From c502c62005ab5e273fcd02c0db3c65748e23a28d Mon Sep 17 00:00:00 2001
From: John van Groningen
Date: Wed, 14 Mar 2012 11:16:23 +0000
Subject: add thread safe version for 64 bit windows

---
 thread/aap.asm                     | 2402 +++++++++++++++
 thread/acompact.asm                | 1375 +++++++++
 thread/acompact_rmark.asm          | 1131 +++++++
 thread/acompact_rmark_prefetch.asm | 1398 +++++++++
 thread/acompact_rmarkr.asm         | 1095 +++++++
 thread/acopy.asm                   | 1394 +++++++++
 thread/afileIO3.asm                | 1021 +++++++
 thread/amark.asm                   | 2156 +++++++++++++
 thread/amark_prefetch.asm          | 1964 ++++++++++++
 thread/astartup.asm                | 5856 ++++++++++++++++++++++++++++++++++++
 thread/athread.asm                 |  371 +++
 11 files changed, 20163 insertions(+)
 create mode 100644 thread/aap.asm
 create mode 100644 thread/acompact.asm
 create mode 100644 thread/acompact_rmark.asm
 create mode 100644 thread/acompact_rmark_prefetch.asm
 create mode 100644 thread/acompact_rmarkr.asm
 create mode 100644 thread/acopy.asm
 create mode 100644 thread/afileIO3.asm
 create mode 100644 thread/amark.asm
 create mode 100644 thread/amark_prefetch.asm
 create mode 100644 thread/astartup.asm
 create mode 100644 thread/athread.asm

(limited to 'thread')

diff --git a/thread/aap.asm b/thread/aap.asm
new file mode 100644
index 0000000..787537b
--- /dev/null
+++ b/thread/aap.asm
@@ -0,0 +1,2402 @@
+
+
+	d2 equ r10
+	d3 equ r11
+
+ ifndef NEW_DESCRIPTORS
+	extern	__cycle__in__spine:near
+	extern	collect_2:near
+ endif
+ 
+_TEXT	segment para 'CODE'
+_TEXT	ends
+_DATA	segment para 'DATA'
+_DATA	ends
+
+	_TEXT segment
+
+	public	ap_2
+	public	ap_3
+	public	ap_4
+	public	ap_5
+	public	ap_6
+	public	ap_7
+	public	ap_8
+	public	ap_9
+	public	ap_10
+	public	ap_11
+	public	ap_12
+	public	ap_13
+	public	ap_14
+	public	ap_15
+	public	ap_16
+	public	ap_17
+	public	ap_18
+	public	ap_19
+	public	ap_20
+	public	ap_21
+	public	ap_22
+	public	ap_23
+	public	ap_24
+	public	ap_25
+	public	ap_26
+	public	ap_27
+	public	ap_28
+	public	ap_29
+	public	ap_30
+	public	ap_31
+	public	ap_32
+
+	public	add_empty_node_2
+	public	add_empty_node_3
+	public	add_empty_node_4
+	public	add_empty_node_5
+	public	add_empty_node_6
+	public	add_empty_node_7
+	public	add_empty_node_8
+	public	add_empty_node_9
+	public	add_empty_node_10
+	public	add_empty_node_11
+	public	add_empty_node_12
+	public	add_empty_node_13
+	public	add_empty_node_14
+	public	add_empty_node_15
+	public	add_empty_node_16
+	public	add_empty_node_17
+	public	add_empty_node_18
+	public	add_empty_node_19
+	public	add_empty_node_20
+	public	add_empty_node_21
+	public	add_empty_node_22
+	public	add_empty_node_23
+	public	add_empty_node_24
+	public	add_empty_node_25
+	public	add_empty_node_26
+	public	add_empty_node_27
+	public	add_empty_node_28
+	public	add_empty_node_29
+	public	add_empty_node_30
+	public	add_empty_node_31
+	public	add_empty_node_32
+
+	public	yet_args_needed_5
+	public	yet_args_needed_6
+	public	yet_args_needed_7
+	public	yet_args_needed_8
+	public	yet_args_needed_9
+	public	yet_args_needed_10
+	public	yet_args_needed_11
+	public	yet_args_needed_12
+	public	yet_args_needed_13
+	public	yet_args_needed_14
+	public	yet_args_needed_15
+	public	yet_args_needed_16
+	public	yet_args_needed_17
+	public	yet_args_needed_18
+	public	yet_args_needed_19
+	public	yet_args_needed_20
+	public	yet_args_needed_21
+	public	yet_args_needed_22
+	public	yet_args_needed_23
+	public	yet_args_needed_24
+	public	yet_args_needed_25
+	public	yet_args_needed_26
+	public	yet_args_needed_27
+	public	yet_args_needed_28
+	public	yet_args_needed_29
+	public	yet_args_needed_30
+	public	yet_args_needed_31
+
+ap_32:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,32*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap32
+
+ap_31:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,31*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap31
+
+ap_30:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,30*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap30
+
+ap_29:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,29*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap29
+
+ap_28:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,28*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap28
+
+ap_27:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,27*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap27
+
+ap_26:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,26*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap26
+
+ap_25:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,25*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap25
+
+ap_24:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,24*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap24
+
+ap_23:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,23*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap23
+
+ap_22:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,22*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap22
+
+ap_21:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,21*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap21
+
+ap_20:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,20*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap20
+
+ap_19:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,19*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap19
+
+ap_18:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,18*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap18
+
+ap_17:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,17*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap17
+
+ap_16:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,16*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap16
+
+ap_15:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,15*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap15
+
+ap_14:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,14*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap14
+
+ap_13:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,13*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap13
+
+ap_12:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,12*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap12
+
+ap_11:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,11*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap11
+
+ap_10:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,10*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap10
+
+ap_9:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,9*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap9
+
+ap_8:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,8*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap8
+
+ap_7:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,7*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap7
+
+ap_6:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,6*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap6
+
+ap_5:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,5*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap5
+
+ap_4:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,4*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap4
+
+ap_3:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,3*8
+	cmp	word ptr [rbp],bx
+	je	fast_ap
+
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap3
+
+ap_2:
+	mov	rbp,qword ptr [r8]
+	mov	rbx,2*8
+	cmp	word ptr [rbp],bx
+	jne	no_fast_ap2_
+
+fast_ap_2_2_:
+	movzx	rax,word ptr -2[rbp]
+	add	rbx,rbp
+	mov	ebp,dword ptr -6[rbx]
+ ifdef PROFILE
+	sub	rbp,24
+ else
+	sub	rbp,12
+ endif
+	cmp	rax,1
+	jb	repl_args_0_2
+	je	repl_args_1
+
+	cmp	rax,3
+	jb	repl_args_2
+
+	mov	qword ptr [rsi],rcx
+	mov	qword ptr 8[rsi],rdx
+	lea	rsi,16[rsi]
+	mov	rdx,qword ptr 16[r8]
+
+	jmp	fast_ap_
+
+no_fast_ap2_:
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	mov	rdx,r8
+	mov	r8d,dword ptr 2[rbp]
+	add	rsi,8
+	jmp	no_fast_ap2
+
+fast_ap_2_2:
+	mov	r8,rdx
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+	jmp	fast_ap_2_2_
+
+fast_ap_2:
+	mov	r8,rdx
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+fast_ap:
+	movzx	rax,word ptr -2[rbp]
+	add	rbx,rbp
+	mov	ebp,dword ptr -6[rbx]
+ ifdef PROFILE
+	sub	rbp,24
+ else
+	sub	rbp,12
+ endif
+	cmp	rax,1
+	jb	repl_args_0
+	je	repl_args_1
+
+	cmp	rax,3
+	jb	repl_args_2
+
+	mov	qword ptr [rsi],rcx
+	mov	qword ptr 8[rsi],rdx
+	lea	rsi,16[rsi]
+	mov	rdx,qword ptr 16[r8]
+
+fast_ap_:
+	mov	r8,qword ptr 8[r8]
+	je	repl_args_3
+
+	cmp	rax,5
+	jb	repl_args_4
+	je	repl_args_5
+
+	cmp	rax,7
+	jb	repl_args_6
+
+repl_args_7_:
+	mov	rbx,qword ptr -16[rdx+rax*8]
+	mov	qword ptr [rsi],rbx
+	sub	rax,1
+	add	rsi,8
+	cmp	rax,6
+	jne	repl_args_7_
+
+repl_args_6:
+	mov	rax,qword ptr 32[rdx]
+	mov	qword ptr [rsi],rax
+	mov	rax,qword ptr 24[rdx]
+	mov	qword ptr 8[rsi],rax
+	mov	rax,qword ptr 16[rdx]
+	mov	qword ptr 16[rsi],rax
+	mov	rcx,qword ptr 8[rdx]
+	mov	rdx,qword ptr [rdx]
+	add	rsi,24
+	jmp	rbp
+
+repl_args_0:
+	mov	r8,rdx
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+repl_args_0_2:
+	jmp	rbp
+
+repl_args_1:
+	mov	r8,qword ptr 8[r8]
+	jmp	rbp
+
+repl_args_2:
+	mov	qword ptr [rsi],rcx
+	mov	rcx,rdx
+	add	rsi,8
+	mov	rdx,qword ptr 16[r8]
+	mov	r8,qword ptr 8[r8]
+	jmp	rbp
+
+repl_args_3:
+	mov	rcx,qword ptr 8[rdx]
+	mov	rdx,qword ptr [rdx]
+	jmp	rbp
+
+repl_args_4:
+	mov	rax,qword ptr 16[rdx]
+	mov	qword ptr [rsi],rax
+	mov	rcx,qword ptr 8[rdx]
+	mov	rdx,qword ptr [rdx]
+	add	rsi,8
+	jmp	rbp
+
+repl_args_5:
+	mov	rax,qword ptr 24[rdx]
+	mov	qword ptr [rsi],rax
+	mov	rax,qword ptr 16[rdx]
+	mov	qword ptr 8[rsi],rax
+	mov	rcx,qword ptr 8[rdx]
+	mov	rdx,qword ptr [rdx]
+	add	rsi,16
+	jmp	rbp
+
+
+no_fast_ap32:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,31*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap31:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,30*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap30:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,29*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap29:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,28*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap28:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,27*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap27:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,26*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap26:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,25*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap25:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,24*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap24:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,23*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap23:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,22*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap22:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,21*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap21:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,20*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap20:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,19*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap19:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,18*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap18:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,17*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap17:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,16*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap16:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,15*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap15:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,14*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap14:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,13*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap13:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,12*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap12:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,11*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap11:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,10*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap10:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,9*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap9:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,8*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap8:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,7*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap7:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,6*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap6:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,5*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap5:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,4*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap4:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,3*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap3:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+
+	mov	rbx,2*8
+ 	cmp	word ptr [rbp],bx
+	je	fast_ap_2
+
+	mov	r8d,dword ptr 2[rbp]
+no_fast_ap2:
+	call	r8
+	mov	rbp,qword ptr [rcx]
+	mov	rdx,rcx
+	mov	rcx,qword ptr -8[rsi]
+	sub	rsi,8
+	mov	r8d,dword ptr 2[rbp]
+	jmp	r8
+
+
+add_empty_node_2:
+	sub	r15,3
+	jb	add_empty_node_2_gc
+add_empty_node_2_gc_:
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	mov	r8,rdx
+	mov	rdx,rcx
+	mov	rcx,rdi
+	add	rdi,24
+	ret
+add_empty_node_2_gc:
+	call	collect_2
+	jmp	add_empty_node_2_gc_
+
+add_empty_node_3:
+	sub	r15,3
+	jb	add_empty_node_3_gc
+add_empty_node_3_gc_:
+	mov	qword ptr [rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_3_gc:
+	call	collect_3
+	jmp	add_empty_node_3_gc_
+
+add_empty_node_4:
+	sub	r15,3
+	jb	add_empty_node_4_gc
+add_empty_node_4_gc_:
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	qword ptr -8[rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_4_gc:
+	call	collect_3
+	jmp	add_empty_node_4_gc_
+
+add_empty_node_5:
+	sub	r15,3
+	jb	add_empty_node_5_gc
+add_empty_node_5_gc_:
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+	mov	qword ptr -16[rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_5_gc:
+	call	collect_3
+	jmp	add_empty_node_5_gc_
+
+add_empty_node_6:
+	sub	r15,3
+	jb	add_empty_node_6_gc
+add_empty_node_6_gc_:
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+	mov	rbp,qword ptr -24[rsi]
+	mov	qword ptr -16[rsi],rbp
+	mov	qword ptr -24[rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_6_gc:
+	call	collect_3
+	jmp	add_empty_node_6_gc_
+
+add_empty_node_7:
+	sub	r15,3
+	jb	add_empty_node_7_gc
+add_empty_node_7_gc_:
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+	mov	rbp,qword ptr -24[rsi]
+	mov	qword ptr -16[rsi],rbp
+	mov	rbp,qword ptr -32[rsi]
+	mov	qword ptr -24[rsi],rbp
+	mov	qword ptr -32[rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_7_gc:
+	call	collect_3
+	jmp	add_empty_node_7_gc_
+
+add_empty_node_8:
+	sub	r15,3
+	jb	add_empty_node_8_gc
+add_empty_node_8_gc_:
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+	mov	rbp,qword ptr -24[rsi]
+	mov	qword ptr -16[rsi],rbp
+	mov	rbp,qword ptr -32[rsi]
+	mov	qword ptr -24[rsi],rbp
+	mov	rbp,qword ptr -40[rsi]
+	mov	qword ptr -32[rsi],rbp
+	mov	qword ptr -40[rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_8_gc:
+	call	collect_3
+	jmp	add_empty_node_8_gc_
+
+add_empty_node_9:
+	sub	r15,3
+	jb	add_empty_node_9_gc
+add_empty_node_9_gc_:
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+	mov	rbp,qword ptr -24[rsi]
+	mov	qword ptr -16[rsi],rbp
+	mov	rbp,qword ptr -32[rsi]
+	mov	qword ptr -24[rsi],rbp
+	mov	rbp,qword ptr -40[rsi]
+	mov	qword ptr -32[rsi],rbp
+	mov	rbp,qword ptr -48[rsi]
+	mov	qword ptr -40[rsi],rbp
+	mov	qword ptr -48[rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_9_gc:
+	call	collect_3
+	jmp	add_empty_node_9_gc_
+
+add_empty_node_10:
+	sub	r15,3
+	jb	add_empty_node_10_gc
+add_empty_node_10_gc_:
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+	mov	rbp,qword ptr -24[rsi]
+	mov	qword ptr -16[rsi],rbp
+	mov	rbp,qword ptr -32[rsi]
+	mov	qword ptr -24[rsi],rbp
+	mov	rbp,qword ptr -40[rsi]
+	mov	qword ptr -32[rsi],rbp
+	mov	rbp,qword ptr -48[rsi]
+	mov	qword ptr -40[rsi],rbp
+	mov	rbp,qword ptr -56[rsi]
+	mov	qword ptr -48[rsi],rbp
+	mov	qword ptr -56[rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_10_gc:
+	call	collect_3
+	jmp	add_empty_node_10_gc_
+
+add_empty_node_11:
+	sub	r15,3
+	jb	add_empty_node_11_gc
+add_empty_node_11_gc_:
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+	mov	rbp,qword ptr -24[rsi]
+	mov	qword ptr -16[rsi],rbp
+	mov	rbp,qword ptr -32[rsi]
+	mov	qword ptr -24[rsi],rbp
+	mov	rbp,qword ptr -40[rsi]
+	mov	qword ptr -32[rsi],rbp
+	mov	rbp,qword ptr -48[rsi]
+	mov	qword ptr -40[rsi],rbp
+	mov	rbp,qword ptr -56[rsi]
+	mov	qword ptr -48[rsi],rbp
+	mov	rbp,qword ptr -64[rsi]
+	mov	qword ptr -56[rsi],rbp
+	mov	qword ptr -64[rsi],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_11_gc:
+	call	collect_3
+	jmp	add_empty_node_11_gc_
+
+add_empty_node_32:
+	mov	rbx,7
+	jmp	add_empty_node_12_
+add_empty_node_28:
+	mov	rbx,6
+	jmp	add_empty_node_12_
+add_empty_node_24:
+	mov	rbx,5
+	jmp	add_empty_node_12_
+add_empty_node_20:
+	mov	rbx,4
+	jmp	add_empty_node_12_
+add_empty_node_16:
+	mov	rbx,3
+	jmp	add_empty_node_12_
+add_empty_node_12:
+	mov	rbx,2
+add_empty_node_12_:
+	sub	r15,3
+	jb	add_empty_node_12_gc
+add_empty_node_12_gc_:
+	mov	rax,rsi
+add_empty_node_12_lp:
+	mov	rbp,qword ptr -8[rax]
+	mov	qword ptr [rax],rbp
+	mov	rbp,qword ptr -16[rax]
+	mov	qword ptr -8[rax],rbp
+	mov	rbp,qword ptr -24[rax]
+	mov	qword ptr -16[rax],rbp
+	mov	rbp,qword ptr -32[rax]
+	mov	qword ptr -24[rax],rbp
+	sub	rax,32
+	sub	rbx,1
+	jne	add_empty_node_12_lp
+	mov	qword ptr [rax],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_12_gc:
+	call	collect_3
+	jmp	add_empty_node_12_gc_
+
+add_empty_node_29:
+	mov	rbx,6
+	jmp	add_empty_node_13_
+add_empty_node_25:
+	mov	rbx,5
+	jmp	add_empty_node_13_
+add_empty_node_21:
+	mov	rbx,4
+	jmp	add_empty_node_13_
+add_empty_node_17:
+	mov	rbx,3
+	jmp	add_empty_node_13_
+add_empty_node_13:
+	mov	rbx,2
+add_empty_node_13_:
+	sub	r15,3
+	jb	add_empty_node_13_gc
+add_empty_node_13_gc_:
+	mov	rax,rsi
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+add_empty_node_13_lp:
+	mov	rbp,qword ptr -16[rax]
+	mov	qword ptr -8[rax],rbp
+	mov	rbp,qword ptr -24[rax]
+	mov	qword ptr -16[rax],rbp
+	mov	rbp,qword ptr -32[rax]
+	mov	qword ptr -24[rax],rbp
+	mov	rbp,qword ptr -40[rax]
+	mov	qword ptr -32[rax],rbp
+	sub	rax,32
+	sub	rbx,1
+	jne	add_empty_node_13_lp
+	mov	qword ptr -8[rax],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_13_gc:
+	call	collect_3
+	jmp	add_empty_node_13_gc_
+
+add_empty_node_30:
+	mov	rbx,6
+	jmp	add_empty_node_14_
+add_empty_node_26:
+	mov	rbx,5
+	jmp	add_empty_node_14_
+add_empty_node_22:
+	mov	rbx,4
+	jmp	add_empty_node_14_
+add_empty_node_18:
+	mov	rbx,3
+	jmp	add_empty_node_14_
+add_empty_node_14:
+	mov	rbx,2
+add_empty_node_14_:
+	sub	r15,3
+	jb	add_empty_node_14_gc
+add_empty_node_14_gc_:
+	mov	rax,rsi
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rsi,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+add_empty_node_14_lp:
+	mov	rbp,qword ptr -24[rax]
+	mov	qword ptr -16[rax],rbp
+	mov	rbp,qword ptr -32[rax]
+	mov	qword ptr -24[rax],rbp
+	mov	rbp,qword ptr -40[rax]
+	mov	qword ptr -32[rax],rbp
+	mov	rbp,qword ptr -48[rax]
+	mov	qword ptr -40[rax],rbp
+	sub	rax,32
+	sub	rbx,1
+	jne	add_empty_node_14_lp
+	mov	qword ptr -16[rax],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_14_gc:
+	call	collect_3
+	jmp	add_empty_node_14_gc_
+
+add_empty_node_31:
+	mov	rbx,6
+	jmp	add_empty_node_15_
+add_empty_node_27:
+	mov	rbx,5
+	jmp	add_empty_node_15_
+add_empty_node_23:
+	mov	rbx,4
+	jmp	add_empty_node_15_
+add_empty_node_19:
+	mov	rbx,3
+	jmp	add_empty_node_15_
+add_empty_node_15:
+	mov	rbx,2
+add_empty_node_15_:
+	sub	r15,3
+	jb	add_empty_node_15_gc
+add_empty_node_15_gc_:
+	mov	rax,rbp
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr [rsi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr -8[rsi],rbp
+	mov	rbp,qword ptr -24[rsi]
+	mov	qword ptr -16[rsi],rbp
+add_empty_node_15_lp:
+	mov	rbp,qword ptr -32[rax]
+	mov	qword ptr -24[rax],rbp
+	mov	rbp,qword ptr -40[rax]
+	mov	qword ptr -32[rax],rbp
+	mov	rbp,qword ptr -48[rax]
+	mov	qword ptr -40[rax],rbp
+	mov	rbp,qword ptr -56[rax]
+	mov	qword ptr -48[rax],rbp
+	sub	rax,32
+	sub	rbx,1
+	jne	add_empty_node_15_lp
+	mov	qword ptr -24[rax],rdi
+	mov	qword ptr [rdi],offset __cycle__in__spine
+	add	rsi,8
+	add	rdi,24
+	ret
+add_empty_node_15_gc:
+	call	collect_3
+	jmp	add_empty_node_15_gc_
+
+
+yet_args_needed_0:
+	sub	r15,2
+	jb	yet_args_needed_0_gc
+yet_args_needed_0_gc_r:
+	mov	qword ptr 8[rdi],rcx
+	mov	rax,qword ptr [rdx]
+	mov	rcx,rdi
+	add	rax,8
+	mov	qword ptr [rdi],rax
+	add	rdi,16
+	ret
+
+yet_args_needed_0_gc:
+	call	collect_2
+	jmp	yet_args_needed_0_gc_r
+
+
+	align	(4)
+	sub	r15,3
+	jae short build_node_2_gc_r
+	jmp short build_node_2_gc
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+ endif
+yet_args_needed_1:
+	sub	r15,3
+	jb	yet_args_needed_1_gc
+yet_args_needed_1_gc_r:
+	mov	qword ptr 16[rdi],rcx
+	mov	rax,qword ptr [rdx]
+	mov	rcx,rdi
+	add	rax,8
+	mov	qword ptr [rdi],rax
+	mov	rbx,qword ptr 8[rdx]
+	mov	qword ptr 8[rdi],rbx
+	add	rdi,24
+	ret
+
+yet_args_needed_1_gc:
+	call	collect_2
+	jmp	yet_args_needed_1_gc_r
+
+build_node_2_gc_r:
+	mov	qword ptr [rdi],rbx
+	mov	qword ptr 8[rdi],rdx
+	mov	qword ptr 16[rdi],rcx
+	mov	rcx,rdi
+	add	rdi,24
+	ret
+
+build_node_2_gc:
+	call	collect_2
+	jmp	build_node_2_gc_r
+
+
+	align	(4)
+	sub	r15,5
+	jae short build_node_3_gc_r
+	jmp short build_node_3_gc
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_2:
+	sub	r15,5
+	jb	gc_22
+gc_r_22:
+	mov	rax,qword ptr [rdx]
+	mov	qword ptr 8[rdi],rcx
+	add	rax,8
+	mov	d2,qword ptr 8[rdx]
+	mov	qword ptr 16[rdi],rax
+	lea	rcx,16[rdi]
+	mov	rbp,qword ptr 16[rdx]
+	mov	qword ptr 24[rdi],d2
+	mov	qword ptr [rdi],rbp
+	mov	qword ptr 32[rdi],rdi
+	add	rdi,40
+	ret
+
+gc_22:	call	collect_2
+	jmp	gc_r_22
+
+build_node_3_gc_r:
+	mov	qword ptr [rdi],rbx
+	lea	rbp,24[rdi]
+	mov	qword ptr 8[rdi],r8
+	mov	qword ptr 16[rdi],rbp
+	mov	qword ptr 24[rdi],rdx
+	mov	qword ptr 32[rdi],rcx
+	mov	rcx,rdi
+	add	rdi,40
+	ret
+
+build_node_3_gc:
+	call	collect_2
+	jmp	build_node_3_gc_r
+
+
+	align	(4)
+	sub	r15,6
+	jae short build_node_4_gc_r
+	jmp short build_node_4_gc
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_3:
+	sub	r15,6
+	jb	gc_23
+gc_r_23:
+	mov	rax,qword ptr [rdx]
+	mov	qword ptr 16[rdi],rcx
+	add	rax,8
+	mov	d2,qword ptr 8[rdx]
+	mov	qword ptr 24[rdi],rax
+	mov	rdx,qword ptr 16[rdx]
+	mov	qword ptr 32[rdi],d2
+	mov	rbp,qword ptr [rdx]
+	mov	qword ptr 40[rdi],rdi
+	mov	qword ptr [rdi],rbp
+	mov	rbp,qword ptr 8[rdx]
+	lea	rcx,24[rdi]
+	mov	qword ptr 8[rdi],rbp
+	add	rdi,48
+	ret
+
+gc_23:	call	collect_2
+	jmp	gc_r_23
+
+build_node_4_gc_r:
+	mov	qword ptr [rdi],rbx
+	lea	rbp,24[rdi]
+	mov	qword ptr 8[rdi],r8
+	mov	qword ptr 16[rdi],rbp
+	mov	qword ptr 24[rdi],rdx
+	mov	qword ptr 32[rdi],rcx
+	mov	rcx,rdi
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr 40[rdi],rbp
+	sub	rsi,8
+	add	rdi,48
+	ret
+
+build_node_4_gc:
+	call	collect_2
+	jmp	build_node_4_gc_r
+
+
+	align	(4)
+	sub	r15,7
+	jae short build_node_5_gc_r
+	jmp 	build_node_5_gc
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_4:
+	sub	r15,7
+	jb	gc_24
+gc_r_24:
+	mov	rax,qword ptr [rdx]
+	mov	qword ptr 24[rdi],rcx
+	add	rax,8
+	mov	d2,qword ptr 8[rdx]
+	mov	qword ptr 32[rdi],rax
+	mov	rdx,qword ptr 16[rdx]
+	mov	qword ptr 40[rdi],d2
+	mov	rbp,qword ptr [rdx]
+	mov	48[rdi],rdi
+	mov	qword ptr [rdi],rbp
+	mov	rbp,qword ptr 8[rdx]
+	lea	rcx,32[rdi]
+	mov	qword ptr 8[rdi],rbp
+	mov	rbp,qword ptr 16[rdx]
+	mov	qword ptr 16[rdi],rbp
+	add	rdi,56
+	ret
+
+gc_24:	call	collect_2
+	jmp	gc_r_24
+
+build_node_5_gc_r:
+	mov	qword ptr [rdi],rbx
+	lea	rbp,24[rdi]
+	mov	qword ptr 8[rdi],r8
+	mov	qword ptr 16[rdi],rbp
+	mov	qword ptr 24[rdi],rdx
+	mov	qword ptr 32[rdi],rcx
+	mov	rcx,rdi
+	mov	rbp,qword ptr -8[rsi]
+	mov	qword ptr 40[rdi],rbp
+	mov	rbp,qword ptr -16[rsi]
+	mov	qword ptr 48[rdi],rbp
+	sub	rsi,16
+	add	rdi,56
+	ret
+
+build_node_5_gc:
+	call	collect_2
+	jmp	build_node_5_gc_r
+
+
+	align	(4)
+	mov	eax,8
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_5:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,8
+	jmp	yet_args_needed_
+
+
+	align	(4)
+	mov	eax,9
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_6:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,9
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,10
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_7:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,10
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,11
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_8:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,11
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,12
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_9:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,12
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,13
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_10:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,13
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,14
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_11:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,14
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,15
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_12:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,15
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,16
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_13:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,16
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,17
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_14:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,17
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,18
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_15:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,18
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,19
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_16:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,19
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,20
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_17:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,20
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,21
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_18:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,21
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,22
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_19:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,22
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,23
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_20:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,23
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,24
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_21:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,24
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,25
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_22:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,25
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,26
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_23:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,26
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,27
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_24:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,27
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,28
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_25:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,28
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,29
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_26:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,29
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,30
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_27:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,30
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,31
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_28:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,31
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,32
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_29:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,32
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,33
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_30:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,33
+	jmp	yet_args_needed_
+
+	align	(4)
+	mov	eax,34
+	jmp	build_node_
+	nop
+	nop
+	align	(4)	
+ ifdef PROFILE
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop	
+	nop
+	nop
+	nop
+	nop	
+ endif
+yet_args_needed_31:
+	mov	rbx,qword ptr [rdx]
+	mov	rax,34
+	jmp	yet_args_needed_
+
+yet_args_needed:
+	mov	rbx,qword ptr [rdx]
+	movzx	rax,word ptr -2[rbx]
+	add	rax,3
+yet_args_needed_:
+	sub	r15,rax
+	jb	yet_args_needed_gc
+yet_args_needed_gc_r:
+	mov	d3,qword ptr 8[rdx]
+	sub	rax,3+1+4
+	mov	rdx,qword ptr 16[rdx]
+	mov	d2,rdi
+	mov	rbp,qword ptr [rdx]
+	mov	qword ptr [rdi],rbp
+	mov	rbp,qword ptr 8[rdx]
+	mov	qword ptr 8[rdi],rbp
+	mov	rbp,qword ptr 16[rdx]
+	mov	qword ptr 16[rdi],rbp
+	add	rdx,24
+	add	rdi,24
+
+yet_args_needed_cp_a:
+	mov	rbp,qword ptr [rdx]
+	add	rdx,8
+	mov	qword ptr [rdi],rbp
+	add	rdi,8
+	sub	rax,1
+	jge	yet_args_needed_cp_a
+
+	mov	qword ptr [rdi],rcx
+	add	rbx,8
+	mov	qword ptr 8[rdi],rbx
+	lea	rcx,8[rdi]
+	mov	qword ptr 16[rdi],d3
+	mov	qword ptr 24[rdi],d2
+	add	rdi,32
+	ret
+
+yet_args_needed_gc:
+	call	collect_2
+	jmp	yet_args_needed_gc_r
+
+build_node_:
+	sub	r15,rax
+	jb	build_node_gc
+build_node_gc_r:
+	mov	qword ptr [rdi],rbx
+	lea	rbp,24[rdi]
+	mov	qword ptr 8[rdi],r8
+	mov	qword ptr 16[rdi],rbp
+	mov	qword ptr 24[rdi],rdx
+	mov	qword ptr 32[rdi],rcx
+	mov	rcx,rdi
+	mov	r8,qword ptr -8[rsi]
+	mov	qword ptr 40[rdi],r8
+	mov	r8,qword ptr -16[rsi]
+	sub	rsi,16
+	mov	qword ptr 48[rdi],r8
+	add	rdi,56
+
+	sub	rax,5+2
+build_node_cp_a:
+	mov	r8,qword ptr -8[rsi]
+	sub	rsi,8
+	mov	qword ptr [rdi],r8
+	add	rdi,8
+	sub	rax,1
+	jne	build_node_cp_a
+
+	ret
+
+build_node_gc:
+	call	collect_3
+	jmp	build_node_gc_r
+	
+_TEXT	ends
+
+ ifndef NEW_DESCRIPTORS
+	end
+ endif
diff --git a/thread/acompact.asm b/thread/acompact.asm
new file mode 100644
index 0000000..202565c
--- /dev/null
+++ b/thread/acompact.asm
@@ -0,0 +1,1375 @@
+
+; mark used nodes and pointers in argument parts and link backward pointers
+
+ if THREAD
+	mov	rax,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rax,qword ptr heap_size_65
+ endif
+	shl	rax,6
+ if THREAD
+	mov	qword ptr heap_size_64_65_offset[r9],rax 
+ else
+	mov	qword ptr heap_size_64_65,rax 
+ endif
+
+	lea	rax,(-16000)[rsp]
+ if THREAD
+	mov	qword ptr end_stack_offset[r9],rax
+ else
+	mov	qword ptr end_stack,rax
+ endif
+
+	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]
+
+ if THREAD
+	mov	qword ptr end_vector_offset[r9],rcx
+ else
+	mov	qword ptr end_vector,rcx 
+ endif
+
+	call	rmark_stack_nodes
+	
+	pop	rax
+	test	rax,rax 
+	jne	mark_cafs_lp
+
+end_mark_cafs:
+ if THREAD
+	mov	rsi,qword ptr stack_p_offset[r9]
+
+	mov	rcx,qword ptr stack_top_offset[r9]
+	mov	qword ptr end_vector_offset[r9],rcx 
+ else
+	mov	rsi,qword ptr stack_p
+
+	mov	rcx,qword ptr stack_top
+	mov	qword ptr end_vector,rcx 
+ endif
+
+	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]
+
+ if THREAD
+	mov	qword ptr end_vector_offset[r9],rcx
+ else
+	mov	qword ptr end_vector,rcx
+ endif
+
+	call	rmarkp_stack_nodes
+	
+	pop	rax
+	test	rax,rax 
+	jne	rmarkp_cafs_lp
+
+end_rmarkp_cafs:
+ if THREAD
+	mov	rsi,qword ptr stack_p_offset[r9]
+
+	mov	rcx,qword ptr stack_top_offset[r9]
+	mov	qword ptr end_vector_offset[r9],rcx
+ else
+	mov	rsi,qword ptr stack_p
+
+	mov	rcx,qword ptr stack_top
+	mov	qword ptr end_vector,rcx
+ endif
+
+	call	rmarkp_stack_nodes
+
+	call	add_mark_compact_garbage_collect_time
+	
+	jmp	compact_heap
+
+	include	acompact_rmark.asm
+
+	include	acompact_rmark_prefetch.asm
+
+	include acompact_rmarkr.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:
+ if THREAD
+ 	lea	rax,__Nil-8
+	cmp	rbp,rax
+ else
+	lea	r9,__Nil-8
+	cmp	rbp,r9
+ endif
+	je	end_finalizers_after_compact1
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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
+ if THREAD
+	lea	rax,__Nil-8
+	cmp	rcx,rax
+ else
+	lea	r9,__Nil-8
+	cmp	rcx,r9
+endif
+	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
+ if THREAD
+	lea	rax,__Nil-8
+	cmp	qword ptr [rbp],rax
+ else
+	lea	r9,__Nil-8
+	cmp	qword ptr [rbp],r9
+ endif
+	je	free_finalizer_list_empty
+
+ if THREAD
+	mov	qword ptr end_vector_offset[r9],offset free_finalizer_list+8
+ else
+	mov	qword ptr end_vector,offset free_finalizer_list+8
+ endif
+
+	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:
+
+ if THREAD
+	mov	rax,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rax,qword ptr heap_size_65
+ endif
+	mov	rbx,rax 
+	shl	rbx,6
+
+ if THREAD
+	add	rbx,qword ptr heap_p3_offset[r9]
+	mov	qword ptr end_heap_p3_offset[r9],rbx 
+ else
+	add	rbx,qword ptr heap_p3
+	mov	qword ptr end_heap_p3,rbx 
+ endif
+
+	add	rax,3
+	shr	rax,2
+	mov	r12,rax
+
+ if THREAD
+	mov	r8,qword ptr heap_vector_offset[r9]
+ else
+	mov	r8,qword ptr heap_vector
+ endif
+
+	lea	rbx,4[r8]
+	neg	rbx
+ if THREAD
+	mov	qword ptr neg_heap_vector_plus_4_offset[r9],rbx 
+
+ 	mov	rdi,qword ptr heap_p3_offset[r9]
