16)[rbp]
+	imul	rbx,rax
+	sub	rbx,1
+	jc	copy_lp1
+	jmp	copy_lp2_lp1
+
+copy_array_21_a:
+	mov	rbx,qword ptr (-16)[rbp]
+	sub	rbx,1
+	jc	copy_lp1
+	jmp	copy_lp2_lp1
+
+;
+;	Copy nodes to the other semi-space
+;
+
+copy_lp2:
+	mov	rdx,qword ptr [rbp]
+
+; selectors:
+continue_after_selector_2:
+	mov	rcx,qword ptr [rdx]
+	test	cl,2
+	je	not_in_hnf_2
+
+in_hnf_2:
+	movzx	rax,word ptr (-2)[rcx]
+	test	rax,rax
+	je	copy_arity_0_node2
+
+	cmp	rax,256
+	jae	copy_record_2
+
+	sub	rax,2
+	mov	[rbp],rdi 
+
+	lea	rbp,8[rbp ]
+	ja	copy_hnf_node2_3
+
+	mov	[rdi],rcx
+	jb	copy_hnf_node2_1
+
+	inc	rdi
+	mov	rcx,8[rdx]
+
+	mov	[rdx],rdi 
+	mov	rax,16[rdx]
+
+	sub	rbx,1
+	mov	(8-1)[rdi],rcx 
+
+	mov	(16-1)[rdi],rax 
+	lea	rdi,(24-1)[rdi]
+
+	jae	copy_lp2
+	ret
+
+copy_hnf_node2_1:
+	inc	rdi 
+	mov	rax,8[rdx]
+
+	sub	rbx,1
+	mov	[rdx],rdi
+
+	mov	(8-1)[rdi],rax 
+	lea	rdi,(16-1)[rdi]
+
+	jae	copy_lp2
+	ret
+
+copy_hnf_node2_3:
+	mov	[rdi],rcx 
+	inc	rdi 
+
+	mov	[rdx],rdi 
+	mov	rcx,8[rdx]
+
+	mov	(8-1)[rdi],rcx 
+	mov	rcx,16[rdx]
+
+	add	rdi,24-1
+	mov	rdx,[rcx]
+	
+	test	dl,1
+	jne	arguments_already_copied_2
+
+	mov	(-8)[rdi],rdi 
+	add	rcx,8
+
+	mov	[rdi],rdx 
+	inc	rdi
+
+	mov	(-8)[rcx],rdi 
+	add	rdi,8-1
+
+cp_hnf_arg_lp2:
+	mov	rdx,[rcx]
+	add	rcx,8
+
+	mov	[rdi],rdx 
+	add	rdi,8
+
+	dec	rax 
+	jne	cp_hnf_arg_lp2
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+arguments_already_copied_2:
+	mov	(-8)[rdi],rdx 
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_arity_0_node2:
+	cmp	rcx,offset dINT+2
+	jb	copy_real_file_or_string_2
+
+	cmp	rcx,offset CHAR+2
+	ja	copy_normal_hnf_0_2
+
+copy_int_bool_or_char_2:
+	mov	rax,8[rdx]
+	je	copy_char_2
+
+	cmp	rcx,offset dINT+2
+	jne	no_small_int_or_char_2
+
+copy_int_2:
+	cmp	rax,33
+	jae	no_small_int_or_char_2
+
+	shl	rax,4
+	add	rbp,8
+
+	add	rax,offset small_integers
+	sub	rbx,1
+
+	mov	(-8)[rbp],rax 
+	jae	copy_lp2
+	ret
+
+copy_char_2:	
+	and	rax,255
+
+	shl	rax,4
+	add	rbp,8
+
+	add	rax,offset static_characters
+	sub	rbx,1
+
+	mov	(-8)[rbp],rax 
+	jae	copy_lp2
+	ret
+	
+no_small_int_or_char_2:
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+copy_record_node2_1_b:
+ endif
+	mov	(-16)[rsi],rcx 
+	add	rbp,8
+
+	mov	(-8)[rsi],rax 
+	sub	rsi,15
+
+	mov	[rdx],rsi 
+	dec	rsi
+
+	mov	(-8)[rbp],rsi 
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_normal_hnf_0_2:
+ ifdef NEW_DESCRIPTORS
+	sub	rcx,2-(-8)
+ else
+	sub	rcx,2-(-12)
+ endif
+	sub	rbx,1
+
+	mov	[rbp],rcx 
+	lea	rbp,8[rbp]
+	jae	copy_lp2
+	ret
+
+copy_real_file_or_string_2:
+	lea	r9,__STRING__+2
+	cmp	rcx,r9
+	jbe	copy_string_or_array_2
+
+copy_real_or_file_2:
+	mov	(-24)[rsi],rcx 
+	sub	rsi,24-1
+
+	mov	[rdx],rsi 
+	dec	rsi 
+
+	mov	rax,8[rdx]
+	mov	rcx,16[rdx]
+
+	mov	[rbp],rsi 
+	add	rbp,8
+
+	mov	8[rsi],rax 
+	sub	rbx,1
+
+	mov	16[rsi],rcx 
+
+	jae	copy_lp2
+	ret
+
+already_copied_2:
+	dec	rcx 
+	sub	rbx,1
+
+	mov	[rbp],rcx 
+	lea	rbp,8[rbp]
+
+	jae	copy_lp2
+	ret
+
+copy_record_2:
+	sub	rax,258
+	ja	copy_record_node2_3
+
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+ 	jb	copy_record_node2_1
+ 
+ 	cmp	word ptr (-2+2)[rcx],0
+	je	copy_real_or_file_2
+
+	mov	qword ptr [rbp],rdi
+	mov	qword ptr [rdi],rcx
+
+	lea	rcx,1[rdi]
+	mov	rax,qword ptr 8[rdx]
+
+	mov	qword ptr [rdx],rcx
+
+	mov	qword ptr 8[rdi],rax
+	mov	rax,qword ptr 16[rdx]
+
+	add	rbp,8
+	mov	qword ptr 16[rdi],rax
+
+	add	rdi,24
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_record_node2_1:
+	mov	rax,qword ptr 8[rdx]
+
+	cmp	word ptr (-2+2)[rcx],0
+	je	copy_record_node2_1_b
+
+	mov	qword ptr [rbp],rdi
+	mov	qword ptr [rdi],rcx
+
+	lea	rcx,1[rdi]
+	mov	qword ptr 8[rdi],rax
+
+	mov	qword ptr [rdx],rcx
+	add	rbp,8
+
+	add	rdi,16
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+ else
+	mov	qword ptr [rbp],rdi
+	mov	qword ptr [rdi],rcx
+
+	lea	rcx,1[rdi]
+	mov	rax,qword ptr 8[rdx]
+
+	mov	qword ptr [rdx],rcx
+	jb	copy_record_node2_1
+
+	mov	qword ptr 8[rdi],rax
+	mov	rax,qword ptr 16[rdx]
+
+	add	rbp,8
+	mov	qword ptr 16[rdi],rax
+
+	add	rdi,24
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_record_node2_1:
+	add	rbp,8
+	mov	qword ptr 8[rdi],rax 
+
+	add	rdi,16
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+ endif
+
+copy_record_node2_3:
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+ 	cmp	word ptr (-2+2)[rcx],1
+ 	jbe	copy_record_node2_3_ab_or_b
+ endif
+
+	push	rax 
+	lea	rax,1[rdi]
+	
+	mov	qword ptr [rdx],rax 
+	mov	rax,qword ptr 16[rdx]
+
+	mov	qword ptr [rdi],rcx 
+	mov	rdx,qword ptr 8[rdx]
+
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+	mov	qword ptr 8[rdi],rdx 
+	mov	qword ptr [rbp],rdi
+	add	rbp,8
+
+	mov	rcx,rax
+	test	byte ptr [rax],1		
+	jne	record_arguments_already_copied_2
+ else
+	mov	rcx,rax 
+	sub	rax,qword ptr heap_p1
+
+	shr	rax,4
+	mov	qword ptr 8[rdi],rdx 
+
+	mov	rdx,rax 
+	and	rax,31
+
+	shr	rdx,3
+	mov	qword ptr [rbp],rdi
+
+	and	rdx,-4
+	mov	eax,dword ptr (bit_set_table)[rax*4]
+
+	add	rdx,qword ptr heap_copied_vector
+	add	rbp,8
+
+	test	eax,[rdx]
+	jne	record_arguments_already_copied_2
+
+	or	[rdx],eax
+ endif
+	lea	rdx,24[rdi]
+
+	pop	rax
+	mov	qword ptr 16[rdi],rdx
+
+	add	rdi,25
+	mov	rdx,qword ptr [rcx]
+
+	mov	qword ptr [rcx],rdi
+	add	rcx,8
+
+	mov	qword ptr (-1)[rdi],rdx
+	add	rdi,7
+
+cp_record_arg_lp2:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+
+	mov	qword ptr [rdi],rdx
+	add	rdi,8
+
+	sub	rax,1
+	jne	cp_record_arg_lp2
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+record_arguments_already_copied_2:
+	mov	rdx,qword ptr [rcx]
+	pop	rax 
+
+	mov	qword ptr 16[rdi],rdx 
+	add	rdi,24
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+copy_record_node2_3_ab_or_b:
+	jb	copy_record_node2_3_b
+
+	push	rax
+	lea	rax,1[rdi]
+
+	mov	qword ptr [rdx],rax 
+	mov	rax,qword ptr 16[rdx]
+
+	mov	qword ptr [rdi],rcx 
+	mov	rdx,qword ptr 8[rdx]
+
+	mov	rcx,rax 
+	sub	rax,qword ptr heap_p1
+
+	shr	rax,4
+	mov	qword ptr 8[rdi],rdx 
+
+	mov	rdx,rax 
+	and	rax,31
+
+	shr	rdx,3
+	mov	qword ptr [rbp],rdi
+
+	and	rdx,-4
+	mov	eax,dword ptr (bit_set_table)[rax*4]
+
+	add	rdx,qword ptr heap_copied_vector
+	add	rbp,8
+
+	test	eax,[rdx]
+	jne	record_arguments_already_copied_2
+
+	or	[rdx],eax
+	pop	rax
+
+	sub	rsi,8
+
+	shl	rax,3
+	sub	rsi,rax
+	
+	push	rsi
+	add	rsi,1
+	
+	mov	qword ptr 16[rdi],rsi
+	add	rdi,24
+	
+	mov	rdx,qword ptr [rcx]
+	jmp	cp_record_arg_lp3_c
+
+copy_record_node2_3_b:
+	push	rax
+	lea	rax,(-24+1)[rsi]
+
+	mov	qword ptr [rdx],rax 
+	mov	rax,qword ptr 16[rdx]
+
+	mov	qword ptr (-24)[rsi],rcx 
+	mov	rdx,qword ptr 8[rdx]
+
+	mov	rcx,rax 
+	sub	rax,qword ptr heap_p1
+
+	shr	rax,4
+	mov	qword ptr (-16)[rsi],rdx 
+
+	mov	rdx,rax 
+	and	rax,31
+	sub	rsi,24
+
+	shr	rdx,3
+	mov	qword ptr [rbp],rsi
+
+	and	rdx,-4
+	mov	eax,dword ptr (bit_set_table)[rax*4]
+
+	add	rdx,qword ptr heap_copied_vector
+	add	rbp,8
+
+	test	eax,[rdx]
+	jne	record_arguments_already_copied_3_b
+
+	or	[rdx],eax
+	pop	rax
+
+	mov	rdx,rsi
+	sub	rsi,8
+
+	shl	rax,3
+	sub	rsi,rax
+
+	mov	qword ptr 16[rdx],rsi
+
+	mov	rdx,qword ptr [rcx]
+
+	push	rsi
+	add	rsi,1
+
+cp_record_arg_lp3_c:
+	mov	qword ptr [rcx],rsi
+	add	rcx,8
+	mov	qword ptr (-1) [rsi],rdx
+	add	rsi,7
+
+cp_record_arg_lp3:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	
+	mov	qword ptr [rsi],rdx
+	add	rsi,8
+	
+	sub	rax,8
+	jne	cp_record_arg_lp3
+	
+	pop	rsi
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+record_arguments_already_copied_3_b:
+	mov	rdx,qword ptr [rcx]
+	pop	rax
+	
+	mov	qword ptr 16 [rsi],rdx
+	
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+ endif
+
+not_in_hnf_2:
+	test	cl,1
+	jne	already_copied_2
+
+	movsxd	rax,dword ptr (-4)[rcx]
+	test	rax,rax 
+	jle	copy_arity_0_node2_
+
+copy_node2_1_:
+	and	rax,255
+	sub	rax,2
+	jl	copy_arity_1_node2
+copy_node2_3:
+	mov	[rbp],rdi 
+	add	rbp,8
+	mov	[rdi],rcx 
+	inc	rdi
+	mov	[rdx],rdi
+	mov	rcx,8[rdx]
+	add	rdx,16
+	mov	(8-1)[rdi],rcx 
+	add	rdi,16-1
+
+cp_arg_lp2:
+	mov	rcx,[rdx]
+	add	rdx,8
+	mov	[rdi],rcx 
+	add	rdi,8
+	sub	rax,1
+	jae	cp_arg_lp2
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_arity_1_node2:
+copy_arity_1_node2_:
+	mov	[rbp],rdi 
+	inc	rdi 
+
+	add	rbp,8
+	mov	[rdx],rdi 
+
+	mov	rax,8[rdx]
+	mov	(-1)[rdi],rcx 
+
+	mov	(8-1)[rdi],rax 
+	add	rdi,24-1
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_indirection_2:
+	mov	rax,rdx 
+	mov	rdx,8[rdx]
+
+	mov	rcx,[rdx]
+	test	cl,2
+	jne	in_hnf_2
+
+	test	cl,1
+	jne	already_copied_2
+
+	cmp	dword ptr (-4)[rcx],-2
+	je	skip_indirections_2
+
+	mov	eax,(-4)[rcx]
+	test	rax,rax 
+	jle	copy_arity_0_node2_
+	jmp	copy_node2_1_
+
+skip_indirections_2:
+	mov	rdx,8[rdx]
+
+	mov	rcx,[rdx]
+	test	cl,2
+	jne	update_indirection_list_2
+	test	cl,1
+	jne	update_indirection_list_2
+
+	cmp	dword ptr (-4)[rcx],-2
+	je	skip_indirections_2
+
+update_indirection_list_2:
+	lea	rcx,8[rax]
+	mov	rax,8[rax]
+	mov	[rcx],rdx 
+	cmp	rdx,rax 
+	jne	update_indirection_list_2
+
+	jmp	continue_after_selector_2
+
+copy_selector_2:
+	cmp	rax,-2
+	je	copy_indirection_2
+	jl	copy_record_selector_2
+
+	mov	rax,8[rdx]
+ ifdef NEW_DESCRIPTORS
+	mov	d2,[rax]
+	test	d2b,2
+ 	je	copy_arity_1_node2_
+
+	mov	d3d,dword ptr (-8)[rcx]
+
+	cmp	word ptr (-2)[d2],2
+	jbe	copy_selector_2_
+ 
+ 	mov	d2,16[rax]
+
+	test	byte ptr [d2],1
+	jne	copy_arity_1_node2_
+
+	movzx	d3,word ptr 4[d3]
+	mov	qword ptr [rdx],offset __indirection
+
+	cmp	d3,16
+	jl	copy_selector_2_1
+	je	copy_selector_2_2
+
+	mov	rcx,qword ptr (-24)[d2+d3]
+	mov	qword ptr 8[rdx],rcx
+	mov	rdx,rcx
+	jmp	continue_after_selector_2
+
+copy_selector_2_1:
+	mov	rcx,qword ptr 8[rax]
+	mov	qword ptr 8[rdx],rcx
+	mov	rdx,rcx
+	jmp	continue_after_selector_2
+
+copy_selector_2_2:
+	mov	rcx,qword ptr [d2]
+	mov	qword ptr 8[rdx],rcx
+	mov	rdx,rcx
+	jmp	continue_after_selector_2
+
+copy_selector_2_:
+	movzx	d3,word ptr 4[d3]
+	mov	qword ptr [rdx],offset __indirection
+
+	mov	rcx,qword ptr [rax+d3]
+	mov	qword ptr 8[rdx],rcx
+	mov	rdx,rcx
+	jmp	continue_after_selector_2
+ else
+	mov	rax,[rax]
+ 	test	al,2
+	je	copy_arity_1_node2_
+
+	cmp	word ptr (-2)[rax],2
+	jbe	copy_selector_2_
+
