diff --git a/Makefile b/Makefile
index 1ec14c4..b6c1bf2 100644
--- a/Makefile
+++ b/Makefile
@@ -11,7 +11,7 @@ kpc: vm/Makefile
 	@cd vm && make --no-print-directory -s verbose=yes
 
 kas: kpc as/regs.lst as/k-as.py
-	@cp vm/in/instrs.lst as
+	@cp vm/ob/instrs.lst as
 
 dos: kas dosclean
 	@cd ka && make --no-print-directory -s
@@ -24,7 +24,7 @@ dosclean:
 	@rm -f $(KODIR)/*.com $(KODIR)/*.sym 
 
 inclean:
-	@rm -f vm/in/arch_i.h vm/in/instrs.lst vm/ob/in/*.o
+	@rm -f vm/ob/arch_i.h vm/ob/instrs.lst vm/ob/in/*.o
 
 clean: dosclean
 	@cd vm && make clean --no-print-directory -s verbose=no
diff --git a/ka/command.k b/ka/command.k
index d2bce5f..cf7337c 100644
--- a/ka/command.k
+++ b/ka/command.k
@@ -13,7 +13,6 @@ start:
     mov rbp, zero
 
     cls
-    cld
     call main
 
     mov rax, Sys.EnterHaltMode
diff --git a/ka/crt/crt.k b/ka/crt/crt.k
index bbcf983..81a369a 100644
--- a/ka/crt/crt.k
+++ b/ka/crt/crt.k
@@ -10,4 +10,5 @@ include "crt/err/errno.k"
 include "crt/fmt/format.k"
 include "crt/str/string.k"
 include "crt/lib/time.k"
+include "crt/lib/arith128.k"
 
diff --git a/ka/crt/fmt/doprnt.k b/ka/crt/fmt/doprnt.k
index 3ee0ba7..4156a26 100644
--- a/ka/crt/fmt/doprnt.k
+++ b/ka/crt/fmt/doprnt.k
@@ -34,7 +34,7 @@ doprnt:
     mov ax0, b[rbx]
     call .doput
 
-    inc rbx
+    add rbx, rbx, 1
     jmp .print_regular
 
 .check_modf:
@@ -71,7 +71,7 @@ doprnt:
     cmp ax0, zero
     jmp.z .main_loop
 
-    inc rdi
+    add rdi, rdi, 1
     call .doput
     jmp .print_string
 
@@ -82,10 +82,8 @@ doprnt:
     jmp .main_loop
 
 .modf_p:
-    mov ax0, '0'
-    call .doput
-    mov ax0, 'x'
-    call .doput
+    call .doput, '0'
+    call .doput, 'x'
     ; Fallthrough
 
 .modf_x:
@@ -122,44 +120,29 @@ doprnt:
     add.z rsp, rsp, 80
     jmp.z .main_loop
 
-    inc rdi
+    add rdi, rdi, 1
     call .doput
     jmp .print_itoa_buf
 
 .modf_percent:
-    mov ax0, '%'
-    call .doput
+    call .doput, '%'
     jmp .main_loop
 
 .bad_modifier:
     ; print "%?" to clearly indicate that something is wrong
-    mov ax0, '%'
-    call .doput
-
-    mov ax0, '?'
-    call .doput
+    call .doput, '%'
+    call .doput, '?'
 
     jmp .main_loop
 
 .nullstring:
     ; %s was passed a NULL
-    mov ax0, '('
-    call .doput
-
-    mov ax0, 'n'
-    call .doput
-
-    mov ax0, 'u'
-    call .doput
-
-    mov ax0, 'l'
-    call .doput
-
-    mov ax0, 'l'
-    call .doput
-
-    mov ax0, ')'
-    call .doput
+    call .doput, '('
+    call .doput, 'n'
+    call .doput, 'u'
+    call .doput, 'l'
+    call .doput, 'l'
+    call .doput, ')'
 
     jmp .main_loop
 
@@ -178,7 +161,7 @@ doprnt:
 ;
 .doput:
     ; update print count
-    inc nx1
+    add nx1, nx1, 1
 
     ; if n==0, don't print
     ; we follow the C convention that sprintf()-like functions
@@ -188,7 +171,7 @@ doprnt:
     ret.z
 
     ; if n>0, decrement n and print
-    dec nx0
+    sub nx0, nx0, 1
     call nx3
     
     ; did putc fail?
diff --git a/ka/crt/fmt/ltostr.k b/ka/crt/fmt/ltostr.k
index 6bd1686..57270a3 100644
--- a/ka/crt/fmt/ltostr.k
+++ b/ka/crt/fmt/ltostr.k
@@ -34,10 +34,9 @@ ltostr:
     b.z ax3, zero, .conv
     b.nz ax2, 10, .conv     ; base 10
 