+ else
+	mov	qword ptr neg_heap_vector_plus_4,rbx 
+
+	mov	rdi,qword ptr heap_p3
+ endif
+	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:
+ if THREAD
+	mov	rbp,qword ptr neg_heap_vector_plus_4_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_vector_plus_4
+ endif
+
+	add	rbp,r8
+
+	shl	rbp,6
+ if THREAD
+	add	rbp,qword ptr heap_p3_offset[r9]
+ else
+	add	rbp,qword ptr heap_p3
+ endif
+
+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
+ if THREAD
+	mov	rbx,qword ptr heap_vector_offset[r9]
+ else
+	mov	rbx,qword ptr heap_vector
+ endif
+
+update_up_list_1r:
+	mov	rdx,rax
+ if THREAD
+	add	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	add	rax,qword ptr neg_heap_p3
+ endif
+
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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
+
+ if THREAD
+	mov	rbp,qword ptr neg_heap_vector_plus_4_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_vector_plus_4
+ endif
+	add	rbp,r8
+	shl	rbp,6
+ if THREAD
+	add	rbp,qword ptr heap_p3_offset[r9]
+ else
+	add	rbp,qword ptr heap_p3
+ endif
+
+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
+
+ if THREAD
+	cmp	rax,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rax,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rax,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rax,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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_
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	jae	move_record_3_2
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+
+	push	rbp 
+
+	add	rax,rdx 
+
+ if THREAD
+	mov	rbx,qword ptr heap_vector_offset[r9]
+ else
+	mov	rbx,qword ptr heap_vector
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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 
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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_record_2bb
+
+move_record_2_ab:
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	cmp	rdx,rcx 
+	jb	move_record_2_1
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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_real_int_bool_or_char
+
+move_record_2bb:
+	mov	rax,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+move_real_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:
+ if THREAD
+	lea	rbx,__STRING__+2
+	cmp	rax,rbx
+ else
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+ endif
+	jbe	move_string_or_array
+	cmp	rax,offset CHAR+2
+	jbe	move_real_int_bool_or_char
+
+	test	rsi,rsi
+	jne	bsf_and_copy_nodes
+	jmp	find_non_zero_long
+
+move_string_or_array:
+	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
+
+ if THREAD
+	mov	rbp,qword ptr neg_heap_vector_plus_4_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_vector_plus_4
+ endif
+	add	rbp,r8
+
+	shl	rbp,6
+
+ if THREAD
+	add	rbp,qword ptr heap_p3_offset[r9]
+ else
+	add	rbp,qword ptr heap_p3
+ endif
+
+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 dINT+2
+	jle	cp_s_arg_lp3
+	cmp	rbx,offset BOOL+2
+	je	move_bool_array
+
+move_int32_or_real32_array:
+	add	rax,1
+	shr	rax,1
+	jmp	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
+
+ if THREAD
+	cmp	rbx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rbx,qword ptr end_heap_p3
+ endif
+	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
+
+ if THREAD
+	cmp	rax,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rax,qword ptr end_heap_p3
+ endif
+	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_
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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_
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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_
+
+ if THREAD
+	cmp	rdx,qword ptr end_heap_p3_offset[r9]
+ else
+	cmp	rdx,qword ptr end_heap_p3
+ endif
+	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:
+ if THREAD
+	lea	rbx,__Nil-8
+	cmp	rcx,rbx
+ else
+	lea	r9,__Nil-8
+	cmp	rcx,r9
+ endif
+	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/thread/acompact_rmark.asm b/thread/acompact_rmark.asm
new file mode 100644
index 0000000..90469b8
--- /dev/null
+++ b/thread/acompact_rmark.asm
@@ -0,0 +1,1131 @@
+
+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
+
+rmark_stack_nodes:
+ if THREAD
+	cmp	rsi,qword ptr end_vector_offset[r9]
+ else
+	cmp	rsi,qword ptr end_vector
+ endif
+	je	end_rmark_nodes
+
+rmark_more_stack_nodes:
+	mov	rcx,qword ptr [rsi]
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
+ if THREAD
+	cmp	rsi,qword ptr end_vector_offset[r9]
+ else
+	cmp	rsi,qword ptr end_vector
+ endif
+	jne	rmark_more_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:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
+
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_using_reversal
+
+rmark_node:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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_:
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_using_reversal_
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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 
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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:
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_using_reversal_
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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 
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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:
+ if THREAD
+	mov	rdx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rdx,qword ptr neg_heap_p3
+ endif
+	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	r12,rbx
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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]
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+	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,r12
+
+	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:
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+	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,r12
+
+	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]
+
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+	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:
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+	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,r12
+
+	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]
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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,r12
+	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
+
+rmark_next_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]
+
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jae	rmark_node
+
+	jmp	rmark_using_reversal
+
+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
+
+ if THREAD
+	mov	rbp,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_p3
+ endif
+	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]
+ if THREAD
+	mov	rbp,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_p3
+ endif
+	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]
+ if THREAD
+	mov	rbp,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_p3
+ endif
+
+	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:
+ if THREAD
+	cmp	rax,offset __ARRAY__+2
+ else
+	lea	r9,__ARRAY__+2
+	cmp	rax,r9
+ endif
+	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
+
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_array_using_reversal
+
+	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:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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:
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_array_using_reversal
+
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+
+rmark_lr_array:
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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]
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
diff --git a/thread/acompact_rmark_prefetch.asm b/thread/acompact_rmark_prefetch.asm
new file mode 100644
index 0000000..e921c8f
--- /dev/null
+++ b/thread/acompact_rmark_prefetch.asm
@@ -0,0 +1,1398 @@
+
+_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
+ if THREAD
+	cmp	rsi,qword ptr end_vector_offset[r9]
+ else
+	cmp	rsi,qword ptr end_vector
+ endif
+	je	end_rmarkp_nodes
+
+rmarkp_stack_nodes:
+	mov	rcx,qword ptr [rsi]
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
+ if THREAD
+	cmp	rsi,qword ptr end_vector_offset[r9]
+ else
+	cmp	rsi,qword ptr end_vector
+ endif
+	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:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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 
+
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_using_reversal
+
+rmarkp_node:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
+ if THREAD
+	cmp	rsi,qword ptr end_vector_offset[r9]
+ else
+	cmp	rsi,qword ptr end_vector
+ endif
+	je	rmarkp_last_item_in_queue
+
+	mov	rcx,[rsi]
+	mov	8[rsp],rsi
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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]
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+
+	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_:
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_using_reversal_
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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 
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	jnc	rmarkp_next_node
+
+	mov	rbx,rdx 
+	jmp	rmarkp_node_
+
+rmarkp_no_reverse_argument_pointer:
+	mov	rsi,rdx 
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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:
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_using_reversal_
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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 
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx 
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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:
+ if THREAD
+	mov	rdx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rdx,qword ptr neg_heap_p3
+ endif
+	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	r12,rbx 
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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]
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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,r12
+
+	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:
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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,r12
+
+	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]
+
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+	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:
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+
+	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,r12
+
+	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]
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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,r12
+	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]
+
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jae	rmarkp_node
+
+	jmp	rmark_using_reversal
+
+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
+
+ if THREAD
+	mov	rbp,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_p3
+ endif
+	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]
+ if THREAD
+	mov	rbp,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_p3
+ endif
+	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]
+ if THREAD
+	mov	rbp,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbp,qword ptr neg_heap_p3
+ endif
+
+	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:
+ if THREAD
+	cmp	rax,offset __ARRAY__+2
+ else
+	lea	r9,__ARRAY__+2
+	cmp	rax,r9
+ endif
+	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
+
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_array_using_reversal
+
+	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:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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:
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jb	rmark_array_using_reversal
+
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+
+rmarkp_lr_array:
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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]
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	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/thread/acompact_rmarkr.asm b/thread/acompact_rmarkr.asm
new file mode 100644
index 0000000..56a91a0
--- /dev/null
+++ b/thread/acompact_rmarkr.asm
@@ -0,0 +1,1095 @@
+
+rmark_using_reversal:
+	push	rsi
+	push	rsi
+	mov	rsi,1
+	jmp	rmarkr_node
+
+rmark_using_reversal_:
+	sub	rcx,8
+	push	rbx
+	push	rsi
+	cmp	rcx,rbx
+	ja	rmark_no_undo_reverse_1
+	mov	qword ptr [rsi],rcx
+	mov	qword ptr [rcx],rax
+rmark_no_undo_reverse_1:
+	mov	rsi,1
+	jmp	rmarkr_arguments
+
+rmark_array_using_reversal:
+	push	rbx
+	push	rsi
+	cmp	rcx,rbx
+	ja	rmark_no_undo_reverse_2
+	mov	qword ptr [rsi],rcx
+	mov	qword ptr [rcx],offset __ARRAY__+2
+rmark_no_undo_reverse_2:
+	mov	rsi,1
+	jmp	rmarkr_arguments
+
+rmarkr_hnf_2:
+	or	qword ptr [rcx],2
+	mov	rbp,qword ptr 8[rcx]
+	mov	qword ptr 8[rcx],rsi
+	lea	rsi,8[rcx]
+	mov	rcx,rbp
+
+rmarkr_node:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx
+
+ if THREAD
+	cmp	rax,qword ptr heap_size_64_65_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size_64_65
+ endif
+	jnc	rmarkr_next_node_after_static
+
+	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	rmarkr_next_node
+
+	or	rbp,rax
+	mov	dword ptr [rdi+rbx*4],ebp
+
+rmarkr_arguments:
+	mov	rax,qword ptr [rcx]
+	test	al,2
+	je	rmarkr_lazy_node
+
+	movzx	rbp,word ptr (-2)[rax]
+	test	rbp,rbp
+	je	rmarkr_hnf_0
+
+	add	rcx,8
+
+	cmp	rbp,256
+	jae	rmarkr_record
+
+	sub	rbp,2
+	je	rmarkr_hnf_2
+	jc	rmarkr_hnf_1
+
+rmarkr_hnf_3:
+	mov	rdx,qword ptr 8[rcx]
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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	rmarkr_shared_argument_part
+
+	or	dword ptr [rdi+rbx*4],eax
+
+rmarkr_no_shared_argument_part:
+	or	qword ptr [rcx],2
+	mov	qword ptr 8[rcx],rsi
+	add	rcx,8
+
+	or	qword ptr [rdx],1
+	lea	rdx,[rdx+rbp*8]
+
+	mov	rbp,qword ptr [rdx]
+	mov	qword ptr [rdx],rcx
+	mov	rsi,rdx
+	mov	rcx,rbp
+	jmp	rmarkr_node
+
+rmarkr_shared_argument_part:
+	cmp	rdx,rcx
+	ja	rmarkr_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	rmarkr_hnf_1
+
+rmarkr_record:
+	sub	rbp,258
+	je	rmarkr_record_2
+	jb	rmarkr_record_1
+
+rmarkr_record_3:
+	movzx	rbp,word ptr (-2+2)[rax]
+	sub	rbp,1
+	jb	rmarkr_record_3_bb
+	je	rmarkr_record_3_ab
+	dec	rbp
+	je	rmarkr_record_3_aab
+	jmp	rmarkr_hnf_3
+
+rmarkr_record_3_bb:
+	mov	rdx,qword ptr (16-8)[rcx]
+	sub	rcx,8
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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	rmarkr_next_node
+
+	add	eax,eax
+	jne	rmarkr_bit_in_same_word1
+	inc	rbp
+	mov	rax,1
+rmarkr_bit_in_same_word1:
+	test	eax,dword ptr [rdi+rbp*4]
+	je	rmarkr_not_yet_linked_bb
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rcx
+
+	add	rax,2*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,(16+2+1)[rcx]
+	mov	qword ptr 16[rcx],rbp 
+	mov	qword ptr [rdx],rax 
+	jmp	rmarkr_next_node
+
+rmarkr_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	rmarkr_next_node
+
+rmarkr_record_3_ab:
+	mov	rdx,qword ptr 8[rcx]
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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	rmarkr_hnf_1
+
+	add	eax,eax 
+	jne	rmarkr_bit_in_same_word2
+	inc	rbp
+	mov	rax,1
+rmarkr_bit_in_same_word2:
+	test	eax,dword ptr [rdi+rbp*4]
+	je	rmarkr_not_yet_linked_ab
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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	rmarkr_hnf_1
+
+rmarkr_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	rmarkr_hnf_1
+
+rmarkr_record_3_aab:
+	mov	rdx,qword ptr 8[rcx]
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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	rmarkr_shared_argument_part
+	or	dword ptr [rdi+rbp*4],eax
+
+	add	qword ptr [rcx],2
+	mov	qword ptr 8[rcx],rsi
+	add	rcx,8
+
+	mov	rsi,qword ptr [rdx]
+	mov	qword ptr [rdx],rcx 
+	mov	rcx,rsi
+	lea	rsi,1[rdx]
+	jmp	rmarkr_node
+
+rmarkr_record_2:
+	cmp	word ptr (-2+2)[rax],1
+	ja	rmarkr_hnf_2
+	je	rmarkr_hnf_1
+	sub	rcx,8
+	jmp	rmarkr_next_node
+
+rmarkr_record_1:
+	cmp	word ptr (-2+2)[rax],0
+	jne	rmarkr_hnf_1
+	sub	rcx,8
+	jmp	rmarkr_next_node
+
+rmarkr_lazy_node_1:
+	jne	rmarkr_selector_node_1
+
+rmarkr_hnf_1:
+	mov	rbp,qword ptr [rcx]
+	mov	qword ptr [rcx],rsi 
+
+	lea	rsi,2[rcx]
+	mov	rcx,rbp 
+	jmp	rmarkr_node
+
+rmarkr_indirection_node:
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	lea	rbx,(-8)[rcx+rbx]
+
+	mov	rax,rbx
+	and	rax,31*8
+	shr	rbx,8
+	mov	eax,dword ptr (bit_clear_table2)[rax]
+	and	dword ptr [rdi+rbx*4],eax
+
+	mov	rcx,qword ptr [rcx]
+	jmp	rmarkr_node
+
+rmarkr_selector_node_1:
+	add	rbp,3
+	je	rmarkr_indirection_node
+
+	mov	rdx,qword ptr [rcx]
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	add	rbx,rdx
+	shr	rbx,3
+
+	add	rbp,1
+	jle	rmarkr_record_selector_node_1
+
+	push	rax
+	mov	rax,rbx
+
+	shr	rbx,5
+	and	rax,31
+
+	mov	eax,dword ptr (bit_set_table)[rax*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rax 
+
+	pop	rax
+	jne	rmarkr_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmarkr_hnf_1
+
+	cmp	word ptr (-2)[rbx],2
+	jbe	rmarkr_small_tuple_or_record
+
+rmarkr_large_tuple_or_record:
+	mov	rbx,qword ptr 16[rdx]
+ if THREAD
+	add	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	add	rbx,qword ptr neg_heap_p3
+ endif
+	shr	rbx,3
+
+	push	rax
+	mov	rax,rbx
+
+	shr	rbx,5
+	and	rax,31
+
+	mov	eax,dword ptr (bit_set_table)[rax*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rax
+
+	pop	rax
+	jne	rmarkr_hnf_1
+
+ ifdef NEW_DESCRIPTORS
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+	lea	rbx,(-8)[rcx+rbx]
+
+	push	rcx
+
+	mov	eax,dword ptr (-8)[rax]
+
+	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
+
+	movzx	eax,word ptr 4[rax]
+	cmp	rax,16
+	jl	rmarkr_tuple_or_record_selector_node_2
+	mov	rdx,qword ptr 16[rdx]
+	je	rmarkr_tuple_selector_node_2
+	mov	rcx,qword ptr (-24)[rdx+rax]
+	pop	rdx
+	mov	qword ptr (-8)[rdx],offset __indirection
+	mov	qword ptr [rdx],rcx
+	jmp	rmarkr_node
+
+rmarkr_tuple_selector_node_2:
+	mov	rcx,qword ptr [rdx]
+	pop	rdx
+	mov	qword ptr (-8)[rdx],offset __indirection
+	mov	qword ptr [rdx],rcx
+	jmp	rmarkr_node
+ else
+rmarkr_small_tuple_or_record:
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+	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	rbp
+	mov	eax,4[rax]
+	call	near ptr rax
+	pop	rbp
+	pop	rdx
+
+	mov	qword ptr (-8)[rdx],offset __indirection
+	mov	qword ptr [rdx],rcx
+	jmp	rmarkr_node
+ endif
+
+rmarkr_record_selector_node_1:
+	je	rmarkr_strict_record_selector_node_1
+
+	push	rax
+	mov	rax,rbx
+
+	shr	rbx,5
+	and	rax,31
+
+	mov	eax,dword ptr (bit_set_table)[rax*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rax 
+
+	pop	rax
+	jne	rmarkr_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmarkr_hnf_1
+
+	cmp	word ptr (-2)[rbx],258
+ ifdef NEW_DESCRIPTORS
+	jbe	rmarkr_small_tuple_or_record
+
+	mov	rbx,qword ptr 16[rdx]
+  if THREAD
+	add	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	add	rbx,qword ptr neg_heap_p3
+  endif
+	shr	rbx,3
+
+	push	rax
+	mov	rax,rbx
+	shr	rbx,5
+	and	rax,31
+	mov	eax,dword ptr (bit_set_table)[rax*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rax
+	pop	rax
+	jne	rmarkr_hnf_1
+
+rmarkr_small_tuple_or_record:
+  if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+  else
+	mov	rbx,qword ptr neg_heap_p3
+  endif
+	lea	rbx,(-8)[rcx+rbx]
+
+	push	rcx
+
+	mov	eax,(-8)[rax]
+
+	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 
+
+	movzx	eax,word ptr 4[rax]
+	cmp	rax,16
+	jle	rmarkr_tuple_or_record_selector_node_2
+	mov	rdx,qword ptr 16[rdx]
+	sub	rax,24
+rmarkr_tuple_or_record_selector_node_2:
+	mov	rcx,qword ptr [rdx+rax]
+	pop	rdx
+	mov	qword ptr (-8)[rdx],offset __indirection
+	mov	qword ptr [rdx],rcx
+	jmp	rmarkr_node
+ else
+	jbe	rmarkr_small_tuple_or_record
+	jmp	rmarkr_large_tuple_or_record
+ endif
+
+rmarkr_strict_record_selector_node_1:
+	push	rax
+	mov	rax,rbx
+
+	shr	rbx,5
+	and	rax,31
+
+	mov	eax,dword ptr (bit_set_table)[rax*4]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rax 
+
+	pop	rax 
+	jne	rmarkr_hnf_1
+
+	mov	rbx,qword ptr [rdx]
+	test	bl,2
+	je	rmarkr_hnf_1
+
+	cmp	word ptr (-2)[rbx],258
+	jbe	rmarkr_select_from_small_record
+
+	mov	rbx,qword ptr 16[rdx]
+ if THREAD
+	add	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	add	rbx,qword ptr neg_heap_p3
+ endif
+
+	push	rax
+	mov	rax,rbx
+
+	shr	rbx,8
+	and	rax,31*8
+
+	mov	eax,dword ptr (bit_set_table2)[rax]
+	mov	ebx,dword ptr [rdi+rbx*4]
+	and	rbx,rax
+
+	pop	rax
+	jne	rmarkr_hnf_1
+
+rmarkr_select_from_small_record:
+	mov	eax,(-8)[rax]
+	sub	rcx,8
+
+ ifdef NEW_DESCRIPTORS
+	movzx	ebx,word ptr 4[rax]
+	cmp	rbx,16
+	jle	rmarkr_strict_record_selector_node_2
+	add	rbx,qword ptr 16[rdx]
+	mov	rbx,qword ptr (-24)[rbx]
+	jmp	rmarkr_strict_record_selector_node_3
+rmarkr_strict_record_selector_node_2:
+	mov	rbx,qword ptr [rdx+rbx]
+rmarkr_strict_record_selector_node_3:
+	mov	qword ptr 8[rcx],rbx
+
+	movzx	ebx,word ptr 6[rax]
+	test	rbx,rbx
+	je	rmarkr_strict_record_selector_node_5
+	cmp	rbx,16
+	jle	rmarkr_strict_record_selector_node_4
+	mov	rdx,qword ptr 16[rdx]
+	sub	rbx,24
+rmarkr_strict_record_selector_node_4:
+	mov	rbx,qword ptr [rdx+rbx]
+	mov	qword ptr 16[rcx],rbx
+rmarkr_strict_record_selector_node_5:
+
+	mov	rax,qword ptr (-8)[rbx]
+	mov	qword ptr [rcx],rax
+ else
+	mov	eax,4[rax]
+	call	near ptr rax
+ endif
+	jmp	rmarkr_next_node
+
+; a2,d1: free
+
+rmarkr_next_node:
+	test	rsi,3
+	jne	rmarkr_parent
+
+	mov	rbp,qword ptr (-8)[rsi]
+	mov	rbx,3
+	
+	and	rbx,rbp
+	sub	rsi,8
+
+	cmp	rbx,3
+	je	rmarkr_argument_part_cycle1
+
+	mov	rdx,qword ptr 8[rsi]
+	mov	qword ptr [rsi],rdx
+
+rmarkr_c_argument_part_cycle1:
+	cmp	rcx,rsi 
+	ja	rmarkr_no_reverse_1
+
+	mov	rdx,qword ptr [rcx]
+	lea	rax,(8+1)[rsi]
+	mov	qword ptr 8[rsi],rdx
+	mov	qword ptr [rcx],rax
+
+	or	rsi,rbx
+	mov	rcx,rbp
+	xor	rcx,rbx
+	jmp	rmarkr_node
+
+rmarkr_no_reverse_1:
+	mov	qword ptr 8[rsi],rcx
+	mov	rcx,rbp
+	or	rsi,rbx
+	xor	rcx,rbx
+	jmp	rmarkr_node
+
+rmarkr_lazy_node:
+	movsxd	rbp,dword ptr (-4)[rax]
+	test	rbp,rbp
+	je	rmarkr_next_node
+
+	add	rcx,8
+
+	sub	rbp,1
+	jle	rmarkr_lazy_node_1
+
+	cmp	rbp,255
+	jge	rmarkr_closure_with_unboxed_arguments
+
+rmarkr_closure_with_unboxed_arguments_:
+	or	qword ptr [rcx],2
+	lea	rcx,[rcx+rbp*8]
+
+	mov	rbp,qword ptr [rcx]
+	mov	qword ptr [rcx],rsi
+	mov	rsi,rcx
+	mov	rcx,rbp
+	jmp	rmarkr_node
+
+rmarkr_closure_with_unboxed_arguments:
+; (a_size+b_size)+(b_size<<8)
+;	add	rbp,1
+	mov	rax,rbp
+	and	rbp,255
+	shr	rax,8
+	sub	rbp,rax
+;	sub	rbp,1
+	jg	rmarkr_closure_with_unboxed_arguments_
+	je	rmarkr_hnf_1
+	sub	rcx,8
+	jmp	rmarkr_next_node
+
+rmarkr_hnf_0:
+	cmp	rax,offset dINT+2
+	je	rmarkr_int_3
+
+	cmp	rax,offset CHAR+2
+ 	je	rmarkr_char_3
+
+	jb	rmarkr_no_normal_hnf_0
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	add	rbx,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 
+
+ ifdef NEW_DESCRIPTORS
+	lea	rcx,((-8)-2)[rax]
+ else
+	lea	rcx,((-12)-2)[rax]
+ endif
+	jmp	rmarkr_next_node_after_static
+
+rmarkr_int_3:
+	mov	rbp,qword ptr 8[rcx]
+	cmp	rbp,33
+	jnc	rmarkr_next_node
+
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	add	rbx,rcx
+
+	mov	rcx,rbx
+	and	rcx,31*8
+	shr	rbx,8
+	mov	ecx,dword ptr (bit_clear_table2)[rcx]
+	shl	rbp,4
+	and	dword ptr [rdi+rbx*4],ecx 
+
+	lea	rcx,(small_integers)[rbp]
+	jmp	rmarkr_next_node_after_static
+
+rmarkr_char_3:
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+
+	movzx	rax,byte ptr 8[rcx]
+	add	rbx,rcx
+
+	mov	rbp,rbx
+	and	rbp,31*8
+	shr	rbx,8
+	mov	ebp,dword ptr (bit_clear_table2)[rbp]
+	and	dword ptr [rdi+rbx*4],ebp
+
+	shl	rax,4
+	lea	rcx,static_characters[rax]
+	jmp	rmarkr_next_node_after_static
+
+rmarkr_no_normal_hnf_0:
+ if THREAD
+	cmp	rax,offset __ARRAY__+2
+ else
+	lea	r9,__ARRAY__+2
+	cmp	rax,r9
+ endif
+	jne	rmarkr_next_node
+
+	mov	rax,qword ptr 16[rcx]
+	test	rax,rax
+	je	rmarkr_lazy_array
+
+	movzx	rbx,word ptr (-2+2)[rax]
+	test	rbx,rbx
+	je	rmarkr_b_array
+
+	movzx	rax,word ptr (-2)[rax]
+	test	rax,rax
+	je	rmarkr_b_array
+
+	sub	rax,256
+	cmp	rbx,rax
+	je	rmarkr_a_record_array
+
+rmarkr_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	rmarkr_lr_array
+
+rmarkr_b_array:
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	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	rmarkr_next_node
+
+rmarkr_a_record_array:
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+	cmp	rbx,2
+	jb	rmarkr_lr_array
+
+	imul	rax,rbx
+	jmp	rmarkr_lr_array
+
+rmarkr_lazy_array:
+	mov	rax,qword ptr 8[rcx]
+	add	rcx,16
+
+rmarkr_lr_array:
+ if THREAD
+	mov	rbx,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rbx,qword ptr neg_heap_p3
+ endif
+	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	rmarkr_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
+	add	rax,2
+	mov	rbx,qword ptr (-8)[rdx]
+	sub	rdx,8
+	mov	qword ptr [rcx],rbx
+	mov	qword ptr [rdx],rax
+
+	mov	rax,qword ptr (-8)[rcx]
+	sub	rcx,8
+	mov	qword ptr [rcx],rsi
+	mov	rsi,rcx
+	mov	rcx,rax
+	jmp	rmarkr_node
+
+rmarkr_array_length_0_1:
+	lea	rcx,-16[rcx]
+	jb	rmarkr_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	rmarkr_hnf_1
+
+; a2: free
+
+rmarkr_parent:
+	mov	rbx,rsi
+	and	rbx,3
+
+	and	rsi,-4
+	je	end_rmarkr
+
+	sub	rbx,1
+	je	rmarkr_argument_part_parent
+
+	mov	rbp,qword ptr [rsi]
+
+	cmp	rcx,rsi
+	ja	rmarkr_no_reverse_2
+
+	mov	rdx,rcx
+	lea	rax,1[rsi]
+	mov	rcx,qword ptr [rdx]
+	mov	qword ptr [rdx],rax
+
+rmarkr_no_reverse_2:
+	mov	qword ptr [rsi],rcx 
+	lea	rcx,(-8)[rsi]
+	mov	rsi,rbp
+	jmp	rmarkr_next_node
+
+rmarkr_argument_part_parent:
+	mov	rbp,qword ptr [rsi]
+
+	mov	rdx,rsi
+	mov	rsi,rcx
+	mov	rcx,rdx
+
+rmarkr_skip_upward_pointers:
+	mov	rax,rbp
+	and	rax,3
+	cmp	rax,3
+	jne	rmarkr_no_upward_pointer
+
+	lea	rdx,(-3)[rbp]
+	mov	rbp,qword ptr (-3)[rbp]
+	jmp	rmarkr_skip_upward_pointers
+
+rmarkr_no_upward_pointer:
+	cmp	rsi,rcx
+	ja	rmarkr_no_reverse_3
+
+	mov	rbx,rsi
+	mov	rsi,qword ptr [rsi]
+	lea	rax,1[rcx]
+	mov	qword ptr [rbx],rax
+
+rmarkr_no_reverse_3:
+	mov	qword ptr [rdx],rsi
+	lea	rsi,(-8)[rbp]
+
+	and	rsi,-4
+
+	mov	rdx,rsi
+	mov	rbx,3
+
+	mov	rbp,qword ptr [rsi]
+
+	and	rbx,rbp
+	mov	rax,qword ptr 8[rdx]
+
+	or	rsi,rbx
+	mov	qword ptr [rdx],rax
+
+	cmp	rcx,rdx
+	ja	rmarkr_no_reverse_4
+
+	mov	rax,qword ptr [rcx]
+	mov	qword ptr 8[rdx],rax
+	lea	rax,(8+2+1)[rdx]
+	mov	qword ptr [rcx],rax
+	mov	rcx,rbp
+	and	rcx,-4
+	jmp	rmarkr_node
+
+rmarkr_no_reverse_4:
+	mov	qword ptr 8[rdx],rcx
+	mov	rcx,rbp
+	and	rcx,-4
+	jmp	rmarkr_node
+
+rmarkr_argument_part_cycle1:
+	mov	rax,qword ptr 8[rsi]
+	push	rdx
+
+rmarkr_skip_pointer_list1:
+	mov	rdx,rbp
+	and	rdx,-4
+	mov	rbp,qword ptr [rdx]
+	mov	rbx,3
+	and	rbx,rbp
+	cmp	rbx,3
+	je	rmarkr_skip_pointer_list1
+
+	mov	qword ptr [rdx],rax
+	pop	rdx
+	jmp	rmarkr_c_argument_part_cycle1
+
+rmarkr_next_node_after_static:
+	test	rsi,3
+	jne	rmarkr_parent_after_static
+
+	mov	rbp,qword ptr (-8)[rsi]
+	mov	rbx,3
+	
+	and	rbx,rbp
+	sub	rsi,8
+
+	cmp	rbx,3
+	je	rmarkr_argument_part_cycle2
+	
+	mov	rax,qword ptr 8[rsi]
+	mov	qword ptr [rsi],rax
+
+rmarkr_c_argument_part_cycle2:
+	mov	qword ptr 8[rsi],rcx
+	mov	rcx,rbp
+	or	rsi,rbx
+	xor	rcx,rbx
+	jmp	rmarkr_node
+
+rmarkr_parent_after_static:
+	mov	rbx,rsi
+	and	rbx,3
+
+	and	rsi,-4
+	je	end_rmarkr_after_static
+
+	sub	rbx,1
+	je	rmarkr_argument_part_parent_after_static
+
+	mov	rbp,qword ptr [rsi]
+	mov	qword ptr [rsi],rcx
+	lea	rcx,(-8)[rsi]
+	mov	rsi,rbp
+	jmp	rmarkr_next_node
+	
+rmarkr_argument_part_parent_after_static:
+	mov	rbp,qword ptr [rsi]
+
+	mov	rdx,rsi
+	mov	rsi,rcx
+	mov	rcx,rdx
+
+;	movl	rbp,qword ptr [rdx]
+rmarkr_skip_upward_pointers_2:
+	mov	rax,rbp
+	and	rax,3
+	cmp	rax,3
+	jne	rmarkr_no_reverse_3
+
+	lea	rdx,(-3)[rbp]
+	mov	rbp,qword ptr (-3)[rbp]
+	jmp	rmarkr_skip_upward_pointers_2
+
+rmarkr_argument_part_cycle2:
+	mov	rax,qword ptr 8[rsi]
+	push	rdx
+
+rmarkr_skip_pointer_list2:
+	mov	rdx,rbp 
+	and	rdx,-4
+	mov	rbp,qword ptr [rdx]
+	mov	rbx,3
+	and	rbx,rbp
+	cmp	rbx,3
+	je	rmarkr_skip_pointer_list2
+
+	mov	qword ptr [rdx],rax
+	pop	rdx
+	jmp	rmarkr_c_argument_part_cycle2
+
+end_rmarkr_after_static:
+	mov	rsi,qword ptr [rsp]
+	add	rsp,16
+	mov	qword ptr [rsi],rcx
+	jmp	rmarkr_next_stack_node
+
+end_rmarkr:
+	pop	rsi
+	pop	rbx
+
+	cmp	rcx,rbx
+	ja	rmarkr_no_reverse_5
+
+	mov	rdx,rcx
+	lea	rax,1[rsi]
+	mov	rcx,qword ptr [rcx]
+	mov	qword ptr [rdx],rax
+
+rmarkr_no_reverse_5:
+	mov	qword ptr [rsi],rcx
+
+rmarkr_next_stack_node:
+ if THREAD
+	cmp	rsp,qword ptr end_stack_offset[r9]
+ else
+	cmp	rsp,qword ptr end_stack
+ endif
+	jae	rmarkr_end
+
+	mov	rcx,qword ptr [rsp]
+	mov	rsi,qword ptr 8[rsp]
+	add	rsp,16
+
+	cmp	rcx,1
+	ja	rmark_using_reversal
+
+	test	qword ptr flags,4096
+	je	rmark_next_node_
+	jmp	rmarkp_next_node_
+
+rmarkr_end:
+	test	qword ptr flags,4096
+	je	rmark_next_node
+	jmp	rmarkp_next_node
diff --git a/thread/acopy.asm b/thread/acopy.asm
new file mode 100644
index 0000000..b0213d0
--- /dev/null
+++ b/thread/acopy.asm
@@ -0,0 +1,1394 @@
+
+COPY_RECORDS_WITHOUT_POINTERS_TO_END_OF_HEAP = 1
+
+	push	rsi
+
+ if THREAD
+	mov	rdi,heap_p2_offset[r9]
+
+	mov	rax,heap_size_257_offset[r9]
+ else
+	mov	rdi,heap_p2
+
+	mov	rax,heap_size_257
+ endif
+	shl	rax,7
+ if THREAD
+	mov	semi_space_size_offset[r9],rax 
+ else
+	mov	semi_space_size,rax 
+ endif
+	lea	rsi,[rdi+rax]
+
+ if THREAD
+	mov	qword ptr (heap2_begin_and_end_offset+8)[r9],rsi 
+ else
+	mov	qword ptr (heap2_begin_and_end+8),rsi 
+ endif
+
+	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]
+ if THREAD
+	mov	rbp,stack_p_offset[r9]
+ else
+	mov	rbp,stack_p
+ endif
+	sub	rbx,rbp 
+	shr	rbx,3
+
+	sub	rbx,1
+	jb	end_copy0
+	call	copy_lp2
+end_copy0:
+ if THREAD
+	mov	rbp,heap_p2_offset[r9]
+ else
+	mov	rbp,heap_p2
+ endif
+
+	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 __STRING__+2
+	jbe	copy_string_or_array_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
+
+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_record_node2_bb
+
+	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
+
+copy_record_node2_bb:
+	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
+ 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
+ if THREAD
+	sub	rax,qword ptr heap_p1_offset[r9]
+ else
+	sub	rax,qword ptr heap_p1
+ endif
+
+	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]
+
+ if THREAD
+	add	rdx,qword ptr heap_copied_vector_offset[r9]
+ else
+	add	rdx,qword ptr heap_copied_vector
+ endif
+	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
+ if THREAD
+	sub	rax,qword ptr heap_p1_offset[r9]
+ else
+	sub	rax,qword ptr heap_p1
+ endif
+
+	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]
+
+ if THREAD
+	add	rdx,qword ptr heap_copied_vector_offset[r9]
+ else
+	add	rdx,qword ptr heap_copied_vector
+ endif
+	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
+ if THREAD
+	sub	rax,qword ptr heap_p1_offset[r9]
+ else
+	sub	rax,qword ptr heap_p1
+ endif
+
+	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]
+
+ if THREAD
+	add	rdx,qword ptr heap_copied_vector_offset[r9]
+ else
+	add	rdx,qword ptr heap_copied_vector
+ endif
+	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
+
+	sub	rdx,1
+	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]
+ if THREAD
+	sub	d4,qword ptr heap_p1_offset[r9]
+ else
+	sub	d4,qword ptr heap_p1
+ endif
+
+	mov	d5,d4 
+	and	d4,31*16
+
+	shr	d5,7
+	
+	shr	d4,2
+	and	d5,-4
+
+ if THREAD
+	add	d5,qword ptr heap_copied_vector_offset[r9]
+ else
+	add	d5,qword ptr heap_copied_vector
+ endif
+
+	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]
+ if THREAD
+	sub	d4,qword ptr heap_p1_offset[r9]
+ else
+	sub	d4,qword ptr heap_p1
+ endif
+
+	mov	d5,d4 
+	and	d4,31*16
+
+	shr	d5,7
+
+	shr	d4,2
+	and	d5,-4
+
+ if THREAD
+	add	d5,qword ptr heap_copied_vector_offset[r9]
+ else
+	add	d5,qword ptr heap_copied_vector
+ endif
+
+	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
+
+ if THREAD
+	sub	rdx,heap_p1_offset[r9]
+	cmp	rdx,semi_space_size_offset[r9]
+ else
+	sub	rdx,heap_p1
+	cmp	rdx,semi_space_size
+ endif
+	jae	copy_string_or_array_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_array_2:
+ if THREAD
+	sub	rdx,heap_p1_offset[r9]
+	cmp	rdx,semi_space_size_offset[r9]
+ else
+	sub	rdx,heap_p1
+	cmp	rdx,semi_space_size
+ endif
+	jae	copy_string_or_array_constant
+
+	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 dINT+2
+	jle	copy_int_or_real_array_2
+
+	cmp	rax,offset BOOL+2
+	je	copy_bool_array_2
+
+copy_int32_or_real32_array_2:
+	add	rbx,1
+	shr	rbx,1
+
+copy_int_or_real_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:
+	add	rbx,7
+	shr	rbx,3
+	jmp	copy_int_or_real_array_2
+
+copy_string_or_array_constant:
+	mov	qword ptr [rbp],rcx
+	add	rbp,8
+
+	sub	rbx,1
+	jae	copy_lp2
+	ret
+
+end_copy1:
+ if THREAD
+	mov	heap_end_after_gc_offset[r9],rsi
+ else
+	mov	heap_end_after_gc,rsi
+ endif
+
+	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:
+ if THREAD
+	lea	rax,__Nil-8
+	cmp	rbp,rax
+ else
+	lea	r9,__Nil-8
+	cmp	rbp,r9
+ endif
+	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/thread/afileIO3.asm b/thread/afileIO3.asm
new file mode 100644
index 0000000..62efe53
--- /dev/null
+++ b/thread/afileIO3.asm
@@ -0,0 +1,1021 @@
+
+; nonvolatile registers: rbx rsi rdi rbp r12 r13 r14 r15
+; volatile    registers: rax rcx rdx r8 r9 r10 r11
+
+	THREAD equ 1
+
+_TEXT	segment para 'CODE'
+_TEXT	ends
+_DATA	segment para 'DATA'
+_DATA	ends
+
+	_DATA segment
+	align	8
+
+freadstring_error:
+	db	"Error in freadsubstring parameters."