+  if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+copy_selector_2__:
+  endif
+	mov	rax,8[rdx]
+	mov	rax,16[rax]
+	test	byte ptr [rax],1
+	jne	copy_arity_1_node2_
+
+copy_selector_2_:
+	mov	eax,(-8)[rcx]
+
+	mov	rcx,8[rdx]
+	push	rdx 
+	push	rbp
+	mov	eax,4[rax]
+	call	near ptr rax
+	pop	rbp 
+	pop	rdx 
+
+	mov	qword ptr [rdx],offset __indirection
+	mov	8[rdx],rcx
+
+	mov	rdx,rcx 
+	jmp	continue_after_selector_2
+ endif
+
+copy_record_selector_2:
+	cmp	rax,-3
+ 	mov	rax,qword ptr 8[rdx]
+	mov	d2,qword ptr [rax]
+	je	copy_strict_record_selector_2
+
+ 	test	d2b,2
+	je	copy_arity_1_node2_
+
+	mov	d3d,dword ptr (-8)[rcx]
+
+	cmp	word ptr (-2)[d2],258
+ ifdef NEW_DESCRIPTORS
+	jbe	copy_record_selector_2_
+ else
+	jbe	copy_selector_2_
+ endif
+
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+	cmp	word ptr (-2+2)[d2],2
+	jae	copy_selector_2__
+ endif
+
+	mov	d4,qword ptr 16[rax]
+
+	lea	d2,(-24)[d4]
+	sub	d4,qword ptr heap_p1
+
+	mov	d5,d4 
+	and	d4,31*16
+
+	shr	d5,7
+	
+	shr	d4,2
+	and	d5,-4
+
+	add	d5,qword ptr heap_copied_vector
+
+	mov	d4d,dword ptr (bit_set_table)[d4]
+
+	and	d4d,dword ptr [d5]
+
+ ifdef NEW_DESCRIPTORS
+  if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+	je	copy_record_selector_2_
+	jmp	copy_arity_1_node2_
+copy_selector_2__:
+	mov	d4,qword ptr 16[rax]
+	lea	d2,(-24)[d4]
+	test	byte ptr [d4],1
+	jne	copy_arity_1_node2_	
+  else
+	jne	copy_arity_1_node2_
+  endif
+copy_record_selector_2_:
+	movzx	d3,word ptr 4[d3]
+	mov	qword ptr [rdx],offset __indirection
+
+	cmp	d3,16
+	jle	copy_record_selector_3
+	mov	rax,d2
+copy_record_selector_3:
+	mov	rcx,qword ptr [rax+d3]
+	mov	qword ptr 8[rdx],rcx
+	mov	rdx,rcx
+	jmp	continue_after_selector_2
+ else
+	jne	copy_arity_1_node2_
+ 	jmp	copy_selector_2_
+ endif
+
+copy_strict_record_selector_2:
+	test	d2b,2
+	je	copy_arity_1_node2_
+
+	mov	d3d,dword ptr (-8)[rcx]
+
+	cmp	word ptr (-2)[d2],258
+	jbe	copy_strict_record_selector_2_
+
+ if COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP
+	cmp	word ptr (-2+2)[d2],2
+	jb	copy_strict_record_selector_2_b
+
+	mov	d4,qword ptr 16[rax]
+	lea	d2,(-24)[d4]
+	test	byte ptr [d4],1
+	jne	copy_arity_1_node2_	
+
+	jmp	copy_strict_record_selector_2_
+
+copy_strict_record_selector_2_b:
+ endif
+
+	mov	d4,qword ptr 16[rax]
+
+	lea	d2,(-24)[d4]
+	sub	d4,qword ptr heap_p1
+
+	mov	d5,d4 
+	and	d4,31*16
+
+	shr	d5,7
+
+	shr	d4,2
+	and	d5,-4
+
+	add	d5,qword ptr heap_copied_vector
+
+	mov	d4d,dword ptr (bit_set_table)[d4]
+	
+	and	d4d,[d5]
+
+	jne	copy_arity_1_node2_
+
+copy_strict_record_selector_2_:
+ ifdef NEW_DESCRIPTORS
+	movzx	rcx,word ptr 4[d3]
+	cmp	rcx,16
+	jle	copy_strict_record_selector_3
+	mov	rcx,qword ptr [d2+rcx]
+	jmp	copy_strict_record_selector_4
+copy_strict_record_selector_3:
+	mov	rcx,qword ptr [rax+rcx]
+copy_strict_record_selector_4:
+	mov	qword ptr 8[rdx],rcx
+
+	movzx	rcx,word ptr 6[d3]
+	test	rcx,rcx
+	je	copy_strict_record_selector_6
+	cmp	rcx,16
+	jle	copy_strict_record_selector_5
+	mov	rax,d2
+copy_strict_record_selector_5:
+	mov	rcx,qword ptr [rax+rcx]
+	mov	qword ptr 16[rdx],rcx
+copy_strict_record_selector_6:
+
+	mov	rcx,qword ptr (-8)[d3]
+	mov	qword ptr [rdx],rcx
+	jmp	in_hnf_2
+ else
+ 	mov	rcx,rdx 
+	mov	rdx,qword ptr 8[rdx]
+	
+	push	rbp
+	mov	eax,4[d3]
+	call	near ptr rax
+	pop	rbp 
+
+	mov	rdx,rcx 
+	mov	rcx,qword ptr [rcx]
+	test	cl,2
+	jne	in_hnf_2
+	hlt
+ endif
+
+copy_arity_0_node2_:
+	jl	copy_selector_2
+
+	mov	(-24)[rsi],rcx 
+	sub	rsi,24
+	mov	[rbp],rsi 
+	lea	rax,1[rsi]
+
+	add	rbp,8
+	mov	[rdx],rax 
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_string_or_array_2:
+	mov	rcx,rdx 
+	jne	copy_array_2
+	mov	rax,rcx 
+
+	sub	rax,heap_p1
+	cmp	rax,semi_space_size
+	jae	copy_string_constant
+
+	mov	rdx,8[rcx]
+	add	rbp,8
+
+	add	rdx,7
+	push	rbx
+
+	mov	rax,rdx 
+	and	rdx,-8
+	
+	shr	rax,3
+	sub	rsi,rdx 
+
+	mov	rbx,[rcx]
+	add	rcx,8
+
+	mov	(-16)[rsi],rbx 
+	sub	rsi,16
+
+	mov	(-8)[rbp],rsi 
+	lea	rdx,1[rsi]
+	
+	mov	(-8)[rcx],rdx 
+	lea	rdx,8[rsi]
+
+cp_s_arg_lp2:
+	mov	rbx,[rcx]
+	add	rcx,8
+
+	mov	[rdx],rbx 
+	add	rdx,8
+
+	sub	rax,1
+	jge	cp_s_arg_lp2
+
+	pop	rbx 
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_string_constant:
+	mov	qword ptr [rbp],rdx 
+	add	rbp,8
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+copy_array_2:
+	push	rbx 
+
+	mov	rax,qword ptr 16[rcx]
+	test	rax,rax 
+	je	copy_array_a2
+
+	movzx 	rbx,word ptr (-2)[rax]
+
+	test	rbx,rbx 
+	je	copy_strict_basic_array_2
+	
+	sub	rbx,256
+	imul	rbx,qword ptr 8[rcx]
+	jmp	copy_array_a3
+
+copy_array_a2:
+	mov	rbx,qword ptr 8[rcx]
+copy_array_a3:
+	mov	rdx,rdi 
+	lea	rdi,24[rdi+rbx*8]
+
+	mov	qword ptr [rbp],rdx 
+	mov	rax,qword ptr [rcx]
+
+	add	rbp,8
+	mov	qword ptr [rdx],rax 
+
+	lea	rax,1[rdx]
+	add	rdx,8
+
+	mov	qword ptr [rcx],rax 
+	add	rcx,8
+
+	lea	rax,1[rbx]
+	jmp	cp_s_arg_lp2
+
+copy_strict_basic_array_2:
+	mov	rbx,qword ptr 8[rcx]
+	cmp	rax,offset BOOL+2
+	je	copy_bool_array_2
+
+copy_int_array_2:
+	shl	rbx,3
+	lea	rdx,(-24)[rsi]
+
+	sub	rdx,rbx 
+	mov	rax,qword ptr [rcx]
+
+	shr	rbx,3
+	mov	qword ptr [rbp],rdx 
+
+	add	rbp,8
+	mov	rsi,rdx 
+	
+	mov	qword ptr [rdx],rax 
+	lea	rax,1[rdx]
+
+	add	rdx,8
+	mov	qword ptr [rcx],rax 
+
+	add	rcx,8
+	lea	rax,1[rbx]
+	jmp	cp_s_arg_lp2
+
+copy_bool_array_2:
+	lea	rax,7[rbx]
+	shr	rax,3
+	jmp	copy_int_array_2
+
+end_copy1:
+	mov	heap_end_after_gc,rsi 
+
+	mov	rcx,offset finalizer_list
+	mov	rdx,offset free_finalizer_list
+	mov	rbp,qword ptr finalizer_list
+
+determine_free_finalizers_after_copy:
+	mov	rax,qword ptr [rbp]
+	test	al,1
+	je	finalizer_not_used_after_copy
+
+	mov	rbp,qword ptr 8[rbp]
+	sub	rax,1
+	mov	qword ptr [rcx],rax 
+	lea	rcx,8[rax]
+	jmp	determine_free_finalizers_after_copy
+
+finalizer_not_used_after_copy:
+	lea	r9,__Nil-8
+	cmp	rbp,r9
+	je	end_finalizers_after_copy
+
+	mov	qword ptr [rdx],rbp 
+	lea	rdx,8[rbp]
+	mov	rbp,qword ptr 8[rbp]
+	jmp	determine_free_finalizers_after_copy	
+
+end_finalizers_after_copy:
+	mov	qword ptr [rcx],rbp 
+	mov	qword ptr [rdx],rbp 
diff --git a/amark.asm b/amark.asm
new file mode 100644
index 0000000..aaff6ab
--- /dev/null
+++ b/amark.asm
@@ -0,0 +1,1927 @@
+
+	mov	rax,qword ptr heap_size_65
+	xor	rbx,rbx 
+	
+	mov	qword ptr n_marked_words,rbx 
+	shl	rax,6
+
+	mov	qword ptr lazy_array_list,rbx 
+	mov	qword ptr heap_size_64_65,rax 
+	
+	lea	rsi,(-4000)[rsp]
+
+	mov	rax,qword ptr caf_list
+
+	mov	qword ptr end_stack,rsi 
+
+	mov	r10,neg_heap_p3
+	mov	r11,heap_size_64_65
+	mov	r13,qword ptr end_stack
+	mov	r14,0
+
+	test	rax,rax 
+	je	_end_mark_cafs
+
+_mark_cafs_lp:
+	mov	rbx,qword ptr [rax]
+	mov	rbp,qword ptr (-8)[rax]
+
+	push	rbp
+	lea	rbp,8[rax]
+	lea	r12,8[rax+rbx*8]
+
+	call	_mark_stack_nodes
+
+	pop	rax 
+	test	rax,rax 
+	jne	_mark_cafs_lp
+
+_end_mark_cafs:
+	mov	rsi,qword ptr stack_top
+	mov	rbp,qword ptr stack_p
+
+	mov	r12,rsi 
+	call	_mark_stack_nodes
+
+continue_mark_after_pmark:
+	mov	qword ptr n_marked_words,r14
+
+	mov	rcx,qword ptr lazy_array_list
+
+	test	rcx,rcx 
+	je	end_restore_arrays
+
+restore_arrays:
+	mov	rbx ,qword ptr [rcx]
+	lea	r9,__ARRAY__+2
+	mov	qword ptr [rcx],r9
+
+	cmp	rbx,1
+	je	restore_array_size_1
+
+	lea	rdx,[rcx+rbx*8]
+	mov	rax,qword ptr 16[rdx]
+	test	rax,rax 
+	je	restore_lazy_array
+
+	mov	rbp,rax 
+	push	rdx 
+
+	xor	rdx,rdx 
+	mov	rax,rbx 
+	movzx	rbx,word ptr (-2+2)[rbp]
+
+	div	rbx 
+	mov	rbx,rax 
+
+	pop	rdx
+	mov	rax,rbp 
+
+restore_lazy_array:
+	mov	rdi,qword ptr 16[rcx]
+	mov	rbp,qword ptr 8[rcx]
+	mov	qword ptr 8[rcx],rbx 
+	mov	rsi,qword ptr 8[rdx]
+	mov	qword ptr 16[rcx],rax 
+	mov	qword ptr 8[rdx],rbp 
+	mov	qword ptr 16[rdx],rdi 
+
+	test	rax,rax
+	je	no_reorder_array
+
+	movzx	rdx,word ptr (-2)[rax]
+	sub	rdx,256
+	movzx	rbp,word ptr (-2+2)[rax]
+	cmp	rbp,rdx 
+	je	no_reorder_array
+
+	add	rcx,24
+	imul	rbx,rdx 
+	mov	rax,rdx 
+	lea	rdx,[rcx+rbx*8]
+	mov	rbx,rbp 
+	sub	rax,rbp 
+
+	call	reorder
+
+no_reorder_array:
+	mov	rcx,rsi 
+	test	rcx,rcx 
+	jne	restore_arrays
+
+	jmp	end_restore_arrays
+
+restore_array_size_1:
+	mov	rbp,qword ptr 8[rcx]
+	mov	rdx,qword ptr 16[rcx]
+	mov	qword ptr 8[rcx],rbx 
+	mov	rax,qword ptr 24[rcx]
+	mov	qword ptr 24[rcx],rbp 
+	mov	qword ptr 16[rcx],rax 
+
+	mov	rcx,rdx 
+	test	rcx,rcx 
+	jne	restore_arrays
+
+end_restore_arrays:
+	mov	rdi,qword ptr heap_vector
+	lea	rcx,finalizer_list
+	lea	rdx,free_finalizer_list
+
+	mov	rbp,qword ptr [rcx]
+determine_free_finalizers_after_mark:
+	lea	r9,__Nil-8
+	cmp	rbp,r9
+	je	end_finalizers_after_mark
+
+	lea	rax,[r10+rbp]
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	esi,dword ptr (bit_set_table2)[rax]
+	test	esi,dword ptr [rdi+rbx*4]
+	je	finalizer_not_used_after_mark
+
+	lea	rcx,8[rbp]
+	mov	rbp,qword ptr 8[rbp]
+	jmp	determine_free_finalizers_after_mark
+
+finalizer_not_used_after_mark:
+	mov	qword ptr [rdx],rbp 
+	lea	rdx,8[rbp]
+
+	mov	rbp,qword ptr 8[rbp]
+	mov	qword ptr [rcx],rbp 
+	jmp	determine_free_finalizers_after_mark
+
+end_finalizers_after_mark:
+	mov	qword ptr [rdx],rbp 
+
+	call	add_garbage_collect_time
+
+	mov	rax,qword ptr bit_vector_size
+
+	mov	rdi,qword ptr n_allocated_words
+	add	rdi,qword ptr n_marked_words
+	shl	rdi,3
+
+	mov	rsi,rax
+	shl	rsi,3
+
+	push	rdx 
+	push	rax 
+
+	mov	rax,rdi 
+	mul	qword ptr heap_size_multiple
+	shrd	rax,rdx,8
+	shr	rdx,8
+
+	mov	rbx,rax 
+	test	rdx,rdx 
+	
+	pop	rax 
+	pop	rdx 
+	
+	je	not_largest_heap
+
+	mov	rbx,qword ptr heap_size_65
+	shl	rbx,6
+
+not_largest_heap:
+	cmp	rbx,rsi 
+	jbe	no_larger_heap
+	
+	mov	rsi,qword ptr heap_size_65
+	shl	rsi,6
+	cmp	rbx,rsi
+	jbe	not_larger_than_heap
+	mov	rbx,rsi 
+not_larger_than_heap:
+	mov	rax,rbx 
+	shr	rax,3
+	mov	qword ptr bit_vector_size,rax 
+no_larger_heap:
+
+	mov	rbp,rax
+
+	mov	rdi,qword ptr heap_vector
+
+	shr	rbp,5
+
+	test	al,31
+	je	no_extra_word
+
+	mov	dword ptr [rdi+rbp*4],0
+
+no_extra_word:
+	sub	rax,qword ptr n_marked_words
+	shl	rax,3
+	mov	qword ptr n_last_heap_free_bytes,rax 
+
+	mov	rax,qword ptr n_marked_words
+	shl	rax,3
+	add	qword ptr total_gc_bytes,rax 
+
+	test	qword ptr flags,2
+	je	_no_heap_use_message2
+
+	mov	r12,rsp
+	and	rsp,-16
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+