-    sgn r11, ax1    ; extract ax1 sign
-
-    cmp r11, -1     ; negative?
-    neg.z ax1, ax1
+    shr r11, ax1, 63        ; extract ax1 sign
+    cmp r11, zero           ; negative?
+    sub.nz ax1, zero, ax1   ; yes
 
     ; main loop
 .conv:
@@ -55,7 +54,7 @@ ltostr:
 
 .next:
     mov b[ax0], r10
-    inc ax0
+    add ax0, ax0, 1
 
     div ax1, ax1, ax2
     jmp .conv
@@ -64,12 +63,11 @@ ltostr:
 .fini:
     cmp r11, -1
     mov.z b[ax0], '-'
-    inc.z ax0
+    add.z ax0, ax0, 1
 
     mov b[ax0], zero
 
-    mov ax0, rax
-    call strrev2
+    call strrev2, rax
 
     ret
 
diff --git a/ka/crt/fmt/strtol.k b/ka/crt/fmt/strtol.k
index ad0e290..c3d1361 100644
--- a/ka/crt/fmt/strtol.k
+++ b/ka/crt/fmt/strtol.k
@@ -39,12 +39,12 @@ strtoq:
 
 .skip_spc:
     cmp b[rdx], ' '
-    inc.z rdx
+    add.z rdx, rdx, 1
     jmp.z .skip_spc
 
     ; skip +
     cmp b[rdx], '+'
-    inc.z rdx
+    add.z rdx, rdx, 1
 
     ; signed?
     cmp ax2, zero
@@ -53,7 +53,7 @@ strtoq:
 
     ; parse '-'
     cmp b[rdx], '-'
-    inc.z rdx
+    add.z rdx, rdx, 1
     mov.z r10, 1
     mov.nz r10, zero
 
@@ -64,7 +64,7 @@ strtoq:
     ; base prefix?
     b.nz b[rdx], '0', .main_loop
 
-    inc rdx
+    add rdx, rdx, 1
     movzx rcx, b[rdx]
 
     ; "0x"/"0b" prefix
@@ -81,7 +81,7 @@ strtoq:
     ; if not, leave rax = 0 and *rdx = 'x'
     b.nz ax1, 16, .done
     ; else
-    inc rdx
+    add rdx, rdx, 1
     jmp .main_loop
 
 .parsed_0b:
@@ -89,7 +89,7 @@ strtoq:
     ; if not, leave rax = 0 and *rdx = 'b'
     b.nz ax1, 2, .done
     ; else
-    inc rdx
+    add rdx, rdx, 1
     jmp .main_loop
 
 .base_0:
@@ -98,15 +98,15 @@ strtoq:
     cmp b[rdx], '0'
     mov.nz ax1, 10
     jmp.nz .main_loop
-    inc rdx
+    add rdx, rdx, 1
 
     cmp b[rdx], 'x'
-    inc.z rdx
+    add.z rdx, rdx, 1
     mov.z ax1, 16
     jmp.z .main_loop
 
     cmp b[rdx], 'b'
-    inc.z rdx
+    add.z rdx, rdx, 1
     mov.z ax1, 2
     jmp.z .main_loop
 
@@ -114,7 +114,7 @@ strtoq:
 
 .main_loop:
     movzx r12, b[rdx]
-    inc rdx
+    add rdx, rdx, 1
 
     cmp r12, '0'
     jmp.b .done
@@ -147,7 +147,7 @@ strtoq:
     ret.z
 
     ; yes
-    neg rax, rax
+    sub rax, zero, rax
     ret
 
 .bad:
diff --git a/ka/crt/lib/arith128.k b/ka/crt/lib/arith128.k
new file mode 100644
index 0000000..6023875
--- /dev/null
+++ b/ka/crt/lib/arith128.k
@@ -0,0 +1,59 @@
+; The OS/K Team licenses this file to you under the MIT license.
+; See the LICENSE file in the project root for more information.
+
+;
+; int:int add_lq_q(int:int x, int y)
+;
+add_lq_q:
+    addf rax, ax0, ax2
+    adcx rdx, ax1, zero
+    ret
+
+;
+; int:int add_lq_lq(int:int x, int:int y)
+;
+add_lq_lq:
+    addf rax, ax0, ax2
+    adcx rdx, ax1, ax3
+    ret
+
+;
+; int:int sub_lq_q(int:int x, int y)
+;
+sub_lq_q:
+    subf rax, ax0, ax2
+    sbbx rdx, ax1, zero
+    ret
+
+;
+; int:int sub_q_lq(int x, int:int y)
+;
+sub_q_lq:
+    subf rax, ax0, ax1
+    sbbx rdx, zero, ax2
+    ret
+
+;
+; int:int sub_lq_lq(int:int x, int:int y)
+;
+sub_lq_lq:
+    subf rax, ax0, ax2
+    sbbx rdx, ax1, ax3
+    ret
+
+;
+; int:int mul_lq_q(int x, int y)
+;
+mul_q_q:
+    mov rax, ax0
+    mulhi rdx, rax, ax1
+    ret
+
+;
+; int:int imul_lq_q(int x, int y)
+;
+imul_q_q:
+    mov rax, ax0
+    imulhi rdx, rax, ax1
+    ret
+
diff --git a/ka/crt/lib/time.k b/ka/crt/lib/time.k
index b56edcb..055da65 100644
--- a/ka/crt/lib/time.k
+++ b/ka/crt/lib/time.k
@@ -36,7 +36,7 @@ DaysInYear:
     b.nz rcx, zero, .end
 