+	db	10,0
+	db	0,0,0
+fwritestring_error:
+	db	"Error in fwritesubstring parameters."
+	db	10,0
+	db	0,0
+
+	extrn	clean_exception_handler:near
+	public	clean_unwind_info
+clean_unwind_info:
+	DD	000000009H
+	DD	imagerel(clean_exception_handler)
+
+_DATA	ends
+
+	_TEXT segment
+
+	public	stdioF
+	public	stderrF
+	public	openF
+	public	closeF
+	public	reopenF
+	public	readFC
+	public	readFI
+	public	readFR
+	public	readFS
+	public	readFString
+	public	readLineF
+	public	writeFC
+	public	writeFI
+	public	writeFR
+	public	writeFS
+	public	writeFString
+	public	endF
+	public	errorF
+	public	positionF
+	public	seekF
+	public	shareF
+	public	flushF
+	public	openSF
+	public	readSFC
+	public	readSFI
+	public	readSFR
+	public	readSFS
+	public	readLineSF
+	public	endSF
+	public	positionSF
+	public	seekSF
+
+; imports
+
+	extrn	open_file:near
+	extrn	open_stdio:near
+	extrn	open_stderr:near
+	extrn	re_open_file:near
+	extrn	close_file:near
+	extrn	file_read_char:near
+	extrn	file_read_int:near
+	extrn	file_read_real:near
+	extrn	file_read_characters:near
+	extrn	file_read_line:near
+	extrn	file_write_char:near
+	extrn	file_write_int:near
+	extrn	file_write_real:near
+	extrn	file_write_characters:near
+	extrn	file_end:near
+	extrn	file_error:near
+	extrn	file_position:near
+	extrn	file_seek:near
+	extrn	file_share:near
+	extrn	flush_file_buffer:near
+	extrn	open_s_file:near
+	extrn	file_read_s_char:near
+	extrn	file_read_s_int:near
+	extrn	file_read_s_real:near
+	extrn	file_read_s_string:near
+	extrn	file_read_s_line:near
+	extrn	file_s_end:near
+	extrn	file_s_position:near
+	extrn	file_s_seek:near
+
+	extrn	collect_0:near
+	extrn	collect_1:near
+
+	extrn	print_error:near
+
+	extrn	__STRING__:near
+
+stdioF:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,32
+	call	open_stdio
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	rbx,rax
+	mov	rax,-1
+	ret
+
+stderrF:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,32
+	call	open_stderr
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	rbx,rax
+	mov	rax,-1
+	ret
+
+openF:
+	mov	rbp,rsp
+	and	rsp,-16
+
+	mov	rdx,rax
+	add	rcx,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,32
+	call	open_file
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	xor	r10,r10
+	test	rax,rax
+	setns	r10b
+	mov	rbx,rax
+	mov	rax,-1
+	ret
+
+closeF:	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	close_file
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	ret
+
+reopenF:
+	mov	rdx,rax
+	mov	rcx,r10
+	mov	rbx,r10
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	re_open_file
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r10d,eax
+	mov	rax,-1
+	ret
+
+readFC:
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_read_char
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	cmp	eax,-1
+	je	readFC_eof
+
+	mov	r10,rax
+	mov	rax,-1
+	mov	r11,1
+	ret
+
+readFC_eof:
+	xor	r10,r10
+	mov	rax,-1
+	xor	r11,r11
+	ret
+
+readFI:
+	mov	rbp,rsp
+	sub	rsp,40+8
+ if THREAD
+	mov	r14,r9
+ endif
+	and	rsp,-16
+
+	lea	rdx,32[rsp]
+	mov	rcx,rbx
+	call	file_read_int
+
+	mov	r10,32[rsp]
+
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r11,rax
+	mov	rax,-1
+	ret
+
+readFR:
+	mov	rbp,rsp
+	sub	rsp,40+8
+ if THREAD
+	mov	r14,r9
+ endif
+	and	rsp,-16
+
+	lea	rdx,32[rsp]
+	mov	rcx,rbx
+	call	file_read_real
+
+	movlpd	xmm0,qword ptr 32[rsp]
+	
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r10,rax
+	mov	rax,-1	
+	ret
+
+readFString:
+	mov	rbp,8[rcx]
+	cmp	r11,rbp
+	jae	readFString_error
+
+	sub	rbp,r11
+	cmp	r10,rbp
+	ja	readFString_error
+
+	push	rcx
+
+	push	r11
+
+	lea	r8,16[rcx+r11]
+	mov	rdx,rsp
+	mov	rcx,rbx
+	
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_read_characters
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	add	rsp,8
+	pop	rcx
+	
+	mov	r10,rax
+	mov	rax,-1
+	ret
+
+readFString_error:
+	mov	rbp,offset freadstring_error
+	jmp	print_error
+
+readFS:	lea	rbp,16+7[rax]
+	shr	rbp,3
+	sub	r15,rbp
+	jb	readFS_gc
+readFS_r_gc:
+	add	r15,rbp
+
+	lea	r8,__STRING__+2
+	mov	qword ptr [rdi],r8
+;	mov	qword ptr [rdi],offset __STRING__+2
+
+	lea	r8,16[rdi]
+	mov	8[rdi],rax
+	lea	rdx,8[rdi]
+	mov	rcx,r10
+
+	mov	rbx,r10
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_read_characters
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+
+readFS_end:
+	add	rax,16+7
+	mov	rcx,rdi
+	shr	rax,3
+	sub	r15,rax
+	lea	rdi,[rdi+rax*8]
+	mov	rax,-1
+	ret
+
+readFS_gc:	push	rbp
+	call	collect_0
+	pop	rbp
+	jmp	readFS_r_gc
+
+readLineF:
+	cmp	r15,32+2
+	jb	readLineF_gc
+
+readLineF_r_gc:
+	lea	r8,__STRING__+2
+	mov	qword ptr [rdi],r8
+;	mov	qword ptr [rdi],offset __STRING__+2
+
+	lea	r8,16[rdi]
+	lea	rdx,-16[r15*8]
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_read_line
+	mov	rsp,rbp
+
+	mov	8[rdi],rax
+ if THREAD
+	mov	r9,r14
+ endif
+	test	rax,rax
+	jns	readFS_end
+
+	lea	rax,-16[r15*8]
+	mov	r12,rdi
+	mov	8[rdi],rax
+	add	rdi,16
+
+readLineF_lp:
+	add	rdi,rax
+
+	mov	r13,8[r12]
+	mov	rcx,r12
+	shr	r13,3
+	xor	r15,r15
+	add	r13,2+32
+	sub	r15,r13
+	
+	call	collect_1
+
+	add	r15,r13
+	mov	rax,8[rcx]
+	lea	rdx,16[rcx]
+	lea	rcx,7[rax]
+	shr	rcx,3
+	sub	r15,2
+	sub	r15,rcx
+
+	mov	r12,rdi
+
+	lea	rbp,__STRING__+2
+	mov	qword ptr [rdi],rbp
+;	mov	qword ptr [rdi],offset __STRING__+2
+
+	mov	8[rdi],rax
+	add	rdi,16
+	jmp	st_copy_string1
+
+copy_st_lp1:
+	mov	rbp,[rdx]
+	add	rdx,8
+	mov	[rdi],rbp
+	add	rdi,8
+st_copy_string1:
+	sub	rcx,1
+	jnc	copy_st_lp1
+
+	mov	r8,rdi
+	lea	rdx,[r15*8]
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_read_line
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	test	rax,rax
+	js	readLineF_again
+
+	add	8[r12],rax
+	add	rax,7
+
+	mov	rcx,r12
+
+	shr	rax,3
+	sub	r15,rax
+	lea	rdi,[rdi+rax*8]
+
+	mov	rax,-1
+	ret
+
+readLineF_gc:
+	sub	r15,32+2
+	call	collect_0
+	add	r15,32+2
+	jmp	readLineF_r_gc
+
+readLineF_again:
+	mov	rcx,8[r12]
+	lea	rax,[r15*8]
+	add	rcx,rax
+	mov	8[r12],rcx
+	jmp	readLineF_lp
+
+writeFC:
+	mov	rcx,r10
+	mov	rdx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_write_char
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	rax,-1
+	ret
+
+writeFI:
+	mov	rcx,r10
+	mov	rdx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_write_int
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	rax,-1
+	ret
+
+writeFR:
+	mov	rdx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+ 	call	file_write_real
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	rax,-1
+	ret
+
+writeFS:
+	mov	r8,rbx
+	mov	rdx,8[rcx]
+	add	rcx,16
+	
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_write_characters
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	rax,-1
+	ret
+
+writeFString:
+	mov	rbp,8[rcx]
+	cmp	r11,rbp
+	jae	writeFString_error
+
+	sub	rbp,r11
+	cmp	r10,rbp
+	ja	writeFString_error
+
+	mov	r8,rbx
+	mov	rdx,r10
+	lea	rcx,16[rcx+r11]
+	
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_write_characters
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	rax,-1
+
+	ret
+
+writeFString_error:
+	mov	rbp,offset fwritestring_error
+	jmp	print_error
+
+endF:
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_end
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r10,rax
+	mov	rax,-1
+	ret
+
+errorF:
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_error
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r10,rax
+	mov	rax,-1
+	ret
+
+positionF:
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_position
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r10,rax
+	mov	rax,-1
+	ret
+
+seekF:
+	mov	r8,rax
+	mov	rdx,rbx
+	mov	rcx,r11
+	
+	mov	rbx,r11
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_seek
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r10,rax
+	mov	rax,-1
+	ret
+
+shareF:
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_share
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	rax,-1
+	ret
+
+flushF:
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	flush_file_buffer
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r10,rax
+	mov	rax,-1
+	ret
+
+openSF:	mov	rdx,rax
+	add	rcx,8
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	open_s_file
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	xor	r10,r10
+	test	rax,rax
+	setns	r10b
+
+	mov	rbx,rax
+	xor	rax,rax
+	ret
+
+readSFC:
+	push	rax
+
+	mov	rdx,rsp
+	mov	rax,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_read_s_char
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	cmp	rax,-1
+	je	readSFC_eof
+
+	pop	rax
+	mov	r10,rax
+	mov	r11,1
+	ret
+
+readSFC_eof:
+	pop	rax
+	xor	r10,r10
+	xor	r11,r11
+	ret
+
+readSFI:
+	push	rax
+	mov	r8,rsp
+	sub	rsp,8
+	mov	rdx,rsp
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_read_s_int
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	pop	r10
+	mov	r11,rax
+	pop	rax
+	
+	ret
+
+readSFR:
+	push	rax
+	mov	r8,rsp
+	sub	rsp,8
+	mov	rdx,rsp
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	sub	rsp,40+8
+ if THREAD
+	mov	r14,r9
+ endif
+	and	rsp,-16
+	call	file_read_s_real
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	movlpd	xmm0,qword ptr [rsp]
+	mov	r10,rax
+	add	rsp,8
+	pop	rax
+
+	ret
+
+readSFS:
+	lea	rbp,16+7[rax]
+	shr	rbp,3
+	sub	r15,rbp
+	jb	readSFS_gc
+readSFS_r_gc:
+	add	r15,rbp
+
+ if THREAD
+	lea	rbp,__STRING__+2
+	mov	qword ptr [rdi],rbp
+ else
+	lea	r9,__STRING__+2
+	mov	qword ptr [rdi],r9
+ endif
+;	mov	qword ptr [rdi],offset __STRING__+2
+
+	push	rbx
+
+ if THREAD
+	mov	r14,r9
+ endif
+	mov	r9,rsp
+	lea	r8,8[rdi]
+	mov	rdx,rax
+	mov	rcx,r10
+	
+	mov	rbx,r10
+	
+	mov	rbp,rsp
+	or	rsp,8
+	sub	rsp,40
+	call	file_read_s_string
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+
+readSFS_end:
+	add	rax,16+7
+	mov	rcx,rdi
+	shr	rax,3
+	sub	r15,rax
+	lea	rdi,[rdi+rax*8]
+	pop	rax
+	ret
+
+readSFS_gc:	push	rbp
+	call	collect_0
+	pop	rbp
+	jmp	readSFS_r_gc
+
+readLineSF:
+	cmp	r15,32+2
+	ja	readLineSF_gc
+
+readLineSF_r_gc:
+	push	rax
+
+ if THREAD
+	lea	rbp,__STRING__+2
+	mov	qword ptr [rdi],rbp
+ else
+	lea	r9,__STRING__+2
+	mov	qword ptr [rdi],r9
+ endif
+;	mov	qword ptr [rdi],offset __STRING__+2
+
+ if THREAD
+	mov	r14,r9
+ endif
+	mov	r9,rsp
+	lea	r8,16[rdi]
+	lea	rdx,-16[r15*8]
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+	sub	rsp,40
+	call	file_read_s_line
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	8[rdi],rax
+
+	test	rax,rax
+	jns	readSFS_end
+
+	lea	rax,-16[r15*8]
+	mov	r12,rdi
+	mov	8[rdi],rax
+	add	rdi,16
+
+readLineSF_lp:
+	add	rdi,rax
+
+	mov	r13,8[r12]
+	mov	rcx,r12
+	shr	r13,3
+	xor	r15,r15
+	add	r13,2+32
+	sub	r15,r13
+
+	call	collect_1
+
+	add	r15,r13
+	mov	rax,8[rcx]
+	lea	rdx,16[rcx]
+	lea	rcx,7[rax]
+	shr	rcx,3
+	sub	r15,2
+	sub	r15,rcx
+
+	mov	r12,rdi
+
+	lea	rbp,__STRING__+2
+	mov	qword ptr [rdi],rbp
+;	mov	qword ptr [rdi],offset __STRING__+2
+
+	mov	8[rdi],rax
+	add	rdi,16
+	jmp	st_copy_string2
+
+copy_st_lp2:
+	mov	rbp,[rdx]
+	add	rdx,8
+	mov	[rdi],rbp
+	add	rdi,8
+st_copy_string2:
+	sub	rcx,1
+	jnc	copy_st_lp2
+
+ if THREAD
+	mov	r14,r9
+ endif
+	mov	r9,rsp
+	mov	r8,rdi
+	lea	rdx,[r15*8]
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+	sub	rsp,40
+	call	file_read_s_line
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	test	rax,rax
+	js	readLineSF_again
+
+	add	8[r12],rax
+	add	rax,7
+
+	mov	rcx,r12
+
+	shr	rax,3
+	sub	r15,rax
+	lea	rdi,[rdi+rax*8]
+
+	pop	rax
+	ret
+
+readLineSF_gc:
+	sub	r15,32+2
+	call	collect_0
+	add	r15,32+2
+	jmp	readLineSF_r_gc
+
+readLineSF_again:
+	mov	rcx,8[r12]
+	lea	rax,[r15*8]
+	add	rcx,rax
+	mov	8[r12],rcx
+	jmp	readLineSF_lp
+
+endSF:
+	mov	rdx,rax
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_s_end
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	ret
+
+positionSF:
+	mov	rdx,rax
+	mov	rcx,rbx
+
+	mov	rbp,rsp
+	or	rsp,8
+ if THREAD
+	mov	r14,r9
+ endif
+	sub	rsp,40
+	call	file_s_position
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	ret
+
+seekSF:
+	push	rbx
+ if THREAD
+	mov	r14,r9
+ endif
+	mov	r9,rsp
+	mov	r8,rax
+	mov	rdx,r10
+	mov	rcx,r11
+
+	mov	rbx,r11
+
+	mov	rbp,rsp
+	or	rsp,8
+	sub	rsp,40
+	call	file_s_seek
+	mov	rsp,rbp
+ if THREAD
+	mov	r9,r14
+ endif
+	mov	r10,rax	
+	pop	rax
+
+	ret
+
+_TEXT	ends
+
+
+	end
diff --git a/thread/amark.asm b/thread/amark.asm
new file mode 100644
index 0000000..76bf18f
--- /dev/null
+++ b/thread/amark.asm
@@ -0,0 +1,2156 @@
+
+ if THREAD
+	mov	rax,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rax,qword ptr heap_size_65
+ endif
+	xor	rbx,rbx
+
+ if THREAD
+	mov	qword ptr n_marked_words_offset[r9],rbx
+ else
+	mov	qword ptr n_marked_words,rbx
+ endif
+	shl	rax,6
+
+ if THREAD
+	mov	qword ptr lazy_array_list_offset[r9],rbx
+ else
+	mov	qword ptr lazy_array_list,rbx
+ endif
+	mov	r11,rax
+	
+	lea	rsi,(-4000)[rsp]
+
+	mov	rax,qword ptr caf_list
+
+ if THREAD
+	mov	qword ptr end_stack_offset[r9],rsi 
+
+	mov	r10,neg_heap_p3_offset[r9]
+	mov	r13,qword ptr end_stack_offset[r9]
+ else
+	mov	qword ptr end_stack,rsi 
+
+	mov	r10,neg_heap_p3
+	mov	r13,qword ptr end_stack
+ endif
+	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:
+ if THREAD
+	mov	rsi,qword ptr stack_top_offset[r9]
+	mov	rbp,qword ptr stack_p_offset[r9]
+ else
+	mov	rsi,qword ptr stack_top
+	mov	rbp,qword ptr stack_p
+ endif
+
+	mov	r12,rsi 
+	call	_mark_stack_nodes
+
+continue_mark_after_pmark:
+ if THREAD
+	mov	qword ptr n_marked_words_offset[r9],r14
+
+	mov	rcx,qword ptr lazy_array_list_offset[r9]
+ else
+	mov	qword ptr n_marked_words,r14
+
+	mov	rcx,qword ptr lazy_array_list
+ endif
+
+	test	rcx,rcx 
+	je	end_restore_arrays
+
+restore_arrays:
+	mov	rbx ,qword ptr [rcx]
+ if THREAD
+	lea	rbp,__ARRAY__+2
+	mov	qword ptr [rcx],rbp
+ else
+	lea	r9,__ARRAY__+2
+	mov	qword ptr [rcx],r9
+ endif
+
+	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:
+ if THREAD
+	mov	rdi,qword ptr heap_vector_offset[r9]
+ else
+	mov	rdi,qword ptr heap_vector
+ endif
+	lea	rcx,finalizer_list
+	lea	rdx,free_finalizer_list
+
+	mov	rbp,qword ptr [rcx]
+determine_free_finalizers_after_mark:
+ if THREAD
+	lea	rax,__Nil-8
+	cmp	rbp,rax
+ else
+	lea	r9,__Nil-8
+	cmp	rbp,r9
+ endif
+	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
+
+ if THREAD
+	mov	rax,qword ptr bit_vector_size_offset[r9]
+
+	mov	rdi,qword ptr n_allocated_words_offset[r9]
+	add	rdi,qword ptr n_marked_words_offset[r9]
+ else
+	mov	rax,qword ptr bit_vector_size
+
+	mov	rdi,qword ptr n_allocated_words
+	add	rdi,qword ptr n_marked_words
+ endif
+	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
+
+ if THREAD
+	mov	rbx,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rbx,qword ptr heap_size_65
+ endif
+	shl	rbx,6
+
+not_largest_heap:
+	cmp	rbx,rsi 
+	jbe	no_larger_heap
+
+ if THREAD
+	mov	rsi,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rsi,qword ptr heap_size_65
+ endif
+	shl	rsi,6
+	cmp	rbx,rsi
+	jbe	not_larger_than_heap
+	mov	rbx,rsi 
+not_larger_than_heap:
+	mov	rax,rbx 
+	shr	rax,3
+ if THREAD
+	mov	qword ptr bit_vector_size_offset[r9],rax
+ else
+	mov	qword ptr bit_vector_size,rax
+ endif
+no_larger_heap:
+
+	mov	rbp,rax
+
+ if THREAD
+	mov	rdi,qword ptr heap_vector_offset[r9]
+ else
+	mov	rdi,qword ptr heap_vector
+ endif
+	shr	rbp,5
+
+	test	al,31
+	je	no_extra_word
+
+	mov	dword ptr [rdi+rbp*4],0
+
+no_extra_word:
+ if THREAD
+	sub	rax,qword ptr n_marked_words_offset[r9]
+ else
+	sub	rax,qword ptr n_marked_words
+ endif
+	shl	rax,3
+ if THREAD
+	mov	qword ptr n_last_heap_free_bytes_offset[r9],rax
+
+	mov	rax,qword ptr n_marked_words_offset[r9]
+ else
+	mov	qword ptr n_last_heap_free_bytes,rax 
+
+	mov	rax,qword ptr n_marked_words
+ endif
+	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
+ if THREAD
+	mov	rbx,r9
+ endif
+ 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
+ if THREAD
+	mov	r9,rbx
+ endif
+
+ ifdef LINUX
+  if THREAD
+	mov	rdi,qword ptr n_marked_words_offset[r9]
+  else
+	mov	rdi,qword ptr n_marked_words
+  endif
+	shl	rdi,3
+ else
+  if THREAD
+	mov	rcx,qword ptr n_marked_words_offset[r9]
+  else
+	mov	rcx,qword ptr n_marked_words
+  endif
+	shl	rcx,3
+ endif
+	call	ew_print_int
+ if THREAD
+	mov	r9,rbx
+ endif
+
+ 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
+ if THREAD
+	mov	r9,rbx
+ endif
+
+_no_heap_use_message2:
+	call	call_finalizers
+
+ if THREAD
+	mov	rsi,qword ptr n_allocated_words_offset[r9]
+ else
+	mov	rsi,qword ptr n_allocated_words
+ endif
+	xor	rbx,rbx 
+
+	mov	rcx,rdi
+ if THREAD
+	mov	qword ptr n_free_words_after_mark_offset[r9],rbx
+ else
+	mov	qword ptr n_free_words_after_mark,rbx 
+ endif
+
+_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
+ if THREAD
+	add	qword ptr n_free_words_after_mark_offset[r9],rax
+ else
+	add	qword ptr n_free_words_after_mark,rax 
+ endif
+	mov	dword ptr (-4)[rcx],ebx 
+
+	cmp	rax,rsi
+	jb	_scan_bits
+
+_found_free_memory:
+ if THREAD
+	mov	qword ptr bit_counter_offset[r9],rbp
+	mov	qword ptr bit_vector_p_offset[r9],rcx
+ else
+	mov	qword ptr bit_counter,rbp
+	mov	qword ptr bit_vector_p,rcx
+ endif
+
+	lea	rbx,(-4)[rdx]
+	sub	rbx,rdi 
+	shl	rbx,6
+ if THREAD
+	mov	rdi,qword ptr heap_p3_offset[r9]
+ else
+	mov	rdi,qword ptr heap_p3
+ endif
+	add	rdi,rbx 
+
+	mov	r15,rax
+	lea	rbx,[rdi+rax*8]
+
+	sub	r15,rsi
+ if THREAD
+	mov	rsi,qword ptr stack_top_offset[r9]
+
+	mov	qword ptr heap_end_after_gc_offset[r9],rbx
+ else
+	mov	rsi,qword ptr stack_top
+
+	mov	qword ptr heap_end_after_gc,rbx 
+ endif
+
+	jmp	restore_registers_after_gc_and_return
+
+_end_bits:
+	mov	rax,rcx 
+	sub	rax,rdx 
+	add	rax,4
+_end_bits2:
+	shl	rax,3
+ if THREAD
+	add	qword ptr n_free_words_after_mark_offset[r9],rax
+ else
+	add	qword ptr n_free_words_after_mark,rax 
+ endif
+	cmp	rax,rsi 
+	jae	_found_free_memory
+
+_end_scan:
+ if THREAD
+	mov	qword ptr bit_counter_offset[r9],rbp
+ else
+	mov	qword ptr bit_counter,rbp
+ endif
+	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]
+ if THREAD
+	mov	r15,rbp
+ else
+	mov	r9,rbp
+ endif
+
+	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
+ if THREAD
+	mov	rdx,r15
+ else
+	mov	rdx,r9
+ endif
+	je	_mark_tuple_selector_node_2
+ if THREAD
+	mov	rcx,qword ptr (-24)[r15+rax]
+ else
+	mov	rcx,qword ptr (-24)[r9+rax]
+ endif
+	mov	qword ptr [rbp],rcx
+	jmp	_mark_node
+
+_mark_tuple_selector_node_2:
+ if THREAD
+	mov	rcx,qword ptr [r15]
+ else
+	mov	rcx,qword ptr [r9]
+ endif
+	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 