+	lea	rdi,marked_gc_string_1
+ else
+	sub	rsp,32
+
+	lea	rcx,marked_gc_string_1
+ endif
+	call	ew_print_string
+
+ ifdef LINUX
+	mov	rdi,qword ptr n_marked_words
+	shl	rdi,3
+ else
+	mov	rcx,qword ptr n_marked_words
+	shl	rcx,3
+ endif
+	call	ew_print_int
+
+ ifdef LINUX
+	lea	rdi,heap_use_after_gc_string_2
+ else
+	lea	rcx,heap_use_after_gc_string_2
+ endif
+	call	ew_print_string
+
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,r12
+
+_no_heap_use_message2:
+	call	call_finalizers
+
+	mov	rsi,qword ptr n_allocated_words
+	xor	rbx,rbx 
+
+	mov	rcx,rdi 
+	mov	qword ptr n_free_words_after_mark,rbx 
+
+_scan_bits:
+	cmp	ebx,dword ptr [rcx]
+	je	_zero_bits
+	mov	dword ptr [rcx],ebx 
+	add	rcx,4
+	sub	rbp,1
+	jne	_scan_bits
+
+	jmp	_end_scan
+
+_zero_bits:
+	lea	rdx,4[rcx]
+	add	rcx,4
+	sub	rbp,1
+	jne	_skip_zero_bits_lp1
+	jmp	_end_bits
+
+_skip_zero_bits_lp:
+	test	rax,rax 
+	jne	_end_zero_bits
+_skip_zero_bits_lp1:
+	mov	eax,dword ptr [rcx]
+	add	rcx,4
+	sub	rbp,1
+	jne	_skip_zero_bits_lp
+
+	test	rax,rax
+	je	_end_bits
+	mov	rax,rcx
+	mov	dword ptr (-4)[rcx],ebx
+	sub	rax,rdx
+	jmp	_end_bits2	
+
+_end_zero_bits:
+	mov	rax,rcx 
+	sub	rax,rdx 
+	shl	rax,3
+	add	qword ptr n_free_words_after_mark,rax 
+	mov	dword ptr (-4)[rcx],ebx 
+
+	cmp	rax,rsi
+	jb	_scan_bits
+
+_found_free_memory:
+	mov	qword ptr bit_counter,rbp 
+	mov	qword ptr bit_vector_p,rcx 
+
+	lea	rbx,(-4)[rdx]
+	sub	rbx,rdi 
+	shl	rbx,6
+	mov	rdi,qword ptr heap_p3
+	add	rdi,rbx 
+
+	mov	r15,rax
+	lea	rbx,[rdi+rax*8]
+
+	sub	r15,rsi
+	mov	rsi,qword ptr stack_top
+
+	mov	qword ptr heap_end_after_gc,rbx 
+
+	jmp	restore_registers_after_gc_and_return
+
+_end_bits:
+	mov	rax,rcx 
+	sub	rax,rdx 
+	add	rax,4
+_end_bits2:
+	shl	rax,3
+	add	qword ptr n_free_words_after_mark,rax 
+	cmp	rax,rsi 
+	jae	_found_free_memory
+
+_end_scan:
+	mov	qword ptr bit_counter,rbp 
+	jmp	compact_gc
+
+; %rbp : pointer to stack element
+; %rdi : heap_vector
+; %rax ,%rbx ,%rcx ,%rdx ,%rsi : free
+
+_mark_stack_nodes:
+	cmp	rbp,r12
+	je	_end_mark_nodes
+_mark_stack_nodes_:
+	mov	rcx,qword ptr [rbp]
+
+	add	rbp,8
+	lea	rdx,[r10+rcx]
+
+	cmp	rdx,r11
+	jnc	_mark_stack_nodes
+
+	mov	rbx,rdx 
+	and	rdx,31*8
+	shr	rbx,8
+	mov	esi,dword ptr (bit_set_table2)[rdx]
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	_mark_stack_nodes
+
+	push	rbp
+	push	0
+	jmp	_mark_arguments
+
+_mark_hnf_2:
+	cmp	rsi,20000000h
+	jbe	fits_in_word_6
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_6:
+	add	r14,3
+
+_mark_record_2_c:
+	mov	rbx,qword ptr 8[rcx]
+	push	rbx 
+
+	cmp	rsp,r13
+	jb	__mark_using_reversal
+
+_mark_node2:
+_shared_argument_part:
+	mov	rcx,qword ptr [rcx]
+
+_mark_node:
+	lea	rdx,[r10+rcx]
+	cmp	rdx,r11
+	jnc	_mark_next_node
+
+	mov	rbx,rdx 
+	and	rdx,31*8
+	shr	rbx,8
+	mov	esi,dword ptr (bit_set_table2)[rdx]
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	_mark_next_node
+
+_mark_arguments:
+	mov	rax,qword ptr [rcx]
+	test	rax,2
+	je	_mark_lazy_node
+	
+	movzx	rbp,word ptr (-2)[rax]
+
+	test	rbp,rbp
+	je	_mark_hnf_0
+
+	or	dword ptr [rdi+rbx*4],esi 
+	add	rcx,8
+
+	cmp	rbp,256
+	jae	_mark_record
+
+	sub	rbp,2
+	je	_mark_hnf_2
+	jb	_mark_hnf_1
+
+_mark_hnf_3:
+	mov	rdx,qword ptr 8[rcx]
+
+	cmp	rsi,20000000h
+	jbe	fits_in_word_1
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_1:	
+
+	add	r14,3
+	lea	rax,[r10+rdx]
+	mov	rbx,rax
+	
+	and	rax,31*8
+	shr	rbx,8
+
+	mov	esi,dword ptr (bit_set_table2)[rax]
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	_shared_argument_part
+
+_no_shared_argument_part:
+	or	dword ptr [rdi+rbx*4],esi 
+	add	rbp,1
+
+	add	r14,rbp 
+	lea	rax,[rax+rbp*8]
+	lea	rdx,(-8)[rdx+rbp*8]
+
+	cmp	rax,32*8
+	jbe	fits_in_word_2
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_2:
+
+	mov	rbx,qword ptr [rdx]
+	sub	rbp,2
+	push	rbx 
+
+_push_hnf_args:
+	mov	rbx,qword ptr (-8)[rdx]
+	sub	rdx,8
+	push	rbx 
+	sub	rbp,1
+	jge	_push_hnf_args
+
+	cmp	rsp,r13
+	jae	_mark_node2
+
+	jmp	__mark_using_reversal
+
+_mark_hnf_1:
+	cmp	rsi,40000000h
+	jbe	fits_in_word_4
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_4:
+	add	r14,2
+	mov	rcx,qword ptr [rcx]
+	jmp	_mark_node
+
+_mark_lazy_node_1:
+	add	rcx,8
+	or	dword ptr [rdi+rbx*4],esi 
+	cmp	rsi,20000000h
+	jbe	fits_in_word_3
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_3:
+	add	r14,3
+
+	cmp	rbp,1
+	je	_mark_node2
+
+_mark_selector_node_1:
+	add	rbp,2
+	mov	rdx,qword ptr [rcx]
+	je	_mark_indirection_node
+
+	lea	rsi,[r10+rdx]
+	mov	rbx,rsi 
+
+	shr	rbx,8
+	and	rsi,31*8
+
+	add	rbp,1
+
+	mov	esi,dword ptr (bit_set_table2)[rsi]
+	jle	_mark_record_selector_node_1
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	_mark_node3
+
+	mov	rbp,qword ptr [rdx]
+	test	rbp,2
+	je	_mark_node3
+
+	cmp	word ptr (-2)[rbp],2
+	jbe	_small_tuple_or_record
+
+_large_tuple_or_record:
+	mov	rbp,qword ptr 16[rdx]
+	mov	r9,rbp
+
+	add	rbp,r10
+	mov	rbx,rbp
+	and	rbp,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_set_table2)[rbp]
+	test	ebp,dword ptr [rdi+rbx*4]
+	jne	_mark_node3
+
+ ifdef NEW_DESCRIPTORS
+	mov	eax,(-8)[rax]
+	lea	rbp,__indirection
+	mov	qword ptr (-8)[rcx],rbp
+	movzx	eax,word ptr 4[rax]
+	mov	rbp,rcx
+
+	cmp	rax,16
+	jl	_mark_tuple_selector_node_1
+	mov	rdx,r9
+	je	_mark_tuple_selector_node_2
+	mov	rcx,qword ptr (-24)[r9+rax]
+	mov	qword ptr [rbp],rcx
+	jmp	_mark_node
+
+_mark_tuple_selector_node_2:
+	mov	rcx,qword ptr [r9]
+	mov	qword ptr [rbp],rcx
+	jmp	_mark_node	
+ endif
+
+_small_tuple_or_record:
+ ifdef NEW_DESCRIPTORS
+	mov	eax,(-8)[rax]
+	lea	rbp,__indirection
+	mov	qword ptr (-8)[rcx],rbp
+	movzx	eax,word ptr 4[rax]
+	mov	rbp,rcx
+_mark_tuple_selector_node_1:
+	mov	rcx,qword ptr [rdx+rax]
+	mov	qword ptr [rbp],rcx	
+ else
+	mov	eax,(-8)[rax]
+	push	rcx
+	mov	rcx,rdx
+	mov	eax,4[rax]
+	call	near ptr rax
+	pop	rdx 
+
+	lea	r9,__indirection
+	mov	qword ptr (-8)[rdx],r9
+	mov	qword ptr [rdx],rcx 
+ endif
+	jmp	_mark_node
+
+_mark_record_selector_node_1:
+	je	_mark_strict_record_selector_node_1
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	_mark_node3
+
+	mov	rbp,qword ptr [rdx]
+	test	rbp,2
+	je	_mark_node3
+
+	cmp	word ptr (-2)[rbp],258
+	jbe	_small_tuple_or_record
+ ifdef NEW_DESCRIPTORS
+	mov	rbp,qword ptr 16[rdx]
+	mov	r9,rbp
+
+	add	rbp,r10
+	mov	rbx,rbp
+	and	rbp,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_set_table2)[rbp]
+	test	ebp,dword ptr [rdi+rbx*4]
+	jne	_mark_node3
+
+	mov	eax,(-8)[rax]
+	lea	rbp,__indirection
+	mov	qword ptr (-8)[rcx],rbp
+	movzx	eax,word ptr 4[rax]
+	mov	rbp,rcx
+
+	cmp	rax,16
+	jle	_mark_record_selector_node_2
+	mov	rdx,r9
+	sub	rax,24
+_mark_record_selector_node_2:
+	mov	rcx,qword ptr [rdx+rax]
+	mov	qword ptr [rbp],rcx
+	jmp	_mark_node
+ else
+	jmp	_large_tuple_or_record
+ endif
+
+_mark_strict_record_selector_node_1:
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	_mark_node3
+
+	mov	rbp,qword ptr [rdx]
+	test	rbp,2
+	je	_mark_node3
+
+	cmp	word ptr (-2)[rbp],258
+	jbe	_select_from_small_record
+
+	mov	rbp,qword ptr 16[rdx]
+	mov	r9,rbp
+
+	add	rbp,r10
+	mov	rbx,rbp 
+	and	rbp,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_set_table2)[rbp]
+	test	ebp,dword ptr [rdi+rbx*4]
+	jne	_mark_node3
+
+_select_from_small_record:
+	mov	eax,dword ptr (-8)[rax]
+	sub	rcx,8
+
+ ifdef NEW_DESCRIPTORS
+	movzx	ebx,word ptr 4[rax]
+	cmp	rbx,16
+	jle	_mark_strict_record_selector_node_2
+	mov	rbx,qword ptr (-24)[r9+rbx]
+	jmp	_mark_strict_record_selector_node_3
+_mark_strict_record_selector_node_2:
+	mov	rbx,qword ptr [rdx+rbx]
+_mark_strict_record_selector_node_3:
+	mov	qword ptr 8[rcx],rbx
+
+	movzx	ebx,word ptr 6[rax]
+	test	rbx,rbx
+	je	_mark_strict_record_selector_node_5
+	cmp	rbx,16
+	jle	_mark_strict_record_selector_node_4
+	mov	rdx,r9
+	sub	rbx,24
+_mark_strict_record_selector_node_4:
+	mov	rbx,qword ptr [rdx+rbx]
+	mov	qword ptr 16[rcx],rbx
+_mark_strict_record_selector_node_5:
+
+	mov	rax,qword ptr (-8)[rax]
+	mov	qword ptr [rcx],rax
+ else
+	mov	eax,4[rax]
+	call	near ptr rax
+ endif
+	jmp	_mark_next_node
+
+_mark_indirection_node:
+_mark_node3:
+	mov	rcx,rdx 
+	jmp	_mark_node
+
+_mark_next_node:
+	pop	rcx 
+	test	rcx,rcx 
+	jne	_mark_node
+
+	pop	rbp 
+	cmp	rbp,r12
+	jne	_mark_stack_nodes_
+
+_end_mark_nodes:
+	ret
+
+_mark_lazy_node:
+	movsxd	rbp,dword ptr (-4)[rax]
+	test	rbp,rbp 
+	je	_mark_real_or_file
+
+	cmp	rbp,1
+	jle	_mark_lazy_node_1
+
+	cmp	rbp,256
+	jge	_mark_closure_with_unboxed_arguments
+	inc	rbp
+	or	dword ptr [rdi+rbx*4],esi 
+
+	add	r14,rbp 
+	lea	rdx,[rdx+rbp*8]
+	lea	rcx,[rcx+rbp*8]
+
+	cmp	rdx,32*8
+	jbe	fits_in_word_7
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_7:
+	sub	rbp,3
+_push_lazy_args:
+	mov	rbx,qword ptr (-8)[rcx]
+	sub	rcx,8
+	push	rbx
+	sub	rbp,1
+	jge	_push_lazy_args
+
+	sub	rcx,8
+
+	cmp	rsp,r13
+	jae	_mark_node2
+	
+	jmp	__mark_using_reversal
+
+_mark_closure_with_unboxed_arguments:
+	mov	rax,rbp 
+	and	rbp,255
+	sub	rbp,1
+	je	_mark_real_or_file
+
+	shr	rax,8
+	add	rbp,2
+
+	or	dword ptr [rdi+rbx*4],esi 
+	add	r14,rbp 
+	lea	rdx,[rdx+rbp*8]
+
+	sub	rbp,rax
+
+	cmp	rdx,32*8
+	jbe	fits_in_word_7_
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_7_:
+	sub	rbp,2
+	jl	_mark_next_node
+
+	lea	rcx,16[rcx+rbp*8]
+	jne	_push_lazy_args
+
+_mark_closure_with_one_boxed_argument:
+	mov	rcx,qword ptr (-8)[rcx]
+	jmp	_mark_node
+
+_mark_hnf_0:
+	lea	r9,dINT+2
+	cmp	rax,r9
+	jb	_mark_real_file_or_string
+
+	or	dword ptr [rdi+rbx*4],esi 
+
+	lea	r9,CHAR+2
+	cmp	rax,r9
+	ja	_mark_normal_hnf_0
+
+_mark_bool:
+	add	r14,2
+
+	cmp	rsi,40000000h
+	jbe	_mark_next_node
+
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	_mark_next_node
+
+_mark_normal_hnf_0:
+	inc	r14
+	jmp	_mark_next_node
+
+_mark_real_file_or_string:
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+	jbe	_mark_string_or_array
+
+_mark_real_or_file:
+	or	dword ptr [rdi+rbx*4],esi 
+	add	r14,3
+
+	cmp	rsi,20000000h
+	jbe	_mark_next_node
+
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	_mark_next_node
+
+_mark_record:
+	sub	rbp,258
+	je	_mark_record_2
+	jl	_mark_record_1
+
+_mark_record_3:
+	add	r14,3
+