 .leap:
-    inc  rax
+    add rax, rax, 1
 
 .end:
     ret
diff --git a/ka/crt/str/strcmp.k b/ka/crt/str/strcmp.k
index e2dfb01..ebfe69f 100644
--- a/ka/crt/str/strcmp.k
+++ b/ka/crt/str/strcmp.k
@@ -11,12 +11,27 @@
 ;
 strcmp:
     mov rcx, STRLEN_MAX
-    cmpzsb.rep.z ax0, ax1
 
-    mov rax, b[ax0-1]
-    mov rcx, b[ax1-1]
-    sub rax, rax, rcx
+.1:
+    jmp.cxz .2
 
+    movzx rax, b[ax0]
+    movzx rdx, b[ax1]
+
+    cmp rax, rdx
+    jmp.nz .2
+
+    ; both zero?
+    add r11, rax, rdx
+    b.z r11, zero, .2
+
+    add ax0, ax0, 1
+    add ax1, ax1, 1
+    sub rcx, rcx, 1
+    jmp .1
+
+.2:
+    sub rax, rax, rdx
     ret
 
 ;
@@ -24,14 +39,22 @@ strcmp:
 ;
 strncmp:
     mov rcx, ax2
-    mov.cxz rax, zero
-    ret.cxz
 
-    cmpzsb.rep.z ax0, ax1
+.1:
+    jmp.cxz .2
 
-    mov rax, b[ax0-1]
-    mov rcx, b[ax1-1]
-    sub rax, rax, rcx
+    mov rax, b[ax0]
+    mov rdx, b[ax1]
 
+    cmp rax, rdx
+    jmp.nz .2
+
+    add ax0, ax0, 1
+    add ax1, ax1, 1
+    sub rcx, rcx, 1
+    jmp .1
+
+.2:
+    sub rax, rax, rdx
     ret
 
diff --git a/ka/crt/str/strcpy.k b/ka/crt/str/strcpy.k
index e662549..52093c9 100644
--- a/ka/crt/str/strcpy.k
+++ b/ka/crt/str/strcpy.k
@@ -5,8 +5,17 @@
 ; void strcpy(char *, const char *)
 ;
 strcpy:
-    mov rcx, STRLEN_MAX
-    movsb.rep.nz ax0, ax1
+    mov rax, b[ax1]
+    mov b[ax0], rax
+
+    b.z rax, zero, .1
+
+    add ax0, ax0, 1
+    add ax1, ax1, 1
+
+    jmp strcpy
+
+.1:
     ret
 
 ;
@@ -14,24 +23,42 @@ strcpy:
 ;
 strncpy:
     mov rcx, ax2
+
+.1:
     ret.cxz
 
-    movsb.rep.nz ax0, ax1
-    ret
+    mov rax, b[ax1]
+    mov b[ax0], rax
+
+    add ax0, ax0, 1
+    add ax1, ax1, 1
+    sub rcx, rcx, 1
+
+    jmp .1
 
 ;
 ; void strnzcpy(char *, const char *, int)
 ;
 strnzcpy:
     mov rcx, ax2
-    ret.cxz
-
-    dec rcx
-    jmp.cxz .1
-
-    movsb.rep.nz ax0, ax1
 