+
+ if THREAD
+	lea	rbp,__indirection
+	mov	qword ptr (-8)[rdx],rbp
+ else
+	lea	r9,__indirection
+	mov	qword ptr (-8)[rdx],r9
+ endif
+	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]
+  if THREAD
+	mov	r15,rbp
+  else
+	mov	r9,rbp
+  endif
+	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
+  if THREAD
+	mov	rdx,r15
+  else
+	mov	rdx,r9
+  endif
+	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]
+ if THREAD
+	mov	r15,rbp
+ else
+	mov	r9,rbp
+ endif
+
+	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
+ if THREAD
+	mov	rbx,qword ptr (-24)[r15+rbx]
+ else
+	mov	rbx,qword ptr (-24)[r9+rbx]
+ endif
+	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
+ if THREAD
+	mov	rdx,r15
+ else
+	mov	rdx,r9
+ endif
+	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_node2_bb
+
+	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_node2_bb
+
+	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:
+ if THREAD
+	lea	rbp,__STRING__+2
+	cmp	rax,rbp
+ else
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+ endif
+	jbe	_mark_string_or_array
+
+	or	dword ptr [rdi+rbx*4],esi 
+
+ if THREAD
+	lea	rbp,CHAR+2
+	cmp	rax,rbp
+ else
+	lea	r9,CHAR+2
+	cmp	rax,r9
+ endif
+	ja	_mark_normal_hnf_0
+
+_mark_real_int_bool_or_char:
+	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_node2_bb:
+	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_real_int_bool_or_char
+
+_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 dINT+2
+	jle	_mark_strict_int_or_real_array
+	cmp	rbp,offset BOOL+2
+	je	_mark_strict_bool_array
+_mark_strict_int32_or_real32_array:
+	add	rax,6+1
+	shr	rax,1
+	jmp	_mark_basic_array_
+_mark_strict_int_or_real_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]
+ if THREAD
+	mov	r15,r8
+ else
+	mov	r9,r8
+ endif
+
+	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
+ if THREAD
+	mov	rcx,qword ptr (-24)[r15+rax]
+ else
+	mov	rcx,qword ptr (-24)[r9+rax]
+ endif
+	mov	qword ptr [r8],rcx
+	jmp	__mark_node
+
+__mark_tuple_selector_node_2:
+ if THREAD
+	mov	rcx,qword ptr [r15]
+ else
+	mov	rcx,qword ptr [r9]
+ endif
+	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]
+ if THREAD
+	mov	r15,r8
+ else
+	mov	r9,r8
+ endif
+
+	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
+ if THREAD
+	mov	rbp,r15
+ else
+	mov	rbp,r9
+ endif
+	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]
+ if THREAD
+	mov	r15,r8
+ else
+	mov	r9,r8
+ endif
+
+	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
+ if THREAD
+	mov	rbx,qword ptr (-24)[r15+rbx]
+ else
+	mov	rbx,qword ptr (-24)[r9+rbx]
+ endif
+	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
+ if THREAD
+	mov	rbp,r15
+ else
+	mov	rbp,r9
+ endif
+	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_node2_bb
+
+	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_node2_bb
+
+__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_real_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:
+	cmp	rax,offset __STRING__+2
+	jbe	__mark_string_or_array
+
+ 	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_real_bool_or_small_string
+
+ ifdef NEW_DESCRIPTORS
+	lea	rcx,((-8)-2)[rax]
+ else
+	lea	rcx,((-12)-2)[rax]
+ endif
+	jmp	__mark_next_node
+
+__mark_node2_bb:
+	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_real_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
+ if THREAD
+	mov	rbx,qword ptr lazy_array_list_offset[r9]
+ else
+	mov	rbx,qword ptr lazy_array_list
+ endif
+	add	rbp,2
+	mov	qword ptr [rdx],rbx 
+	mov	qword ptr (-8)[rcx],rbp 
+	mov	qword ptr (-16)[rcx],rax 
+	sub	rcx,16
+ if THREAD
+	mov	qword ptr lazy_array_list_offset[r9],rcx
+ else
+	mov	qword ptr lazy_array_list,rcx 
+ endif
+
+	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]
+ if THREAD
+	mov	rdx,qword ptr lazy_array_list_offset[r9]
+ else
+	mov	rdx,qword ptr lazy_array_list
+ endif
+	mov	qword ptr 24[rcx],rbp 
+	mov	qword ptr 16[rcx],rdx 
+	mov	qword ptr [rcx],rax
+ if THREAD
+	mov	qword ptr lazy_array_list_offset[r9],rcx
+ else
+	mov	qword ptr lazy_array_list,rcx
+ endif
+	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 dINT+2
+	jle	__mark__strict__int__or__real__array
+	cmp	rbp,offset BOOL+2
+	je	__mark__strict__bool__array
+__mark__strict__int32__or__real32__array:
+	add	rax,6+1
+	shr	rax,1
+	jmp	__mark_basic_array
+__mark__strict__int__or__real__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/thread/amark_prefetch.asm b/thread/amark_prefetch.asm
new file mode 100644
index 0000000..86ec2fe
--- /dev/null
+++ b/thread/amark_prefetch.asm
@@ -0,0 +1,1964 @@
+
+_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:
+ if THREAD
+	mov	rax,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rax,qword ptr heap_size_65
+ endif
+	xor	rbx,rbx 
+
+ if THREAD
+	mov	qword ptr n_marked_words_offset[r9],rbx
+ else
+	mov	qword ptr n_marked_words,rbx
+ endif
+	shl	rax,6
+
+	mov	r11,rax
+ if THREAD
+	mov	qword ptr lazy_array_list_offset[r9],rbx 
+ else
+	mov	qword ptr lazy_array_list,rbx 
+ endif
+	
+	lea	rsi,(-4000)[rsp]
+
+	mov	rax,qword ptr caf_list
+
+ if THREAD
+	mov	qword ptr end_stack_offset[r9],rsi
+ else
+	mov	qword ptr end_stack,rsi
+ endif
+
+	mov	r15,0
+	mov	r8,0
+
+ if THREAD
+	mov	r10,neg_heap_p3_offset[r9]
+	mov	r13,qword ptr end_stack_offset[r9]
+ else
+	mov	r10,neg_heap_p3
+	mov	r13,qword ptr end_stack
+ endif
+	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:
+ if THREAD
+	mov	rsi,qword ptr stack_top_offset[r9]
+	mov	rbp,qword ptr stack_p_offset[r9]
+ else
+	mov	rsi,qword ptr stack_top
+	mov	rbp,qword ptr stack_p
+ endif
+
+	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]
+ if THREAD
+	mov	rsi,rbp
+ else
+	mov	r9,rbp
+ endif
+
+	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
+ if THREAD
+	mov	rdx,rsi
+ else
+	mov	rdx,r9
+ endif
+	je	pmark_tuple_selector_node_2
+ if THREAD
+	mov	rcx,qword ptr (-24)[rsi+rax]
+ else
+	mov	rcx,qword ptr (-24)[r9+rax]
+ endif
+	mov	qword ptr [rbp],rcx
+	jmp	pmark_node
+
+pmark_tuple_selector_node_2:
+ if THREAD
+	mov	rcx,qword ptr [rsi]
+ else
+	mov	rcx,qword ptr [r9]
+ endif
+	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
+
+  if THREAD
+	lea	rbp,__indirection
+	mov	qword ptr (-8)[rdx],rbp
+  else
+	lea	r9,__indirection
+	mov	qword ptr (-8)[rdx],r9
+  endif
+	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]
+  if THREAD
+	mov	rsi,rbp
+  else
+	mov	r9,rbp
+  endif
+
+	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
+ if THREAD
+	mov	rdx,rsi
+ else
+	mov	rdx,r9
+ endif
+	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]
+ if THREAD
+	mov	rsi,rbp
+ else
+	mov	r9,rbp
+ endif
+
+	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
+  if THREAD
+	mov	rbx,qword ptr (-24)[rsi+rbx]
+  else
+	mov	rbx,qword ptr (-24)[r9+rbx]
+  endif
+	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
+  if THREAD
+	mov	rdx,rsi
+  else
+	mov	rdx,r9
+  endif
+	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:
+ if THREAD
+	cmp	rax,offset dINT+2
+ else
+	lea	r9,dINT+2
+	cmp	rax,r9
+ endif
+	jb	pmark_real_file_or_string
+
+	or	dword ptr [rdi+rbx*4],esi 
+
+ if THREAD
+	cmp	rax,offset CHAR+2
+ else
+	lea	r9,CHAR+2
+	cmp	rax,r9
+ endif
+	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:
+ if THREAD
+	cmp	rax,offset __STRING__+2
+ else
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+ endif
+	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 dINT+2
+	jle	pmark_strict_int_or_real_array
+	cmp	rbp,offset BOOL+2
+	je	pmark_strict_bool_array
+	add	rax,6+1
+	shr	rax,1
+	jmp	pmark_basic_array_
+pmark_strict_int_or_real_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
+ if THREAD
+	mov	r13,qword ptr end_stack_offset[r9]
+ else
+	mov	r13,qword ptr end_stack
+ endif
+	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:
+ if THREAD
+	mov	r13,qword ptr 16[rbp]
+	add	r13,r10
+	mov	rbx,r13
+	and	r13,31*8
+ else
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8
+	and	r8,31*8
+ endif
+	shr	rbx,8
+ if THREAD
+	mov	r13d,dword ptr (bit_set_table2)[r13]
+	test	r13d,dword ptr [rdi+rbx*4]
+ else
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+ endif
+	jne	pmarkr_no_selector_2
+
+ ifdef NEW_DESCRIPTORS
+  if THREAD
+	mov	rbx,qword ptr 16[rbp]
+  endif
+	mov	eax,dword ptr (-8)[rax]
+	lea	rdx,__indirection
+  if THREAD
+  else
+	pop	rbx 
+  endif
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+ if THREAD
+	mov	r13,rcx
+ else
+	mov	r8,rcx
+ endif
+
+	cmp	rax,16
+	jl	pmarkr_tuple_selector_node_1
+  if THREAD
+  else
+	mov	rdx,r9
+  endif
+	je	pmarkr_tuple_selector_node_2
+  if THREAD
+	mov	rcx,qword ptr (-24)[rbx+rax]
+	mov	qword ptr [r13],rcx
+	pop	rbx
+  else
+	mov	rcx,qword ptr (-24)[r9+rax]
+	mov	qword ptr [r8],rcx
+  endif
+	jmp	pmarkr_node
+
+pmarkr_tuple_selector_node_2:
+  if THREAD
+	mov	rcx,qword ptr [rbx]
+	mov	qword ptr [r13],rcx
+	pop	rbx
+  else
+	mov	rcx,qword ptr [r9]
+	mov	qword ptr [r8],rcx
+  endif
+	jmp	pmarkr_node
+ endif
+
+pmarkr_small_tuple_or_record:
+ ifdef NEW_DESCRIPTORS
+	mov	eax,(-8)[rax]
+	lea	rdx,__indirection
+  if THREAD
+  else
+	pop	rbx 
+  endif
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+ if THREAD
+	mov	r13,rcx
+ else
+	mov	r8,rcx
+ endif
+pmarkr_tuple_selector_node_1:
+	mov	rcx,qword ptr [rbp+rax]
+ if THREAD
+	mov	qword ptr [r13],rcx
+	pop	rbx
+ else
+	mov	qword ptr [r8],rcx
+ endif
+	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
+
+  if THREAD
+	mov	r13,qword ptr 16[rbp]
+	add	r13,r10
+	mov	rbx,r13
+	and	r13,31*8
+  else
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8
+	and	r8,31*8
+  endif
+	shr	rbx,8
+  if THREAD
+	mov	r13d,dword ptr (bit_set_table2)[r13]
+	test	r13d,dword ptr [rdi+rbx*4]
+  else
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+  endif
+	jne	pmarkr_no_selector_2
+
+  if THREAD
+	mov	rbx,qword ptr 16[rbp]
+  endif
+
+pmarkr_small_record:
+	mov	eax,dword ptr (-8)[rax]
+	lea	rdx,__indirection
+  if THREAD
+  else
+	pop	rbx 
+  endif
+
+	mov	qword ptr (-8)[rcx],rdx
+	movzx	eax,word ptr 4[rax]
+  if THREAD
+	mov	r13,rcx
+  else
+	mov	r8,rcx
+  endif
+
+	cmp	rax,16
+	jle	pmarkr_record_selector_node_2
+  if THREAD
+	mov	rbp,rbx
+  else
+	mov	rdx,r9
+  endif
+	sub	rax,24
+pmarkr_record_selector_node_2:
+	mov	rcx,qword ptr [rbp+rax]
+  if THREAD
+	mov	qword ptr [r13],rcx
+	pop	rbx
+  else
+	mov	qword ptr [r8],rcx
+  endif
+	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
+
+ if THREAD
+	mov	r13,qword ptr 16[rbp]
+	add	r13,r10
+	mov	rbx,r13 
+	and	r13,31*8
+ else
+	mov	r8,qword ptr 16[rbp]
+	mov	r9,r8
+
+	add	r8,r10
+	mov	rbx,r8 
+	and	r8,31*8
+ endif
+	shr	rbx,8
+ if THREAD
+	mov	r13d,dword ptr (bit_set_table2)[r13]
+	test	r13d,dword ptr [rdi+rbx*4]
+ else
+	mov	r8d,dword ptr (bit_set_table2)[r8]
+	test	r8d,dword ptr [rdi+rbx*4]
+ endif
+	jne	pmarkr_no_selector_2
+ if THREAD
+	mov	r13,qword ptr 16[rbp]
+ endif
+
+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
+ if THREAD
+	mov	rbx,qword ptr (-24)[r13+rbx]
+ else
+	mov	rbx,qword ptr (-24)[r9+rbx]
+ endif
+	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
+ if THREAD
+	mov	rbp,r13
+ else
+	mov	rbp,r9
+ endif
+	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:
+	mov	rax,qword ptr (-8)[rax]
+	mov	qword ptr [rcx],rax
+	pop	rbx
+ 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:
+ if THREAD
+	cmp	rax,offset __STRING__+2 
+ else
+	lea	r9,__STRING__+2
+	cmp	rax,r9
+ endif
+	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
+ if THREAD
+	mov	rbx,qword ptr lazy_array_list_offset[r9]
+ else
+	mov	rbx,qword ptr lazy_array_list
+ endif
+	add	rbp,2
+	mov	qword ptr [rdx],rbx 
+	mov	qword ptr (-8)[rcx],rbp 
+	mov	qword ptr (-16)[rcx],rax 
+	sub	rcx,16
+ if THREAD
+	mov	qword ptr lazy_array_list_offset[r9],rcx 
+ else
+	mov	qword ptr lazy_array_list,rcx 
+ endif
+
+	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]
+ if THREAD
+	mov	rdx,qword ptr lazy_array_list_offset[r9]
+ else
+	mov	rdx,qword ptr lazy_array_list
+ endif
+	mov	qword ptr 24[rcx],rbp 
+	mov	qword ptr 16[rcx],rdx 
+	mov	qword ptr [rcx],rax
+ if THREAD
+	mov	qword ptr lazy_array_list_offset[r9],rcx 
+ else
+	mov	qword ptr lazy_array_list,rcx 
+ endif
+	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 dINT+2
+	jle	pmarkr_strict_int_or_real_array
+	cmp	rbp,offset BOOL+2
+	je	pmarkr_strict_bool_array
+	add	rax,6+1
+	shr	rax,1
+	jmp	pmarkr_basic_array
+pmarkr_strict_int_or_real_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
diff --git a/thread/astartup.asm b/thread/astartup.asm
new file mode 100644
index 0000000..33ac35f
--- /dev/null
+++ b/thread/astartup.asm
@@ -0,0 +1,5856 @@
+
+;	File:	astartup.asm
+;	Author:	John van Groningen
+;	Machine:	amd64
+
+_TEXT	segment para 'CODE'
+_TEXT	ends
+_DATA	segment para 'DATA'
+_DATA	ends
+
+	d2 equ r10
+	d3 equ r11
+	d4 equ r12
+	d5 equ r13
+
+	d3d equ r11d
+	d4d equ r12d
+
+	d2b equ r10b
+
+ ifdef LINUX
+	.intel_syntax noprefix
+ endif
+
+	THREAD equ 1
+
+ ifndef LINUX
+	extrn	convert_real_to_string:near
+ endif
+ ifndef LINUX
+	extrn	write_heap:near
+ endif
+	extrn	return_code:near
+	extrn	execution_aborted:near
+	extrn	e____system__kFinalizerGCTemp:near
+	extrn	e____system__kFinalizer:near
+
+ ifdef LINUX
+	.globl	times
+	.globl	exit
+ else
+	extrn	GetTickCount:near
+	extrn	ExitProcess:near
+ endif
+ if THREAD
+	extrn	TlsAlloc:near
+ endif
+
+ ifdef USE_LIBM
+	extrn	cos:near
+	extrn	sin:near
+	extrn	tan:near
+	extrn	atan:near
+ endif
+
+NEW_DESCRIPTORS = 1
+
+	_DATA segment
+	align	(1 shl 3)
+
+ ife THREAD
+semi_space_size	dq	0
+
+heap_p1	dq	0
+heap_p2	dq	0
+heap_p3	dq	0
+neg_heap_p3	dq	0
+end_heap_p3	dq	0
+
+neg_heap_vector_plus_4	dq	0
+
+heap_size_64_65	dq	0
+heap_vector	dq	0
+stack_top	dq	0
+end_vector	dq	0
+
+heap_size_257	dq	0
+heap_copied_vector	dq	0
+
+heap_end_after_gc	dq	0
+extra_heap	dq	0
+extra_heap_size	dq	0
+stack_p	dq	0
+halt_sp	dq	0
+
+n_allocated_words	dq	0
+ endif
+
+ if THREAD
+	public	tlsp_tls_index
+tlsp_tls_index	dq	0
+ endif
+
+last_time	dq	0
+execute_time	dq	0
+garbage_collect_time	dq	0
+IO_time	dq	0
+
+compact_garbage_collect_time	dq	0
+mark_compact_garbage_collect_time	dq	0
+total_gc_bytes	dq	0
+total_compact_gc_bytes	dq	0
+
+ ife THREAD
+	public	saved_heap_p
+saved_heap_p label ptr
+	dq	0
+	dq	0
+
+	public	saved_a_stack_p
+saved_a_stack_p	dq	0
+
+	public	end_a_stack
+end_a_stack	dq	0
+
+	public	int_to_real_scratch
+int_to_real_scratch	dq	0
+ endif
+
+heap_end_write_heap	dq	0
+d3_flag_write_heap	dq	0
+
+ ife THREAD
+heap2_begin_and_end label ptr
+		dq	0
+		dq	0
+ endif
+
+	public	a_stack_guard_page
+a_stack_guard_page	dq	0
+
+	public	profile_stack_pointer
+profile_stack_pointer	dq	0
+
+dll_initialised	dq	0
+ ife THREAD
+	public	end_b_stack
+end_b_stack	dq	0
+ endif
+basic_only	dq	0
+ ife THREAD
+heap_size_65	dq	0
+heap_copied_vector_size	dq	0
+heap_end_after_copy_gc	dq	0
+heap_mbp	dq	0
+heap_p		dq	0
+stack_mbp	dq	0
+
+bit_counter	label ptr
+	dq	0
+bit_vector_p	label ptr
+	dq	0
+zero_bits_before_mark	label ptr
+	dq	1
+n_free_words_after_mark	label ptr
+	dq	1000
+n_last_heap_free_bytes	label ptr
+	dq	0
+lazy_array_list	label ptr
+	dq	0
+n_marked_words	label ptr
+	dq	0
+end_stack	label ptr
+	dq	0
+
+bit_vector_size	label ptr
+	dq	0
+ endif
+
+ if THREAD
+ 
+	comm	main_thread_local_storage:512
+
+heap_p1_offset			= 0
+heap_p2_offset			= 8
+heap_p3_offset			= 16
+
+saved_heap_p_offset		= 24
+saved_r15_offset		= 32
+saved_a_stack_p_offset		= 40
+
+heap_vector_offset		= 48
+end_vector_offset		= 56 ; temp
+neg_heap_vector_plus_4_offset	= 64 ; temp
+
+heap_size_64_65_offset		= 72 ; temp
+heap_size_257_offset		= 80
+
+heap_copied_vector_offset	= 88
+
+heap_end_after_gc_offset	= 96
+
+extra_heap_offset		= 104
+extra_heap_size_offset		= 112
+
+stack_top_offset		= 120 ; temp
+stack_p_offset			= 128
+
+halt_sp_offset			= 136
+
+n_allocated_words_offset	= 144 ; temp
+
+heap2_begin_and_end_offset	= 152
+heap_copied_vector_size_offset	= 168
+heap_end_after_copy_gc_offset	= 176
+heap_mbp_offset			= 184
+heap_p_offset			= 192
+stack_mbp_offset		= 200
+bit_counter_offset		= 208
+bit_vector_p_offset		= 216
+bit_vector_size_offset		= 224
+zero_bits_before_mark_offset	= 232
+n_free_words_after_mark_offset	= 240
+n_last_heap_free_bytes_offset	= 248
+n_marked_words_offset		= 256 ; temp
+end_stack_offset		= 264 ; temp
+lazy_array_list_offset		= 272 ; temp
+heap_size_offset		= 280
+heap_size_65_offset		= 288
+a_stack_size_offset		= 296
+garbage_collect_flag_offset	= 304
+
+semi_space_size_offset	= 312 ; temp
+neg_heap_p3_offset		= 320 ; temp
+end_heap_p3_offset		= 328 ; temp
+
+ endif
+
+caf_list	label ptr
+	dq	0
+	public	caf_listp
+caf_listp	label ptr
+	dq	0
+	
+zero_length_string	label ptr
+	dq	__STRING__+2
+	dq	0
+true_string	label ptr
+	dq	__STRING__+2
+	dq	4
+true_c_string	label ptr
+	db	"True"
+	db	0,0,0,0
+false_string	label ptr
+	dq	__STRING__+2
+	dq	5
+false_c_string	label ptr
+	db	"False"
+	db	0,0,0
+file_c_string	label ptr
+	db	"File"
+	db	0,0,0,0
+ ife THREAD
+garbage_collect_flag	label ptr
+	db	0
+	db	0,0,0
+	align	(1 shl 3)
+ endif
+
+	comm	sprintf_buffer:32
+
+out_of_memory_string_1	label ptr
+	db	"Not enough memory to allocate heap and stack"
+	db	10,0
+printf_int_string	label ptr
+	db	"%d"
+	db	0
+printf_real_string	label ptr
+	db	"%.15g"
+	db	0
+printf_string_string	label ptr
+	db	"%s"
+	db	0
+printf_char_string	label ptr
+	db	"%c"
+	db	0
+garbage_collect_string_1	label ptr
+	db	"A stack: "
+	db	0
+garbage_collect_string_2	label ptr
+	db	" bytes. BC stack: "
+	db	0
+garbage_collect_string_3	label ptr
+	db	" bytes."
+	db	10,0
+heap_use_after_gc_string_1	label ptr
+	db	"Heap use after garbage collection: "
+	db	0
+heap_use_after_compact_gc_string_1	label ptr
+	db	"Heap use after compacting garbage collection: "
+	db	0
+heap_use_after_gc_string_2	label ptr
+	db	" Bytes."
+	db	10,0
+stack_overflow_string	label ptr
+	db	"Stack overflow."
+	db	10,0
+out_of_memory_string_4	label ptr
+	db	"Heap full."
+	db	10,0
+time_string_1	label ptr
+	db	"Execution: "
+	db	0
+time_string_2	label ptr
+	db	"  Garbage collection: "
+	db	0
+time_string_3	label ptr
+	db	" "
+	db	0
+time_string_4	label ptr
+	db	"  Total: "
+	db	0
+high_index_string	label ptr
+	db	"Index too high in UPDATE string."
+	db	10,0
+low_index_string	label ptr
+	db	"Index negative in UPDATE string."