+	cmp	rsi,20000000h
+	jbe	fits_in_word_13
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_13:
+	mov	rdx,qword ptr 8[rcx]
+
+	movzx	rbx,word ptr (-2+2)[rax]
+	lea	rsi,[r10+rdx]
+
+	mov	rax,rsi 
+	and	rsi,31*8
+
+	shr	rax,8
+	sub	rbx,1
+
+	mov	edx,dword ptr (bit_set_table2)[rsi]
+	jb	_mark_record_3_bb
+
+	test	edx,dword ptr [rdi+rax*4]
+	jne	_mark_node2
+
+	add	rbp,1
+	or	dword ptr [rdi+rax*4],edx 
+	add	r14,rbp 
+	lea	rsi,[rsi+rbp*8]
+
+	cmp	rsi,32*8
+	jbe	_push_record_arguments
+	or	dword ptr 4[rdi+rax*4],1
+_push_record_arguments:
+	mov	rdx,qword ptr 8[rcx]
+	mov	rbp,rbx 
+	shl	rbx,3
+	add	rdx,rbx 
+	sub	rbp,1
+	jge	_push_hnf_args
+
+	jmp	_mark_node2
+
+_mark_record_3_bb:
+	test	edx,dword ptr [rdi+rax*4]
+	jne	_mark_next_node
+
+	add	rbp,1
+	or	dword ptr [rdi+rax*4],edx 
+	add	r14,rbp 
+	lea	rsi,[rsi+rbp*8]
+	
+	cmp	rsi,32*8
+	jbe	_mark_next_node
+
+	or	dword ptr 4[rdi+rax*4],1
+	jmp	_mark_next_node
+
+_mark_record_2:
+	cmp	rsi,20000000h
+	jbe	fits_in_word_12
+	or	dword ptr 4[rdi+rbx*4],1
+fits_in_word_12:
+	add	r14,3
+
+	cmp	word ptr (-2+2)[rax],1
+	ja	_mark_record_2_c
+	je	_mark_node2
+	jmp	_mark_next_node
+
+_mark_record_1:
+	cmp	word ptr (-2+2)[rax],0
+	jne	_mark_hnf_1
+
+	jmp	_mark_bool
+
+_mark_string_or_array:
+	je	_mark_string_
+
+_mark_array:
+	mov	rbp,qword ptr 16[rcx]
+	test	rbp,rbp
+	je	_mark_lazy_array
+
+	movzx	rax,word ptr (-2)[rbp]
+
+	test	rax,rax 
+	je	_mark_strict_basic_array
+
+	movzx	rbp,word ptr (-2+2)[rbp]
+	test	rbp,rbp 
+	je	_mark_b_record_array
+
+	cmp	rsp,r13
+	jb	_mark_array_using_reversal
+
+	sub	rax,256
+	cmp	rax,rbp 
+	je	_mark_a_record_array
+
+_mark_ab_record_array:
+	or	dword ptr [rdi+rbx*4],esi 
+	mov	rbp,qword ptr 8[rcx]
+
+	imul	rax,rbp 
+	add	rax,3
+
+	add	r14,rax 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+
+	cmp	rbx,rax 
+	jae	_end_set_ab_array_bits
+
+	inc	rbx 
+	mov	rbp,1
+	cmp	rbx,rax 
+	jae	_last_ab_array_bits
+
+_mark_ab_array_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx
+	cmp	rbx,rax 
+	jb	_mark_ab_array_lp
+
+_last_ab_array_bits:
+	or	dword ptr [rdi+rbx*4],ebp 
+
+_end_set_ab_array_bits:
+	mov	rax,qword ptr 8[rcx]
+	mov	rdx,qword ptr 16[rcx]
+	movzx	rbx,word ptr (-2+2)[rdx]
+	movzx	rdx,word ptr (-2)[rdx]
+	shl	rbx,3
+	lea	rdx,(-2048)[rdx*8]
+	push	rbx
+	push	rdx
+	lea	rbp,24[rcx]
+	push	r12
+	jmp	_mark_ab_array_begin
+	
+_mark_ab_array:
+	mov	rbx,qword ptr 16[rsp]
+	push	rax
+	push	rbp
+	lea	r12,[rbp+rbx]
+
+	call	_mark_stack_nodes
+
+	mov	rbx,qword ptr (8+16)[rsp]
+	pop	rbp
+	pop	rax
+	add	rbp,rbx 
+_mark_ab_array_begin:
+	sub	rax,1
+	jnc	_mark_ab_array
+
+	pop	r12
+	add	rsp,16
+	jmp	_mark_next_node
+
+_mark_a_record_array:
+	or	dword ptr [rdi+rbx*4],esi 
+	mov	rbp,qword ptr 8[rcx]
+
+	imul	rax,rbp 
+	push	rax
+
+	add	rax,3
+
+	add	r14,rax 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+	
+	cmp	rbx,rax 
+	jae	_end_set_a_array_bits
+
+	inc	rbx 
+	mov	rbp,1
+	cmp	rbx,rax 
+	jae	_last_a_array_bits
+
+_mark_a_array_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx 
+	cmp	rbx,rax
+	jb	_mark_a_array_lp
+
+_last_a_array_bits:
+	or	dword ptr [rdi+rbx*4],ebp
+
+_end_set_a_array_bits:
+	pop	rax 
+	lea	rbp,24[rcx]
+
+	push	r12
+	lea	r12,24[rcx+rax*8]
+
+	call	_mark_stack_nodes
+
+	pop	r12
+	jmp	_mark_next_node
+
+_mark_lazy_array:
+	cmp	rsp,r13
+	jb	_mark_array_using_reversal
+
+	or	dword ptr [rdi+rbx*4],esi 
+	mov	rax,qword ptr 8[rcx]
+
+	add	rax,3
+
+	add	r14,rax 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+	
+	cmp	rbx,rax 
+	jae	_end_set_lazy_array_bits
+
+	inc	rbx 
+	mov	rbp,1
+	cmp	rbx,rax 
+	jae	_last_lazy_array_bits
+
+_mark_lazy_array_lp:
+	or	dword ptr [rdi+rbx*4],ebp
+	inc	rbx
+	cmp	rbx,rax
+	jb	_mark_lazy_array_lp
+
+_last_lazy_array_bits:
+	or	dword ptr [rdi+rbx*4],ebp
+
+_end_set_lazy_array_bits:
+	mov	rax,qword ptr 8[rcx]
+	lea	rbp,24[rcx]
+
+	push	r12
+	lea	r12,24[rcx+rax*8]
+
+	call	_mark_stack_nodes
+
+	pop	r12
+	jmp	_mark_next_node
+
+_mark_array_using_reversal:
+	push	0
+	mov	rsi,1
+	jmp	__mark_node
+
+_mark_strict_basic_array:
+	mov	rax,qword ptr 8[rcx]
+	cmp	rbp,offset BOOL+2
+	je	_mark_strict_bool_array
+_mark_strict_real_array:
+_mark_strict_int_array:
+	add	rax,3
+	jmp	_mark_basic_array_
+_mark_strict_bool_array:
+	add	rax,24+7
+	shr	rax,3
+	jmp	_mark_basic_array_
+
+_mark_b_record_array:
+	mov	rbp,qword ptr 8[rcx]
+	sub	rax,256
+	imul	rax,rbp 
+	add	rax,3
+	jmp	_mark_basic_array_
+
+_mark_string_:
+	mov	rax,qword ptr 8[rcx]
+	add	rax,16+7
+	shr	rax,3
+
+_mark_basic_array_:
+	or	dword ptr [rdi+rbx*4],esi 
+
+	add	r14,rax 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+	
+	cmp	rbx,rax 
+	jae	_mark_next_node
+
+	inc	rbx 
+	mov	rbp,1
+	cmp	rbx,rax 
+	jae	_last_string_bits
+
+_mark_string_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx
+	cmp	rbx,rax 
+	jb	_mark_string_lp
+
+_last_string_bits:
+	or	dword ptr [rdi+rbx*4],ebp 
+	jmp	_mark_next_node
+
+__end_mark_using_reversal:
+	pop	rdx 
+	test	rdx,rdx 
+	je	_mark_next_node
+	mov	qword ptr [rdx],rcx 
+	jmp	_mark_next_node
+
+__mark_using_reversal:
+	push	rcx 
+	mov	rsi,1
+	mov	rcx,qword ptr [rcx]
+	jmp	__mark_node
+
+__mark_arguments:
+	mov	rax,qword ptr [rcx]
+	test	al,2
+	je	__mark_lazy_node
+
+	movzx	rbp,word ptr (-2)[rax]
+	test	rbp,rbp 
+	je	__mark_hnf_0
+
+	add	rcx,8
+
+	cmp	rbp,256
+	jae	__mark__record
+
+	sub	rbp,2
+	je	__mark_hnf_2
+	jb	__mark_hnf_1
+
+__mark_hnf_3:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+
+	or	dword ptr [rdi+rbx*4],edx
+
+	cmp	rdx,20000000h
+
+	mov	rax,qword ptr 8[rcx]
+
+	jbe	fits__in__word__1
+	or	dword ptr 4[rdi+rbx*4],1
+fits__in__word__1:
+	add	rax,r10
+
+	mov	rbx,rax 
+	and	rax,31*8
+
+	shr	rbx,8
+
+	mov	edx,dword ptr (bit_set_table2)[rax]
+	test	edx,dword ptr [rdi+rbx*4]
+	jne	__shared_argument_part
+
+__no_shared_argument_part:
+	or	dword ptr [rdi+rbx*4],edx 
+	mov	rdx,qword ptr 8[rcx]
+
+	add	rbp,1
+	mov	qword ptr 8[rcx],rsi 
+
+	add	r14,rbp 
+	add	rcx,8
+
+	shl	rbp,3
+	or	qword ptr [rdx],1
+
+	add	rax,rbp
+	add	rdx,rbp
+
+	cmp	rax,32*8
+	jbe	fits__in__word__2
+	or	dword ptr 4[rdi+rbx*4],1
+fits__in__word__2:
+
+	mov	rbp ,qword ptr (-8)[rdx]
+	mov	qword ptr (-8)[rdx],rcx 
+	lea	rsi,(-8)[rdx]
+	mov	rcx,rbp 
+	jmp	__mark_node
+
+__mark_hnf_1:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,2
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,40000000h
+	jbe	__shared_argument_part
+	or	dword ptr 4[rdi+rbx*4],1
+__shared_argument_part:
+	mov	rbp,qword ptr [rcx]
+	mov	qword ptr [rcx],rsi
+	lea	rsi,2[rcx]
+	mov	rcx,rbp 
+	jmp	__mark_node
+
+__mark_no_selector_2:
+	pop	rbx
+__mark_no_selector_1:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,20000000h
+	jbe	__shared_argument_part
+
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	__shared_argument_part
+
+__mark_lazy_node_1:
+	je	__mark_no_selector_1
+
+__mark_selector_node_1:
+	add	rbp,2
+	je	__mark_indirection_node
+
+	add	rbp,1
+
+	push	rbx
+	mov	rbp,qword ptr [rcx]
+	push	rax 
+	lea	rax,[r10+rbp] 
+
+	jle	__mark_record_selector_node_1
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rbx*4]
+	pop	rax 
+	jne	__mark_no_selector_2
+
+	mov	rbx,qword ptr [rbp]
+	test	bl,2
+	je	__mark_no_selector_2
+
+	cmp	word ptr (-2)[rbx],2
+	jbe	__small_tuple_or_record
+
+__large_tuple_or_record:
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8
+	and	r8,31*8
+	shr	rbx,8
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+	jne	__mark_no_selector_2
+
+ ifdef NEW_DESCRIPTORS
+	mov	eax,dword ptr (-8)[rax]
+	lea	rdx,__indirection
+	pop	rbx
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+	mov	r8,rcx
+
+	cmp	rax,16
+	jl	__mark_tuple_selector_node_1
+	je	__mark_tuple_selector_node_2
+	mov	rcx,qword ptr (-24)[r9+rax]
+	mov	qword ptr [r8],rcx
+	jmp	__mark_node
+
+__mark_tuple_selector_node_2:
+	mov	rcx,qword ptr [r9]
+	mov	qword ptr [r8],rcx
+	jmp	__mark_node
+ endif
+
+__small_tuple_or_record:
+ ifdef NEW_DESCRIPTORS
+	mov	eax,dword ptr (-8)[rax]
+	lea	rdx,__indirection
+	pop	rbx 
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+	mov	r8,rcx
+__mark_tuple_selector_node_1:
+	mov	rcx,qword ptr [rbp+rax]
+	mov	qword ptr [r8],rcx
+	jmp	__mark_node
+ else
+	mov	eax,(-8)[rax]
+	pop	rbx 
+
+	push	rcx 
+	mov	rcx,qword ptr [rcx]
+	mov	eax,4[rax]
+	call	near ptr rax
+	pop	rdx
+
+	mov	qword ptr (-8)[rdx],offset __indirection
+	mov	qword ptr [rdx],rcx 
+ endif
+	jmp	__mark_node
+
+__mark_record_selector_node_1:
+	je	__mark_strict_record_selector_node_1
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rbx*4]
+	pop	rax
+	jne	__mark_no_selector_2
+
+	mov	rbx,qword ptr [rbp]
+	test	bl,2
+	je	__mark_no_selector_2
+
+	cmp	word ptr (-2)[rbx],258
+ ifdef NEW_DESCRIPTORS
+	jbe	__small_record
+
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8
+	and	r8,31*8
+	shr	rbx,8
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+	jne	__mark_no_selector_2
+
+__small_record:
+	mov	eax,(-8)[rax]
+	lea	rdx,__indirection
+	pop	rbx 
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+	mov	r8,rcx
+
+	cmp	rax,16
+	jle	__mark_record_selector_node_2
+	mov	rbp,r9
+	sub	rax,24
+__mark_record_selector_node_2:
+	mov	rcx,qword ptr [rbp+rax]
+	mov	qword ptr [r8],rcx
+	jmp	__mark_node
+ else
+	jbe	__small_tuple_or_record
+	jmp	__large_tuple_or_record
+ endif
+
+__mark_strict_record_selector_node_1:
+	mov	rbx,rax
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rbx *4]
+	pop	rax 
+	jne	__mark_no_selector_2
+
+	mov	rbx,qword ptr [rbp]
+	test	bl,2
+	je	__mark_no_selector_2
+
+	cmp	word ptr (-2)[rbx],258
+	jle	__select_from_small_record
+
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8 
+	and	r8,31*8
+	shr	rbx,8
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+	jne	__mark_no_selector_2
+	
+__select_from_small_record:
+ ifdef NEW_DESCRIPTORS
+	mov	eax,(-8)[rax]
+	sub	rcx,8
+
+	movzx	ebx,word ptr 4[rax]
+	cmp	rbx,16
+	jle	__mark_strict_record_selector_node_2
+	mov	rbx,qword ptr (-24)[r9+rbx]
+	jmp	__mark_strict_record_selector_node_3
+__mark_strict_record_selector_node_2:
+	mov	rbx,qword ptr [rbp+rbx]
+__mark_strict_record_selector_node_3:
+	mov	qword ptr 8[rcx],rbx
+
+	movzx	ebx,word ptr 6[rax]