 .1:
+    jmp.cxz .2
+
+    mov rax, b[ax1]
+    mov b[ax0], rax
+
+    b.z rax, zero, .3
+
+    add ax0, ax0, 1
+    add ax1, ax1, 1
+    sub rcx, rcx, 1
+
+    jmp .1
+
+.2:
     mov b[ax0], zero
+
+.3:
     ret
 
diff --git a/ka/crt/str/strlen.k b/ka/crt/str/strlen.k
index 880d34d..355a4b4 100644
--- a/ka/crt/str/strlen.k
+++ b/ka/crt/str/strlen.k
@@ -15,6 +15,10 @@ strnlen:
 ; int strlen(char *)
 ;
 strlen:
-    mov ax1, STRLEN_MAX
-    jmp strnlen
+    mov rcx, STRLEN_MAX
+    mov rdx, rcx
+    scasb.rep.nz ax0, zero
+
+    sub rax, rdx, rcx
+    ret
 
diff --git a/ka/crt/str/strrev.k b/ka/crt/str/strrev.k
index b3d81ca..bfb0425 100644
--- a/ka/crt/str/strrev.k
+++ b/ka/crt/str/strrev.k
@@ -18,7 +18,7 @@ strrev:
     ; the null terminator
     mov rcx, STRLEN_MAX
     scasb.rep.nz ax1, zero
-    dec ax1
+    sub ax1, ax1, 1
 
 .2:
     ; copy, going backward though str
@@ -30,8 +30,8 @@ strrev:
     mov.z b[ax0+1], zero
     ret.z
 
-    inc ax0
-    dec ax1
+    add ax0, ax0, 1
+    sub ax1, ax1, 1
     jmp .2
 
 
@@ -50,7 +50,7 @@ strrev2:
     ; the null terminator
     mov rcx, STRLEN_MAX
     scasb.rep.nz ax1, zero
-    dec ax1
+    sub ax1, ax1, 1
 
     ; increase ax0 while decreasing ax1, performing exchanges
 .2:
@@ -60,8 +60,8 @@ strrev2:
     xchg rax, b[ax0]
     mov b[ax1], rax
 
-    inc ax0
-    dec ax1
+    add ax0, ax0, 1
+    sub ax1, ax1, 1
     jmp .2
 
 .3:
diff --git a/ka/sys/intr/common.k b/ka/sys/intr/common.k
index 230b3ab..922154f 100644
--- a/ka/sys/intr/common.k
+++ b/ka/sys/intr/common.k
@@ -28,17 +28,13 @@ TrapHandlers.epilog:
 
     ; TRAP return values: RAX-RDX
 
-    mov ax0, r12
-    mov ax1, $rax
     mov ax2, rax
-    call RFS.StoreReg
+    call RFS.StoreReg, r12, $rax
 
-    mov ax1, $rdx
     mov ax2, rdx
-    call RFS.StoreReg
+    call RFS.StoreReg, r12, $rdx
 
-    mov ax0, r11
-    call IDT.DoneHandling
+    call IDT.DoneHandling, r11
 
     iret
 
diff --git a/ka/sys/intr/trap0.k b/ka/sys/intr/trap0.k
index f87ea9b..d0f11bb 100644
--- a/ka/sys/intr/trap0.k
+++ b/ka/sys/intr/trap0.k
@@ -37,18 +37,16 @@ trap0_handler:
 .handle_Exit:
 
     ; Open COMMAND.COM
-    mov ax0, .cmdcom
-    call DISK.OpenFile
+    call DISK.OpenFile, .cmdcom
 
     ; Crash on failure
     cmp rax, zero
     crash.l
 
     ; Load at CMDCOM_LOADP
-    mov ax0, rax
     mov ax1, CMDCOM_LOADP
     mov ax2, CMDCOM_MAXSZ
-    call DISK.ReadFile
+    call DISK.ReadFile, rax
 
     ; Assume that COMMAND.COM being
     ; less then 4KB means something
@@ -57,32 +55,26 @@ trap0_handler:
     crash.b
 
     ; Close the handle
-    mov ax0, rax
-    call DISK.CloseFile
+    call DISK.CloseFile, rax
 
     ; Code address
-    mov ax0, zero
-    mov ax1, $rip
     mov ax2, 0x100000
-    call RFS.StoreReg
+    call RFS.StoreReg, zero, $rip
 
     ; Usermode
-    mov ax1, $cr0
     mov ax2, 3
-    call RFS.StoreReg
+    call RFS.StoreReg, zero, $cr0
 
     mov rcx, CMDCOM_LOADP
     sub rcx, rcx, 0x100000
 
     ; Code offset
-    mov ax1, $cr1
     mov ax2, rcx
-    call RFS.StoreReg
+    call RFS.StoreReg, zero, $cr1
 
     ; Data offset
-    mov ax1, $cr2
     mov ax2, rcx
-    call RFS.StoreReg
+    call RFS.StoreReg, zero, $cr2
 
     ; Return frame
     mov q[rbp-16], zero
@@ -128,7 +120,9 @@ trap0_handler:
 .handle_HaltMode:
     hlt
 .HLT.loop:
-    xpause
+    pause
+    pause
+    pause
 
     scan rax
     b.z rax, zero, .HLT.loop
diff --git a/ka/sys/tests.k b/ka/sys/tests.k
index ce85f98..6c3a3d8 100644
--- a/ka/sys/tests.k
+++ b/ka/sys/tests.k
@@ -1,6 +1,61 @@
 ; The OS/K Team licenses this file to you under the MIT license.
 ; See the LICENSE file in the project root for more information.
 