+	db	10,0
+IO_error_string	label ptr
+	db	"IO error: "
+	db	0
+new_line_string	label ptr
+	db	10,0
+sprintf_time_string	label ptr
+	db	"%d.%02d"
+	db	0
+marked_gc_string_1	label ptr
+	db	"Marked: "
+	db	0
+ if THREAD
+tls_alloc_error_string	label ptr
+	db	"Could not allocate thread local storage index"
+	db	10,0
+ endif
+
+ ifdef PROFILE
+	align	8
+m_system:
+	dd	6
+	db	"System"
+	db	0
+	db	0
+
+	dd	m_system
+garbage_collector_name:
+	dq	0
+	db	"garbage_collector"
+	db	0
+	align	8
+ endif
+
+	align	16
+	public	sign_real_mask
+sign_real_mask	label ptr
+	dq	8000000000000000h,8000000000000000h
+	public	abs_real_mask
+abs_real_mask	label ptr
+	dq	7fffffffffffffffh,7fffffffffffffffh
+
+	align (1 shl 3) 
+NAN_real	label ptr
+	dd	0ffffffffh,7fffffffh
+one_real	label ptr
+	dd	00000000h,3ff00000h
+zero_real	label ptr
+	dd	00000000h,00000000h
+
+	align (1 shl 2) 
+bit_set_table	label ptr
+	dd	00000001h,00000002h,00000004h,00000008h
+	dd	00000010h,00000020h,00000040h,00000080h
+	dd	00000100h,00000200h,00000400h,00000800h
+	dd	00001000h,00002000h,00004000h,00008000h
+	dd	00010000h,00020000h,00040000h,00080000h
+	dd	00100000h,00200000h,00400000h,00800000h
+	dd	01000000h,02000000h,04000000h,08000000h
+	dd	10000000h,20000000h,40000000h,80000000h
+	dd	0
+bit_set_table2	label ptr
+	dd	00000001h,0,00000002h,0,00000004h,0,00000008h,0
+	dd	00000010h,0,00000020h,0,00000040h,0,00000080h,0
+	dd	00000100h,0,00000200h,0,00000400h,0,00000800h,0
+	dd	00001000h,0,00002000h,0,00004000h,0,00008000h,0
+	dd	00010000h,0,00020000h,0,00040000h,0,00080000h,0
+	dd	00100000h,0,00200000h,0,00400000h,0,00800000h,0
+	dd	01000000h,0,02000000h,0,04000000h,0,08000000h,0
+	dd	10000000h,0,20000000h,0,40000000h,0,80000000h,0
+	dd	0,0
+bit_clear_table	label ptr
+	dd	0fffffffeh,0fffffffdh,0fffffffbh,0fffffff7h
+	dd	0ffffffefh,0ffffffdfh,0ffffffbfh,0ffffff7fh
+	dd	0fffffeffh,0fffffdffh,0fffffbffh,0fffff7ffh
+	dd	0ffffefffh,0ffffdfffh,0ffffbfffh,0ffff7fffh
+	dd	0fffeffffh,0fffdffffh,0fffbffffh,0fff7ffffh
+	dd	0ffefffffh,0ffdfffffh,0ffbfffffh,0ff7fffffh
+	dd	0feffffffh,0fdffffffh,0fbffffffh,0f7ffffffh
+	dd	0efffffffh,0dfffffffh,0bfffffffh,7fffffffh
+	dd	0ffffffffh
+bit_clear_table2	label ptr
+	dd	0fffffffeh,-1,0fffffffdh,-1,0fffffffbh,-1,0fffffff7h,-1
+	dd	0ffffffefh,-1,0ffffffdfh,-1,0ffffffbfh,-1,0ffffff7fh,-1
+	dd	0fffffeffh,-1,0fffffdffh,-1,0fffffbffh,-1,0fffff7ffh,-1
+	dd	0ffffefffh,-1,0ffffdfffh,-1,0ffffbfffh,-1,0ffff7fffh,-1
+	dd	0fffeffffh,-1,0fffdffffh,-1,0fffbffffh,-1,0fff7ffffh,-1
+	dd	0ffefffffh,-1,0ffdfffffh,-1,0ffbfffffh,-1,0ff7fffffh,-1
+	dd	0feffffffh,-1,0fdffffffh,-1,0fbffffffh,-1,0f7ffffffh,-1
+	dd	0efffffffh,-1,0dfffffffh,-1,0bfffffffh,-1,7fffffffh,-1
+	dd	0ffffffffh,-1
+first_one_bit_table	label ptr
+	db	-1,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	5,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	6,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	5,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	7,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	5,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	6,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	5,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+	db	4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0
+
+	align(1 shl  2) 
+	comm	sprintf_time_buffer:20
+
+	align(1 shl  3)
+
+;	public	small_integers
+	comm	small_integers:33*16
+;	public	static_characters
+	comm	static_characters:256*16
+
+;	extrn	clean_exception_handler:near
+;	public	clean_unwind_info
+;clean_unwind_info:
+;	DD	000000009H
+;	DD	imagerel(clean_exception_handler)
+
+_DATA	ends
+	_TEXT segment
+
+	public	abc_main
+	public	print
+	public	print_char
+	public	print_int
+	public	print_real
+	public	print__string__
+	public	print__chars__sc
+	public	print_sc
+	public	print_symbol
+	public	print_symbol_sc
+	public	printD
+	public	DtoAC
+	public	push_t_r_args
+	public	push_a_r_args
+	public	halt
+	public	dump
+
+	public	catAC
+	public	sliceAC
+	public	updateAC
+	public	eqAC
+	public	cmpAC
+
+	public	string_to_string_node
+	public	int_array_to_node
+	public	real_array_to_node
+
+	public	_create_arrayB
+	public	_create_arrayC
+	public	_create_arrayI
+	public	_create_arrayI32
+	public	_create_arrayR
+	public	_create_arrayR32
+	public	_create_r_array
+	public	create_array
+	public	create_arrayB
+	public	create_arrayC
+	public	create_arrayI
+	public	create_arrayI32
+	public	create_arrayR
+	public	create_arrayR32
+	public	create_R_array
+
+	public	BtoAC
+	public	ItoAC
+	public	RtoAC
+	public	eqD
+
+	public	collect_0
+	public	collect_1
+	public	collect_2
+	public	collect_3
+
+	public	yet_args_needed
+	public	yet_args_needed_0
+	public	yet_args_needed_1
+	public	yet_args_needed_2
+	public	yet_args_needed_3
+	public	yet_args_needed_4
+
+	public	_c3,_c4,_c5,_c6,_c7,_c8,_c9,_c10,_c11,_c12
+	public	_c13,_c14,_c15,_c16,_c17,_c18,_c19,_c20,_c21,_c22
+	public	_c23,_c24,_c25,_c26,_c27,_c28,_c29,_c30,_c31,_c32
+
+	public	e__system__nind
+	public	e__system__eaind
+; old names of the previous two labels for compatibility, remove later
+	public	__indirection,__eaind
+	extrn	e__system__dind:near
+	public	eval_fill
+
+	public	eval_upd_0,eval_upd_1,eval_upd_2,eval_upd_3,eval_upd_4
+	public	eval_upd_5,eval_upd_6,eval_upd_7,eval_upd_8,eval_upd_9
+	public	eval_upd_10,eval_upd_11,eval_upd_12,eval_upd_13,eval_upd_14
+	public	eval_upd_15,eval_upd_16,eval_upd_17,eval_upd_18,eval_upd_19
+	public	eval_upd_20,eval_upd_21,eval_upd_22,eval_upd_23,eval_upd_24
+	public	eval_upd_25,eval_upd_26,eval_upd_27,eval_upd_28,eval_upd_29
+	public	eval_upd_30,eval_upd_31,eval_upd_32
+
+	public	repl_args_b
+	public	push_arg_b
+	public	del_args
+
+	public	add_IO_time
+	public	add_execute_time
+	public	IO_error
+	public	stack_overflow
+
+	public	out_of_memory_4
+	public	print_error
+
+ ifdef LINUX
+	.globl	__start
+ else
+	extrn	_start:near
+ endif
+
+ ifdef PROFILE
+;	extrn	init_profiler:near
+;	extrn	profile_n:near
+;	extrn	profile_s:near
+;	extrn	profile_r:near
+;	extrn	write_profile_information:near
+;	extrn	write_profile_stack:near
+ endif
+
+  ifdef USE_LIBM
+	public	cos_real
+	public	sin_real
+	public	tan_real
+	public	asin_real
+	public	acos_real
+	public	atan_real
+	public	ln_real
+	public	log10_real
+	public	exp_real
+	public	pow_real
+ endif
+	public	entier_real
+	public	r_to_i_real
+ ifdef USE_LIBM
+	public	_c_pow
+	public	_c_log10
+	public	_c_entier
+ endif
+
+	public	__driver
+
+; from system.abc:
+	extrn	dINT:near
+	extrn	INT32:near
+	extrn	CHAR:near
+	extrn	BOOL:near
+	extrn	REAL:near
+	extrn	REAL32:near
+	extrn	FILE:near
+	extrn	__STRING__:near
+	extrn	__ARRAY__:near
+	extrn	__cycle__in__spine:near
+	extrn	__print__graph:near
+	extrn	__eval__to__nf:near
+
+; from wcon.c:
+	extrn	w_print_char:near
+	extrn	w_print_string:near
+	extrn	w_print_text:near
+	extrn	w_print_int:near
+	extrn	w_print_real:near
+
+	extrn	ew_print_char:near
+	extrn	ew_print_text:near
+	extrn	ew_print_string:near
+	extrn	ew_print_int:near
+	extrn	ew_print_real:near
+
+	extrn	ab_stack_size:near
+	extrn	heap_size:near
+	extrn	flags:near
+
+; from standard c library:
+
+ ifndef LINUX
+	extrn	allocate_memory:near
+  ife THREAD
+	extrn	allocate_memory_with_guard_page_at_end:near
+  endif
+	extrn	free_memory:near
+ endif
+ 
+	extrn	heap_size_multiple:near
+	extrn	initial_heap_size:near
+
+	extrn	min_write_heap_size:near
+
+	extrn	__Nil:near
+;	public	finalizer_list
+	comm	finalizer_list:qword
+;	public	free_finalizer_list
+	comm	free_finalizer_list:qword
+
+abc_main:
+	push	rbx
+	push	rcx
+	push	rdx 
+	push	rbp 
+	push	rsi 
+	push	rdi 
+
+	call	init_clean
+	test	rax,rax
+	jne	init_error
+
+	call	init_timer
+
+ if THREAD
+	mov	halt_sp_offset[r9],rsp
+ else
+	mov	halt_sp,rsp
+ endif
+
+ ifdef PROFILE
+	call	init_profiler
+ endif
+
+ ifdef LINUX
+	call	__start
+
+exit_:
+ else
+	call	_start
+
+exit:
+ endif
+
+	call	exit_clean
+
+init_error:
+	pop	rdi
+	pop	rsi
+	pop	rbp
+	pop	rdx
+	pop	rcx
+	pop	rbx
+
+ ifdef LINUX
+	mov	eax,dword ptr return_code
+	jne	return_code_set_1
+	mov	eax,-1
+return_code_set_1:
+ endif
+	ret
+
+
+	public	DllMain
+DllMain:
+	cmp	edx,1
+	je	DLL_PROCESS_ATTACH
+	jb	DLL_PROCESS_DETACH
+	ret
+
+DLL_PROCESS_ATTACH:
+	push	rbx 
+	push	rbp 
+	push	rsi 
+	push	rdi
+ ifndef LINUX
+	db	49h
+	push	rsp
+	db	49h
+	push	rbp
+	db	49h
+	push	rsi
+	db	49h
+	push	rdi
+ else
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+ endif
+	mov	qword ptr dll_initialised,1
+
+	call	init_clean
+	test	rax,rax
+	jne	init_dll_error
+
+	call	init_timer
+
+ if THREAD
+	mov	halt_sp_offset[r9],rsp 
+ else
+	mov	halt_sp,rsp 
+ endif
+
+ ifdef PROFILE
+	call	init_profiler
+ endif
+
+ if THREAD
+	mov	qword ptr saved_heap_p_offset[r9],rdi
+	mov	qword ptr saved_r15_offset[r9],r15
+	mov	qword ptr saved_a_stack_p_offset[r9],rsi
+ else
+	mov	qword ptr saved_heap_p,rdi
+	mov	qword ptr saved_heap_p+8,r15
+	mov	saved_a_stack_p,rsi
+ endif
+
+	mov	rax,1
+	jmp	exit_dll_init
+
+init_dll_error:
+	xor	rax,rax
+	jmp	exit_dll_init
+	
+DLL_PROCESS_DETACH:
+	push	rbx 
+	push	rbp 
+	push	rsi 
+	push	rdi
+ ifndef LINUX
+	db	49h
+	push	rsp
+	db	49h
+	push	rbp
+	db	49h
+	push	rsi
+	db	49h
+	push	rdi
+ else
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+ endif
+
+ if THREAD
+	mov	rdi,qword ptr saved_heap_p_offset[r9]
+	mov	r15,qword ptr saved_r15_offset[r9]
+	mov	rsi,qword ptr saved_a_stack_p_offset[r9]
+ else
+	mov	rdi,qword ptr saved_heap_p
+	mov	r15,qword ptr saved_heap_p+8
+	mov	rsi,saved_a_stack_p
+ endif
+
+	call	exit_clean
+
+exit_dll_init:
+ ifndef LINUX
+	db	49h
+	pop	rdi
+	db	49h
+	pop	rsi
+	db	49h
+	pop	rbp
+	db	49h
+	pop	rsp
+ else
+	pop	r15
+	pop	r14
+	pop	r13
+	pop	r12
+ endif
+	pop	rdi 
+	pop	rsi 
+	pop	rbp 
+	pop	rbx 
+	ret
+
+init_clean:
+ if THREAD
+  ifdef LINUX
+	sub	rsp,8
+
+	mov	rdi,rsp
+	sub	rsi,rsi
+
+	mov	rbp,rsp
+	and	rsp,-16
+	call	pthread_key_create
+	mov	rsp,rbp
+
+ 	lea	r9,main_thread_local_storage
+
+	test	eax,eax
+	jne	tls_alloc_error
+
+	mov	rdi,qword ptr [rsp]
+	mov	rsi,r9
+
+	mov	qword ptr tlsp_tls_index,rdi
+
+	mov	rbp,rsp
+	and	rsp,-16
+ 	call	pthread_setspecific
+	mov	rsp,rbp
+
+ 	lea	r9,main_thread_local_storage
+
+	test	eax,eax
+	jne	tls_alloc_error
+
+	add	rsp,8
+
+ 	lea	r9,main_thread_local_storage
+  else
+ 	sub	rsp,32
+	call	TlsAlloc
+	add	rsp,32
+
+	cmp	rax,64
+	jae	tls_alloc_error
+
+	mov	qword ptr tlsp_tls_index,rax
+
+ 	lea	r9,main_thread_local_storage
+
+	mov	qword ptr gs:[1480h+rax*8],r9
+  endif
+ endif
+
+	lea	rax,128[rsp]
+	sub	rsp,32+8
+
+ ife THREAD
+	sub	rax,qword ptr ab_stack_size
+	mov	end_b_stack,rax
+ endif
+	mov	rax,qword ptr flags
+	and	rax,1
+	mov	basic_only,rax 
+
+;	call	allow_prefetch_for_athlon
+
+	mov	rax,qword ptr heap_size
+ if THREAD
+	mov	heap_size_offset[r9],rax
+ endif
+	sub	rax,7
+	xor	rdx,rdx 
+	mov	rbx,65
+	div	rbx
+ if THREAD
+	mov	qword ptr heap_size_65_offset[r9],rax 
+
+	mov	rax,qword ptr heap_size_offset[r9]
+ else
+	mov	qword ptr heap_size_65,rax 
+
+	mov	rax,qword ptr heap_size
+ endif
+
+	sub	rax,7
+	xor	rdx,rdx 
+	mov	rbx,257
+	div	rbx
+ if THREAD
+	mov	heap_size_257_offset[r9],rax
+ else
+	mov	heap_size_257,rax
+ endif
+	add	rax,7
+	and	rax,-8
+ if THREAD
+	mov	qword ptr heap_copied_vector_size_offset[r9],rax
+	mov	qword ptr heap_end_after_copy_gc_offset[r9],0
+
+	mov	rax,qword ptr heap_size_offset[r9]
+ else
+	mov	qword ptr heap_copied_vector_size,rax
+	mov	qword ptr heap_end_after_copy_gc,0
+
+	mov	rax,qword ptr heap_size
+ endif
+	add	rax,7
+	and	rax,-8
+ if THREAD
+	mov	qword ptr heap_size_offset[r9],rax
+ else
+	mov	qword ptr heap_size,rax
+ endif
+	add	rax,7
+
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	rdi,rax
+	call	malloc
+ else
+	mov	rcx,rax
+	call	allocate_memory
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+	test	rax,rax 
+	je	no_memory_2
+
+ if THREAD
+	mov	heap_mbp_offset[r9],rax
+ else
+	mov	heap_mbp,rax
+ endif
+	lea	rdi,7[rax]
+	and	rdi,-8
+ if THREAD
+	mov	heap_p_offset[r9],rdi 
+ else
+	mov	heap_p,rdi 
+ endif
+
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r14,rdi
+	mov	rdi,qword ptr ab_stack_size
+  if THREAD
+	mov	qword ptr a_stack_size_offset[r9],rdi
+  endif
+	add	rdi,7
+	call	malloc
+	mov	rdi,r14
+ else
+	mov	rcx,qword ptr ab_stack_size
+  if THREAD
+	mov	qword ptr a_stack_size_offset[r9],rcx
+  endif
+	add	rcx,7
+  if THREAD
+	call	allocate_memory
+  else
+	call	allocate_memory_with_guard_page_at_end
+  endif
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	
+	test	rax,rax 
+	je	no_memory_3
+
+ if THREAD
+	mov	stack_mbp_offset[r9],rax 
+
+	add	rax,qword ptr a_stack_size_offset[r9]
+ else
+	mov	stack_mbp,rax 
+
+	add	rax,qword ptr ab_stack_size
+ endif
+	add	rax,7+4095
+	and	rax,-4096
+	mov	qword ptr a_stack_guard_page,rax
+ if THREAD
+	sub	rax,qword ptr a_stack_size_offset[r9]
+ else
+	sub	rax,qword ptr ab_stack_size
+ endif
+
+	add	rax,7
+	and	rax,-8
+
+	mov	rsi,rax
+ if THREAD
+	mov	stack_p_offset[r9],rax
+ else
+	mov	stack_p,rax
+ endif
+ ife THREAD
+	add	rax,qword ptr ab_stack_size
+	sub	rax,64
+	mov	qword ptr end_a_stack,rax 
+ endif
+	lea	rcx,small_integers
+	xor	rax,rax 
+	lea	rbx,(dINT+2)
+
+make_small_integers_lp:
+	mov	[rcx],rbx 
+	mov	8[rcx],rax 
+	inc	rax 
+	add	rcx,16
+	cmp	rax,33
+	jne	make_small_integers_lp
+
+	lea	rcx,static_characters
+	xor	rax,rax 
+	lea	rbx,(CHAR+2)
+
+make_static_characters_lp:
+	mov	[rcx],rbx 
+	mov	8[rcx],rax 
+	inc	rax 
+	add	rcx,16
+	cmp	rax,256
+	jne	make_static_characters_lp
+
+	lea	rcx,(caf_list+8)
+	mov	qword ptr caf_listp,rcx 
+
+	lea	rcx,__Nil-8
+	mov	qword ptr finalizer_list,rcx
+	mov	qword ptr free_finalizer_list,rcx
+
+ if THREAD
+	mov	heap_p1_offset[r9],rdi
+
+	mov	rbp,qword ptr heap_size_257_offset[r9]
+ else
+	mov	heap_p1,rdi
+
+	mov	rbp,qword ptr heap_size_257
+ endif
+	shl	rbp,4
+	lea	rax,[rdi+rbp*8]
+ if THREAD
+	mov	heap_copied_vector_offset[r9],rax
+	add	rax,heap_copied_vector_size_offset[r9]
+	mov	heap_p2_offset[r9],rax
+
+	mov	byte ptr garbage_collect_flag_offset[r9],0
+ else
+	mov	heap_copied_vector,rax
+	add	rax,heap_copied_vector_size
+	mov	heap_p2,rax
+
+	mov	byte ptr garbage_collect_flag,0
+ endif
+
+	test	byte ptr flags,64
+	je	no_mark1
+
+ if THREAD
+	mov	rax,qword ptr heap_size_65_offset[r9]
+	mov	qword ptr heap_vector_offset[r9],rdi 
+ else
+	mov	rax,qword ptr heap_size_65
+	mov	qword ptr heap_vector,rdi 
+ endif
+	add	rdi,rax
+
+	add	rdi,7
+	and	rdi,-8
+
+ if THREAD
+	mov	qword ptr heap_p3_offset[r9],rdi
+ else
+	mov	qword ptr heap_p3,rdi
+ endif
+	lea	rbp,[rax*8]
+ if THREAD
+	mov	byte ptr garbage_collect_flag_offset [r9],-1
+ else
+	mov	byte ptr garbage_collect_flag,-1
+ endif
+
+no_mark1:
+	mov	rax,qword ptr initial_heap_size
+
+	mov	rbx,4000
+	test	byte ptr flags,64
+	jne	no_mark9
+	add	rbx,rbx 
+no_mark9:
+
+	cmp	rax,rbx 
+	jle	too_large_or_too_small
+	shr	rax,3
+	cmp	rax,rbp 
+	jge	too_large_or_too_small
+	mov	rbp,rax 
+too_large_or_too_small:
+
+	lea	rax,[rdi+rbp*8]
+ if THREAD
+	mov	heap_end_after_gc_offset[r9],rax
+ else
+	mov	heap_end_after_gc,rax
+ endif
+
+	test	byte ptr flags,64
+	je	no_mark2
+ if THREAD
+	mov	qword ptr bit_vector_size_offset[r9],rbp
+ else
+	mov	qword ptr bit_vector_size,rbp
+ endif
+no_mark2:
+
+	mov	r15,rbp
+
+	add	rsp,32+8
+	xor	rax,rax
+	ret
+
+ if THREAD
+tls_alloc_error:
+	mov	rbp,rsp
+	and	rsp,-16
+  ifdef LINUX
+	lea	rdi,tls_alloc_error_string
+  else
+	lea	rcx,tls_alloc_error_string
+  endif
+	call	ew_print_string
+	mov	rsp,rbp
+
+	add	rsp,8
+	mov	rax,1
+	ret
+ endif
+
+no_memory_2:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	lea	rdi,out_of_memory_string_1
+ else
+	lea	rcx,out_of_memory_string_1
+ endif
+	call	ew_print_string
+	mov	rsp,rbp
+
+	mov	qword ptr execution_aborted,1
+
+	add	rsp,32
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	mov	rax,1
+	ret
+
+no_memory_3:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	lea	rdi,out_of_memory_string_1
+ else
+	lea	ecx,out_of_memory_string_1
+ endif
+	call	ew_print_string
+
+	mov	qword ptr execution_aborted,1
+
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+ ifdef LINUX
+  if THREAD
+	mov	rdi,heap_mbp_offset[r9]
+  else
+	mov	rdi,heap_mbp
+  endif
+	call	free
+ else
+  if THREAD
+	mov	rcx,heap_mbp_offset[r9]
+  else
+	mov	rcx,heap_mbp
+  endif
+	call	free_memory
+ endif
+
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	add	rsp,32
+	mov	rax,1
+	ret
+
+exit_clean:
+	call	add_execute_time
+
+	mov	rax,qword ptr flags
+	test	al,8
+	je	no_print_execution_time
+
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifndef LINUX
+	sub	rsp,32
+ endif
+
+ ifdef LINUX
+	lea	rdi,time_string_1
+ else
+	lea	rcx,time_string_1
+ endif
+	call	ew_print_string
+
+	mov	rax,execute_time
+	call	print_time
+	
+ ifdef LINUX
+	lea	rdi,time_string_2
+ else
+	lea	rcx,time_string_2
+ endif
+	call	ew_print_string
+
+	mov	rax,garbage_collect_time
+ ifdef MEASURE_GC
+ else
+	add	rax,mark_compact_garbage_collect_time
+	add	rax,compact_garbage_collect_time
+ endif
+	call	print_time
+
+ ifdef MEASURE_GC
+
+  ifdef LINUX
+	lea	rdi,time_string_3
+  else
+	lea	rcx,time_string_3
+  endif
+	call	ew_print_string
+
+	mov	rax,mark_compact_garbage_collect_time
+	call	print_time
+
+  ifdef LINUX
+	lea	rdi,time_string_3
+  else
+	lea	rcx,time_string_3
+  endif
+	call	ew_print_string
+
+	mov	rax,compact_garbage_collect_time
+	call	print_time
+ 
+ endif
+
+ ifdef LINUX
+	lea	rdi,time_string_4
+ else
+	lea	rcx,time_string_4
+ endif
+	call	ew_print_string
+
+	mov	rax,execute_time
+	add	rax,garbage_collect_time
+	add	rax,IO_time
+
+	add	rax,mark_compact_garbage_collect_time
+	add	rax,compact_garbage_collect_time
+
+	call	print_time
+
+ ifdef LINUX
+	mov	rdi,10
+ else
+	mov	rcx,10
+ endif
+	call	ew_print_char
+
+ ifdef MEASURE_GC
+
+  ifdef LINUX
+	mov	rdi,total_gc_bytes
+  else
+	mov	rcx,total_gc_bytes
+  endif
+	call	ew_print_int
+
+  ifdef LINUX
+	mov	rdi,32
+  else
+	mov	rcx,32
+  endif
+	call	ew_print_char
+
+  ifdef LINUX
+	mov rdi,total_compact_gc_bytes
+  else
+	mov rcx,total_compact_gc_bytes
+  endif
+	call	ew_print_int
+
+  ifdef LINUX
+	mov	rdi,32
+  else
+	mov	rcx,32
+  endif
+	call	ew_print_char
+
+	mov	rax,1000
+	cvtsi2sd	xmm1,rax
+	cvtsi2sd	xmm0,qword ptr garbage_collect_time
+	divsd	xmm0,xmm1
+	call	ew_print_real
+
+  ifdef LINUX
+	mov	rdi,32
+  else
+	mov	rcx,32
+  endif
+	call	ew_print_char
+
+	mov	rax,1000
+	cvtsi2sd	xmm1,rax
+	cvtsi2sd	xmm0,qword ptr mark_compact_garbage_collect_time
+	divsd	xmm0,xmm1
+	call	ew_print_real
+
+  ifdef LINUX
+	mov	rdi,32
+  else
+	mov	rcx,32
+  endif
+	call	ew_print_char
+
+	mov	rax,1000
+	cvtsi2sd	xmm1,rax
+	cvtsi2sd	xmm0,qword ptr compact_garbage_collect_time
+	divsd	xmm0,xmm1
+	call	ew_print_real
+
+  ifdef LINUX
+	mov	rdi,10
+  else
+	mov	rcx,10
+  endif
+	call	ew_print_char
+
+	mov	rax,1000
+	cvtsi2sd	xmm1,rax
+	cvtsi2sd	xmm2,qword ptr garbage_collect_time
+	divsd	xmm2,xmm1
+	mov	rax,qword ptr total_gc_bytes
+	cvtsi2sd	xmm0,rax
+	divsd	xmm0,xmm2
+	call	ew_print_real
+
+  ifdef LINUX
+	mov	rdi,32
+  else
+	mov	rcx,32
+  endif
+	call	ew_print_char
+
+	mov	rax,1000
+	cvtsi2sd	xmm1,rax
+	cvtsi2sd	xmm2,qword ptr mark_compact_garbage_collect_time
+	divsd	xmm2,xmm1
+	mov	rax,qword ptr total_compact_gc_bytes
+	cvtsi2sd	xmm0,rax
+	divsd	xmm0,xmm2
+	call	ew_print_real
+
+  ifdef LINUX
+	mov	rdi,32
+  else
+	mov	rcx,32
+  endif
+	call	ew_print_char
+
+	mov	rax,1000
+	cvtsi2sd	xmm1,rax
+	cvtsi2sd	xmm2,qword ptr compact_garbage_collect_time
+	divsd	xmm2,xmm1
+	mov	rax,qword ptr total_compact_gc_bytes
+	cvtsi2sd	xmm0,rax
+	divsd	xmm0,xmm2
+	call	ew_print_real
+
+  ifdef LINUX
+	mov	rdi,32
+  else
+	mov	rcx,32
+  endif
+	call	ew_print_char
+
+	mov	rax,1000
+	cvtsi2sd	xmm1,rax
+	cvtsi2sd	xmm2,qword ptr mark_compact_garbage_collect_time
+	cvtsi2sd	xmm3,qword ptr compact_garbage_collect_time
+	addsd	xmm2,xmm3
+	divsd	xmm2,xmm1
+	mov	rax,qword ptr total_compact_gc_bytes
+	cvtsi2sd	xmm0,rax
+	divsd	xmm0,xmm2
+	call	ew_print_real
+
+  ifdef LINUX
+	mov	rdi,10
+  else
+	mov	rcx,10
+  endif
+	call	ew_print_char
+ 
+ endif
+
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+no_print_execution_time:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+  if THREAD
+	mov	rdi,stack_mbp_offset[r9]
+  else
+	mov	rdi,stack_mbp
+  endif
+	call	free
+  if THREAD
+ 	mov	r9,rbx
+  endif
+
+  if THREAD
+	mov	rdi,heap_mbp_offset[r9]
+  else
+	mov	rdi,heap_mbp
+  endif
+	call	free
+ else
+   if THREAD
+	mov	rcx,stack_mbp_offset[r9]
+   else
+	mov	rcx,stack_mbp
+   endif
+	sub	rsp,32
+	call	free_memory
+  if THREAD
+ 	mov	r9,rbx
+  endif
+
+  if THREAD
+	mov	rcx,heap_mbp_offset[r9]
+  else
+	mov	rcx,heap_mbp
+  endif
+	call	free_memory
+	add	rsp,32
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+ ifdef PROFILE
+  ifndef TRACE
+	call	write_profile_information
+  endif
+ endif
+
+	ret
+
+__driver:
+	mov	rbp,qword ptr flags
+	test	rbp,16
+	je	__print__graph
+	jmp	__eval__to__nf
+
+print_time:
+	push	rbp
+ if THREAD
+	mov	r14,r9
+ 	push	rbx
+ endif
+
+	xor	rdx,rdx 
+	mov	rbx,1000
+	div	rbx
+	mov	rcx,rax 
+	mov	rax,rdx 
+	xor	rdx,rdx 
+	mov	rbx,10
+	div	rbx
+
+	push	rax
+
+	mov	rbp,rsp
+	and	rsp,-16
+ ifdef LINUX
+	mov	rdi,rcx
+ else
+	sub	rsp,32
+ endif
+	call	ew_print_int
+	mov	rsp,rbp
+
+	lea	rcx,sprintf_time_buffer
+
+	xor	rdx,rdx 
+	mov	rbx,10
+
+;	movb	$'.',(%rcx)
+	mov	byte ptr [rcx],46
+
+	pop	rax
+
+	div	rbx 
+	add	rax,48
+	add	rdx,48
+	mov	byte ptr 1[rcx],al 
+	mov	byte ptr 2[rcx],dl 
+
+	mov	rbp,rsp
+	and	rsp,-16
+ ifdef LINUX
+	mov	rsi,3
+	mov	rdi,rcx
+ else
+	mov	rdx,3
+	sub	rsp,32
+ endif
+	call	ew_print_text
+	mov	rsp,rbp
+
+ if THREAD
+ 	pop	rbx
+ 	mov	r9,r14
+ endif
+	pop	rbp
+	ret
+
+print_sc:
+	mov	rbp,basic_only
+	test	rbp,rbp 
+	jne	end_print
+
+print:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	mov	rdi,rax
+ else
+	mov	rcx,rax
+	sub	rsp,32
+ endif
+	call	w_print_string
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+end_print:
+	ret
+
+dump:
+	call	print
+	jmp	halt
+
+printD:	test	al,2
+	jne	printD_
+
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+
+	lea	rdi,4[rax]
+	mov	esi,0[rax]
+ else
+	lea	rcx,4[rax]
+	mov	edx,dword ptr [rax]
+	sub	rsp,32
+ endif
+	call	w_print_text
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rsi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+
+DtoAC_record:
+ ifdef NEW_DESCRIPTORS
+	movsxd	rbp,dword ptr (-6)[rax]
+ else
+	movsx	rbp,dword ptr (-4)[rbp]
+ endif
+	jmp	DtoAC_string_a2
+
+DtoAC:	test	al,2
+	jne	DtoAC_
+
+	mov	rbp,rax 
+	jmp	DtoAC_string_a2
+
+DtoAC_:
+ ifdef NEW_DESCRIPTORS
+	cmp	word ptr (-2)[rax],256
+	jae	DtoAC_record
+
+	movzx	rbx,word ptr [rax]
+	lea	rbp,10[rax+rbx]
+ else
+	lea	rbp,(-2)[rax]
+	movsx	rbx,word ptr [rbp]
+	cmp	rbx,256
+	jae	DtoAC_record
+
+	shl	rbx,3
+	sub	rbp,rbx 
+
+ 	movzx	rbx,word ptr (-2)[rbp]
+	lea	rbp,4[rbp+rbx*8]
+ endif
+
+DtoAC_string_a2:
+	mov	eax,dword ptr [rbp]
+	lea	rcx,4[rbp]
+	jmp	build_string
+
+print_symbol:
+	xor	rbx,rbx 
+	jmp	print_symbol_2
+
+print_symbol_sc:
+	mov	rbx,basic_only
+print_symbol_2:
+	mov	rax,[rcx]
+
+	cmp	rax,offset dINT+2