+	test	rbx,rbx
+	je	__mark_strict_record_selector_node_5
+	cmp	rbx,16
+	jle	__mark_strict_record_selector_node_4
+	mov	rbp,r9
+	sub	rbx,24
+__mark_strict_record_selector_node_4:
+	mov	rbx,qword ptr [rbp+rbx]
+	mov	qword ptr 16[rcx],rbx
+__mark_strict_record_selector_node_5:
+	pop	rbx
+
+	mov	rax,qword ptr (-8)[rax]
+	mov	qword ptr [rcx],rax
+ else
+	mov	eax,(-8)[rax]
+	pop	rbx
+	mov	rdx,qword ptr [rcx]
+	sub	rcx,8
+	mov	eax,4[rax]
+	call	near ptr rax
+ endif
+	jmp	__mark_node
+
+__mark_indirection_node:
+	mov	rcx,qword ptr [rcx]
+	jmp	__mark_node
+
+__mark_hnf_2:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+	or	dword ptr [rdi+rbx*4],edx
+	cmp	rdx,20000000h
+	jbe	fits__in__word__6
+	or	dword ptr 4[rdi+rbx*4],1
+fits__in__word__6:
+
+__mark_record_2_c:
+	mov	rax,qword ptr [rcx]
+	mov	rbp,qword ptr 8[rcx]
+	or	rax,2
+	mov	qword ptr 8[rcx],rsi 
+	mov	qword ptr [rcx],rax 
+	lea	rsi,8[rcx]
+	mov	rcx,rbp
+
+__mark_node:
+	lea	rdx,[r10+rcx]
+	cmp	rdx,r11
+	jae	__mark_next_node
+
+	mov	rbx,rdx 
+	and	rdx,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_set_table2)[rdx]
+	test	ebp,dword ptr [rdi+rbx*4]
+	je	__mark_arguments
+
+__mark_next_node:
+	test	rsi,3
+	jne	__mark_parent
+
+	mov	rbp,qword ptr (-8)[rsi]
+	mov	rdx,qword ptr [rsi]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr (-8)[rsi],rdx 
+	sub	rsi,8
+
+	mov	rcx,rbp 
+	and	rbp,3
+	and	rcx,-4
+	or	rsi,rbp 
+	jmp	__mark_node
+
+__mark_parent:
+	mov	rbx,rsi 
+	and	rsi,-4
+	je	__end_mark_using_reversal
+
+	and	rbx,3
+	mov	rbp,qword ptr [rsi]
+	mov	qword ptr [rsi],rcx 
+
+	sub	rbx,1
+	je	__argument_part_parent
+	
+	lea	rcx,(-8)[rsi]
+	mov	rsi,rbp 
+	jmp	__mark_next_node
+
+__argument_part_parent:
+	and	rbp,-4
+	mov	rdx,rsi 
+	mov	rcx,qword ptr (-8)[rbp]
+	mov	rbx,qword ptr [rbp]
+	mov	qword ptr (-8)[rbp],rbx 
+	mov	qword ptr [rbp],rdx 
+	lea	rsi,(2-8)[rbp]
+	jmp	__mark_node
+
+__mark_lazy_node:
+	movsxd	rbp,dword ptr(-4)[rax]
+	test	rbp,rbp 
+	je	__mark_real_or_file
+
+	add	rcx,8
+	cmp	rbp,1
+	jle	__mark_lazy_node_1
+	cmp	rbp,256
+	jge	__mark_closure_with_unboxed_arguments
+
+	add	rbp,1
+	mov	rax,rdx 
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,rbp
+
+	lea	rax,[rax+rbp*8]
+	sub	rbp,2
+
+	or	dword ptr [rdi+rbx*4],edx 
+
+	cmp	rax,32*8
+	jbe	fits__in__word__7
+	or	dword ptr 4[rdi+rbx*4],1
+fits__in__word__7:
+__mark_closure_with_unboxed_arguments__2:
+	lea	rdx,[rcx+rbp*8]
+	mov	rax,qword ptr [rcx]
+	or	rax,2
+	mov	qword ptr [rcx],rax 
+	mov	rcx,qword ptr [rdx]
+	mov	qword ptr [rdx],rsi 
+	mov	rsi,rdx 
+	jmp	__mark_node
+
+__mark_closure_with_unboxed_arguments:
+	mov	rax,rbp 
+	and	rbp,255
+
+	sub	rbp,1
+	je	__mark_closure_1_with_unboxed_argument
+	add	rbp,2
+
+	shr	rax,8
+	add	r14,rbp 
+
+	push	rcx 
+	lea	rcx,[rdx+rbp*8]
+
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	sub	rbp,rax 
+
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rcx,32*8
+	jbe	fits__in_word_7_
+	or	dword ptr 4[rdi+rbx*4],1
+fits__in_word_7_:
+	pop	rcx 
+	sub	rbp,2
+	jg	__mark_closure_with_unboxed_arguments__2
+	je	__shared_argument_part
+	sub	rcx,8
+	jmp	__mark_next_node
+
+__mark_closure_1_with_unboxed_argument:
+	sub	rcx,8
+	jmp	__mark_real_or_file
+
+__mark_hnf_0:
+	cmp	rax,offset dINT+2
+	jne	__no_int_3
+
+	mov	rbp,qword ptr 8[rcx]
+	cmp	rbp,33
+	jb	____small_int
+
+__mark_bool_or_small_string:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,2
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,40000000h
+	jbe	__mark_next_node
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	__mark_next_node
+
+____small_int:
+	shl	rbp,4
+	lea	rcx,(small_integers)[rbp]
+	jmp	__mark_next_node
+
+__no_int_3:
+	jb	__mark_real_file_or_string
+
+ 	cmp	rax,offset CHAR+2
+ 	jne	__no_char_3
+
+	movzx	rbp,byte ptr 8[rcx]
+	shl	rbp,4
+	lea	rcx,(static_characters)[rbp]
+	jmp	__mark_next_node
+
+__no_char_3:
+	jb	__mark_bool_or_small_string
+
+ ifdef NEW_DESCRIPTORS
+	lea	rcx,((-8)-2)[rax]
+ else
+	lea	rcx,((-12)-2)[rax]
+ endif
+	jmp	__mark_next_node
+
+__mark_real_file_or_string:
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+	jbe	__mark_string_or_array
+
+__mark_real_or_file:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+
+	or	dword ptr [rdi+rbx*4],edx
+
+	cmp	rdx,20000000h
+	jbe	__mark_next_node
+
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	__mark_next_node
+
+__mark__record:
+	sub	rbp,258
+	je	__mark_record_2
+	jl	__mark_record_1
+
+__mark_record_3:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,20000000h
+	jbe	fits__in__word__13
+	or	dword ptr 4[rdi+rbx*4],1
+fits__in__word__13:
+	movzx	rbx,word ptr (-2+2)[rax]
+
+	mov	rdx,qword ptr 8[rcx]
+	add	rdx,r10
+	mov	rax,rdx 
+	and	rdx,31*8
+	shr	rax,8
+
+	push	rsi 
+
+	mov	esi,dword ptr (bit_set_table2)[rdx]
+	test	esi,dword ptr [rdi+rax*4]
+	jne	__shared_record_argument_part
+
+	add	rbp,1
+	or	dword ptr [rdi+rax *4],esi
+
+	lea	rdx,[rdx+rbp*8]
+	add	r14,rbp 
+
+	pop	rsi 
+
+	cmp	rdx,32*8
+	jbe	fits__in__word__14
+	or	dword ptr 4[rdi+rax*4],1
+fits__in__word__14:
+	sub	rbx,1
+	mov	rdx,qword ptr 8[rcx]
+	jl	__mark_record_3_bb
+	je	__shared_argument_part
+
+	mov	qword ptr 8[rcx],rsi 
+	add	rcx,8
+
+	sub	rbx,1
+	je	__mark_record_3_aab
+
+	lea	rsi,[rdx+rbx*8]
+	mov	rax,qword ptr [rdx]
+	or	rax,1
+	mov	rbp,qword ptr [rsi]
+	mov	qword ptr [rdx],rax 
+	mov	qword ptr [rsi],rcx 
+	mov	rcx,rbp 
+	jmp	__mark_node
+
+__mark_record_3_bb:
+	sub	rcx,8
+	jmp	__mark_next_node
+
+__mark_record_3_aab:
+	mov	rbp,qword ptr [rdx]
+	mov	qword ptr [rdx],rcx 
+	lea	rsi,1[rdx]
+	mov	rcx,rbp 
+	jmp	__mark_node
+
+__shared_record_argument_part:
+	mov	rdx,qword ptr 8[rcx]
+
+	pop	rsi
+
+	test	rbx,rbx 
+	jne	__shared_argument_part
+	sub	rcx,8
+	jmp	__mark_next_node
+
+__mark_record_2:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,20000000h
+	jbe	fits__in__word_12
+	or	dword ptr 4[rdi+rbx*4],1
+fits__in__word_12:
+	cmp	word ptr (-2+2)[rax],1
+	ja	__mark_record_2_c
+	je	__shared_argument_part
+	sub	rcx,8
+	jmp	__mark_next_node
+
+__mark_record_1:
+	cmp	word ptr (-2+2)[rax],0
+	jne	__mark_hnf_1
+	sub	rcx,8
+	jmp	__mark_bool_or_small_string
+
+__mark_string_or_array:
+	je	__mark_string_
+
+__mark_array:
+	mov	rbp,qword ptr 16[rcx]
+	test	rbp,rbp 
+	je	__mark_lazy_array
+
+	movzx	rax,word ptr (-2)[rbp]
+	test	rax,rax 
+	je	__mark_strict_basic_array
+
+	movzx	rbp,word ptr (-2+2)[rbp]
+	test	rbp,rbp 
+	je	__mark_b_record_array
+
+	sub	rax,256
+	cmp	rax,rbp 
+	je	__mark_a_record_array
+
+__mark__ab__record__array:
+	push	rdx 
+	push	rbx 
+	mov	rbx,rbp 
+
+	mov	rbp,qword ptr 8[rcx]
+	add	rcx,16
+	push	rcx 
+
+	shl	rbp,3
+	mov	rdx,rax 
+	imul	rdx,rbp 
+
+	sub	rax,rbx 
+	add	rcx,8
+	add	rdx,rcx 
+
+	call	reorder
+	
+	pop	rcx 
+
+	xchg	rax,rbx 
+	mov	rbp,qword ptr (-8)[rcx]
+	imul	rax,rbp 
+	imul	rbx,rbp 
+	add	r14,rbx 
+	add	rbx,rax 
+
+	shl	rbx,3
+	lea	rbp,[r10+rcx]
+	add	rbp,rbx 
+
+	pop	rbx 
+	pop	rdx 
+
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	or	dword ptr [rdi+rbx*4],edx
+
+	lea	rdx,[rcx+rax*8]
+	jmp	__mark_r_array
+
+__mark_a_record_array:
+	imul	rax,qword ptr 8[rcx]
+	add	rcx,16
+	jmp	__mark_lr_array
+
+__mark_lazy_array:
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+
+__mark_lr_array:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	mov	rbp,r10
+	or	dword ptr [rdi+rbx*4],edx 
+	lea	rdx,[rcx+rax*8]
+	add	rbp,rdx 
+__mark_r_array:
+	shr	rbp,8
+
+	cmp	rbx,rbp 
+	jae	__skip_mark_lazy_array_bits
+
+	inc	rbx 
+
+__mark_lazy_array_bits:
+	or	dword ptr [rdi+rbx*4],1
+	inc	rbx
+	cmp	rbx,rbp 
+	jbe	__mark_lazy_array_bits
+
+__skip_mark_lazy_array_bits:
+	add	r14,3
+	add	r14,rax 
+
+	cmp	rax,1
+	jbe	__mark_array_length_0_1
+
+	mov	rbp,qword ptr [rdx]
+	mov	rbx,qword ptr [rcx]
+	mov	qword ptr [rdx],rbx 
+	mov	qword ptr [rcx],rbp 
+	
+	mov	rbp,qword ptr (-8)[rdx]
+	sub	rdx,8
+	mov	rbx,qword ptr lazy_array_list
+	add	rbp,2
+	mov	qword ptr [rdx],rbx 
+	mov	qword ptr (-8)[rcx],rbp 
+	mov	qword ptr (-16)[rcx],rax 
+	sub	rcx,16
+	mov	qword ptr lazy_array_list,rcx 
+
+	mov	rcx,qword ptr (-8)[rdx]
+	mov	qword ptr (-8)[rdx],rsi 
+	lea	rsi,(-8)[rdx]
+	jmp	__mark_node
+
+__mark_array_length_0_1:
+	lea	rcx,(-16)[rcx]
+	jb	__mark_next_node
+
+	mov	rbx,qword ptr 24[rcx]
+	mov	rbp,qword ptr 16[rcx]
+	mov	rdx,qword ptr lazy_array_list
+	mov	qword ptr 24[rcx],rbp 
+	mov	qword ptr 16[rcx],rdx 
+	mov	qword ptr [rcx],rax 
+	mov	qword ptr lazy_array_list,rcx 
+	mov	qword ptr 8[rcx],rbx 
+	add	rcx,8
+
+	mov	rbp,qword ptr [rcx]
+	mov	qword ptr [rcx],rsi 
+	lea	rsi,2[rcx]
+	mov	rcx,rbp 
+	jmp	__mark_node
+
+__mark_b_record_array:
+	mov	rbp,qword ptr 8[rcx]
+	sub	rax,256
+	imul	rax,rbp 
+	add	rax,3
+	jmp	__mark_basic_array
+
+__mark_strict_basic_array:
+	mov	rax,qword ptr 8[rcx]
+	cmp	rbp,offset BOOL+2
+	je	__mark__strict__bool__array
+__mark__strict__real__array:
+__mark__strict__int__array:
+	add	rax,3
+	jmp	__mark_basic_array
+__mark__strict__bool__array:
+	add	rax,24+7
+	shr	rax,3
+	jmp	__mark_basic_array
+
+__mark_string_:
+	mov	rax,qword ptr 8[rcx]
+	add	rax,16+7
+	shr	rax,3
+
+__mark_basic_array:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,rax 
+
+	or	dword ptr [rdi+rbx*4],edx 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+
+	cmp	rbx,rax 
+	jae	__mark_next_node
+
+	inc	rbx 
+	mov	rbp,1
+
+	cmp	rbx,rax 
+	jae	__last__string__bits
+
+__mark_string_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx
+	cmp	rbx,rax 
+	jb	__mark_string_lp
+
+__last__string__bits:
+	or	dword ptr [rdi+rbx*4],ebp 
+	jmp	__mark_next_node
diff --git a/amark_prefetch.asm b/amark_prefetch.asm
new file mode 100644
index 0000000..b0e38c4
--- /dev/null
+++ b/amark_prefetch.asm
@@ -0,0 +1,1748 @@
+
+_TEXT	ends
+	_DATA segment
+n_queue_items:
+	dq	0
+queue_first:
+	dq	0
+queue:
+	dq	0,0,0,0,0,0,0,0
+	dq	0,0,0,0,0,0,0,0
+_DATA	ends
+	_TEXT segment
+
+pmark:
+	mov	rax,qword ptr heap_size_65
+	xor	rbx,rbx 
+
+	mov	qword ptr n_marked_words,rbx 
+	shl	rax,6
+
+	mov	qword ptr heap_size_64_65,rax 
+	mov	qword ptr lazy_array_list,rbx 
+	
+	lea	rsi,(-4000)[rsp]
+
+	mov	rax,qword ptr caf_list
+
+	mov	qword ptr end_stack,rsi 
+
+	mov	r15,0
+	mov	r8,0
+
+	mov	r10,neg_heap_p3