+arith128_test:
+    call .test_add
+    call .test_sub
+
+    ret
+
+.print_lq:
+    push rax
+    push rdx
+    mov ax0, .fmt
+    call printf
+    add rsp, rsp, 16
+    prn 10
+    ret
+.fmt = "0x%x:%x"
+
+.test_add:
+
+    mov ax0, 0x99AABBCCDDEEFF00
+    mov ax1, 0x1122334455667788
+    mov ax2, 0xFF
+    call add_lq_q
+    call .print_lq
+
+    mov ax0, 0xFFFFFFFFFFFFFFFF
+    mov ax1, zero
+    mov ax2, 5
+    call add_lq_q
+    call .print_lq
+
+    mov ax0, 0x5500660077008800
+    mov ax1, 0x1100220033004400
+    mov ax2, 0x00AA00BB00CC00DD
+    mov ax3, 0x009900CE00DF00AB
+    call add_lq_lq
+    call .print_lq
+
+    ret
+
+.test_sub:
+
+    mov ax0, 0x99AABBCCDDEEFFFF
+    mov ax1, 0x1122334455667788
+    mov ax2, 0xFF
+    call sub_lq_q
+    call .print_lq
+
+    mov ax0, zero
+    mov ax1, zero
+    mov ax2, 2
+    call sub_lq_q
+    call .print_lq
+
+    ret
+
 file_test:
 .bufsize := 256
 
diff --git a/ka/usr/cmd/dir.k b/ka/usr/cmd/dir.k
index ce86966..48ac4ca 100644
--- a/ka/usr/cmd/dir.k
+++ b/ka/usr/cmd/dir.k
@@ -32,7 +32,7 @@ builtins.dir:
     b.z rax, 0, .end
 
     ; found something
-    inc nx0
+    add nx0, nx0, 1
 
     ; separate extension from file name
     mov rcx, NAME_MAX
@@ -46,13 +46,13 @@ builtins.dir:
 
     ; calculate where to put extension
     sub r11, r10, .buf
-    dec r11
+    sub r11, r11, 1
 
 .ext_pad:
     ; print at least 11 non-space characters before extension
     b.ae r11, 11, .print_ext
     prn ' '
-    inc r11
+    add r11, r11, 1
     jmp .ext_pad
 
 .print_ext:
@@ -61,15 +61,16 @@ builtins.dir:
     prn ' '
 
     cmp b[r10], '.'
-    inc.z r10
+    add.z r10, r10, 1
 
 .print_ext.1:
     b.z b[r10], 0, .print_ext.2
 
     ; print and decrease rcx, unless it's already 0
-    prn b[r10]
-    inc r10
-    dec.cxnz rcx
+    mov r12, b[r10]
+    prn r12
+    add r10, r10, 1
+    sub.cxnz rcx, rcx, 1
 
     jmp .print_ext.1
 
@@ -88,10 +89,9 @@ builtins.dir:
     shr rax, rdx, 10
     and rdx, rdx, 1023
 
-    mov ax0, .bytesstr
     push rdx
     push rax
-    call printf
+    call printf, .bytesstr
     add rsp, rsp, 16
 
 .bytesstr = "%d kilobytes + %d bytes"
@@ -102,14 +102,11 @@ builtins.dir:
     jmp .next
 
 .end:
-    mov ax0, .endstr1
     push nx0
-    call printf
+    call printf, .endstr1
     add rsp, rsp, 8
 
-    mov rcx, STRLEN_MAX
-    mov rdx, .endstr2
-    prns.rep.nz rdx
+    call print, .endstr2
 
     pop nx0
     ret
diff --git a/ka/usr/cmd/main.k b/ka/usr/cmd/main.k
index 4846ff9..32554b3 100644
--- a/ka/usr/cmd/main.k
+++ b/ka/usr/cmd/main.k
@@ -13,9 +13,7 @@ ps1 = "C:\\> "
 main:
 
 .print_prompt:
-    mov rcx, STRLEN_MAX
-    mov rdx, ps1
-    prns.rep.nz rdx
+    call print, ps1
 
     ; empty argbuf
     mov rcx, argbuf.size
@@ -43,7 +41,7 @@ main:
     b.z rcx, zero, .input_loop
 
     ; delete it
-    dec rcx
+    sub rcx, rcx, 1
     add rdx, rcx, argbuf
     mov b[rdx], zero
 