+	je	print_int_node
+
+	cmp	rax,offset CHAR+2
+	je	print_char_denotation
+
+	cmp	rax,offset BOOL+2
+	je	print_bool
+
+	cmp	rax,offset REAL+2
+	je	print_real_node
+	
+	test	rbx,rbx 
+	jne	end_print_symbol
+
+printD_:
+ ifdef NEW_DESCRIPTORS
+	cmp	word ptr (-2)[rax],256
+	jae	print_record
+
+	movzx	rbx,word ptr [rax]
+	lea	rbp,10[rax+rbx]
+	jmp	print_string_a2
+
+print_record:
+	movsxd	rbp,dword ptr (-6)[rax]
+	jmp	print_string_a2
+ else
+	lea	rbp,(-2)[rax]
+	movsx	rbx,word ptr [rbp]
+	cmp	rbx,256
+	jae	print_record
+
+	shl	rbx,3
+	sub	rbp,rbx
+
+  	movzx	rbx,word ptr (-2)[rbp]
+	lea	rbp,4[rbp+rbx*8]
+	jmp	print_string_a2
+
+print_record:
+	mov	ebp,(-4)[rbp]
+	jmp	print_string_a2
+ endif
+
+end_print_symbol:
+	ret
+
+print_int_node:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	mov	rdi,8[rcx]
+ else
+	sub	rsp,32
+	mov	rcx,8[rcx]
+ endif
+	call	w_print_int
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+
+print_int:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	mov	rdi,rax
+ else
+	mov	rcx,rax
+	sub	rsp,32
+ endif
+	call	w_print_int
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+
+print_char_denotation:
+	test	rbx,rbx 
+	jne	print_char_node
+
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	r14,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+ else
+	sub	rsp,32
+ endif
+	mov	rbx,8[rcx]
+
+ ifdef LINUX
+	mov	rdi,0x27
+ else
+	mov	rcx,27h
+ endif
+	call	w_print_char
+
+ ifdef LINUX
+	mov	rdi,rbx
+ else
+	mov	rcx,rbx
+ endif
+	call	w_print_char
+
+ ifdef LINUX
+	mov	rdi,0x27
+ else
+	mov	rcx,27h
+ endif
+	call	w_print_char
+
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,r14
+ endif
+	ret
+
+print_char_node:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+ 
+	mov	rdi,8[rcx]
+else
+	mov	rcx,8[rcx]
+	sub	rsp,32
+ endif
+	call	w_print_char
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+	
+print_char:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	
+	mov	rdi,rax
+ else
+	mov	rcx,rax 
+	sub	rsp,32
+ endif
+	call	w_print_char
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+
+print_bool:
+	movsx	rcx,byte ptr 8[rcx]
+	test	rcx,rcx 
+	je	print_false
+
+print_true:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	lea	rdi,true_c_string
+ else
+	lea	rcx,true_c_string
+	sub	rsp,32
+ endif
+	call	w_print_string
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+
+print_false:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	lea	rdi,false_c_string
+ else
+	lea	rcx,false_c_string
+	sub	rsp,32
+ endif
+	call	w_print_string
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+
+print_real_node:
+	movlpd	xmm0,qword ptr 8[rcx]
+print_real:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+ else
+	sub	rsp,32
+ endif
+	call	w_print_real
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+
+print_string_a2:
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	lea	rdi,4[rbp]
+	mov	esi,0[rbp]
+	mov	rbp,rsp
+	and	rsp,-16
+ else
+	lea	rcx,4[rbp]
+	mov	edx,0[rbp]
+	mov	rbp,rsp
+	and	rsp,-16
+	sub	rsp,32
+ endif
+	call	w_print_text
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	ret
+
+print__chars__sc:
+	mov	rbp,basic_only
+	test	rbp,rbp 
+	jne	no_print_chars
+
+print__string__:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	mov	rsi,8[rcx]
+	lea	rdi,16[rcx]
+ else
+	mov	rdx,8[rcx]
+	lea	rcx,16[rcx]
+	sub	rsp,32
+ endif
+	call	w_print_text
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+no_print_chars:
+	ret
+
+push_a_r_args:
+	push	rdi
+
+	mov	rdx,qword ptr 16[rcx]
+	sub	rdx,2
+	movzx	rdi,word ptr [rdx]
+	sub	rdi,256
+	movzx	rbx,word ptr 2[rdx]
+	add	rdx,4
+	push	rdx 
+		
+	mov	rdx,rdi 
+	sub	rdx,rbx 
+
+	shl	rax,3
+	lea	rcx,24[rcx+rbx*8]
+	dec	rdi 
+mul_array_size_lp:
+	add	rcx,rax 
+	sub	rdi,1
+	jnc	mul_array_size_lp
+
+	lea	rdi,[rcx+rdx*8]
+	jmp	push_a_elements
+push_a_elements_lp:
+	mov	rax,qword ptr (-8)[rcx]
+	sub	rcx,8
+	mov	qword ptr [rsi],rax 
+	add	rsi,8
+push_a_elements:
+	sub	rbx,1
+	jnc	push_a_elements_lp
+
+	mov	rcx,rdi
+	pop	rax
+	pop	rdi
+
+	pop	rbp 
+	jmp	push_b_elements
+push_b_elements_lp:
+	push	(-8)[rcx]
+	sub	rcx,8
+push_b_elements:
+	sub	rdx,1
+	jnc	push_b_elements_lp
+
+	jmp	rbp
+
+push_t_r_args:
+	pop	rbp
+
+	mov	rdx,qword ptr [rcx]
+	add	rcx,8
+	sub	rdx,2
+	movzx	rax,word ptr [rdx]
+	sub	rax,256
+	movzx	rbx,word ptr 2[rdx]
+	add	rdx,4
+
+	mov	qword ptr [rsi],rdx
+	mov	qword ptr 8[rsi],rbx
+
+	sub	rbx,rax
+	neg	rbx 
+
+	lea	rdx,[rcx+rax*8]
+	cmp	rax,2
+	jbe	small_record
+	mov	rdx,qword ptr 8[rcx]
+	lea	rdx,(-8)[rdx+rax*8]
+small_record:
+	jmp	push_r_b_elements
+
+push_r_b_elements_lp:
+	dec	rax 
+	jne	not_first_arg_b
+	
+	push	[rcx]
+	jmp	push_r_b_elements
+not_first_arg_b:
+	push	(-8)[rdx]
+	sub	rdx,8
+push_r_b_elements:
+	sub	rbx,1
+	jnc	push_r_b_elements_lp
+
+	mov	rbx,qword ptr 8[rsi]
+	push	rbp
+	push	[rsi]
+	jmp	push_r_a_elements
+
+push_r_a_elements_lp:
+	dec	rax 
+	jne	not_first_arg_a
+	
+	mov	rbp,qword ptr [rcx]
+	mov	qword ptr [rsi],rbp
+	add	rsi,8
+	jmp	push_r_a_elements
+not_first_arg_a:
+	mov	rbp,qword ptr (-8)[rdx]
+	sub	rdx,8
+	mov	qword ptr [rsi],rbp
+	add	rsi,8
+push_r_a_elements:
+	sub	rbx,1
+	jnc	push_r_a_elements_lp
+
+	pop	rax 
+	ret
+
+BtoAC:
+	test	al,al 
+	je	BtoAC_false
+BtoAC_true:
+	mov	rcx,offset true_string
+	ret
+BtoAC_false:
+	mov	rcx,offset false_string
+	ret
+
+RtoAC:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	lea	rsi,printf_real_string
+	lea	rdi,sprintf_buffer
+	mov	rax,1
+	call	sprintf
+	mov	rsi,r13
+	mov	rdi,r14
+ else
+	lea	rdx,sprintf_buffer
+	sub	rsp,32
+	call	convert_real_to_string
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	jmp	return_sprintf_buffer
+
+ItoAC:
+	mov	rcx,offset sprintf_buffer
+	call	int_to_string
+	
+	mov	rax,rcx
+	sub	rax,offset sprintf_buffer
+
+	jmp	sprintf_buffer_to_string
+
+	public	convert_int_to_string
+convert_int_to_string:
+	push	rbp
+	push	rbx
+	mov	rax,rdx
+	call	int_to_string
+	mov	rax,rcx 
+	pop	rbx
+	pop	rbp
+	ret
+
+int_to_string:
+	test	rax,rax
+	jns	no_minus
+	mov	byte ptr [rcx],45
+	inc	rcx
+	neg	rax 
+no_minus:
+	mov	rbp,rcx
+
+	je	zero_digit
+	
+calculate_digits:
+	cmp	rax,10
+	jb	last_digit
+
+	mov	rdx,0cccccccccccccccdh
+	mov	rbx,rax 
+
+	mul	rdx  
+
+	mov	rax,rdx 
+	and	rdx,-8
+	add	rbx,48
+
+	shr	rax,3
+	sub	rbx,rdx
+	shr	rdx,2
+
+	sub	rbx,rdx
+	mov	byte ptr [rcx],bl 
+
+	inc	rcx
+	jmp	calculate_digits
+
+last_digit:
+	test	rax,rax 
+	je	no_zero
+zero_digit:
+	add	rax,48
+	mov	byte ptr [rcx],al 
+	inc	rcx
+no_zero:
+	mov	rdx,rcx
+
+reverse_digits:
+	dec	rdx
+	cmp	rbp,rdx
+	jae	end_reverse_digits
+	mov	bl,byte ptr [rbp]
+	mov	al,byte ptr [rdx] 
+	mov	byte ptr [rdx],bl
+	mov	byte ptr [rbp],al
+	inc	rbp
+	jmp	reverse_digits
+
+end_reverse_digits:
+	mov	byte ptr [rcx],0
+	ret
+
+return_sprintf_buffer:
+	mov	rax,offset sprintf_buffer-1
+skip_characters:
+	inc	rax 
+	cmp	byte ptr [rax],0
+	jne	skip_characters
+
+	sub	rax,offset sprintf_buffer
+
+sprintf_buffer_to_string:
+	mov	rcx,offset sprintf_buffer
+build_string:
+
+	lea	rbx,16+7[rax]
+	shr	rbx,3
+
+	sub	r15,rbx
+	jge	D_to_S_no_gc
+
+	push	rcx 
+	call	collect_0
+	pop	rcx 
+
+D_to_S_no_gc:
+ if THREAD
+	lea	rbp,__STRING__+2
+	sub	rbx,2
+	mov	qword ptr [rdi],rbp
+	mov	8[rdi],rax 
+	mov	rbp,rdi 
+ else
+	sub	rbx,2
+	mov	rbp,rdi
+	lea	r9,__STRING__+2
+	mov	qword ptr [rdi],r9
+	mov	8[rdi],rax 
+ endif
+	add	rdi,16
+	jmp	D_to_S_cp_str_2
+
+D_to_S_cp_str_1:
+	mov	rax,[rcx]
+	add	rcx,8
+	mov	[rdi],rax 
+	add	rdi,8
+D_to_S_cp_str_2:
+	sub	rbx,1
+	jnc	D_to_S_cp_str_1
+	
+	mov	rcx,rbp 
+	ret
+
+eqD:	mov	rax,[rcx]
+	cmp	rax,[rdx]
+	jne	eqD_false
+
+	cmp	rax,offset dINT+2
+	je	eqD_INT
+	cmp	rax,offset CHAR+2
+	je	eqD_CHAR
+	cmp	rax ,offset BOOL+2
+	je	eqD_BOOL
+	cmp	rax ,offset REAL+2
+	je	eqD_REAL
+
+	mov	rax ,1
+	ret
+
+eqD_CHAR:
+eqD_INT:
+	mov	rbx,8[rcx]
+	xor	rax,rax 
+	cmp	rbx,8[rdx]
+	sete	al
+	ret
+
+eqD_BOOL:
+	mov	bl,byte ptr 8[rcx]
+	xor	rax,rax 
+	cmp	bl,byte ptr 8[rdx]
+	sete	al 
+	ret
+
+eqD_REAL:
+	movlpd	xmm0,qword ptr 8[rcx]
+	comisd	xmm0,qword ptr 8[rdx]
+	fnstsw	ax
+	and	ah,68
+	xor	ah,64
+	sete	al
+	and	rax,1
+	ret
+
+eqD_false:
+	xor	rax ,rax 
+	ret
+;
+;	the timer
+;
+
+
+init_timer:
+	mov	rbp,rsp
+	and	rsp,-16
+	sub	rsp,32
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	mov	rdi,rsp
+	call    times
+	mov	rsi,r13
+	mov	rdi,r14
+	mov	eax,[rsp]
+	imul	eax,10
+ else
+	call	GetTickCount
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+	mov	last_time,rax
+	xor	rax,rax
+	mov	execute_time,rax
+	mov	garbage_collect_time,rax
+	mov	IO_time,rax
+
+	mov	mark_compact_garbage_collect_time,rax 
+	mov	compact_garbage_collect_time,rax 
+
+	ret
+
+get_time_diff:
+	mov	rbp,rsp
+	and	rsp,-16
+	sub	rsp,32
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi
+	mov	rdi,rsp
+	call    times
+	mov	rsi,r13
+	mov	rdi,r14
+	mov	eax,[rsp]
+	imul	eax,10
+ else
+	call	GetTickCount
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+	lea	rcx,last_time
+	mov	rdx,[rcx]
+	mov	[rcx],rax 
+	sub	rax,rdx
+	ret
+	
+add_execute_time:
+	call	get_time_diff
+	lea	rcx,execute_time
+
+add_time:
+	add	rax,[rcx]
+	mov	[rcx],rax 
+	ret
+
+add_garbage_collect_time:
+	call	get_time_diff
+	mov	rcx,offset garbage_collect_time
+	jmp	add_time
+
+add_IO_time:
+	call	get_time_diff
+	mov	rcx,offset IO_time
+	jmp	add_time
+
+add_mark_compact_garbage_collect_time:
+	call	get_time_diff
+	mov	rcx,offset mark_compact_garbage_collect_time
+	jmp	add_time
+
+add_compact_garbage_collect_time:
+	call	get_time_diff
+	mov	rcx,offset compact_garbage_collect_time
+	jmp	add_time
+;
+;	the garbage collector
+;
+
+collect_3:
+ ifdef PROFILE
+	lea	rbp,garbage_collector_name
+	call	profile_s
+ endif
+	mov	[rsi],rcx 
+	mov	8[rsi],rdx 
+	mov	16[rsi],r8
+	add	rsi,24
+	call	collect_0_
+	mov	r8,(-8)[rsi]
+	mov	rdx,(-16)[rsi]
+	mov	rcx,(-24)[rsi]
+	sub	rsi,24
+ ifdef PROFILE
+	jmp	profile_r
+ else
+	ret
+ endif
+
+collect_2:
+ ifdef PROFILE
+	lea	rbp,garbage_collector_name
+	call	profile_s
+ endif
+	mov	[rsi],rcx 
+	mov	8[rsi],rdx 
+	add	rsi,16
+	call	collect_0_
+	mov	rdx,(-8)[rsi]
+	mov	rcx,(-16)[rsi]
+	sub	rsi,16
+ ifdef PROFILE
+	jmp	profile_r
+ else
+	ret
+ endif
+
+collect_1:
+ ifdef PROFILE
+	lea	rbp,garbage_collector_name
+	call	profile_s
+ endif
+	mov	[rsi],rcx 
+	add	rsi,8
+	call	collect_0_
+	mov	rcx,(-8)[rsi]
+	sub	rsi,8
+ ifdef PROFILE
+	jmp	profile_r
+ else
+	ret
+ endif
+
+collect_0:
+ ifdef PROFILE
+	lea	rbp,garbage_collector_name
+	call	profile_s
+ endif
+	call	collect_0_
+ ifdef PROFILE
+	jmp	profile_r
+ else
+	ret
+ endif
+
+collect_0_:
+	mov	rbp,rdi
+
+	push	rax
+	push	rbx
+
+ if THREAD
+	mov	rbx,qword ptr heap_end_after_gc_offset[r9]
+ else
+	mov	rbx,qword ptr heap_end_after_gc
+ endif
+	sub	rbx,rdi
+
+	shr	rbx,3
+	sub	rbx,r15
+ if THREAD
+	mov	qword ptr n_allocated_words_offset[r9],rbx
+ else
+	mov	qword ptr n_allocated_words,rbx
+ endif
+
+	test	byte ptr flags,64
+	je	no_mark3
+
+ if THREAD
+	mov	rbp,qword ptr bit_counter_offset[r9]
+ else
+	mov	rbp,qword ptr bit_counter
+ endif
+	test	rbp,rbp 
+	je	no_scan
+
+	push	rsi
+	mov	rsi,rbx
+
+	xor	rbx,rbx
+ if THREAD
+	mov	rcx,qword ptr bit_vector_p_offset[r9]
+ else
+	mov	rcx,qword ptr bit_vector_p
+ endif
+
+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	dword ptr (-4)[rcx],ebx 
+	mov	rax,rcx 
+	sub	rax,rdx 
+	jmp	end_bits2
+
+end_zero_bits:
+	mov	rax,rcx 
+	sub	rax,rdx 
+	shl	rax,3
+ if THREAD
+	add	qword ptr n_free_words_after_mark_offset[r9],rax 
+ else
+	add	qword ptr n_free_words_after_mark,rax
+ endif
+	mov	dword ptr (-4)[rcx],ebx
+
+	cmp	rax,rsi
+	jb	scan_bits
+
+found_free_memory:
+ if THREAD
+	mov	qword ptr bit_counter_offset[r9],rbp
+	mov	qword ptr bit_vector_p_offset[r9],rcx
+ else
+	mov	qword ptr bit_counter,rbp
+	mov	qword ptr bit_vector_p,rcx
+ endif
+
+	lea	rbp,(-4)[rdx]
+ if THREAD
+	sub	rbp,qword ptr heap_vector_offset[r9]
+ else
+	sub	rbp,qword ptr heap_vector
+ endif
+	shl	rbp,6
+ if THREAD
+	mov	rdi,qword ptr heap_p3_offset[r9]
+ else
+	mov	rdi,qword ptr heap_p3
+ endif
+	add	rdi,rbp 
+
+	lea	rbp,[rdi+rax*8]
+ if THREAD
+	mov	qword ptr heap_end_after_gc_offset[r9],rbp
+ else
+	mov	qword ptr heap_end_after_gc,rbp
+ endif
+
+	mov	r15,rax
+	sub	r15,rsi
+
+	pop	rsi 
+	pop	rbx 
+	pop	rax 
+	ret
+
+end_bits:
+	mov	rax,rcx 
+	sub	rax,rdx 
+	add	rax,4
+end_bits2:
+	shl	rax,3
+ if THREAD
+	add	qword ptr n_free_words_after_mark_offset[r9],rax 
+ else
+	add	qword ptr n_free_words_after_mark,rax 
+ endif
+	cmp	rax,rsi 
+	jae	found_free_memory
+
+end_scan:
+	pop	rsi
+ if THREAD
+	mov	qword ptr bit_counter_offset[r9],rbp 
+ else
+	mov	qword ptr bit_counter,rbp 
+ endif
+
+no_scan:
+
+no_mark3:
+ if THREAD
+	movsx	rax,byte ptr garbage_collect_flag_offset[r9]
+ else
+	movsx	rax,byte ptr garbage_collect_flag
+ endif
+	test	rax,rax 
+	jle	collect
+
+	sub	rax,2
+ if THREAD
+	mov	byte ptr garbage_collect_flag_offset[r9],al 
+
+	mov	rbp,qword ptr extra_heap_size_offset[r9]
+ else
+	mov	byte ptr garbage_collect_flag,al 
+
+	mov	rbp,qword ptr extra_heap_size
+ endif
+	cmp	rbx,rbp 
+	ja	collect
+
+ if THREAD
+	mov	rdi,qword ptr extra_heap_offset[r9]
+ else
+	mov	rdi,qword ptr extra_heap
+ endif
+
+	mov	r15,rbp
+
+	lea	rbp,[rdi+rbp*8]
+ if THREAD
+	mov	qword ptr heap_end_after_gc_offset[r9],rbp
+ else
+	mov	qword ptr heap_end_after_gc,rbp
+ endif
+
+	sub	r15,rbx
+
+	pop	rbx 
+	pop	rax 
+	ret
+
+collect:
+ ifdef LINUX
+	sub	rsp,104
+ else
+	sub	rsp,88
+ endif
+	mov	32[rsp],r10
+	mov	24[rsp],r11
+	mov	16[rsp],r12
+	mov	8[rsp],r13
+	mov	[rsp],r14
+	movsd	40[rsp],xmm0
+	movsd	48[rsp],xmm1
+	movsd	56[rsp],xmm2
+	movsd	64[rsp],xmm3
+	movsd	72[rsp],xmm4
+	movsd	80[rsp],xmm5
+ ifdef LINUX
+	movsd	88[rsp],xmm6
+	movsd	96[rsp],xmm7
+ endif
+
+	call	add_execute_time
+
+	test	qword ptr flags,4
+	je	no_print_stack_sizes
+
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+ 	mov	r13,rsi
+	mov	r14,rdi
+ else
+	sub	rsp,32
+ endif
+
+ if 0
+  ifdef LINUX
+	mov	rdi,qword ptr 64[rsp]
+  else
+	mov	rcx,qword ptr 96[rsp]
+  endif
+	call	ew_print_int
+
+  ifdef LINUX
+	mov	rdi,32
+  else
+	mov	rcx,32
+  endif
+	call	ew_print_char
+ endif
+
+ ifdef LINUX
+	lea	rdi,garbage_collect_string_1
+ else
+	lea	rcx,garbage_collect_string_1
+ endif
+	call	ew_print_string
+
+ ifdef LINUX
+	mov	rdi,r13
+  if THREAD
+	mov	r9,rbx
+	sub	rdi,stack_p_offset[r9]
+  else
+	sub	rdi,stack_p
+  endif
+ else
+	mov	rcx,rsi
+  if THREAD
+	mov	r9,rbx
+	sub	rcx,stack_p_offset[r9]
+  else
+	sub	rcx,stack_p
+  endif
+ endif
+	call	ew_print_int
+
+ ifdef LINUX
+	lea	rdi,garbage_collect_string_2
+ else
+	lea	rcx,garbage_collect_string_2
+ endif
+	call	ew_print_string
+
+ ifdef LINUX
+  if THREAD
+	mov	r9,rbx
+	mov	rdi,halt_sp_offset[r9]
+  else
+	mov	rdi,halt_sp
+  endif
+	sub	rdi,rsp 
+ else
+  if THREAD
+	mov	r9,rbx
+	mov	rcx,halt_sp_offset[r9]
+  else
+	mov	rcx,halt_sp
+  endif
+	sub	rcx,rsp 
+ endif
+	call	ew_print_int
+
+ ifdef LINUX
+	lea	rdi,garbage_collect_string_3
+ else
+	lea	rcx,garbage_collect_string_3
+ endif
+	call	ew_print_string
+
+ ifdef LINUX
+	mov	rsi,r13
+	mov	rdi,r14
+ endif
+	mov	rsp,rbp
+  if THREAD
+ 	mov	r9,rbx
+  endif
+
+no_print_stack_sizes:
+ if THREAD
+	mov	rax,stack_p_offset[r9]
+	add	rax,qword ptr a_stack_size_offset[r9]
+ else
+	mov	rax,stack_p
+	add	rax,qword ptr ab_stack_size
+ endif
+	cmp	rsi,rax 
+	ja	stack_overflow
+
+	test	byte ptr flags,64
+	jne	compacting_collector
+
+ if THREAD
+	cmp	byte ptr garbage_collect_flag_offset[r9],0
+ else
+	cmp	byte ptr garbage_collect_flag,0
+ endif
+	jne	compacting_collector
+
+ if THREAD
+	mov	rbp,heap_copied_vector_offset[r9]
+
+	cmp	qword ptr heap_end_after_copy_gc_offset[r9],0
+ else
+	mov	rbp,heap_copied_vector
+
+	cmp	qword ptr heap_end_after_copy_gc,0
+ endif
+
+	je	zero_all
+
+	mov	rax,rdi
+ if THREAD
+	sub	rax,qword ptr heap_p1_offset[r9]
+ else
+	sub	rax,qword ptr heap_p1
+ endif
+	add	rax,127*8
+	shr	rax,9
+	call	zero_bit_vector
+
+ if THREAD
+	mov	rdx,qword ptr heap_end_after_copy_gc_offset[r9]
+	sub	rdx,qword ptr heap_p1_offset[r9]
+ else
+	mov	rdx,qword ptr heap_end_after_copy_gc
+	sub	rdx,qword ptr heap_p1
+ endif
+	shr	rdx,7
+	and	rdx,-4
+
+ if THREAD
+	mov	rbp,qword ptr heap_copied_vector_offset[r9]
+	mov	rax,qword ptr heap_copied_vector_size_offset[r9]
+ else
+	mov	rbp,qword ptr heap_copied_vector
+	mov	rax,qword ptr heap_copied_vector_size
+ endif
+	add	rbp,rdx 
+	sub	rax,rdx 
+	shr	rax,2
+
+ if THREAD
+	mov	qword ptr heap_end_after_copy_gc_offset[r9],0
+ else
+	mov	qword ptr heap_end_after_copy_gc,0
+ endif
+
+	call	zero_bit_vector
+	jmp	end_zero_bit_vector
+
+zero_all:
+ if THREAD
+	mov	rax,heap_copied_vector_size_offset[r9]
+ else
+	mov	rax,heap_copied_vector_size
+ endif
+	shr	rax,2
+	call	zero_bit_vector
+
+end_zero_bit_vector:
+
+	include	acopy.asm
+
+ if THREAD
+	mov	qword ptr heap2_begin_and_end_offset[r9],rsi 
+ else
+	mov	qword ptr heap2_begin_and_end,rsi 
+ endif
+	mov	r15,rsi
+	sub	r15,rdi
+
+ if THREAD
+	mov	rax,heap_size_257_offset[r9]
+ else
+	mov	rax,heap_size_257
+ endif
+	shl	rax,7
+	sub	rax,r15
+	add	qword ptr total_gc_bytes,rax
+
+	shr	r15,3
+
+	pop	rsi 
+
+	call	add_garbage_collect_time
+
+ if THREAD
+	sub	r15,qword ptr n_allocated_words_offset[r9]
+ else
+	sub	r15,qword ptr n_allocated_words
+ endif
+	jc	switch_to_mark_scan
+
+	lea	rax,[r15+r15*4]
+	shl	rax,6
+ if THREAD
+	mov	rbx,qword ptr heap_size_offset[r9]
+ else
+	mov	rbx,qword ptr heap_size
+ endif
+	mov	rcx,rbx 
+	shl	rbx,2
+	add	rbx,rcx 
+	add	rbx,rbx 
+	add	rbx,rcx
+	cmp	rax,rbx
+	jnc	no_mark_scan
+
+switch_to_mark_scan:
+ if THREAD
+	mov	rax,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rax,qword ptr heap_size_65
+ endif
+	shl	rax,6
+ if THREAD
+	mov	rbx,qword ptr heap_p_offset[r9]
+
+	mov	rcx,qword ptr heap_p1_offset[r9]
+	cmp	rcx,qword ptr heap_p2_offset[r9]
+ else
+	mov	rbx,qword ptr heap_p
+
+	mov	rcx,qword ptr heap_p1
+	cmp	rcx,qword ptr heap_p2
+ endif
+	jc	vector_at_begin
+	
+vector_at_end:
+ if THREAD
+	mov	qword ptr heap_p3_offset[r9],rbx 
+ else
+	mov	qword ptr heap_p3,rbx
+ endif
+	add	rbx,rax
+ if THREAD
+	mov	qword ptr heap_vector_offset[r9],rbx 
+
+	mov	rax,qword ptr heap_p1_offset[r9]
+	mov	qword ptr extra_heap_offset[r9],rax 
+ else
+	mov	qword ptr heap_vector,rbx
+
+	mov	rax,qword ptr heap_p1
+	mov	qword ptr extra_heap,rax
+ endif
+	sub	rbx,rax
+	shr	rbx,3
+ if THREAD
+	mov	qword ptr extra_heap_size_offset[r9],rbx
+ else
+	mov	qword ptr extra_heap_size,rbx
+ endif
+	jmp	switch_to_mark_scan_2
+
+vector_at_begin:
+ if THREAD
+	mov	qword ptr heap_vector_offset[r9],rbx
+	add	rbx,qword ptr heap_size_offset[r9]
+ else
+	mov	qword ptr heap_vector,rbx
+	add	rbx,qword ptr heap_size
+ endif
+	sub	rbx,rax
+ if THREAD
+	mov	qword ptr heap_p3_offset[r9],rbx 
+ else
+	mov	qword ptr heap_p3,rbx
+ endif
+
+ if THREAD
+	mov	qword ptr extra_heap_offset[r9],rbx 
+	mov	rcx,qword ptr heap_p2_offset[r9]
+ else
+	mov	qword ptr extra_heap,rbx 
+	mov	rcx,qword ptr heap_p2
+ endif
+	sub	rcx,rbx 
+	shr	rcx,3
+ if THREAD
+	mov	qword ptr extra_heap_size_offset[r9],rcx 
+ else
+	mov	qword ptr extra_heap_size,rcx 
+ endif
+
+switch_to_mark_scan_2:
+ if THREAD
+	mov	rax,heap_size_257_offset[r9]
+ else
+	mov	rax,heap_size_257
+ endif
+	shl	rax,7-3
+	sub	rax,r15
+	shl	rax,3
+
+ if THREAD
+	mov	byte ptr garbage_collect_flag_offset[r9],1
+ else
+	mov	byte ptr garbage_collect_flag,1
+ endif
+
+	lea	rcx,heap_use_after_gc_string_1
+
+	test	r15,r15
+	jns	end_garbage_collect
+
+ if THREAD
+	mov	byte ptr garbage_collect_flag_offset[r9],-1
+
+	mov	rbx,qword ptr extra_heap_size_offset[r9]
+ else
+	mov	byte ptr garbage_collect_flag,-1
+
+	mov	rbx,qword ptr extra_heap_size
+ endif
+	mov	r15,rbx
+ if THREAD
+	sub	r15,qword ptr n_allocated_words_offset[r9]
+ else
+	sub	r15,qword ptr n_allocated_words
+ endif
+	js	out_of_memory_4_3
+
+ if THREAD
+	mov	rdi,qword ptr extra_heap_offset[r9]
+ else
+	mov	rdi,qword ptr extra_heap
+ endif
+	shl	rbx,3
+	add	rbx,rdi
+ if THREAD
+	mov	qword ptr heap_end_after_gc_offset[r9],rbx
+ else
+	mov	qword ptr heap_end_after_gc,rbx
+ endif
+
+	mov	qword ptr heap_end_write_heap,rdi 
+
+	mov	qword ptr d3_flag_write_heap,1
+	jmp	end_garbage_collect_
+
+no_mark_scan:
+; exchange the semi_spaces
+ if THREAD
+	mov	rax,heap_p1_offset[r9]
+	mov	rbx,heap_p2_offset[r9]
+	mov	heap_p2_offset[r9],rax 
+	mov	heap_p1_offset[r9],rbx 
+
+	mov	rax,heap_size_257_offset[r9]
+ else
+	mov	rax,heap_p1
+	mov	rbx,heap_p2
+	mov	heap_p2,rax 
+	mov	heap_p1,rbx 
+
+	mov	rax,heap_size_257
+ endif
+	shl	rax,7-3
+	mov	rbx,rax
+	sub	rax,r15
+
+	mov	rcx,rax
+	imul	qword ptr heap_size_multiple
+	shrd	rax,rdx,9
+	shr	rdx,9
+	jne	no_small_heap1
+
+	cmp	rax,4000
+	jge	not_too_small1
+	mov	rax,4000
+not_too_small1:
+	sub	rbx,rax 
+	jb	no_small_heap1
+
+	sub	r15,rbx
+	shl	rbx,3
+ if THREAD
+	mov	rbp,qword ptr heap_end_after_gc_offset[r9]
+	mov	qword ptr heap_end_after_copy_gc_offset[r9],rbp 
+ else
+	mov	rbp,qword ptr heap_end_after_gc
+	mov	qword ptr heap_end_after_copy_gc,rbp 
+ endif
+	sub	rbp,rbx
+ if THREAD
+	mov	qword ptr heap_end_after_gc_offset[r9],rbp
+ else
+	mov	qword ptr heap_end_after_gc,rbp
+ endif
+
+no_small_heap1:
+	mov	rax,rcx
+	shl	rax,3
+
+	lea	rcx,heap_use_after_gc_string_1
+
+end_garbage_collect:
+
+	mov	qword ptr heap_end_write_heap,rdi 
+	mov	qword ptr d3_flag_write_heap,0
+
+end_garbage_collect_:
+	test	qword ptr flags,2
+	je	no_heap_use_message
+
+	push	rax
+
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+
+ ifdef LINUX
+	mov	r13,rsi
+	mov	r14,rdi	
+
+	mov	rdi,rcx
+ else
+	sub	rsp,32
+ endif
+	call	ew_print_string
+
+ ifdef LINUX
+	mov	rdi,[rbp]
+ else
+	mov	rcx,[rbp]
+ 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
+ else
+ 	add	rsp,32
+ endif
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+	pop	rax
+
+no_heap_use_message:
+	call	call_finalizers
+	 
+	test	byte ptr flags,32
+	je	no_write_heap
+
+	cmp	rax,qword ptr min_write_heap_size
+	jb	no_write_heap
+
+	push	rcx 
+	push 	rdx 
+	push	rbp 
+	push	rsi 
+	push	rdi 
+
+	sub	rsp,128
+
+	mov	rax,qword ptr d3_flag_write_heap
+	test	rax,rax 
+	jne	copy_to_compact_with_alloc_in_extra_heap
+
+ if THREAD
+	movsx	rax,byte ptr garbage_collect_flag_offset[r9]
+
+	mov	rcx,qword ptr heap2_begin_and_end_offset[r9]
+	mov	rdx,qword ptr (heap2_begin_and_end_offset+8)[r9]
+
+	lea	rbx,heap_p1_offset[r9]
+ else
+	movsx	rax,byte ptr garbage_collect_flag
+
+	mov	rcx,qword ptr heap2_begin_and_end
+	mov	rdx,qword ptr (heap2_begin_and_end+8)
+
+	mov	rbx,offset heap_p1
+ endif
+	
+	test	rax,rax 
+	je	gc0
+
+ if THREAD
+	lea	rbx,heap_p2_offset[r9]
+ else
+	mov	rbx,offset heap_p2
+ endif
+	jg	gc1
+
+ if THREAD
+	lea	rbx,heap_p3_offset[r9]
+ else
+	mov	rbx,offset heap_p3
+ endif
+	xor	rcx,rcx 
+	xor	rdx,rdx 
+
+gc0:
+gc1:
+	mov	rbx,qword ptr [rbx]
+	
+	mov	rax,rsp 
+	
+	mov	qword ptr [rax],rbx 
+	mov	qword ptr 8[rax],rdi 
+	
+	mov	qword ptr 16[rax],rcx 
+	mov	qword ptr 24[rax],rdx 
+	
+ if THREAD
+	mov	rbx ,qword ptr stack_p_offset[r9]
+ else
+	mov	rbx ,qword ptr stack_p
+ endif
+	mov	qword ptr 32[rax],rbx 
+
+	mov	qword ptr 40[rax],rsi 
+	mov	qword ptr 48[rax],0
+	mov	qword ptr 56[rax],0
+	