+	mov	r11,heap_size_64_65
+	mov	r13,qword ptr end_stack
+	mov	r14,0
+
+	test	rax,rax
+	je	end_pmark_cafs
+
+pmark_cafs_lp:
+	mov	rbx,qword ptr [rax]
+	mov	rbp,qword ptr (-8)[rax]
+
+	push	rbp
+	lea	rbp,8[rax]
+	lea	r12,8[rax+rbx*8]
+
+	call	pmark_stack_nodes
+
+	pop	rax
+	test	rax,rax 
+	jne	pmark_cafs_lp
+
+end_pmark_cafs:
+	mov	rsi,qword ptr stack_top
+	mov	rbp,qword ptr stack_p
+
+	mov	r12,rsi 
+	call	pmark_stack_nodes
+	jmp	continue_mark_after_pmark
+
+; %rbp : pointer to stack element
+; %rdi : heap_vector
+; %rax ,%rbx ,%rcx ,%rdx ,%rsi : free
+
+pmark_stack_nodes:
+	cmp	rbp,r12
+	je	end_pmark_nodes
+pmark_stack_nodes_:
+	mov	rcx,qword ptr [rbp]
+
+	add	rbp,8
+	lea	rdx,[r10+rcx] 
+
+	cmp	rdx,r11
+	jnc	pmark_stack_nodes
+
+	mov	rbx,rdx 
+	and	rdx,31*8
+
+	shr	rbx,8
+	mov	esi,dword ptr (bit_set_table2)[rdx]
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	pmark_stack_nodes
+
+	push	rbp 
+
+	push	0
+
+	jmp	pmark_node_
+
+pmark_hnf_2:
+	cmp	rsi,20000000h
+	jbe	pmark_fits_in_word_6
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_6:
+	add	r14,3
+
+pmark_record_2_c:
+	mov	rbx,qword ptr 8[rcx]
+	push	rbx 
+
+	cmp	rsp,r13
+	jb	pmarkr_using_reversal
+
+pmark_node2:
+pmark_shared_argument_part:
+	mov	rcx,qword ptr [rcx]
+
+pmark_node:
+	lea	rdx,[r10+rcx]
+	cmp	rdx,r11
+	jnc	pmark_next_node
+
+	mov	rbx,rdx 
+	and	rdx,31*8
+
+	shr	rbx,8
+	mov	esi,dword ptr (bit_set_table2)[rdx]
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	pmark_next_node
+
+pmark_node_:
+
+	prefetch	[rcx]
+
+	mov	qword ptr (queue)[r8],rcx 
+	lea	rdx,[r8+r15*8]
+	add	r8,8
+
+	and	r8,15*8
+	and	rdx,15*8
+
+	cmp	r15,-4
+	je	pmark_last_item_in_queue
+
+pmark_add_items:
+	mov	rcx,qword ptr [rsp]
+	test	rcx,rcx 
+	jne	pmark_add_stacked_item
+
+pmark_add_items2:
+	mov	rbp,qword ptr 8[rsp]
+	cmp	rbp,r12
+	je	pmark_last_item_in_queue
+
+	mov	rcx,qword ptr [rbp]
+	add	rbp,8
+	mov	qword ptr 8[rsp],rbp 
+
+	lea	rbp,[r10+rcx]
+	cmp	rbp,r11
+	jnc	pmark_add_items2
+	mov	rax,rbp 
+	and	rbp,31*8
+	shr	rax,8
+	mov	esi,dword ptr (bit_set_table2)[rbp]
+	test	esi,dword ptr [rdi+rax*4]
+	jne	pmark_add_items2
+
+	prefetch	[rcx]
+
+	mov	qword ptr (queue)[r8],rcx 
+	add	r8,8
+	and	r8,15*8
+
+	sub	r15,1
+	
+	cmp	r15,-4
+	jne	pmark_add_items2
+	jmp	pmark_last_item_in_queue
+
+pmark_add_stacked_item:
+	add	rsp ,8
+
+	lea	rbp,[r10+rcx]
+	cmp	rbp,r11
+	jnc	pmark_add_items
+	mov	rax,rbp 
+	and	rbp,31*8
+	shr	rax,8
+	mov	esi,dword ptr (bit_set_table2)[rbp]
+	test	esi,dword ptr [rdi+rax*4]
+	jne	pmark_add_items
+
+	prefetch	[rcx]
+
+	mov	qword ptr (queue)[r8],rcx 
+	add	r8,8
+	and	r8,15*8
+
+	sub	r15,1
+
+	cmp	r15,-4 
+	jne	pmark_add_items
+
+pmark_last_item_in_queue:
+	mov	rcx,qword ptr (queue)[rdx]
+
+	lea	rdx,[r10+rcx]
+
+	mov	rbx,rdx 
+	and	rdx,31*8
+	shr	rbx,8
+	mov	esi,dword ptr (bit_set_table2)[rdx]
+		
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	pmark_next_node
+
+pmark_arguments:
+	mov	rax,qword ptr [rcx]
+	test	rax,2
+	je	pmark_lazy_node
+	
+	movzx	rbp,word ptr (-2)[rax]
+
+	test	rbp,rbp 
+	je	pmark_hnf_0
+
+	or	dword ptr [rdi+rbx*4],esi 
+	add	rcx,8
+
+	cmp	rbp,256
+	jae	pmark_record
+
+	sub	rbp,2
+	je	pmark_hnf_2
+	jb	pmark_hnf_1
+
+pmark_hnf_3:
+	mov	rdx,qword ptr 8[rcx]
+
+	cmp	rsi,20000000h
+	jbe	pmark_fits_in_word_1
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_1:	
+
+	add	r14,3
+	lea	rax,[r10+rdx]
+	mov	rbx,rax 
+
+	and	rax,31*8
+	shr	rbx,8
+
+	mov	esi,dword ptr (bit_set_table2)[rax]
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	pmark_shared_argument_part
+
+pmark_no_shared_argument_part:
+	or	dword ptr [rdi+rbx*4],esi 
+	add	rbp,1
+
+	add	r14,rbp 
+	lea	rax,[rax+rbp*8]
+	lea	rdx,(-8)[rdx+rbp*8]
+
+	cmp	rax,32*8
+	jbe	pmark_fits_in_word_2
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_2:
+
+	mov	rbx,qword ptr [rdx]
+	sub	rbp,2
+	push	rbx 
+
+pmark_push_hnf_args:
+	mov	rbx,qword ptr (-8)[rdx]
+	sub	rdx,8
+	push	rbx 
+	sub	rbp,1
+	jge	pmark_push_hnf_args
+
+	cmp	rsp,r13
+	jae	pmark_node2
+
+	jmp	pmarkr_using_reversal
+
+pmark_hnf_1:
+	cmp	rsi,40000000h
+	jbe	pmark_fits_in_word_4
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_4:
+	add	r14,2
+	mov	rcx,qword ptr [rcx]
+	jmp	pmark_node
+
+pmark_lazy_node_1:
+	add	rcx,8
+	or	dword ptr [rdi+rbx*4],esi 
+	cmp	rsi,20000000h
+	jbe	pmark_fits_in_word_3
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_3:
+	add	r14,3
+
+	cmp	rbp,1
+	je	pmark_node2
+
+pmark_selector_node_1:
+	add	rbp,2
+	mov	rdx,qword ptr [rcx]
+	je	pmark_indirection_node
+
+	lea	rsi,[r10+rdx]
+	mov	rbx,rsi 
+
+	shr	rbx,8
+	and	rsi,31*8
+
+	add	rbp,1
+
+	mov	esi,dword ptr (bit_set_table2)[rsi]
+	jle	pmark_record_selector_node_1
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	pmark_node3
+
+	mov	rbp,qword ptr [rdx]
+	test	rbp,2
+	je	pmark_node3
+
+	cmp	word ptr (-2)[rbp],2
+	jbe	pmark_small_tuple_or_record
+
+pmark_large_tuple_or_record:
+	mov	rbp,qword ptr 16[rdx]
+	mov	r9,rbp
+
+	add	rbp,r10
+	mov	rbx,rbp
+	and	rbp,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_set_table2)[rbp]
+	test	ebp,dword ptr [rdi+rbx*4]
+	jne	pmark_node3
+
+ ifdef NEW_DESCRIPTORS
+	mov	eax,(-8)[rax]
+	lea	rbp,__indirection
+	mov	qword ptr (-8)[rcx],rbp
+	movzx	eax,word ptr 4[rax]
+	mov	rbp,rcx
+
+	cmp	rax,16
+	jl	pmark_tuple_selector_node_1
+	mov	rdx,r9
+	je	pmark_tuple_selector_node_2
+	mov	rcx,qword ptr (-24)[r9+rax]
+	mov	qword ptr [rbp],rcx
+	jmp	pmark_node
+
+pmark_tuple_selector_node_2:
+	mov	rcx,qword ptr [r9]
+	mov	qword ptr [rbp],rcx
+	jmp	pmark_node	
+ endif
+
+pmark_small_tuple_or_record:
+ ifdef NEW_DESCRIPTORS
+	mov	eax,(-8)[rax]
+	lea	rbp,__indirection
+	mov	qword ptr (-8)[rcx],rbp
+	movzx	eax,word ptr 4[rax]
+	mov	rbp,rcx
+pmark_tuple_selector_node_1:
+	mov	rcx,qword ptr [rdx+rax]
+	mov	qword ptr [rbp],rcx	
+ else
+	mov	eax,(-8)[rax]
+	push	rcx
+	mov	rcx,rdx
+	mov	eax,4[rax]
+	call	near ptr rax
+	pop	rdx 
+	
+	lea	r9,__indirection
+	mov	qword ptr (-8)[rdx],r9
+	mov	qword ptr [rdx],rcx
+ endif
+	jmp	pmark_node
+
+pmark_record_selector_node_1:
+	je	pmark_strict_record_selector_node_1
+
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	pmark_node3
+
+	mov	rbp,qword ptr [rdx]
+	test	rbp,2
+	je	pmark_node3
+
+	cmp	word ptr (-2)[rbp],258
+	jbe	pmark_small_tuple_or_record
+ ifdef NEW_DESCRIPTORS
+	mov	rbp,qword ptr 16[rdx]
+	mov	r9,rbp
+
+	add	rbp,r10
+	mov	rbx,rbp
+	and	rbp,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_set_table2)[rbp]
+	test	ebp,dword ptr [rdi+rbx*4]
+	jne	pmark_node3
+
+	mov	eax,(-8)[rax]
+	lea	rbp,__indirection
+	mov	qword ptr (-8)[rcx],rbp
+	movzx	eax,word ptr 4[rax]
+	mov	rbp,rcx
+
+	cmp	rax,16
+	jle	pmark_record_selector_node_2
+	mov	rdx,r9
+	sub	rax,24
+pmark_record_selector_node_2:
+	mov	rcx,qword ptr [rdx+rax]
+	mov	qword ptr [rbp],rcx
+	jmp	pmark_node
+ else
+	jmp	pmark_large_tuple_or_record
+ endif
+
+pmark_strict_record_selector_node_1:
+	test	esi,dword ptr [rdi+rbx*4]
+	jne	pmark_node3
+
+	mov	rbp,qword ptr [rdx]
+	test	rbp,2
+	je	pmark_node3
+
+	cmp	word ptr (-2)[rbp],258
+	jbe	pmark_select_from_small_record
+
+	mov	rbp,qword ptr 16[rdx]
+	add	rbp,r10
+	mov	rbx,rbp 
+	and	rbp,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_set_table2)[rbp]
+	test	ebp,dword ptr [rdi+rbx*4]
+	jne	pmark_node3
+	
+pmark_select_from_small_record:
+	mov	eax,(-8)[rax]
+	sub	rcx,8
+
+ ifdef NEW_DESCRIPTORS
+	movzx	ebx,word ptr 4[rax]
+	cmp	rbx,16
+	jle	pmark_strict_record_selector_node_2
+	mov	rbx,qword ptr (-24)[r9+rbx]
+	jmp	pmark_strict_record_selector_node_3
+pmark_strict_record_selector_node_2:
+	mov	rbx,qword ptr [rdx+rbx]
+pmark_strict_record_selector_node_3:
+	mov	qword ptr 8[rcx],rbx
+
+	movzx	ebx,word ptr 6[rax]
+	test	rbx,rbx
+	je	pmark_strict_record_selector_node_5
+	cmp	rbx,16
+	jle	pmark_strict_record_selector_node_4
+	mov	rdx,r9
+	sub	rbx,24
+pmark_strict_record_selector_node_4:
+	mov	rbx,qword ptr [rdx+rbx]
+	mov	qword ptr 16[rcx],rbx
+pmark_strict_record_selector_node_5:
+
+	mov	rax,qword ptr (-8)[rax]
+	mov	qword ptr [rcx],rax
+ else
+	mov	eax,4[rax]
+	call	near ptr rax
+ endif
+	jmp	pmark_next_node
+
+pmark_indirection_node:
+pmark_node3:
+	mov	rcx,rdx 
+	jmp	pmark_node
+
+pmark_next_node:
+	pop	rcx 
+	test	rcx,rcx 
+	jne	pmark_node
+
+	pop	rbp 
+	cmp	rbp,r12
+	jne	pmark_stack_nodes_
+
+end_pmark_nodes:
+	test	r15,r15
+	je	end_pmark_nodes_
+
+	push	rbp 
+
+	push	0
+
+	lea	rdx,[r8+r15*8]
+	add	r15,1
+
+	and	rdx,15*8
+
+	jmp	pmark_last_item_in_queue
+
+end_pmark_nodes_:
+	ret
+
+pmark_lazy_node:
+	movsxd	rbp,dword ptr (-4)[rax]
+	test	rbp,rbp 
+	je	pmark_real_or_file
+
+	cmp	rbp,1
+	jle	pmark_lazy_node_1
+
+	cmp	rbp,256
+	jge	pmark_closure_with_unboxed_arguments
+	inc	rbp 
+	or	dword ptr [rdi+rbx*4],esi 
+
+	add	r14,rbp 
+	lea	rdx,[rdx+rbp*8]
+	lea	rcx,[rcx+rbp*8]
+
+	cmp	rdx,32*8
+	jbe	pmark_fits_in_word_7
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_7:
+	sub	rbp,3
+pmark_push_lazy_args:
+	mov	rbx,qword ptr (-8)[rcx]
+	sub	rcx,8
+	push	rbx 
+	sub	rbp,1
+	jge	pmark_push_lazy_args
+
+	sub	rcx,8
+
+	cmp	rsp,r13
+	jae	pmark_node2
+	
+	jmp	pmarkr_using_reversal
+
+pmark_closure_with_unboxed_arguments:
+	mov	rax,rbp 
+	and	rbp,255
+	sub	rbp,1
+	je	pmark_real_or_file
+
+	shr	rax,8
+	add	rbp,2
+	
+	or	dword ptr [rdi+rbx*4],esi 
+	add	r14,rbp 
+	lea	rdx,[rdx+rbp*8]
+
+	sub	rbp,rax 
+
+	cmp	rdx,32*8
+	jbe	pmark_fits_in_word_7_
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_7_:
+	sub	rbp,2
+	jl	pmark_next_node
+
+	lea	rcx,16[rcx+rbp*8]
+	jne	pmark_push_lazy_args
+
+pmark_closure_with_one_boxed_argument:
+	mov	rcx,qword ptr (-8)[rcx]
+	jmp	pmark_node
+
+pmark_hnf_0:
+	lea	r9,dINT+2
+	cmp	rax,r9
+	jb	pmark_real_file_or_string
+
+	or	dword ptr [rdi+rbx*4],esi 
+
+	lea	r9,CHAR+2
+	cmp	rax,r9
+	ja	pmark_normal_hnf_0
+
+pmark_bool:
+	add	r14,2
+
+	cmp	rsi,40000000h
+	jbe	pmark_next_node
+
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	pmark_next_node
+
+pmark_normal_hnf_0:
+	inc	r14
+	jmp	pmark_next_node
+
+pmark_real_file_or_string:
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+	jbe	pmark_string_or_array
+
+pmark_real_or_file:
+	or	dword ptr [rdi+rbx*4],esi 