@@ -65,7 +63,7 @@ main:
     ; add character to buffer and increase iterator (rcx)
     add rdx, rcx, argbuf
     mov b[rdx], rax
-    inc rcx
+    add rcx, rcx, 1
 
     ; another one
     jmp .input_loop
@@ -83,7 +81,7 @@ main:
     ; argv1 exists? if so, save its position
     mov r11, rdx
     b.z b[r11], zero, .no_argv1
-    inc r11
+    add r11, r11, 1
 
 .next_space:
     mov r10, b[r11]
@@ -91,7 +89,7 @@ main:
 
     ; skip spaces
     cmp r10, ' '
-    inc.z r11
+    add.z r11, r11, 1
     jmp.z .next_space
 
     mov q[argv1pos], r11
@@ -108,68 +106,59 @@ main:
     sub rcx, rdx, argbuf
     mov rdx, argbuf
     mov rax, argv0
-    movsb.rep rax, rdx
+
+.mmove:
+    jmp.cxz .detect_builtin
+
+    mov r11, b[rdx]
+    mov b[rax], r11
+
+    add rdx, rdx, 1
+    add rax, rax, 1
+    sub rcx, rcx, 1
+
+    jmp .mmove
 
 .detect_builtin:
 
 .builtin_cls = "cls"
-    mov ax0, argv0
-    mov ax1, .builtin_cls
-    call strcmp
+    call strcmp, argv0, .builtin_cls
     b.z rax, 0, .handle_CLS
 
 .builtin_date = "date"
-    mov ax0, argv0
-    mov ax1, .builtin_date
-    call strcmp
+    call strcmp, argv0, .builtin_date
     b.z rax, 0, .handle_DATE
 
 .builtin_dir = "dir"
-    mov ax0, argv0
-    mov ax1, .builtin_dir
-    call strcmp
+    call strcmp, argv0, .builtin_dir
     b.z rax, 0, .handle_DIR
 
 .builtin_dump = "dump"
-    mov ax0, argv0
-    mov ax1, .builtin_dump
-    call strcmp
+    call strcmp, argv0, .builtin_dump
     b.z rax, 0, .handle_DUMP
 
 .builtin_echo = "echo"
-    mov ax0, argv0
-    mov ax1, .builtin_echo
-    call strcmp
+    call strcmp, argv0, .builtin_echo
     b.z rax, 0, .handle_ECHO
 
 .builtin_exit = "exit"
-    mov ax0, argv0
-    mov ax1, .builtin_exit
-    call strcmp
+    call strcmp, argv0, .builtin_exit
     b.z rax, 0, .handle_EXIT
 
 .builtin_help = "help"
-    mov ax0, argv0
-    mov ax1, .builtin_help
-    call strcmp
+    call strcmp, argv0, .builtin_help
     b.z rax, 0, .handle_HELP
 
 .builtin_print = "print"
-    mov ax0, argv0
-    mov ax1, .builtin_print
-    call strcmp
+    call strcmp, argv0, .builtin_print
     b.z rax, 0, .handle_PRINT
 
 .builtin_time = "time"
-    mov ax0, argv0
-    mov ax1, .builtin_time
-    call strcmp
+    call strcmp, argv0, .builtin_time
     b.z rax, 0, .handle_TIME
 
 .builtin_ver = "ver"
-    mov ax0, argv0
-    mov ax1, .builtin_ver
-    call strcmp
+    call strcmp, argv0, .builtin_ver
     b.z rax, 0, .handle_VER
 
     jmp .command_not_found
@@ -182,19 +171,17 @@ main:
     jmp .print_prompt
 
 .handle_DATE:
-    time ax0
     call GetTimeUTC
 
     push b[rax+3]
     mov rcx, b[rax+4]
-    inc rcx
+    add rcx, rcx, 1
     push rcx
     push w[rax+6]
 
-    mov ax0, .datefmt
-    call printf
+    call printf, .datefmt
 
-    add rsp, rsp, 5*8
+    add rsp, rsp, 40
 
     jmp .print_prompt
 
@@ -252,53 +239,36 @@ main:
     jmp .print_prompt
 
 .handle_TIME:
-    time ax0
     call GetTimeUTC
 
     push b[rax]
     push b[rax+1]
     push b[rax+2]
-    mov ax0, .timefmt
-    call printf
-    add rsp, rsp, 3*8
+    call printf, .timefmt
+    add rsp, rsp, 24
 
     jmp .print_prompt
 
 .timefmt = "%d:%d:%d\n"
 