+	mov	qword ptr 64[rax],offset small_integers
+	mov	qword ptr 72[rax],offset static_characters
+	
+	mov	qword ptr 80[rax],offset dINT+2
+	mov	qword ptr 88[rax],offset CHAR+2
+	mov	qword ptr 96[rax],offset REAL+2
+	mov	qword ptr 104[rax],offset BOOL+2
+	mov	qword ptr 112[rax],offset __STRING__+2
+	mov	qword ptr 120[rax],offset __ARRAY__+2
+
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+ ifdef LINUX
+	mov	rdi,rax
+ else
+	mov	rcx,rax
+	sub	rsp,32
+ endif
+ ifndef LINUX
+	call	write_heap
+ endif
+	mov	rsp,rbp
+	add	rsp,128
+ if THREAD
+ 	mov	r9,rbx
+ endif
+	
+	pop	rdi 
+	pop	rsi 
+	pop	rbp 
+	pop	rdx 
+	pop	rcx 
+no_write_heap:
+
+restore_registers_after_gc_and_return:
+	mov	r10,32[rsp]
+	mov	r11,24[rsp]
+	mov	r12,16[rsp]
+	mov	r13,8[rsp]
+	mov	r14,[rsp]
+	movlpd	xmm0,40[rsp]
+	movlpd	xmm1,48[rsp]
+	movlpd	xmm2,56[rsp]
+	movlpd	xmm3,64[rsp]
+	movlpd	xmm4,72[rsp]
+	movlpd	xmm5,80[rsp]
+ ifdef LINUX
+	movlpd	xmm6,88[rsp]
+	movlpd	xmm7,96[rsp]
+	add	rsp,104
+ else
+	add	rsp,88
+ endif
+	pop	rbx 
+	pop	rax 
+	ret
+
+call_finalizers:
+	mov	rax,qword ptr free_finalizer_list
+
+call_finalizers_lp:
+ if THREAD
+	lea	rbx,__Nil-8
+	cmp	rax,rbx
+ else
+	lea	r9,__Nil-8
+	cmp	rax,r9
+ endif
+	je	end_call_finalizers
+	push	8[rax]
+	mov	rbx,qword ptr 16[rax]
+	push	8[rbx]
+	call	qword ptr [rbx]
+	add	rsp,8
+	pop	rax 
+	jmp	call_finalizers_lp
+end_call_finalizers:
+
+ if THREAD
+	lea	rbx,__Nil-8
+	mov	qword ptr free_finalizer_list,rbx
+ else
+	lea	r9,__Nil-8
+	mov	qword ptr free_finalizer_list,r9
+ endif
+	ret
+
+copy_to_compact_with_alloc_in_extra_heap:
+ if THREAD
+	mov	rcx,qword ptr heap2_begin_and_end_offset[r9]
+	mov	rdx,qword ptr (heap2_begin_and_end_offset+8)[r9]
+	lea	rbx,heap_p2_offset[r9]
+ else
+	mov	rcx,qword ptr heap2_begin_and_end
+	mov	rdx,qword ptr (heap2_begin_and_end+8)
+	mov	rbx,offset heap_p2
+ endif
+	jmp	gc1
+
+allow_prefetch_for_athlon:
+	test	qword ptr flags,4096
+	jne	no_prefetch_flag
+
+	xor	rax,rax
+	cpuid
+	test	rax,rax
+	jz	disable_prefetch_flag
+
+ ifdef LINUX
+	cmp	rbx,'A'+('u'*0x100)+('t'*0x10000)+('h'*0x1000000)
+	jne	disable_prefetch_flag
+	cmp	rdx,'e'+('n'*0x100)+('t'*0x10000)+('i'*0x1000000)
+	jne	disable_prefetch_flag
+	cmp	rcx,'c'+('A'*0x100)+('M'*0x10000)+('D'*0x1000000)
+	jne	disable_prefetch_flag
+ else
+	cmp	rbx,'A'+('u' shl 8)+('t' shl 16)+('h' shl 24)
+	jne	disable_prefetch_flag
+	cmp	rdx,'e'+('n' shl 8)+('t' shl 16)+('i' shl 24)
+	jne	disable_prefetch_flag
+	cmp	rcx,'c'+('A' shl 8)+('M' shl 16)+('D' shl 24)
+	jne	disable_prefetch_flag
+ endif
+
+;	mov	rax,1
+;	cpuid
+;	and	rax,0f00h
+;	cmp	rax,600h
+;	je	keep_prefetch_flag
+
+	ret
+
+disable_prefetch_flag:
+	and	qword ptr flags,-4097
+keep_prefetch_flag:
+no_prefetch_flag:
+	ret
+
+out_of_memory_4_3:
+out_of_memory_4_2:
+out_of_memory_4_1:
+out_of_memory_4:
+	call	add_garbage_collect_time
+	
+	mov	rbp,offset out_of_memory_string_4
+	jmp	print_error
+
+zero_bit_vector:
+	xor	rdx,rdx 
+	test	al,1
+	je	zero_bits1_1
+	mov	dword ptr [rbp],edx 
+	add	rbp,4
+zero_bits1_1:
+	shr	rax,1
+
+	mov	rbx,rax 
+	shr	rax,1
+	test	bl,1
+	je	zero_bits1_5
+
+	sub	rbp,8
+	jmp	zero_bits1_2
+
+zero_bits1_4:
+	mov	dword ptr [rbp],edx 
+	mov	dword ptr 4[rbp],edx 
+zero_bits1_2:
+	mov	dword ptr 8[rbp],edx 
+	mov	dword ptr 12[rbp],edx 
+	add	rbp,16
+zero_bits1_5:
+	sub	rax,1
+	jae	zero_bits1_4
+	ret
+
+reorder:
+	push	rsi 
+	push	rbp 
+
+	mov	rbp,rax 
+	shl	rbp,3
+	mov	rsi,rbx 
+	shl	rsi,3
+	add	rcx,rsi 
+	sub	rdx,rbp 
+
+	push	rsi 
+	push	rbp 
+	push	rbx 
+	push	rax 
+	jmp	st_reorder_lp
+
+reorder_lp:
+	mov	rbp,qword ptr [rcx]
+	mov	rsi,qword ptr (-8)[rdx]
+	mov	qword ptr (-8)[rdx],rbp 
+	sub	rdx,8
+	mov	qword ptr [rcx],rsi 
+	add	rcx,8
+	
+	dec	rax
+	jne	next_b_in_element
+	mov	rax,qword ptr [rsp]
+	add	rcx,qword ptr 24[rsp]
+next_b_in_element:
+	dec	rbx 
+	jne	next_a_in_element
+	mov	rbx,qword ptr 8[rsp]
+	sub	rdx,qword ptr 16[rsp]
+next_a_in_element:
+st_reorder_lp:
+	cmp	rdx,rcx 
+	ja	reorder_lp
+
+	pop	rax 
+	pop	rbx 
+	add	rsp,16
+	pop	rbp 
+	pop	rsi 
+	ret
+
+;
+;	the sliding compacting garbage collector
+;
+
+compacting_collector:
+; zero all mark bits
+
+ if THREAD
+	mov	rax,qword ptr heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr heap_p3
+ endif
+	neg	rax
+ if THREAD
+	mov	qword ptr neg_heap_p3_offset[r9],rax 
+
+	mov	qword ptr stack_top_offset[r9],rsi
+
+	mov	rdi,qword ptr heap_vector_offset[r9]
+ else
+	mov	qword ptr neg_heap_p3,rax
+
+	mov	qword ptr stack_top,rsi
+
+	mov	rdi,qword ptr heap_vector
+ endif
+
+	test	byte ptr flags,64
+	je	no_mark4
+
+ if THREAD
+	cmp	qword ptr zero_bits_before_mark_offset[r9],0
+ else
+	cmp	qword ptr zero_bits_before_mark,0
+ endif
+	je	no_zero_bits
+
+ if THREAD
+	mov	qword ptr zero_bits_before_mark_offset[r9],0
+ else
+	mov	qword ptr zero_bits_before_mark,0
+ endif
+
+no_mark4:
+	mov	rbp,rdi
+ if THREAD
+	mov	rax,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rax,qword ptr heap_size_65
+ endif
+	add	rax,3
+	shr	rax,2
+
+	xor	rbx,rbx
+
+	test	al,1
+	je	zero_bits_1
+	mov	dword ptr [rbp],ebx 
+	add	rbp,4
+zero_bits_1:
+	mov	rcx,rax 
+	shr	rax,2
+
+	test	cl,2
+	je	zero_bits_5
+
+	sub	rbp,8
+	jmp	zero_bits_2
+
+zero_bits_4:
+	mov	dword ptr [rbp],ebx 
+	mov	dword ptr 4[rbp],ebx 
+zero_bits_2:
+	mov	dword ptr 8[rbp],ebx 
+	mov	dword ptr 12[rbp],ebx 
+	add	rbp,16
+zero_bits_5:
+	sub	rax,1
+	jnc	zero_bits_4
+
+	test	byte ptr flags,64
+	je	no_mark5
+
+no_zero_bits:
+ if THREAD
+	mov	rax,qword ptr n_last_heap_free_bytes_offset[r9]
+	mov	rbx,qword ptr n_free_words_after_mark_offset[r9]
+ else
+	mov	rax,qword ptr n_last_heap_free_bytes
+	mov	rbx,qword ptr n_free_words_after_mark
+ endif
+	shl	rbx,3
+
+	mov	rbp,rbx 
+	shl	rbp,3
+	add	rbp,rbx 
+	shr	rbp,2
+
+	cmp	rax,rbp 
+	jg	compact_gc
+
+ if THREAD
+	mov	rbx,qword ptr bit_vector_size_offset[r9]
+ else
+	mov	rbx,qword ptr bit_vector_size
+ endif
+	shl	rbx,3
+
+	sub	rax,rbx 
+	neg	rax
+
+	imul	qword ptr heap_size_multiple
+	shrd	rax,rdx,7
+	shr	rdx,7
+	jne	no_smaller_heap
+	
+	cmp	rax,rbx 
+	jae	no_smaller_heap
+	
+	cmp	rbx,8000
+	jbe	no_smaller_heap
+	
+	jmp	compact_gc
+no_smaller_heap:
+	test	qword ptr flags,4096
+	jne	pmark
+
+	include	amark.asm
+
+	include	amark_prefetch.asm
+
+compact_gc:
+ if THREAD
+	mov	qword ptr zero_bits_before_mark_offset[r9],1
+	mov	qword ptr n_last_heap_free_bytes_offset[r9],0
+	mov	qword ptr n_free_words_after_mark_offset[r9],1000
+ else
+	mov	qword ptr zero_bits_before_mark,1
+	mov	qword ptr n_last_heap_free_bytes,0
+	mov	qword ptr n_free_words_after_mark,1000
+ endif
+
+no_mark5:
+
+	include	acompact.asm
+
+ if THREAD
+	mov	rsi,qword ptr stack_top_offset[r9]
+ else
+	mov	rsi,qword ptr stack_top
+ endif
+
+ if THREAD
+	mov	rbx,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rbx,qword ptr heap_size_65
+ endif
+	shl	rbx,6
+ if THREAD
+	add	rbx,qword ptr heap_p3_offset[r9]
+ else
+	add	rbx,qword ptr heap_p3
+ endif
+
+ if THREAD
+	mov	qword ptr heap_end_after_gc_offset[r9],rbx
+ else
+	mov	qword ptr heap_end_after_gc,rbx
+ endif
+
+	sub	rbx,rdi
+	shr	rbx,3
+
+ if THREAD
+	sub	rbx,qword ptr n_allocated_words_offset[r9]
+ else
+	sub	rbx,qword ptr n_allocated_words
+ endif
+	mov	r15,rbx
+	jc	out_of_memory_4_1
+
+	mov	rax,rbx
+	shl	rax,2
+	add	rax,rbx
+	shl	rax,4
+ if THREAD
+	cmp	rax,qword ptr heap_size_offset[r9]
+ else
+	cmp	rax,qword ptr heap_size
+ endif
+	jc	out_of_memory_4_2
+
+ 	test	byte ptr flags,64
+	je	no_mark_6
+
+ if THREAD
+	mov	rax,qword ptr neg_heap_p3_offset[r9]
+ else
+	mov	rax,qword ptr neg_heap_p3
+ endif
+	add	rax,rdi
+ if THREAD
+	mov	rbx,qword ptr n_allocated_words_offset[r9]
+ else
+	mov	rbx,qword ptr n_allocated_words
+ endif
+	lea	rax,[rax+rbx*8]
+
+ if THREAD
+	mov	rbx,qword ptr heap_size_65_offset[r9]
+ else
+	mov	rbx,qword ptr heap_size_65
+ endif
+	shl	rbx,6
+	
+	imul	qword ptr heap_size_multiple
+	shrd	rax,rdx,8
+	shr	rdx,8
+	jne	no_small_heap2
+
+	and	rax,-4
+
+	cmp	rax,8000
+	jae	not_too_small2
+	mov	rax,8000
+not_too_small2:
+	mov	rcx,rbx 
+	sub	rcx,rax 
+	jb	no_small_heap2
+
+ if THREAD
+	sub	qword ptr heap_end_after_gc_offset[r9],rcx
+ else
+	sub	qword ptr heap_end_after_gc,rcx
+ endif
+	shr	rcx,3
+	sub	r15,rcx
+
+	mov	rbx,rax
+
+no_small_heap2:
+	shr	rbx,3
+ if THREAD
+	mov	qword ptr bit_vector_size_offset[r9],rbx
+ else
+	mov	qword ptr bit_vector_size,rbx
+ endif
+
+no_mark_6:
+	jmp	no_copy_garbage_collection
+
+no_copy_garbage_collection:
+	call	add_compact_garbage_collect_time
+
+	mov	rax,rdi
+ if THREAD
+	sub	rax,qword ptr heap_p3_offset[r9]
+ else
+	sub	rax,qword ptr heap_p3
+ endif
+
+	add	qword ptr total_compact_gc_bytes,rax 
+
+	mov	rax,rdi
+ if THREAD
+	sub	rax,qword ptr heap_p3_offset[r9]
+	mov	rbx,qword ptr n_allocated_words_offset[r9]
+ else
+	sub	rax,qword ptr heap_p3
+	mov	rbx,qword ptr n_allocated_words
+ endif
+	lea	rax,[rax+rbx*8]
+
+	lea	rcx,heap_use_after_compact_gc_string_1
+	jmp	end_garbage_collect
+
+ if 0
+	public	clean_exception_handler_
+
+clean_exception_handler_:
+
+	jmp	clean_exception_handler_
+ endif
+ if 0
+	mov	rax,qword ptr [rcx]
+	cmp	dword ptr [rax],0c00000fdh
+	je  	stack_overflow_exception
+
+	cmp	dword ptr [rax],80000001h
+	je  	guard_page_or_access_violation_exception
+
+	cmp	dword ptr [rax] ,0c0000005h
+	je  	guard_page_or_access_violation_exception
+
+no_stack_overflow_exception:
+	mov	rax,0
+	ret
+
+guard_page_or_access_violation_exception:
+	mov	rax,qword ptr 16[rax]
+	and	rax,-4096
+	cmp	qword ptr a_stack_guard_page,rax
+	jne 	no_stack_overflow_exception
+	
+	cmp	qword ptr a_stack_guard_page,0
+	je  	no_stack_overflow_exception
+
+stack_overflow_exception:
+	mov	rax,qword ptr 8[rcx]
+	mov	qword ptr (0F8h)[rax],offset stack_overflow
+
+	mov	rax,-1
+	ret
+ endif
+
+stack_overflow:
+	call	add_execute_time
+
+	mov	rbp,offset stack_overflow_string
+	jmp	print_error
+
+IO_error:
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+
+	mov	rbx,rcx
+ ifdef LINUX
+	lea	rdi,IO_error_string
+ else
+	sub	rsp,32
+	lea	rcx,IO_error_string
+ endif
+	call	ew_print_string
+
+ ifdef LINUX
+	mov	rdi,rbx
+ else
+	mov	rcx,rbx
+ endif
+	call	ew_print_string
+
+ ifdef LINUX
+	lea	rdi,new_line_string
+ else
+	lea	rcx,new_line_string
+ endif
+	call	ew_print_string
+
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+	jmp	halt
+
+print_error:
+ ifdef LINUX
+	mov	rdi,rbp
+ else
+	mov	rcx,rbp 
+ endif
+	mov	rbp,rsp
+	and	rsp,-16
+ if THREAD
+	mov	rbx,r9
+ endif
+	call	ew_print_string
+	mov	rsp,rbp
+ if THREAD
+ 	mov	r9,rbx
+ endif
+
+halt:
+ if THREAD
+	mov	rsp,halt_sp_offset[r9]
+ else
+	mov	rsp,halt_sp
+ endif
+
+ ifdef PROFILE
+	call	write_profile_stack
+ endif
+
+	mov	qword ptr execution_aborted,1
+
+	cmp	qword ptr dll_initialised,0
+ ifdef LINUX
+	je	exit_
+ else
+	je	exit
+ endif
+ ifdef LINUX
+	cmp	dword ptr return_code,0
+ else
+	cmp	qword ptr return_code,0
+ endif
+	jne	return_code_set
+ ifdef LINUX
+	mov	dword ptr return_code,-1
+ else
+	mov	qword ptr return_code,-1
+ endif
+return_code_set:
+ ifdef LINUX
+	mov	edi,dword ptr return_code
+	and	rsp,-16
+	call	exit
+ else
+	push	qword ptr return_code
+	call	(ExitProcess)
+ endif
+	jmp	return_code_set
+
+e__system__eaind:
+__eaind:
+eval_fill:
+	mov	[rsi],rcx 
+	add	rsi,8
+	mov	rcx,rdx 
+	call	qword ptr [rdx]
+	mov	rdx,rcx 
+	mov	rcx,(-8)[rsi]
+	sub	rsi,8
+	
+	mov	rbp,[rdx]
+	mov	[rcx],rbp 
+	mov	rbp,8[rdx]
+	mov	8[rcx],rbp 
+	mov	rbp,16[rdx]
+	mov	16[rcx],rbp 
+	ret
+
+	align	(1 shl 2) 
+	lea	rax,e__system__eaind
+	jmp	rax
+ ifdef LINUX
+; pc relative lea instruction is one byte longer
+	db	0,0
+ else
+	db	0,0,0
+ endif
+	dd	e__system__dind
+	dd	-2
+e__system__nind:
+__indirection:
+	mov	rdx,8[rcx]
+	mov	rax,[rdx]
+	test	al,2
+
+	je	eval_fill2
+
+	mov	[rcx],rax 
+	mov	rbp,8[rdx]
+	mov	8[rcx],rbp 
+	mov	rbp,16[rdx]
+	mov	16[rcx],rbp 
+	ret
+
+eval_fill2:
+ if THREAD
+	lea	rbp,__cycle__in__spine
+	mov	qword ptr [rcx],rbp
+ else
+	lea	r9,__cycle__in__spine
+	mov	qword ptr [rcx],r9
+ endif
+	mov	qword ptr [rsi],rcx 
+
+	test	byte ptr flags,64
+	je	__cycle__in__spine	
+
+	add	rsi,8
+	mov	rcx,rdx 
+	call	rax 
+	mov	rdx,rcx
+	mov	rcx,qword ptr (-8)[rsi]
+	sub	rsi,8
+	
+	mov	rbp,[rdx]
+	mov	[rcx],rbp
+	mov	rbp,8[rdx]
+	mov	8[rcx],rbp
+	mov	rbp,16[rdx]
+	mov	16[rcx],rbp 
+	ret
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_0:
+	mov	qword ptr [rdx],offset __indirection
+	mov	8[rdx],rcx 
+	jmp	rbp
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_1:
+	mov	qword ptr [rdx],offset __indirection
+	mov	rax,8[rdx]
+	mov	8[rdx],rcx 
+	mov	rdx,rax 
+	jmp	rbp
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_2:
+	mov	qword ptr [rdx],offset __indirection
+	mov	r8,8[rdx]
+	mov	8[rdx],rcx
+	mov	rdx,16[rdx]
+	jmp	rbp 
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_3:
+	mov	qword ptr [rdx],offset __indirection
+	mov	r8,8[rdx]
+	mov	8[rdx],rcx
+	mov	[rsi],rcx
+	mov	rcx,24[rdx]
+	add	rsi,8
+	mov	rdx,16[rdx]
+	jmp	rbp
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_4:
+	mov	qword ptr [rdx],offset __indirection
+	mov	r8,8[rdx]
+	mov	8[rdx],rcx 
+	mov	[rsi],rcx 
+	mov	rbx,32[rdx]
+	mov	8[rsi],rbx 
+	mov	rcx,24[rdx]
+	add	rsi,16
+	mov	rdx,16[rdx]
+	jmp	rbp
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_5:
+	mov	qword ptr [rdx],offset __indirection
+	mov	r8,8[rdx]
+	mov	[rsi],rcx 
+	mov	8[rdx],rcx 
+	mov	rbx,40[rdx]
+	mov	8[rsi],rbx 
+	mov	rbx,32[rdx]
+	mov	16[rsi],rbx 
+	mov	rcx,24[rdx]
+	add	rsi,24
+	mov	rdx,16[rdx]
+	jmp	rbp
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_6:
+	mov	qword ptr [rdx],offset __indirection
+	mov	r8,8[rdx]
+	mov	[rsi],rcx 
+	mov	8[rdx],rcx 
+	mov	rbx,48[rdx]
+	mov	8[rsi],rbx
+	mov	rbx,40[rdx]
+	mov	16[rsi],rbx 
+	mov	rbx,32[rdx]
+	mov	24[rsi],rbx 
+	mov	rcx,24[rdx]
+	add	rsi,32
+	mov	rdx,16[rdx]
+	jmp	rbp 
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_7:
+	mov	rax,0
+	mov	rbx,40
+eval_upd_n:
+	mov	qword ptr [rdx],offset __indirection
+	mov	r8,8[rdx]
+	mov	[rsi],rcx 
+	mov	8[rdx],rcx 
+	add	rdx,rbx 
+	mov	rbx,16[rdx ]
+	mov	8[rsi],rbx 
+	mov	rbx,8[rdx]
+	mov	16[rsi],rbx 
+	mov	rbx,[rdx]
+	mov	24[rsi],rbx 
+	add	rsi,32
+
+eval_upd_n_lp:
+	mov	rbx,(-8)[rdx]
+	sub	rdx,8
+	mov	[rsi],rbx 
+	add	rsi,8
+	sub	rax,1
+	jnc	eval_upd_n_lp
+
+	mov	rcx,(-8)[rdx]
+	mov	rdx,(-16)[rdx ]
+	jmp	rbp 
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_8:
+	mov	rax,1
+	mov	rbx,48
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_9:
+	mov	rax,2
+	mov	rbx,56
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_10:
+	mov	rax,3
+	mov	rbx,64
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_11:
+	mov	rax,4
+	mov	rbx,72
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_12:
+	mov	rax,5
+	mov	rbx,80
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_13:
+	mov	rax,6
+	mov	rbx,88
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_14:
+	mov	rax,7
+	mov	rbx,96
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_15:
+	mov	rax,8
+	mov	rbx,104
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_16:
+	mov	rax,9
+	mov	rbx,112
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_17:
+	mov	rax,10
+	mov	rbx,120
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_18:
+	mov	rax,11
+	mov	rbx,128
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_19:
+	mov	rax,12
+	mov	rbx,136
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_20:
+	mov	rax,13
+	mov	rbx,144
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_21:
+	mov	rax,14
+	mov	rbx,152
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_22:
+	mov	rax,15
+	mov	rbx,160
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_23:
+	mov	rax,16
+	mov	rbx,168
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_24:
+	mov	rax,17
+	mov	rbx,176
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_25:
+	mov	rax,18
+	mov	rbx,184
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_26:
+	mov	rax,19
+	mov	rbx,192
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_27:
+	mov	rax,20
+	mov	rbx,200
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_28:
+	mov	rax,21
+	mov	rbx,208
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_29:
+	mov	rax,22
+	mov	rbx,216
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_30:
+	mov	rax,23
+	mov	rbx,224
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_31:
+	mov	rax,24
+	mov	rbx,232
+	jmp	eval_upd_n
+
+ ifdef PROFILE
+	call	profile_n
+	mov	rbp,rax
+ endif
+eval_upd_32:
+	mov	rax,25
+	mov	rbx,240
+	jmp	eval_upd_n
+
+;
+;	STRINGS
+;
+
+catAC:
+	mov	rax,8[rcx]
+	mov	rbx,8[rdx]
+	lea	rbp,16+7[rax+rbx]
+	shr	rbp,3
+	sub	r15,rbp
+	jl	gc_3
+gc_r_3:
+	add	rcx,16
+	add	rdx,16
+
+; fill_node
+
+	mov	r8,rdi
+	mov	qword ptr [rdi],offset __STRING__+2
+
+; store length
+
+	lea	rbp,[rax+rbx]
+	mov	8[rdi],rbp 
+	add	rdi,16
+
+; copy string 1
+
+	lea	rbp,7[rbx]
+	shr	rbp,3
+	add	rbx,rdi 
+
+	xchg	rcx,rbp 
+	xchg	rsi,rdx 
+	cld
+	rep movsq
+	mov	rsi,rdx 
+	mov	rcx,rbp 
+
+	mov	rdi,rbx 
+
+; copy_string 2
+
+cat_string_6:
+	mov	rbp,rax 
+	shr	rbp,3
+	je	cat_string_9
+
+cat_string_7:
+	mov	rbx,[rcx]
+	add	rcx,8
+	mov	[rdi],rbx 
+	add	rdi,8
+	dec	rbp 
+	jne	cat_string_7
+	
+cat_string_9:
+	test	al,4
+	je	cat_string_10
+	mov	ebx,dword ptr [rcx]
+	add	rcx,4
+	mov	dword ptr [rdi],ebx 
+	add	rdi,4
+cat_string_10:
+	test	al,2
+	je	cat_string_11
+	mov	bx,word ptr [rcx]
+	add	rcx,2
+	mov	word ptr [rdi],bx 
+	add	rdi,2
+cat_string_11:
+	test	al,1
+	je	cat_string_12
+	mov	bl,byte ptr [rcx]
+	mov	byte ptr [rdi],bl 
+	inc	rdi
+cat_string_12:
+
+	mov	rcx,r8
+; align heap pointer
+	add	rdi,7
+	and	rdi,-8
+	ret
+
+gc_3:	call	collect_2
+	jmp	gc_r_3
+
+empty_string:
+	mov	rcx,offset zero_length_string
+	ret
+
+sliceAC:
+	mov	rbp,8[rcx]
+	test	rbx,rbx 
+	jns	slice_string_1
+	xor	rbx,rbx 
+slice_string_1:
+	cmp	rbx,rbp 
+	jge	empty_string
+	cmp	rax,rbx 
+	jl	empty_string
+	inc	rax 
+	cmp	rax,rbp 
+	jle	slice_string_2
+	mov	rax,rbp 
+slice_string_2:
+	sub	rax,rbx 
+
+	lea	rbp,(16+7)[rax]
+	shr	rbp,3
+
+	sub	r15,rbp
+	jl	gc_4
+r_gc_4:
+	sub	rbp,2
+	lea	rdx,16[rcx+rbx]
+
+	mov	qword ptr [rdi],offset __STRING__+2
+	mov	8[rdi],rax 
+
+; copy part of string
+	mov	rcx,rbp 
+	mov	rbp,rdi
+	add	rdi,16
+
+	xchg	rsi,rdx 
+	cld
+	rep movsq
+	mov	rsi,rdx 
+	mov	rcx,rbp 
+	ret
+
+gc_4:
+	mov	rbp,rdx 
+	call	collect_1
+	lea	rbp,(16+7)[rax]
+	shr	rbp,3
+	jmp	r_gc_4
+
+updateAC:
+	mov	rbp,8[rcx]
+	cmp	rbx,rbp 
+	jae	update_string_error
+
+	add	rbp,16+7
+	shr	rbp,3
+
+	sub	r15,rbp
+	jl	gc_5
+r_gc_5:
+	mov	rbp,8[rcx]
+	add	rbp,7
+	shr	rbp,3
+
+	mov	rdx,rcx 
+	mov	r8,rdi 
+	mov	qword ptr [rdi],offset __STRING__+2
+	mov	rcx,8[rdx]
+	add	rdx,16
+	mov	8[rdi],rcx 
+	add	rdi,16
+
+	add	rbx,rdi
+
+	mov	rcx,rbp 
+	xchg	rsi,rdx 
+	cld
+	rep movsq
+	mov	rsi,rdx 
+
+	mov	byte ptr [rbx],al 
+	mov	rcx,r8 
+	ret
+
+gc_5:	call	collect_1
+	jmp	r_gc_5
+
+update_string_error:
+	mov	rbp,offset high_index_string
+	test	rax,rax 
+	jns	update_string_error_2
+	mov	rbp,offset low_index_string
+update_string_error_2:
+	jmp	print_error
+
+eqAC:
+	mov	rax,8[rcx]
+	cmp	rax,8[rdx]
+	jne	equal_string_ne
+	add	rcx,16
+	add	rdx,16
+	mov	rbx,rax 
+	and	rbx,7
+	shr	rax,3
+	je	equal_string_d
+equal_string_1:
+	mov	rbp,[rcx]
+	cmp	rbp,[rdx]
+	jne	equal_string_ne
+	add	rcx,8
+	add	rdx,8
+	dec	rax
+	jne	equal_string_1
+equal_string_d:
+	test	bl,4
+	je	equal_string_w
+	mov	eax,dword ptr [rcx]
+	cmp	eax,dword ptr [rdx]
+	jne	equal_string_ne
+	add	rcx,4
+	add	rdx,4
+equal_string_w:
+	test	bl,2
+	je	equal_string_b
+	mov	ax,word ptr [rcx]
+	cmp	ax,word ptr [rdx]
+	jne	equal_string_ne
+	add	rcx,2
+	add	rdx,2
+equal_string_b:
+	test	bl,1
+	je	equal_string_eq
+	mov	bl,byte ptr [rcx]
+	cmp	bl,byte ptr [rdx]
+	jne	equal_string_ne
+equal_string_eq:
+	mov	rax,1
+	ret
+equal_string_ne:
+	xor	rax,rax 
+	ret
+
+cmpAC:
+	mov	rbx,8[rcx]
+	mov	rbp,8[rdx]
+	add	rcx,16
+	add	rdx,16
+	cmp	rbp,rbx 
+	jb	cmp_string_less
+	ja	cmp_string_more
+	xor	rax,rax 
+	jmp	cmp_string_chars
+cmp_string_more:
+	mov	rax,1
+	jmp	cmp_string_chars
+cmp_string_less:
+	mov	rax,-1
+	mov	rbx,rbp 
+	jmp	cmp_string_chars
+
+cmp_string_1:
+	mov	rbp,[rdx]
+	cmp	rbp,[rcx]
+	jne	cmp_string_ne_q
+	add	rdx,8
+	add	rcx,8
+cmp_string_chars:
+	sub	rbx,8
+	jnc	cmp_string_1
+cmp_string_d:
+	test	bl,4
+	je	cmp_string_w
+	mov	ebp,dword ptr [rdx]
+	cmp	ebp,dword ptr [rcx]
+	jne	cmp_string_ne_d
+	add	rdx,4
+	add	rcx,4
+cmp_string_w:
+	test	bl,2
+	je	cmp_string_b
+	mov	bpl,byte ptr [rdx]
+	cmp	bpl,byte ptr [rcx]
+	jne	cmp_string_ne
+	mov	bpl,byte ptr 1[rdx]
+	cmp	bpl,byte ptr 1[rcx]
+	jne	cmp_string_ne
+	add	rdx,2
+	add	rcx,2
+cmp_string_b:
+	test	bl,1
+	je	cmp_string_eq
+	mov	bl,byte ptr [rdx]
+	cmp	bl,byte ptr [rcx]
+	jne	cmp_string_ne
+cmp_string_eq:
+	ret
+cmp_string_ne_d:
+	mov	r10d,[rcx]
+	bswap	ebp
+	bswap	r10d
+	cmp	ebp,r10d
+	jmp	cmp_string_ne
+cmp_string_ne_q:
+	mov	r10,[rcx]
+	bswap	rbp
+	bswap	r10
+	cmp	rbp,r10
+cmp_string_ne:
+	ja	cmp_string_r1
+	mov	rax,-1
+	ret
+cmp_string_r1:
+	mov	rax,1
+	ret
+
+string_to_string_node:
+	mov	rax,qword ptr [rcx]
+	add	rcx,8
+
+	lea	rbx,16+7[rax]
+	shr	rbx,3
+
+	sub	r15,rbx
+	jl	string_to_string_node_gc
+
+string_to_string_node_r:
+	sub	rbx,2
+	mov	qword ptr [rdi],offset __STRING__+2
+	mov	qword ptr 8[rdi],rax 
+	mov	rbp,rdi
+	add	rdi,16
+	jmp	string_to_string_node_4
+	
+string_to_string_node_2:
+	mov	rax,qword ptr [rcx]
+	add	rcx,8
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+string_to_string_node_4:
+	sub	rbx,1
+	jge	string_to_string_node_2
+
+	mov	rcx,rbp 
+	ret
+
+string_to_string_node_gc:
+	push	rcx
+	call	collect_0
+	pop	rcx
+	jmp	string_to_string_node_r
+
+
+int_array_to_node:
+	mov	rax,qword ptr -16[rcx]
+	lea	rbx,3[rax]
+	sub	r15,rbx
+	jl	int_array_to_node_gc
+
+int_array_to_node_r:
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	rdx,rcx
+	mov	qword ptr 8[rdi],rax
+	mov	rcx,rdi
+	mov	qword ptr 16[rdi],offset dINT+2
+	add	rdi,24
+	jmp	int_or_real_array_to_node_4
+
+int_or_real_array_to_node_2:
+	mov	rbx,qword ptr [rdx]
+	add	rdx,8
+	mov	qword ptr [rdi],rbx
+	add	rdi,8
+int_or_real_array_to_node_4:
+	sub	rax,1
+	jge	int_or_real_array_to_node_2
+
+	ret
+
+int_array_to_node_gc:
+	push	rcx
+	call	collect_0
+	pop	rcx
+	jmp	int_array_to_node_r
+
+
+real_array_to_node:
+	mov	rax,qword ptr -16[rcx]
+	lea	rbx,3[rax]
+	sub	r15,rbx
+	jl	real_array_to_node_gc
+
+real_array_to_node_r:
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	rdx,rcx
+	mov	qword ptr 8[rdi],rax
+	mov	rcx,rdi
+	mov	qword ptr 16[rdi],offset REAL+2
+	add	rdi,24
+	jmp	int_or_real_array_to_node_4
+
+real_array_to_node_gc:
+	push	rcx
+	call	collect_0
+	pop	rcx
+	jmp	real_array_to_node_r
+
+
+	align	(1 shl 2)
+	dd	3
+_c3:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+
+	dd	4
+_c4:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	5
+_c5:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	6
+_c6:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	7
+_c7:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	8
+_c8:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	9
+_c9:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	10
+_c10:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	11