+	add	r14,3
+
+	cmp	rsi,20000000h
+	jbe	pmark_next_node
+
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	pmark_next_node
+
+pmark_record:
+	sub	rbp,258
+	je	pmark_record_2
+	jl	pmark_record_1
+
+pmark_record_3:
+	add	r14,3
+
+	cmp	rsi,20000000h
+	jbe	pmark_fits_in_word_13
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_13:
+	mov	rdx,qword ptr 8[rcx]
+
+	movzx	rbx,word ptr (-2+2)[rax]
+	lea	rsi,[r10+rdx] 
+
+	mov	rax,rsi 
+	and	rsi,31*8
+
+	shr	rax,8
+	sub	rbx,1
+
+	mov	edx,dword ptr (bit_set_table2)[rsi]
+	jb	pmark_record_3_bb
+
+	test	edx,dword ptr [rdi+rax*4]
+	jne	pmark_node2
+
+	add	rbp,1
+	or	dword ptr [rdi+rax*4],edx 
+	add	r14,rbp 
+	lea	rsi,[rsi+rbp*8]
+
+	cmp	rsi,32*8
+	jbe	pmark_push_record_arguments
+	or	dword ptr 4[rdi+rax*4],1
+pmark_push_record_arguments:
+	mov	rdx,qword ptr 8[rcx]
+	mov	rbp,rbx 
+	shl	rbx,3
+	add	rdx,rbx 
+	sub	rbp,1
+	jge	pmark_push_hnf_args
+
+	jmp	pmark_node2
+
+pmark_record_3_bb:
+	test	edx,dword ptr [rdi+rax*4]
+	jne	pmark_next_node
+
+	add	rbp,1
+	or	dword ptr [rdi+rax*4],edx 
+	add	r14,rbp 
+	lea	rsi,[rsi+rbp*8]
+	
+	cmp	rsi,32*8
+	jbe	pmark_next_node
+
+	or	dword ptr 4[rdi+rax*4],1
+	jmp	pmark_next_node
+
+pmark_record_2:
+	cmp	rsi,20000000h
+	jbe	pmark_fits_in_word_12
+	or	dword ptr 4[rdi+rbx*4],1
+pmark_fits_in_word_12:
+	add	r14,3
+
+	cmp	word ptr (-2+2)[rax],1
+	ja	pmark_record_2_c
+	je	pmark_node2
+	jmp	pmark_next_node
+
+pmark_record_1:
+	cmp	word ptr (-2+2)[rax],0
+	jne	pmark_hnf_1
+
+	jmp	pmark_bool
+
+pmark_string_or_array:
+	je	pmark_string_
+
+pmark_array:
+	mov	rbp,qword ptr 16[rcx]
+	test	rbp,rbp 
+	je	pmark_lazy_array
+
+	movzx	rax,word ptr (-2)[rbp]
+
+	test	rax,rax 
+	je	pmark_strict_basic_array
+
+	movzx	rbp,word ptr (-2+2)[rbp]
+	test	rbp,rbp 
+	je	pmark_b_record_array
+
+	cmp	rsp,r13
+	jb	pmark_array_using_reversal
+
+	sub	rax,256
+	cmp	rax,rbp 
+	je	pmark_a_record_array
+
+pmark_ab_record_array:
+	or	dword ptr [rdi+rbx*4],esi 
+	mov	rbp,qword ptr 8[rcx]
+
+	imul	rax,rbp 
+	add	rax,3
+
+	add	r14,rax 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+	
+	cmp	rbx,rax 
+	jae	pmark_end_set_ab_array_bits
+
+	inc	rbx 
+	mov	rbp,1
+	cmp	rbx,rax 
+	jae	pmark_last_ab_array_bits
+
+pmark_ab_array_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx 
+	cmp	rbx,rax
+	jb	pmark_ab_array_lp
+
+pmark_last_ab_array_bits:
+	or	dword ptr [rdi+rbx*4],ebp 
+
+pmark_end_set_ab_array_bits:
+	mov	rax,qword ptr 8[rcx]
+	mov	rdx,qword ptr 16[rcx]
+	movzx	rbx,word ptr (-2+2)[rdx]
+	movzx	rdx,word ptr (-2)[rdx]
+	shl	rbx,3
+	lea	rdx,(-2048)[rdx*8]
+	push	rbx 
+	push	rdx 
+	lea	rbp,24[rcx]
+	push	r12
+	jmp	pmark_ab_array_begin
+
+pmark_ab_array:
+	mov	rbx,qword ptr 16[rsp]
+	push	rax 
+	push	rbp 
+	lea	r12,[rbp+rbx]
+
+	call	pmark_stack_nodes
+
+	mov	rbx,qword ptr (8+16)[rsp]
+	pop	rbp 
+	pop	rax 
+	add	rbp,rbx 
+pmark_ab_array_begin:
+	sub	rax,1
+	jnc	pmark_ab_array
+
+	pop	r12
+	add	rsp,16
+	jmp	pmark_next_node
+
+pmark_a_record_array:
+	or	dword ptr [rdi+rbx*4],esi 
+	mov	rbp,qword ptr 8[rcx]
+
+	imul	rax,rbp 
+	push	rax 
+
+	add	rax,3
+
+	add	r14,rax 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+	
+	cmp	rbx,rax 
+	jae	pmark_end_set_a_array_bits
+
+	inc	rbx 
+	mov	rbp,1
+	cmp	rbx,rax 
+	jae	pmark_last_a_array_bits
+
+pmark_a_array_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx 
+	cmp	rbx,rax 
+	jb	pmark_a_array_lp
+
+pmark_last_a_array_bits:
+	or	dword ptr [rdi+rbx*4],ebp 
+
+pmark_end_set_a_array_bits:
+	pop	rax 
+	lea	rbp,24[rcx]
+
+	push	r12
+	lea	r12,24[rcx+rax*8]
+
+	call	pmark_stack_nodes
+
+	pop	r12
+	jmp	pmark_next_node
+
+pmark_lazy_array:
+	cmp	rsp,r13
+	jb	pmark_array_using_reversal
+
+	or	dword ptr [rdi+rbx*4],esi 
+	mov	rax,qword ptr 8[rcx]
+
+	add	rax,3
+
+	add	r14,rax 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+	
+	cmp	rbx,rax 
+	jae	pmark_end_set_lazy_array_bits
+
+	inc	rbx 
+	mov	rbp,1
+	cmp	rbx,rax 
+	jae	pmark_last_lazy_array_bits
+
+pmark_lazy_array_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx 
+	cmp	rbx,rax
+	jb	pmark_lazy_array_lp
+
+pmark_last_lazy_array_bits:
+	or	dword ptr [rdi+rbx*4],ebp 
+
+pmark_end_set_lazy_array_bits:
+	mov	rax,qword ptr 8[rcx]
+	lea	rbp,24[rcx]
+
+	push	r12
+	lea	r12,24[rcx+rax*8]
+
+	call	pmark_stack_nodes
+
+	pop	r12
+	jmp	pmark_next_node
+
+pmark_array_using_reversal:
+	push	0
+	mov	rsi,1
+	jmp	pmarkr_node
+
+pmark_strict_basic_array:
+	mov	rax,qword ptr 8[rcx]
+	cmp	rbp,offset BOOL+2
+	je	pmark_strict_bool_array
+pmark_strict_real_array:
+pmark_strict_int_array:
+	add	rax,3
+	jmp	pmark_basic_array_
+pmark_strict_bool_array:
+	add	rax,24+7
+	shr	rax,3
+	jmp	pmark_basic_array_
+
+pmark_b_record_array:
+	mov	rbp,qword ptr 8[rcx]
+	sub	rax,256
+	imul	rax,rbp 
+	add	rax,3
+	jmp	pmark_basic_array_
+
+pmark_string_:
+	mov	rax,qword ptr 8[rcx]
+	add	rax,16+7
+	shr	rax,3
+
+pmark_basic_array_:
+	or	dword ptr [rdi+rbx*4],esi 
+
+	add	r14,rax 
+	lea	rax,(-8)[rcx+rax*8]
+
+	add	rax,r10
+	shr	rax,8
+	
+	cmp	rbx,rax 
+	jae	pmark_next_node
+
+	inc	rbx 
+	mov	rbp,1
+	cmp	rbx,rax 
+	jae	pmark_last_string_bits
+
+pmark_string_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx
+	cmp	rbx,rax 
+	jb	pmark_string_lp
+
+pmark_last_string_bits:
+	or	dword ptr [rdi+rbx*4],ebp 
+	jmp	pmark_next_node
+
+end_pmarkr_using_reversal:
+	pop	rdx 
+	test	rdx,rdx 
+	je	pmark_next_node
+	mov	qword ptr [rdx],rcx 
+	jmp	pmark_next_node
+
+
+pmarkr_using_reversal:
+	push	rcx 
+	mov	rsi,1
+	mov	rcx,qword ptr [rcx]
+	jmp	pmarkr_node
+
+pmarkr_arguments:
+	mov	rax,qword ptr [rcx]
+	test	al,2
+	je	pmarkr_lazy_node
+
+	movzx	rbp,word ptr (-2)[rax]
+	test	rbp,rbp 
+	je	pmarkr_hnf_0
+
+	add	rcx,8
+
+	cmp	rbp,256
+	jae	pmarkr_record
+
+	sub	rbp,2
+	je	pmarkr_hnf_2
+	jb	pmarkr_hnf_1
+
+pmarkr_hnf_3:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+
+	or	dword ptr [rdi+rbx*4],edx 
+
+	cmp	rdx,20000000h
+
+	mov	rax,qword ptr 8[rcx]
+
+	jbe	pmarkr_fits_in_word_1
+	or	dword ptr 4[rdi+rbx*4],1
+pmarkr_fits_in_word_1:
+	add	rax,r10
+
+	mov	rbx,rax 
+	and	rax,31*8
+
+	shr	rbx,8
+
+	mov	edx,dword ptr (bit_set_table2)[rax]
+	test	edx,dword ptr [rdi+rbx*4]
+	jne	pmarkr_shared_argument_part
+
+pmarkr_no_shared_argument_part:
+	or	dword ptr [rdi+rbx*4],edx 
+	mov	rdx,qword ptr 8[rcx]
+
+	add	rbp,1
+	mov	qword ptr 8[rcx],rsi 
+
+	add	r14,rbp 
+	add	rcx,8
+
+	shl	rbp,3
+	or	qword ptr [rdx],1
+
+	add	rax,rbp 
+	add	rdx,rbp 
+
+	cmp	rax,32*8
+	jbe	pmarkr_fits_in_word_2
+	or	dword ptr 4[rdi+rbx*4],1
+pmarkr_fits_in_word_2:
+
+	mov	rbp ,qword ptr (-8)[rdx]
+	mov	qword ptr (-8)[rdx],rcx 
+	lea	rsi,(-8)[rdx]
+	mov	rcx,rbp 
+	jmp	pmarkr_node
+
+pmarkr_hnf_1:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,2
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,40000000h
+	jbe	pmarkr_shared_argument_part
+	or	dword ptr 4[rdi+rbx*4],1
+pmarkr_shared_argument_part:
+	mov	rbp,qword ptr [rcx]
+	mov	qword ptr [rcx],rsi 
+	lea	rsi,2[rcx]
+	mov	rcx,rbp
+	jmp	pmarkr_node
+
+pmarkr_no_selector_2:
+	pop	rbx 
+pmarkr_no_selector_1:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,20000000h
+	jbe	pmarkr_shared_argument_part
+
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	pmarkr_shared_argument_part
+
+pmarkr_lazy_node_1:
+	je	pmarkr_no_selector_1
+
+pmarkr_selector_node_1:
+	add	rbp,2
+	je	pmarkr_indirection_node
+
+	add	rbp,1
+
+	push	rbx 
+	mov	rbp,qword ptr [rcx]
+	push	rax 
+	lea	rax,[r10+rbp]
+
+	jle	pmarkr_record_selector_node_1
+
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rbx*4]
+	pop	rax 
+	jne	pmarkr_no_selector_2
+
+	mov	rbx,qword ptr [rbp]
+	test	bl,2
+	je	pmarkr_no_selector_2
+
+	cmp	word ptr (-2)[rbx],2
+	jbe	pmarkr_small_tuple_or_record
+
+pmarkr_large_tuple_or_record:
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8
+	and	r8,31*8
+	shr	rbx,8
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+	jne	pmarkr_no_selector_2
+
+ ifdef NEW_DESCRIPTORS
+	mov	eax,dword ptr (-8)[rax]
+	lea	rdx,__indirection
+	pop	rbx 
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+	mov	r8,rcx
+
+	cmp	rax,16
+	jl	pmarkr_tuple_selector_node_1
+	mov	rdx,r9
+	je	pmarkr_tuple_selector_node_2
+	mov	rcx,qword ptr (-24)[r9+rax]
+	mov	qword ptr [r8],rcx
+	jmp	pmarkr_node
+
+pmarkr_tuple_selector_node_2:
+	mov	rcx,qword ptr [r9]
+	mov	qword ptr [r8],rcx
+	jmp	pmarkr_node
+ endif
+
+pmarkr_small_tuple_or_record:
+ ifdef NEW_DESCRIPTORS
+	mov	eax,(-8)[rax]
+	lea	rdx,__indirection
+	pop	rbx 
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+	mov	r8,rcx
+pmarkr_tuple_selector_node_1:
+	mov	rcx,qword ptr [rbp+rax]
+	mov	qword ptr [r8],rcx
+	jmp	pmarkr_node
+ else
+	mov	eax,(-8)[rax]
+	pop	rbx 
+
+	push	rcx
+	mov	rcx,qword ptr [rcx]
+	mov	eax,4[rax]
+	call	near ptr rax
+	pop	rdx 
+
+	mov	qword ptr (-8)[rdx],offset __indirection
+	mov	qword ptr [rdx],rcx 
+ endif
+	jmp	pmarkr_node
+
+pmarkr_record_selector_node_1:
+	je	pmarkr_strict_record_selector_node_1
+
+	mov	rbx,rax
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rbx*4]
+	pop	rax 
+	jne	pmarkr_no_selector_2
+
+	mov	rbx,qword ptr [rbp]
+	test	bl,2
+	je	pmarkr_no_selector_2
+	
+	cmp	word ptr (-2)[rbx],258
+ ifdef NEW_DESCRIPTORS
+	jbe	pmarkr_small_record
+
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8
+	and	r8,31*8
+	shr	rbx,8
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+	jne	pmarkr_no_selector_2
+
+pmarkr_small_record:
+	mov	eax,dword ptr (-8)[rax]
+	lea	rdx,__indirection
+	pop	rbx 
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+	mov	r8,rcx
+
+	cmp	rax,16
+	jle	pmarkr_record_selector_node_2
+	mov	rdx,r9
+	sub	rax,24
+pmarkr_record_selector_node_2:
+	mov	rcx,qword ptr [rbp+rax]
+	mov	qword ptr [r8],rcx
+	jmp	pmarkr_node
+ else
+	jbe	pmarkr_small_tuple_or_record
+	jmp	pmarkr_large_tuple_or_record
+ endif
+
+pmarkr_strict_record_selector_node_1:
+	mov	rbx,rax 
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	test	eax,dword ptr [rdi+rbx*4]
+	pop	rax 
+	jne	pmarkr_no_selector_2
+
+	mov	rbx,qword ptr [rbp]
+	test	bl,2
+	je	pmarkr_no_selector_2
+
+	cmp	word ptr (-2)[rbx],258
+	jle	pmarkr_select_from_small_record
+
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8 