 .handle_VER:
-    mov rcx, STRLEN_MAX
-    mov rdx, cmd.versionstr
-    prns.rep.nz rdx
+    call print, cmd.versionstr
     prn 10
 
     jmp .print_prompt
 
 .handle_HELP:
-    mov rcx, STRLEN_MAX
-
-    mov rdx, .helpmsg
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.cls
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.date
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.dir
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.dump
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.echo
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.exit
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.help
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.print
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.time
-    prns.rep.nz rdx
-    mov rdx, .helpmsg.ver
-    prns.rep.nz rdx
+    call print, .helpmsg
+    call print, .helpmsg.cls
+    call print, .helpmsg.date
+    call print, .helpmsg.dir
+    call print, .helpmsg.dump
+    call print, .helpmsg.echo
+    call print, .helpmsg.exit
+    call print, .helpmsg.help
+    call print, .helpmsg.print
+    call print, .helpmsg.time
+    call print, .helpmsg.ver
 
     jmp .print_prompt
 
@@ -315,22 +285,17 @@ main:
 .helpmsg.ver   = "    VER        Display current COMMAND.COM and DOS kernel versions\n"
 
 .command_not_found:
-    mov rcx, STRLEN_MAX
-    mov rdx, argv0
-    prns.rep.nz rdx
-
-    mov rdx, .cnf_errmsg
-    prns.rep.nz rdx
+    call print, argv0
+    call print, .cnf_errmsg
 
     jmp .print_prompt
 
 .cnf_errmsg = ": command not found\n"
 
 .file_not_found:
-    mov ax0, .fnf_errmsg
     push q[argv1pos]
     push argv0
-    call printf
+    call printf, .fnf_errmsg
     add rsp, rsp, 16
 
     jmp .print_prompt
@@ -338,10 +303,9 @@ main:
 .fnf_errmsg = "%s: %s: file not found\n"
 
 .empty_file:
-    mov ax0, .ef_errmsg
     push q[argv1pos]
     push argv0
-    call printf
+    call printf, .ef_errmsg
     add rsp, rsp, 16
 
     jmp .print_prompt
@@ -349,10 +313,9 @@ main:
 .ef_errmsg = "%s: %s: file is empty\n"
 
 .couldnt_read:
-    mov ax0, .cno_errmsg
     push q[argv1pos]
     push argv0
-    call printf
+    call printf, .cno_errmsg
     add rsp, rsp, 16
 
     jmp .print_prompt
@@ -360,12 +323,8 @@ main:
 .cno_errmsg = "%s: %s: an error occured while reading file\n"
 
 .need_params:
-    mov rcx, STRLEN_MAX
-    mov rdx, argv0
-    prns.rep.nz rdx
-
-    mov rdx, .np_errmsg
-    prns.rep.nz rdx
+    call print, argv0
+    call print, .np_errmsg
 
     jmp .print_prompt
 
diff --git a/vm/Makefile b/vm/Makefile
index 7436d4c..6aee436 100644
--- a/vm/Makefile
+++ b/vm/Makefile
@@ -46,14 +46,14 @@ $(OBJDIR)/%.d: %.c
 #		echo ${CL2}[$@] ${CL}dependencies generated.${CL3};\
 #	fi
 
-in/instrs.lst: in/INSTRS in/arch_i.py
+ob/instrs.lst: in/arch_i.py
 	@cd in && python3 arch_i.py
 
 clean:
-	@rm -f $(OBJDIR)/*/*.o in/arch_i.h in/instrs.lst
+	@rm -f $(OBJDIR)/*/*.o ob/arch_i.h ob/instrs.lst
 	@rm -f $(OBJDIR)/*/*.d
 
-$(KEXE): in/instrs.lst $(obj)
+$(KEXE): ob/instrs.lst $(obj)
 	@gcc -O2 -lSDL2 -lSDL2_ttf -Wall $(obj) -o $(KEXE)
 	@echo ${CL2}[$@] ${CL}made successfully.${CL3}
 
diff --git a/vm/in/ALU b/vm/in/ALU
index b3d07e9..9282191 100644
--- a/vm/in/ALU
+++ b/vm/in/ALU
@@ -5,15 +5,6 @@
 # Logical instructions                                                      #
 #---------------------------------------------------------------------------#
 
-#
-# Bitwise NOT operation
-#
-#       $1 = NOT($2)
-#
-# Preserves all flags
-#
-not r r
-
 #
 # Bitwise OR operation
 #
@@ -22,10 +13,6 @@ not r r
 # Preserves all flags
 #
 or r r ri
-# $dest = $src1 OR NOT($src2)
-orn r r ri
-# $dest = NOT($src1 OR $src2)
-nor r r ri
 
 #
 # Bitwise AND operation
@@ -35,10 +22,6 @@ nor r r ri
 # Preserves all flags
 #
 and r r ri
-# $dest = $src1 AND NOT($src2)
-andn r r ri
-# $dest = NOT($src1 AND $src2)
-nand r r ri
 