+_c11:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	12
+_c12:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	13
+_c13:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	14
+_c14:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	15
+_c15:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	16
+_c16:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	17
+_c17:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	18
+_c18:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	19
+_c19:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	20
+_c20:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	21
+_c21:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	22
+_c22:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	23
+_c23:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	24
+_c24:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	25
+_c25:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	26
+_c26:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	27
+_c27:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	28
+_c28:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	29
+_c29:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	30
+_c30:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	31
+_c31:	jmp	__cycle__in__spine
+	align	(1 shl 2)
+	dd	32
+_c32:	jmp	__cycle__in__spine
+
+;
+;	ARRAYS
+;
+
+_create_arrayB:
+	mov	rbx,rax 
+	add	rax,24+7
+	shr	rax,3
+	sub	r15,rax
+	jge	no_collect_4574
+	call	collect_0
+no_collect_4574:
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rbx 
+	mov	qword ptr 16[rdi],offset BOOL+2
+	lea	rdi,[rdi+rax*8]
+	ret
+
+_create_arrayC:
+	mov	rbx,rax
+	add	rax,16+7
+	shr	rax,3
+	sub	r15,rax
+	jge	no_collect_4573
+	call	collect_0
+no_collect_4573:
+	mov	rcx,rdi 
+	mov	qword ptr [rdi],offset __STRING__+2
+	mov	qword ptr 8[rdi],rbx 
+	lea	rdi,[rdi+rax*8]
+	ret
+
+_create_arrayI:
+	lea	rbp,3[rax]
+	sub	r15,rbp
+	jge	no_collect_4572
+	call	collect_0
+no_collect_4572:
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax
+	lea	rbp,dINT+2
+	mov	qword ptr 16[rdi],rbp
+	lea	rdi,24[rdi+rax*8]
+	ret
+
+_create_arrayI32:
+	mov	rbx,rax
+	add	rax,6+1
+	shr	rax,1
+	sub	r15,rax
+	jge	no_collect_3572
+	call	collect_0
+no_collect_3572:
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rbx
+	mov	qword ptr 16[rdi],offset INT32+2
+	lea	rdi,[rdi+rax*8]
+	ret
+
+_create_arrayR:
+	lea	rbp,3[rax]
+	sub	r15,rbp
+	jge	no_collect_4580
+	call	collect_0
+no_collect_4580:
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax
+	mov	qword ptr 16[rdi],offset REAL+2
+	lea	rdi,24[rdi+rax*8]
+	ret
+
+_create_arrayR32:
+	mov	rbx,rax
+	add	rax,6+1
+	shr	rax,1
+	sub	r15,rax
+	jge	no_collect_3580
+	call	collect_0
+no_collect_3580:
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax
+	mov	qword ptr 16[rdi],offset REAL32+2
+	lea	rdi,[rdi+rax*8]
+	ret
+
+; rax : number of elements, rbx: element descriptor
+; r10 : element size, r11 : element a size, rcx :a_element-> rcx : array
+
+_create_r_array:
+	mov	rbp,rax
+	imul	rbp,r10
+	add	rbp,3
+	sub	r15,rbp
+	jge	no_collect_4586
+	call	collect_1
+no_collect_4586:
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	rdx,rcx
+	mov	qword ptr 8[rdi],rax
+	mov	rcx,rdi 
+	mov	qword ptr 16[rdi],rbx
+	add	rdi,24
+
+	test	r11,r11
+	je	_create_r_array_0
+	sub	r11,2
+	jc	_create_r_array_1
+	je	_create_r_array_2
+	sub	r11,2
+	jc	_create_r_array_3
+	je	_create_r_array_4
+	jmp	_create_r_array_5
+
+_create_r_array_0:
+	imul	r10,rax
+	lea	rdi,[rdi+r10*8]
+	ret
+
+_create_r_array_1:
+	shl	r10,3
+	jmp	_st_fillr1_array
+_fillr1_array:
+	mov	qword ptr [rdi],rdx
+	add	rdi,r10 
+_st_fillr1_array:
+	sub	rax,1
+	jnc	_fillr1_array
+	ret
+
+_create_r_array_2:
+	shl	r10,3
+	jmp	_st_fillr2_array
+_fillr2_array:
+	mov	qword ptr [rdi],rdx
+	mov	qword ptr 8[rdi],rdx
+	add	rdi,r10 
+_st_fillr2_array:
+	sub	rax,1
+	jnc	_fillr2_array
+	ret
+
+_create_r_array_3:
+	shl	r10,3
+	jmp	_st_fillr3_array
+_fillr3_array:
+	mov	qword ptr [rdi],rdx
+	mov	qword ptr 8[rdi],rdx
+	mov	qword ptr 16[rdi],rdx
+	add	rdi,r10 
+_st_fillr3_array:
+	sub	rax,1
+	jnc	_fillr3_array
+	ret
+
+_create_r_array_4:
+	shl	r10,3
+	jmp	_st_fillr4_array
+_fillr4_array:
+	mov	qword ptr [rdi],rdx
+	mov	qword ptr 8[rdi],rdx
+	mov	qword ptr 16[rdi],rdx
+	mov	qword ptr 24[rdi],rdx
+	add	rdi,r10 
+_st_fillr4_array:
+	sub	rax,1
+	jnc	_fillr4_array
+	ret
+
+_create_r_array_5:
+	sub	r10,4
+	sub	r10,r11
+	sub	r11,1
+	shl	r10,3
+	jmp	_st_fillr5_array
+
+_fillr5_array:
+	mov	qword ptr [rdi],rdx
+	mov	qword ptr 8[rdi],rdx
+	mov	qword ptr 16[rdi],rdx
+	mov	qword ptr 24[rdi],rdx
+	add	rdi,32
+
+	mov	rbx,r11
+_copy_elem_5_lp:
+	mov	qword ptr [rdi],rdx
+	add	rdi,8
+	sub	rbx,1
+	jnc	_copy_elem_5_lp
+
+	add	rdi,r10 
+_st_fillr5_array:
+	sub	rax,1
+	jnc	_fillr5_array
+	
+	ret
+
+create_arrayB:
+	mov	r10,rbx
+	add	rbx,24+7
+	shr	rbx,3
+	sub	r15,rbx
+	jge	no_collect_4575
+	call	collect_0
+no_collect_4575:
+	mov	rbp,rax
+	sub	rbx,3
+	shl	rbp,8
+	or	rax,rbp
+	mov	rbp,rax
+	shl	rbp,16
+	or	rax,rbp
+	mov	rbp,rax 
+	shl	rbp,32
+	or	rax,rbp
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],r10
+	mov	qword ptr 16[rdi],offset BOOL+2
+	add	rdi,24
+	jmp	create_arrayBCI
+
+create_arrayC:
+	mov	r10,rbx
+	add	rbx,16+7
+	shr	rbx,3
+	sub	r15,rbx
+	jge	no_collect_4578
+	call	collect_0
+no_collect_4578:
+	mov	rbp,rax
+	sub	rbx,2
+	shl	rbp,8
+	or	rax,rbp
+	mov	rbp,rax
+	shl	rbp,16
+	or	rax,rbp
+	mov	rbp,rax
+	shl	rbp,32
+	or	rax,rbp
+	mov	rcx,rdi 
+	mov	qword ptr [rdi],offset __STRING__+2
+	mov	qword ptr 8[rdi],r10
+	add	rdi,16
+	jmp	create_arrayBCI
+
+create_arrayI32:
+	mov	r10,rbx
+	add	rbx,6+1
+	shr	rbx,1
+	sub	r15,rbx
+	jge	no_collect_3577
+	call	collect_0
+no_collect_3577:
+	mov	rcx,rdi 
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],r10
+	mov	qword ptr 16[rdi],offset INT32+2
+	add	rdi,24
+
+	sub	rbx,3
+
+	mov	ebp,eax
+	shl	rax,32
+	or	rax,rbp
+	jmp	create_arrayBCI
+
+create_arrayI:
+	lea	rbp,3[rbx]
+	sub	r15,rbp
+	jge	no_collect_4577
+	call	collect_0
+no_collect_4577:
+	mov	rcx,rdi 
+	lea	rbp,__ARRAY__+2
+	mov	qword ptr [rdi],rbp
+	mov	qword ptr 8[rdi],rbx 
+	lea	rbp,dINT+2
+	mov	qword ptr 16[rdi],rbp
+	add	rdi,24
+create_arrayBCI:
+	mov	rdx,rbx
+	shr	rbx,1
+	test	dl,1
+	je	st_filli_array
+
+	mov	qword ptr [rdi],rax 
+	add	rdi,8
+	jmp	st_filli_array
+
+filli_array:
+	mov	qword ptr [rdi],rax 
+	mov	qword ptr 8[rdi],rax 
+	add	rdi,16
+st_filli_array:
+	sub	rbx,1
+	jnc	filli_array
+
+	ret
+
+create_arrayR32:
+	cvtsd2ss	xmm0,xmm0
+	movss	dword ptr (-8)[rsp],xmm0
+	mov	r10,rax
+	add	rax,6+1
+	shr	rax,1
+	mov	ebx,dword ptr (-8)[rsp]
+	sub	r15,rax
+	jge	no_collect_3579
+	call	collect_0
+no_collect_3579:
+	mov	rcx,rdi 
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],r10
+	mov	qword ptr 16[rdi],offset REAL32+2
+	add	rdi,24
+
+	sub	rax,3
+
+	mov	edx,ebx
+	shl	rbx,32
+	or	rbx,rdx
+	jmp	st_fillr_array
+
+create_arrayR:
+	movsd	qword ptr (-8)[rsp],xmm0
+	lea	rbp,3[rax]
+
+	mov	rbx,qword ptr (-8)[rsp]
+
+	sub	r15,rbp
+	jge	no_collect_4579
+	call	collect_0
+no_collect_4579:
+	mov	rcx,rdi 
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax
+	mov	qword ptr 16[rdi],offset REAL+2
+	add	rdi,24
+	jmp	st_fillr_array
+fillr_array:
+	mov	qword ptr [rdi],rbx 
+	add	rdi,8
+st_fillr_array:
+	sub	rax,1
+	jnc	fillr_array
+
+	ret
+
+create_array:
+	lea	rbp,3[rax]
+	sub	r15,rbp
+	jge	no_collect_4576
+	call	collect_1
+no_collect_4576:
+	mov	rbx,rcx 
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax 
+	mov	qword ptr 16[rdi],0
+	add	rdi,24
+
+	jmp	fillr1_array
+
+
+
+
+; in rax: number of elements, rbx: element descriptor
+; r10 : element size, r11 : element a size -> rcx : array
+
+create_R_array:
+	sub	r10,2
+	jc	create_R_array_1
+	je	create_R_array_2
+	sub	r10,2
+	jc	create_R_array_3
+	je	create_R_array_4
+	jmp	create_R_array_5
+
+create_R_array_1:
+	lea	rbp,3[rax]
+	sub	r15,rbp
+	jge	no_collect_4581
+	call	collect_0
+no_collect_4581:
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax 
+	mov	qword ptr 16[rdi],rbx 
+	add	rdi,24
+
+	test	r11,r11
+	je	r_array_1_b
+
+	mov	rbx,qword ptr (-8)[rsi]
+	jmp	fillr1_array
+
+r_array_1_b:
+	mov	rbx,qword ptr 8[rsp]
+
+fillr1_array:
+	mov	rdx,rax 
+	shr	rax,1
+	test	dl,1
+	je	st_fillr1_array_1
+
+	mov	qword ptr [rdi],rbx
+	add	rdi,8
+	jmp	st_fillr1_array_1
+
+fillr1_array_lp:
+	mov	qword ptr [rdi],rbx 
+	mov	qword ptr 8[rdi],rbx 
+	add	rdi,16
+st_fillr1_array_1:
+	sub	rax,1
+	jnc	fillr1_array_lp
+
+	ret
+
+create_R_array_2:
+	lea	rbp,3[rax*2]
+	sub	r15,rbp
+	jge	no_collect_4582
+	call	collect_0
+no_collect_4582:
+	mov	rcx,rdi 
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax 
+	mov	qword ptr 16[rdi],rbx 
+	add	rdi,24
+
+	sub	r11,1
+	jc	r_array_2_bb
+	je	r_array_2_ab
+r_array_2_aa:
+	mov	rbx,qword ptr (-8)[rsi]
+	mov	rbp,qword ptr (-16)[rsi]
+	jmp	st_fillr2_array
+r_array_2_ab:
+	mov	rbx,qword ptr (-8)[rsi]
+	mov	rbp,qword ptr 8[rsp]
+	jmp	st_fillr2_array
+r_array_2_bb:
+	mov	rbx,qword ptr 8[rsp]
+	mov	rbp,qword ptr 16[rsp]
+	jmp	st_fillr2_array
+
+fillr2_array_1:
+	mov	qword ptr [rdi],rbx 
+	mov	qword ptr 8[rdi],rbp 
+	add	rdi,16
+st_fillr2_array:
+	sub	rax,1
+	jnc	fillr2_array_1
+
+	ret
+
+create_R_array_3:
+	lea	rbp,3[rax+rax*2]
+	sub	r15,rbp
+	jge	no_collect_4583
+	call	collect_0
+no_collect_4583:
+	mov	rcx,rdi
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax 
+	mov	qword ptr 16[rdi],rbx 
+	add	rdi,24
+
+	pop	rdx 
+	mov	r12,rsp 
+
+	test	r11,r11 
+	je	r_array_3
+	
+	lea	r13,0[r11*8]
+	mov	rbp,rsi
+	sub	rbp,r13
+
+	sub	r11,1
+
+copy_a_to_b_lp3:
+	push	[rbp]
+	add	rbp,8
+	sub	r11,1
+	jnc	copy_a_to_b_lp3
+
+r_array_3:
+	mov	rbx,qword ptr [rsp]
+	mov	r13,qword ptr 8[rsp]
+	mov	rbp,qword ptr 16[rsp]
+	
+	mov	rsp,r12
+	push	rdx
+
+	jmp	st_fillr3_array
+
+fillr3_array_1:
+	mov	qword ptr [rdi],rbx 
+	mov	qword ptr 8[rdi],r13
+	mov	qword ptr 16[rdi],rbp 
+	add	rdi,24
+st_fillr3_array:
+	sub	rax,1
+	jnc	fillr3_array_1
+
+	ret
+
+create_R_array_4:
+	lea	rbp,3[rax+4]
+	sub	r15,rbp
+	jge	no_collect_4584
+	call	collect_0
+no_collect_4584:
+	mov	rcx,rdi 
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax 
+	mov	qword ptr 16[rdi],rbx 
+	add	rdi,24
+
+	pop	rdx 
+	mov	r12,rsp 
+
+	test	r11,r11 
+	je	r_array_4
+
+	lea	r13,0[r11*8]
+	mov	rbp,rsi
+	sub	rbp,r13
+	sub	r11,1
+
+copy_a_to_b_lp4:
+	push	[rbp]
+	add	rbp,8
+	sub	r11,1
+	jnc	copy_a_to_b_lp4
+
+r_array_4:
+	mov	rbx,qword ptr [rsp]
+	mov	r13,qword ptr 8[rsp]
+	mov	r14,qword ptr 16[rsp]
+	mov	rbp,qword ptr 24[rsp]
+
+	mov	rsp,r12
+	push	rdx
+	
+	jmp	st_fillr4_array
+
+fillr4_array:
+	mov	qword ptr [rdi],rbx 
+	mov	qword ptr 8[rdi],r13 
+	mov	qword ptr 16[rdi],r14 
+	mov	qword ptr 24[rdi],rbp 
+	add	rdi,32
+st_fillr4_array:
+	sub	rax,1
+	jnc	fillr4_array
+
+	ret
+
+create_R_array_5:
+	lea	r12,4[r10]
+	mov	rbp,rax
+	imul	rbp,r12
+	add	rbp,3
+	sub	r15,rbp
+	jge	no_collect_4585
+	call	collect_0
+no_collect_4585:
+	mov	qword ptr [rdi],offset __ARRAY__+2
+	mov	qword ptr 8[rdi],rax
+	mov	qword ptr 16[rdi],rbx
+	mov	rcx,rdi
+	add	rdi,24
+
+	pop	rdx
+	mov	r12,rsp
+	
+	test	r11,r11 
+	je	r_array_5
+
+	lea	r13,0[r11*8]
+	mov	rbp,rsi
+	sub	rbp,r13
+	sub	r11,1
+
+copy_a_to_b_lp5:
+	push	[rbp]
+	add	rbp,8
+	sub	r11,1
+	jnc	copy_a_to_b_lp5
+
+r_array_5:
+	mov	r13,qword ptr [rsp]
+	mov	r14,qword ptr 8[rsp]
+ if THREAD
+	lea	rbx,32[rsp+r10*8]
+ endif
+	mov	r8,qword ptr 16[rsp]
+ if THREAD
+	mov	r10,qword ptr 24[rsp]
+ else
+	mov	r9,qword ptr 24[rsp]
+ endif
+	add	rsp,32
+
+ ife THREAD
+	sub	r10,1
+ endif
+	jmp	st_fillr5_array
+
+fillr5_array_1:
+	mov	qword ptr [rdi],r13 
+	mov	qword ptr 8[rdi],r14
+
+	mov	r11,rsp
+ ife THREAD
+	mov	rbx,r10 
+ endif
+
+	mov	qword ptr 16[rdi],r8
+ if THREAD
+	mov	qword ptr 24[rdi],r10
+ else
+	mov	qword ptr 24[rdi],r9
+ endif
+	add	rdi,32
+
+copy_elem_lp5:
+	mov	rbp,qword ptr [r11]
+	add	r11,8
+	mov	qword ptr [rdi],rbp 
+	add	rdi,8
+ if THREAD
+	cmp	r11,rbx
+	jne	copy_elem_lp5
+ else
+	sub	rbx,1
+	jnc	copy_elem_lp5
+ endif
+
+st_fillr5_array:
+	sub	rax,1
+	jnc	fillr5_array_1
+
+	mov	rsp,r12
+	jmp	rdx
+
+ ifndef NEW_DESCRIPTORS
+yet_args_needed:
+; for more than 4 arguments
+	mov	r10,[rdx]
+	movzx	rax,word ptr (-2)[r10]
+	add	rax,3
+	sub	r15,rax
+	jl	gc_1
+gc_r_1:	sub	rax,3+1+4
+	mov	rbx,8[rdx]
+	add	r10,8
+	mov	rdx,16[rdx]
+	mov	rbp,rdi
+	mov	r8,[rdx]
+	mov	[rdi],r8 
+	mov	r8,8[rdx]
+	mov	8[rdi],r8 
+	mov	r8,16[rdx]
+	mov	16[rdi],r8 
+	add	rdx,24
+	add	rdi,24
+
+cp_a:	mov	r8,[rdx]
+	add	rdx,8
+	mov	[rdi],r8 
+	add	rdi,8
+	sub	rax,1
+	jge	cp_a
+
+	mov	[rdi],rcx 
+	mov	8[rdi],r10 
+	lea	rcx,8[rdi]
+	mov	16[rdi],rbx 
+	mov	24[rdi],rbp 
+	add	rdi,32
+	ret
+
+gc_1:
+	call	collect_2
+	jmp	gc_r_1
+
+yet_args_needed_0:
+	sub	r15,2
+	jl	gc_20
+gc_r_20:
+	mov	8[rdi],rcx 
+	mov	rax,[rdx]
+	mov	rcx,rdi
+	add	rax,8
+	mov	[rdi],rax 
+	add	rdi,16
+	ret
+
+gc_20:
+	call	collect_2
+	jmp	gc_r_20
+
+yet_args_needed_1:
+	sub	r15,3
+	jl	gc_21
+gc_r_21:
+	mov	16[rdi],rcx 
+	mov	rax,[rdx]
+	mov	rcx,rdi
+	add	rax,8
+	mov	[rdi],rax 
+	mov	rbx,8[rdx]
+	mov	8[rdi],rbx 
+	add	rdi,24
+	ret
+
+gc_21:
+	call	collect_2
+	jmp	gc_r_21
+
+yet_args_needed_2:
+	sub	r15,5
+	jl	gc_22
+gc_r_22:
+	mov	rax,[rdx]
+	mov	8[rdi],rcx 
+	add	rax,8
+	mov	rbp,8[rdx]
+	mov	16[rdi],rax 
+	lea	rcx,16[rdi]
+	mov	24[rdi],rbp 
+	mov	rbp,16[rdx]
+	mov	[rdi],rbp 
+	mov	32[rdi],rdi 
+	add	rdi,40
+	ret
+
+gc_22:
+	call	collect_2
+	jmp	gc_r_22
+
+yet_args_needed_3:
+	sub	r15,6
+	jl	gc_23
+gc_r_23:
+	mov	rax,[rdx]
+	mov	16[rdi],rcx 
+	add	rax,8
+	mov	rbp,8[rdx]
+	mov	24[rdi],rax 
+	mov	rdx,16[rdx]
+	mov	32[rdi],rbp 
+	mov	rbp,[rdx]
+	mov	40[rdi],rdi 
+	mov	[rdi],rbp 
+	mov	rbp,8[rdx]
+	lea	rcx,24[rdi]
+	mov	8[rdi],rbp 
+	add	rdi,48
+	ret
+
+gc_23:
+	call	collect_2
+	jmp	gc_r_23
+
+yet_args_needed_4:
+	sub	r15,7
+	jl	gc_24
+gc_r_24:
+	mov	rax,[rdx]
+	mov	24[rdi],rcx 
+	add	rax,8
+	mov	rbp,8[rdx]
+	mov	32[rdi],rax 
+	mov	rdx,16[rdx]
+	mov	40[rdi],rbp 
+	mov	rbp,[rdx]
+	mov	48[rdi],rdi 
+	mov	[rdi],rbp 
+	mov	rbp,8[rdx]
+	lea	rcx,32[rdi]
+	mov	8[rdi],rbp 
+	mov	rbp,16[rdx ]
+	mov	16[rdi],rbp 
+	add	rdi,56
+	ret
+
+gc_24:
+	call	collect_2
+	jmp	gc_r_24
+ endif
+
+repl_args_b:
+	test	rax,rax 
+	jle	repl_args_b_1
+
+	dec	rax 
+	je	repl_args_b_4
+
+	mov	rdx,16[rcx]
+	sub	rbx,2
+	jne	repl_args_b_2
+
+	mov	[rsi],rdx 
+	add	rsi,8
+	jmp	repl_args_b_4
+
+repl_args_b_2:
+	lea	rdx,[rdx+rax*8]
+
+repl_args_b_3:
+	mov	rbp,(-8)[rdx]
+	sub	rdx,8
+	mov	[rsi],rbp 
+	add	rsi,8
+	dec	rax 
+	jne	repl_args_b_3
+
+repl_args_b_4:
+	mov	rbp,8[rcx]
+	mov	[rsi],rbp 
+	add	rsi,8
+repl_args_b_1:
+	ret
+
+push_arg_b:
+	cmp	rbx,2
+	jb	push_arg_b_1
+	jne	push_arg_b_2
+	cmp	rbx,rax 
+	je	push_arg_b_1
+push_arg_b_2:
+	mov	rcx,16[rcx]
+	sub	rbx,2
+push_arg_b_1:
+	mov	rcx,[rcx+rbx*8]
+	ret
+
+del_args:
+	mov	rbx,[rcx]
+	sub	rbx,rax 
+	movsx	rax,word ptr (-2)[rbx]
+	sub	rax,2
+	jge	del_args_2
+
+	mov	[rdx],rbx 
+	mov	rbp,8[rcx]
+	mov	8[rdx],rbp 
+	mov	rbp,16[rcx]
+	mov	16[rdx],rbp 
+	ret
+
+del_args_2:
+	jne	del_args_3
+
+	mov	[rdx],rbx 
+	mov	rbp,8[rcx]
+	mov	8[rdx],rbp 
+	mov	rbp,16[rcx]
+	mov	rbp,[rbp]
+	mov	16[rdx],rbp 
+	ret
+
+del_args_3:
+	sub	r15,rax
+	jl	del_args_gc
+del_args_r_gc:
+	mov	[rdx],rbx 
+	mov	16[rdx],rdi 
+	mov	rbp,8[rcx]
+	mov	rcx,16[rcx]
+	mov	8[rdx],rbp 
+
+del_args_copy_args:
+	mov	rbp,[rcx]
+	add	rcx,8
+	mov	[rdi],rbp 
+	add	rdi,8
+	sub	rax,1
+	jg	del_args_copy_args
+
+	ret
+
+del_args_gc:
+	call	collect_2
+	jmp	del_args_r_gc
+
+ ifdef USE_LIBM
+cos_real:
+	mov	rbp,rsp
+	and	rsp,-16
+	mov	r13,rsi
+	mov	r14,rdi
+	call	cos
+	mov	rsp,rbp
+	mov	rsi,r13
+	mov	rdi,r14
+	ret
+
+sin_real:
+	mov	rbp,rsp
+	and	rsp,-16
+	mov	r13,rsi
+	mov	r14,rdi
+	call	sin
+	mov	rsp,rbp
+	mov	rsi,r13
+	mov	rdi,r14
+	ret
+
+tan_real:
+	mov	rbp,rsp
+	and	rsp,-16
+	mov	r13,rsi
+	mov	r14,rdi
+	call	tan
+	mov	rsp,rbp
+	mov	rsi,r13
+	mov	rdi,r14
+	ret
+
+atan_real:
+	mov	rbp,rsp
+	and	rsp,-16
+	mov	r13,rsi
+	mov	r14,rdi
+	call	atan
+	mov	rsp,rbp
+	mov	rsi,r13
+	mov	rdi,r14
+	ret
+
+asin_real:
+acos_real:
+ln_real:
+log10_real:
+exp_real:
+pow_real:
+exp2_real_:
+_c_log10:
+_c_pow:
+_c_entier:
+	int 3
+	ret
+ endif
+	
+entier_real:
+	cvttsd2si rax,xmm0
+	ucomisd	xmm0,qword ptr real_0_0
+	jb	entier_real_m
+	ret
+
+entier_real_m:
+	movsd	qword ptr (-8)[rsp],xmm0
+	mov	rcx,qword ptr (-8)[rsp]
+	mov	rbx,rcx
+	shr rcx,52
+	cmp rcx,0bffh
+	jb	entier_real_m_small
+	cmp	rcx,0bffh+52
+	jae	entier_real_m_large
+	sub	rcx,0bffh-12
+	shl	rbx,cl
+	je	entier_m_exact
+entier_real_m_small:
+	sub	rax,1
+entier_real_m_large:
+entier_m_exact:
+	ret
+
+r_to_i_real:
+	cvtsd2si rax,xmm0
+	ret
+
+	public	getheapend
+
+getheapend:
+	lea	rbx,[rdi+r15*8]
+ if THREAD
+	mov	rax,heap_end_after_gc_offset[r9]
+ else
+	mov	rax,heap_end_after_gc
+ endif
+	ret
+
+_TEXT	ends
+
+	include	..\areals.asm
+
+ ifdef PROFILE
+  ifdef TRACE
+	include	..\atrace.asm
+  else
+	include	..\aprofile.asm
+  endif
+ endif
+
+ ifdef NEW_DESCRIPTORS
+	include	aap.asm
+ endif
+
+ ifdef THREAD
+	include	athread.asm
+ endif
+
+	end
+
diff --git a/thread/athread.asm b/thread/athread.asm
new file mode 100644
index 0000000..38c5bef
--- /dev/null
+++ b/thread/athread.asm
@@ -0,0 +1,371 @@
+
+;	File:	athread.asm
+;	Author:	John van Groningen
+;	Machine:	amd64
+
+	_TEXT segment
+
+ ifdef LINUX
+ else
+	extern	GetProcessHeap:near
+	extern	HeapAlloc:near
+	extern	HeapFree:near
+ endif
+
+	public	clean_new_thread
+
+clean_new_thread:
+ ifdef LINUX
+	push	rbx
+	push	rbp
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+
+	sub	rsp,24
+	mov	qword ptr 8[rsp],rdi
+
+	mov	rdi,768
+	call	malloc
+
+	mov	rbx,rax
+
+	mov	rdi,qword ptr tlsp_tls_index
+	mov	rsi,rax
+	call	pthread_setspecific
+
+	mov	r9,rbx
+
+	mov	rdi,qword ptr 8[rsp]
+
+	mov	rbx,qword ptr 8[rdi]
+ else
+	sub	rsp,56
+	mov	qword ptr 32[rsp],rcx
+
+	call	GetProcessHeap
+
+	mov	rcx,rax
+	xor	rdx,rdx
+	mov	r8,512
+	call	HeapAlloc
+
+	mov	r9,rax
+	mov	rax,qword ptr tlsp_tls_index
+	mov	qword ptr gs:[1480h+rax*8],r9
+
+	mov	rcx,qword ptr 32[rsp]
+
+	mov	rbx,qword ptr 8[rcx]
+ endif
+	test	rbx,rbx
+	cmove	rbx,qword ptr (main_thread_local_storage+heap_size_offset)
+	mov	qword ptr heap_size_offset[r9],rbx
+
+ ifdef LINUX
+	mov	rbx,qword ptr 16[rdi]
+ else
+	mov	rbx,qword ptr 16[rcx]
+ endif
+	test	rbx,rbx
+	cmove	rbx,qword ptr (main_thread_local_storage+a_stack_size_offset)
+	mov	qword ptr a_stack_size_offset[r9],rbx
+	
+	call	init_thread
+
+ ifndef LINUX
+	mov	rcx,qword ptr 32[rsp]
+ endif
+
+	mov	qword ptr saved_heap_p_offset[r9],rdi
+	mov	qword ptr saved_r15_offset[r9],r15
+	mov	qword ptr saved_a_stack_p_offset[r9],rsi
+
+	mov	qword ptr halt_sp_offset[r9],rsp
+
+ ifdef LINUX
+	mov	rdi,qword ptr 8[rsp]
+	call	qword ptr [rdi]
+
+	mov	rdi,qword ptr tlsp_tls_index
+	call	pthread_getspecific
+	mov	r9,rax
+	add	rsp,24
+ else
+	call	qword ptr [rcx]
+
+	mov	r9,qword ptr tlsp_tls_index
+	mov	r9,qword ptr gs:[1480h+r9*8]
+	add	rsp,56
+ endif
+
+	jmp	exit_thread
+
+init_thread:
+	lea	rax,128[rsp]
+	sub	rsp,32+8
+
+	mov	rax,qword ptr heap_size_offset[r9]
+	sub	rax,7
+	xor	rdx,rdx 
+	mov	rbx,65
+	div	rbx
+	mov	qword ptr heap_size_65_offset[r9],rax
+
+	mov	rax,qword ptr heap_size_offset[r9]
+	sub	rax,7
+	xor	rdx,rdx 
+	mov	rbx,257
+	div	rbx
+
+	mov	heap_size_257_offset[r9],rax
+
+	add	rax,7
+	and	rax,-8
+
+	mov	qword ptr heap_copied_vector_size_offset[r9],rax
+	mov	qword ptr heap_end_after_copy_gc_offset[r9],0
+
+	mov	rax,qword ptr heap_size_offset[r9]
+	add	rax,7
+	and	rax,-8
+	mov	qword ptr heap_size_offset[r9],rax 
+	add	rax,7
+
+	mov	rbp,rsp
+	and	rsp,-16
+
+	mov	rbx,r9
+
+ ifdef LINUX
+	mov	rdi,rax
+	call	malloc
+ else
+	mov	rcx,rax
+	call	allocate_memory
+ endif
+	mov	rsp,rbp
+
+	mov	r9,rbx
+
+	test	rax,rax 
+	je	init_thread_no_memory_2
+
+	mov	heap_mbp_offset[r9],rax
+
+	lea	rdi,7[rax]
+	and	rdi,-8
+
+	mov	heap_p_offset[r9],rdi 
+
+	mov	rbp,rsp
+	and	rsp,-16
+
+	mov	rbx,r9
+
+ ifdef LINUX
+	mov	r14,rdi
+	mov	rdi,qword ptr a_stack_size_offset[r9]
+	add	rdi,7
+	call	malloc
+	mov	rdi,r14
+ else
+	mov	rcx,qword ptr a_stack_size_offset[r9]
+	add	rcx,7
+  if 0
+	call	allocate_memory_with_guard_page_at_end
+  else
+	call	allocate_memory
+  endif
+ endif
+	mov	rsp,rbp
+
+	mov	r9,rbx
+	
+	test	rax,rax 
+	je	init_thread_no_memory_3
+
+	mov	stack_mbp_offset[r9],rax 
+
+	add	rax,qword ptr a_stack_size_offset[r9]
+	add	rax,7+4095
+	and	rax,-4096
+;	mov	qword ptr a_stack_guard_page,rax 
+	sub	rax,qword ptr a_stack_size_offset[r9]
+
+	add	rax,7
+	and	rax,-8
+
+	mov	rsi,rax
+
+	mov	stack_p_offset[r9],rax
+
+;	lea	rcx,(caf_list+8)
+;	mov	qword ptr caf_listp,rcx 
+
+;	lea	rcx,__Nil-8
+;	mov	qword ptr finalizer_list,rcx
+;	mov	qword ptr free_finalizer_list,rcx
+
+	mov	heap_p1_offset[r9],rdi
+
+	mov	rbp,qword ptr heap_size_257_offset[r9]
+
+	shl	rbp,4
+	lea	rax,[rdi+rbp*8]
+
+	mov	heap_copied_vector_offset[r9],rax
+	add	rax,heap_copied_vector_size_offset[r9]
+	mov	heap_p2_offset[r9],rax
+
+	mov	byte ptr garbage_collect_flag_offset[r9],0
+
+	test	byte ptr flags,64
+	je	init_thread_no_mark1
+
+	mov	rax,qword ptr heap_size_65_offset[r9]
+
+	mov	qword ptr heap_vector_offset[r9],rdi 
+
+	add	rdi,rax
+
+	add	rdi,7
+	and	rdi,-8
+
+	mov	qword ptr heap_p3_offset[r9],rdi
+
+	lea	rbp,[rax*8]
+
+	mov	byte ptr garbage_collect_flag_offset [r9],-1
+
+init_thread_no_mark1:
+;	mov	rax,qword ptr initial_heap_size
+	mov	rax,qword ptr heap_size_offset[r9]
+
+	mov	rbx,4000
+	test	byte ptr flags,64
+	jne	init_thread_no_mark9
+	add	rbx,rbx 
+init_thread_no_mark9:
+
+	cmp	rax,rbx 
+	jle	init_thread_too_large_or_too_small
+	shr	rax,3
+	cmp	rax,rbp 
+	jge	init_thread_too_large_or_too_small
+	mov	rbp,rax 
+init_thread_too_large_or_too_small:
+
+	lea	rax,[rdi+rbp*8]
+
+	mov	heap_end_after_gc_offset[r9],rax
+
+	test	byte ptr flags,64
+	je	init_thread_no_mark2
+
+	mov	qword ptr bit_vector_size_offset[r9],rbp
+
+init_thread_no_mark2:
+	mov	qword ptr bit_counter_offset[r9],0
+	mov	qword ptr zero_bits_before_mark_offset[r9],0
+
+	mov	r15,rbp
+
+	add	rsp,32+8
+	xor	rax,rax
+	ret
+
+init_thread_no_memory_2:
+	mov	rax,1
+	ret
+
+init_thread_no_memory_3:
+	mov	rbp,rsp
+	and	rsp,-16
+
+	mov	rbx,r9
+
+ ifdef LINUX
+	mov	rdi,heap_mbp_offset[r9]
+	call	free
+ else
+	mov	rcx,heap_mbp_offset[r9]
+	call	free_memory
+ endif
+
+	mov	rsp,rbp
+
+	mov	r9,rbx
+
+	add	rsp,32
+
+	mov	rax,1
+	ret
+
+
+exit_thread:
+	call	add_execute_time
+
+	mov	rbp,rsp
+	and	rsp,-16
+
+	mov	rbx,r9
+
+ ifdef LINUX
+	mov	rdi,stack_mbp_offset[r9]
+	call	free
+
+	mov	r9,rbx
+
+	mov	rdi,heap_mbp_offset[r9]
+	call	free
+
+	mov	rdi,r9
+	call	free
+ else
+	mov	rcx,stack_mbp_offset[r9]
+	sub	rsp,32
+	call	free_memory
+
+	mov	r9,rbx
+
+	mov	rcx,heap_mbp_offset[r9]
+	call	free_memory
+
+	call	GetProcessHeap
+
+	mov	rcx,rax
+	xor	rdx,rdx
+	mov	r8,rbx
+	call	HeapFree
+
+	add	rsp,32
+ endif
+
+	mov	rsp,rbp
+
+ ifdef LINUX
+	pop	r15
+	pop	r14
+	pop	r13
+	pop	r12
+	pop	rbp
+	pop	rbx
+ endif
+
+	xor	rax,rax
+	ret
+
+_TEXT	ends
+
+; bit_counter_offset = 0 ?
+; zero_bits_before_mark_offset = 1 =0 ?
+
+; a_stack_guard_page
+; caf_list
+; caf_listp
+; finalizer_list
+; free_finalizer_list
+; initial_heap_size
+; flags ?
-- 
cgit v1.2.3