+	and	r8,31*8
+	shr	rbx,8
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+	jne	pmarkr_no_selector_2
+
+pmarkr_select_from_small_record:
+ ifdef NEW_DESCRIPTORS
+	mov	eax,(-8)[rax]
+	sub	rcx,8
+
+	movzx	ebx,word ptr 4[rax]
+	cmp	rbx,16
+	jle	pmarkr_strict_record_selector_node_2
+	mov	rbx,qword ptr (-24)[r9+rbx]
+	jmp	pmarkr_strict_record_selector_node_3
+pmarkr_strict_record_selector_node_2:
+	mov	rbx,qword ptr [rdx+rbx]
+pmarkr_strict_record_selector_node_3:
+	mov	qword ptr 8[rcx],rbx
+
+	movzx	ebx,word ptr 6[rax]
+	test	rbx,rbx
+	je	pmarkr_strict_record_selector_node_5
+	cmp	rbx,16
+	jle	pmarkr_strict_record_selector_node_4
+	mov	rbp,r9
+	sub	rbx,24
+pmarkr_strict_record_selector_node_4:
+	mov	rbx,qword ptr [rbp+rbx]
+	mov	qword ptr 16[rcx],rbx
+pmarkr_strict_record_selector_node_5:
+	pop	rbx
+
+	mov	rax,qword ptr (-8)[rax]
+	mov	qword ptr [rcx],rax
+ else
+	mov	eax,(-8)[rax]
+	pop	rbx 
+	mov	rdx,qword ptr [rcx]
+	sub	rcx,8
+	mov	eax,4[rax]
+	call	near ptr rax
+ endif
+	jmp	pmarkr_node
+
+pmarkr_indirection_node:
+	mov	rcx,qword ptr [rcx]
+	jmp	pmarkr_node
+
+pmarkr_hnf_2:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,20000000h
+	jbe	pmarkr_fits_in_word_6
+	or	dword ptr 4[rdi+rbx*4],1
+pmarkr_fits_in_word_6:
+
+pmarkr_record_2_c:
+	mov	rax,qword ptr [rcx]
+	mov	rbp,qword ptr 8[rcx]
+	or	rax,2
+	mov	qword ptr 8[rcx],rsi 
+	mov	qword ptr [rcx],rax 
+	lea	rsi,8[rcx]
+	mov	rcx,rbp 
+
+pmarkr_node:
+	lea	rdx,[r10+rcx] 
+
+	cmp	rdx,r11
+	jae	pmarkr_next_node
+
+	mov	rbx,rdx 
+	and	rdx,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_set_table2)[rdx]
+	test	ebp,dword ptr [rdi+rbx*4]
+	je	pmarkr_arguments
+
+pmarkr_next_node:
+	test	rsi,3
+	jne	pmarkr_parent
+
+	mov	rbp,qword ptr (-8)[rsi]
+	mov	rdx,qword ptr [rsi]
+	mov	qword ptr [rsi],rcx 
+	mov	qword ptr (-8)[rsi],rdx 
+	sub	rsi,8
+
+	mov	rcx,rbp 
+	and	rbp,3
+	and	rcx,-4
+	or	rsi,rbp 
+	jmp	pmarkr_node
+
+pmarkr_parent:
+	mov	rbx,rsi 
+	and	rsi,-4
+	je	end_pmarkr_using_reversal
+
+	and	rbx,3
+	mov	rbp,qword ptr [rsi]
+	mov	qword ptr [rsi],rcx 
+
+	sub	rbx,1
+	je	pmarkr_argument_part_parent
+
+	lea	rcx,(-8)[rsi]
+	mov	rsi,rbp 
+	jmp	pmarkr_next_node
+	
+pmarkr_argument_part_parent:
+	and	rbp,-4
+	mov	rdx,rsi 
+	mov	rcx,qword ptr (-8)[rbp]
+	mov	rbx,qword ptr [rbp]
+	mov	qword ptr (-8)[rbp],rbx 
+	mov	qword ptr [rbp],rdx 
+	lea	rsi,(2-8)[rbp]
+	jmp	pmarkr_node
+
+pmarkr_lazy_node:
+	movsxd	rbp,dword ptr (-4)[rax]
+	test	rbp,rbp 
+	je	pmarkr_real_or_file
+
+	add	rcx,8
+	cmp	rbp,1
+	jle	pmarkr_lazy_node_1
+	cmp	rbp,256
+	jge	pmarkr_closure_with_unboxed_arguments
+
+	add	rbp,1
+	mov	rax,rdx 
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,rbp 
+
+	lea	rax,[rax+rbp*8]
+	sub	rbp,2
+
+	or	dword ptr [rdi+rbx*4],edx 
+
+	cmp	rax,32*8
+	jbe	pmarkr_fits_in_word_7
+	or	dword ptr 4[rdi+rbx*4],1
+pmarkr_fits_in_word_7:
+pmarkr_closure_with_unboxed_arguments_2:
+	lea	rdx,[rcx+rbp*8]
+	mov	rax,qword ptr [rcx]
+	or	rax,2
+	mov	qword ptr [rcx],rax 
+	mov	rcx,qword ptr [rdx]
+	mov	qword ptr [rdx],rsi 
+	mov	rsi,rdx 
+	jmp	pmarkr_node
+
+pmarkr_closure_with_unboxed_arguments:
+	mov	rax,rbp 
+	and	rbp,255
+
+	sub	rbp,1
+	je	pmarkr_closure_1_with_unboxed_argument
+	add	rbp,2
+
+	shr	rax,8
+	add	r14,rbp 
+
+	push	rcx 
+	lea	rcx,[rdx+rbp*8]
+
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	sub	rbp,rax 
+
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rcx,32*8
+	jbe	pmarkr_fits_in_word_7_
+	or	dword ptr 4[rdi+rbx*4],1
+pmarkr_fits_in_word_7_:
+	pop	rcx
+	sub	rbp,2
+	jg	pmarkr_closure_with_unboxed_arguments_2
+	je	pmarkr_shared_argument_part
+	sub	rcx,8
+	jmp	pmarkr_next_node
+
+pmarkr_closure_1_with_unboxed_argument:
+	sub	rcx,8
+	jmp	pmarkr_real_or_file
+
+pmarkr_hnf_0:
+	cmp	rax,offset dINT+2
+	jne	pmarkr_no_int_3
+
+	mov	rbp,qword ptr 8[rcx]
+	cmp	rbp,33
+
+	jb	pmarkr_small_int
+
+pmarkr_bool_or_small_string:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,2
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,40000000h
+	jbe	pmarkr_next_node
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	pmarkr_next_node
+
+pmarkr_small_int:
+	shl	rbp,4
+	lea	rcx,(small_integers)[rbp]
+	jmp	pmarkr_next_node
+
+pmarkr_no_int_3:
+	jb	pmarkr_real_file_or_string
+
+ 	cmp	rax,offset CHAR+2
+ 	jne	pmarkr_no_char_3
+
+	movzx	rbp,byte ptr 8[rcx]
+	shl	rbp,4
+	lea	rcx,(static_characters)[rbp]
+	jmp	pmarkr_next_node
+
+pmarkr_no_char_3:
+	jb	pmarkr_bool_or_small_string
+
+ ifdef NEW_DESCRIPTORS
+	lea	rcx,((-8)-2)[rax]
+ else
+	lea	rcx,((-12)-2)[rax]
+ endif
+	jmp	pmarkr_next_node
+
+pmarkr_real_file_or_string:
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+	jbe	pmarkr_string_or_array
+
+pmarkr_real_or_file:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+
+	or	dword ptr [rdi+rbx*4],edx 
+	
+	cmp	rdx,20000000h
+	jbe	pmarkr_next_node
+
+	or	dword ptr 4[rdi+rbx*4],1
+	jmp	pmarkr_next_node
+
+pmarkr_record:
+	sub	rbp,258
+	je	pmarkr_record_2
+	jl	pmarkr_record_1
+
+pmarkr_record_3:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,20000000h
+	jbe	pmarkr_fits_in_word_13
+	or	dword ptr 4[rdi+rbx*4],1
+pmarkr_fits_in_word_13:
+	movzx	rbx,word ptr (-2+2)[rax]
+
+	mov	rdx,qword ptr 8[rcx]
+	lea	rdx,[r10+rdx]
+	mov	rax,rdx 
+	and	rdx,31*8
+	shr	rax,8
+
+	push	rsi 
+
+	mov	esi,dword ptr (bit_set_table2)[rdx]
+	test	esi,dword ptr [rdi+rax*4]
+	jne	pmarkr_shared_record_argument_part
+
+	add	rbp,1
+	or	dword ptr [rdi+rax*4],esi 
+
+	lea	rdx,[rdx+rbp*8]
+	add	r14,rbp 
+
+	pop	rsi 
+
+	cmp	rdx,32*8
+	jbe	pmarkr_fits_in_word_14
+	or	dword ptr 4[rdi+rax*4],1
+pmarkr_fits_in_word_14:
+	sub	rbx,1
+	mov	rdx,qword ptr 8[rcx]
+	jl	pmarkr_record_3_bb
+	je	pmarkr_shared_argument_part
+
+	mov	qword ptr 8[rcx],rsi 
+	add	rcx,8
+
+	sub	rbx,1
+	je	pmarkr_record_3_aab
+
+	lea	rsi,[rdx+rbx*8]
+	mov	rax,qword ptr [rdx]
+	or	rax,1
+	mov	rbp,qword ptr [rsi]
+	mov	qword ptr [rdx],rax 
+	mov	qword ptr [rsi],rcx 
+	mov	rcx,rbp
+	jmp	pmarkr_node
+
+pmarkr_record_3_bb:
+	sub	rcx,8
+	jmp	pmarkr_next_node
+
+pmarkr_record_3_aab:
+	mov	rbp,qword ptr [rdx]
+	mov	qword ptr [rdx],rcx 
+	lea	rsi,1[rdx]
+	mov	rcx,rbp 
+	jmp	pmarkr_node
+
+pmarkr_shared_record_argument_part:
+	mov	rdx,qword ptr 8[rcx]
+
+	pop	rsi 
+
+	test	rbx,rbx 
+	jne	pmarkr_shared_argument_part
+	sub	rcx,8
+	jmp	pmarkr_next_node
+
+pmarkr_record_2:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,3
+	or	dword ptr [rdi+rbx*4],edx 
+	cmp	rdx,20000000h
+	jbe	pmarkr_fits_in_word_12
+	or	dword ptr 4[rdi+rbx*4],1
+pmarkr_fits_in_word_12:
+	cmp	word ptr (-2+2)[rax],1
+	ja	pmarkr_record_2_c
+	je	pmarkr_shared_argument_part
+	sub	rcx,8
+	jmp	pmarkr_next_node
+
+pmarkr_record_1:
+	cmp	word ptr (-2+2)[rax],0
+	jne	pmarkr_hnf_1
+	sub	rcx,8
+	jmp	pmarkr_bool_or_small_string
+
+pmarkr_string_or_array:
+	je	pmarkr_string_
+
+pmarkr_array:
+	mov	rbp,qword ptr 16[rcx]
+	test	rbp,rbp 
+	je	pmarkr_lazy_array
+
+	movzx	rax,word ptr (-2)[rbp]
+	test	rax,rax 
+	je	pmarkr_strict_basic_array
+
+	movzx	rbp,word ptr (-2+2)[rbp]
+	test	rbp,rbp
+	je	pmarkr_b_record_array
+
+	sub	rax,256
+	cmp	rax,rbp 
+	je	pmarkr_a_record_array
+
+pmarkr_ab_record_array:
+	push	rdx 
+	push	rbx 
+	mov	rbx,rbp 
+
+	mov	rbp,qword ptr 8[rcx]
+	add	rcx,16
+	push	rcx 
+
+	shl	rbp,3
+	mov	rdx,rax 
+	imul	rdx,rbp 
+
+	sub	rax,rbx 
+	add	rcx,8
+	add	rdx,rcx 
+
+	call	reorder
+	
+	pop	rcx 
+
+	xchg	rax,rbx 
+	mov	rbp,qword ptr (-8)[rcx]
+	imul	rax,rbp 
+	imul	rbx,rbp 
+	add	r14,rbx 
+	add	rbx,rax 
+
+	shl	rbx,3
+	lea	rbp,[r10+rcx]
+	add	rbp,rbx 
+
+	pop	rbx 
+	pop	rdx 
+
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	or	dword ptr [rdi+rbx*4],edx 
+
+	lea	rdx,[rcx+rax*8]
+	jmp	pmarkr_r_array
+
+pmarkr_a_record_array:
+	imul	rax,qword ptr 8[rcx]
+	add	rcx,16
+	jmp	pmarkr_lr_array
+
+pmarkr_lazy_array:
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+
+pmarkr_lr_array:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	mov	rbp,r10
+	or	dword ptr [rdi+rbx*4],edx
+	lea	rdx,[rcx+rax*8]
+	add	rbp,rdx 
+pmarkr_r_array:
+	shr	rbp,8
+
+	cmp	rbx,rbp 
+	jae	pmarkr_skip_mark_lazy_array_bits
+
+	inc	rbx 
+
+pmarkr_lazy_array_bits:
+	or	dword ptr [rdi+rbx*4],1
+	inc	rbx 
+	cmp	rbx,rbp 
+	jbe	pmarkr_lazy_array_bits
+
+pmarkr_skip_mark_lazy_array_bits:
+	add	r14,3
+	add	r14,rax 
+
+	cmp	rax,1
+	jbe	pmarkr_array_length_0_1
+
+	mov	rbp,qword ptr [rdx]
+	mov	rbx,qword ptr [rcx]
+	mov	qword ptr [rdx],rbx 
+	mov	qword ptr [rcx],rbp 
+	
+	mov	rbp,qword ptr (-8)[rdx]
+	sub	rdx,8
+	mov	rbx,qword ptr lazy_array_list
+	add	rbp,2
+	mov	qword ptr [rdx],rbx 
+	mov	qword ptr (-8)[rcx],rbp 
+	mov	qword ptr (-16)[rcx],rax 
+	sub	rcx,16
+	mov	qword ptr lazy_array_list,rcx 
+
+	mov	rcx,qword ptr (-8)[rdx]
+	mov	qword ptr (-8)[rdx],rsi 
+	lea	rsi,(-8)[rdx]
+	jmp	pmarkr_node
+
+pmarkr_array_length_0_1:
+	lea	rcx,(-16)[rcx]
+	jb	pmarkr_next_node
+
+	mov	rbx,qword ptr 24[rcx]
+	mov	rbp,qword ptr 16[rcx]
+	mov	rdx,qword ptr lazy_array_list
+	mov	qword ptr 24[rcx],rbp 
+	mov	qword ptr 16[rcx],rdx 
+	mov	qword ptr [rcx],rax 
+	mov	qword ptr lazy_array_list,rcx 
+	mov	qword ptr 8[rcx],rbx
+	add	rcx,8
+
+	mov	rbp,qword ptr [rcx]
+	mov	qword ptr [rcx],rsi 
+	lea	rsi,2[rcx]
+	mov	rcx,rbp 
+	jmp	pmarkr_node
+	
+pmarkr_b_record_array:
+	mov	rbp,qword ptr 8[rcx]
+	sub	rax,256
+	imul	rax,rbp 
+	add	rax,3
+	jmp	pmarkr_basic_array
+
+pmarkr_strict_basic_array:
+	mov	rax,qword ptr 8[rcx]
+	cmp	rbp,offset BOOL+2
+	je	pmarkr_strict_bool_array
+pmarkr_strict_real_array:
+pmarkr_strict_int_array:
+	add	rax,3
+	jmp	pmarkr_basic_array
+pmarkr_strict_bool_array:
+	add	rax,24+7
+	shr	rax,3
+	jmp	pmarkr_basic_array
+
+pmarkr_string_:
+	mov	rax,qword ptr 8[rcx]
+	add	rax,16+7
+	shr	rax,3
+
+pmarkr_basic_array:
+	mov	edx,dword ptr (bit_set_table2)[rdx]
+	add	r14,rax 
+
+	or	dword ptr [rdi+rbx*4],edx 
+	lea	rax,(-8)[rcx+rax*8]
+	
+	add	rax,r10
+	shr	rax,8
+
+	cmp	rbx,rax 
+	jae	pmarkr_next_node
+
+	inc	rbx 
+	mov	rbp,1
+
+	cmp	rbx,rax 
+	jae	pmarkr_last_string_bits
+
+pmarkr_string_lp:
+	or	dword ptr [rdi+rbx*4],ebp 
+	inc	rbx 
+	cmp	rbx,rax 
+	jb	pmarkr_string_lp
+
+pmarkr_last_string_bits:
+	or	dword ptr [rdi+rbx*4],ebp 
+	jmp	pmarkr_next_node