 #
 # Bitwise XOR operation
@@ -48,10 +31,6 @@ nand r r ri
 # Preserves all flags
 #
 xor r r ri
-# $dest = $src1 XOR NOT($src2)
-xorn r r ri
-# $dest = NOT($src1 XOR $src2)
-xnor r r ri
 
 #
 # Logical left/right shift (SHL/SHR)
@@ -91,50 +70,6 @@ sar r r ri
 cmp r ri
 cmp m ri
 
-#
-# Arithmetical SGN operation
-#
-#       IF ($1 == 0) THEN
-#           $2 = 0
-#       ELIF ($1 GT 0) THEN
-#           $2 = 1
-#       ELSE
-#           $2 = LONG_MIN
-#       FI
-#
-# Treats $1 and $2 as signed values
-#
-# Preserves all flags
-#
-sgn r r
-
-#
-# Arithmetical NEG operation
-#
-#       $1 = NOT($2) + 1
-#
-# Preserves all flags
-#
-neg r r
-
-#
-# Arithmetical INC operation
-#
-#       $1 = $1 + 1
-#
-# Preserves all flags
-#
-inc r
-
-#
-# Arithmetical DEC operation
-#
-#       $1 = $1 - 1
-#
-# Preserves all flags
-#
-dec r
-
 #
 # Arithmetical ADD operation
 #
@@ -146,7 +81,6 @@ dec r
 #
 add r r ri
 addf r r ri
-addo r r ri
 
 #
 # Arithmetical SUB operation
@@ -159,26 +93,22 @@ addo r r ri
 #
 sub r r ri
 subf r r ri
-subo r r ri
 
 #
 # Arithmetical ADD/SUB operation, with carry/overflow
 #
 #       $dest = $src1 + $src2 + CF      (ADC)
-#       $dest = $src1 + $src2 + OF      (ADO)
 #       $dest = $src1 - $src2 - CF      (SBB)
-#       $dest = $src1 - $src2 - OF      (SBO)
 #
-# Preserves ZF and SF
-# Flags CF and OF are undefined for now
+# Sets CF if unsigned integer overflow occur, clears it otherwise
+# Sets OF is signed integer overflow occur, clears it otherwise
+# Sets ZF and SF according to the result
 #
 adcx r r ri
-adox r r ri
 sbbx r r ri
-sbox r r ri
 
 #
-# Arithmetical unsigned MUL operation
+# Arithmetical MUL operation
 #
 #       $dest = LO($src1 * $src2)
 #
@@ -187,31 +117,47 @@ sbox r r ri
 #
 mul r r ri
 mulf r r ri
-mulo r r ri
 
 # Arithmetical unsigned MUL operation
 #
 #       $dest1 = HI($dest2 * $src)
 #       $dest2 = LO($dest2 * $src)
 #
-# Sets CF and OF if HI($2 * $3) > 0, clears them otherwise
-# Preserves ZF and SF
+# Preserves all flags
 #
 mulhi r r ri
 
+# Arithmetical signed MUL operation
+#
+#       $dest1 = HI($dest2 ** $src)
+#       $dest2 = LO($dest2 ** $src)
+#
+# Preserves all flags
+#
+imulhi r r ri
+
 #
 # Arithmetical unsigned DIV operation
 #
-#       $dest1 = $1 DIV $2
+#       $dest = $src1 DIV $src2
 #
 # Preserves all flags
 #
 div r r ri
 
 #
-# Arithmetical unsigned modulo operation (REM)
+# Arithmetical signed DIV operation
 #
-#       $1 = $1 MOD $2
+#       $dest = $src1 IDIV $src2
+#
+# Preserves all flags
+#
+idiv r r ri
+
+#
+# Arithmetical modulo operation (REM)
+#
+#       $dest = $src1 MOD $src2
 #
 # Preserves all flags
 #
diff --git a/vm/in/MISC b/vm/in/MISC
index 56d2646..7397baf 100644
--- a/vm/in/MISC
+++ b/vm/in/MISC
@@ -24,36 +24,21 @@ dump
 # Misc. instructions                                                        #
 #---------------------------------------------------------------------------#
 
-#
-# Do nothing (NOOP)
-#
-#       (nothing)
-#
-# Throws:
-#       (nothing)
-#
-# Preserves all flags
-#
-nop
-
 #
 # Pause the CPU for a very short amount of time (PAUSE)
-# Used in spin-like loops
+# Used in spin-type loops
 #
 pause
 
 #
-# Pause the CPU for a relatively long time (XPAUSE)
+# Throw #OVF when RFX.OF=1 (INTO)
 #
-# Throws:
-#       #SYS if not in supervisor mode
-#
-xpause
+into
 
 #
-# Get timestamp in seconds
+# Get timestamp in µseconds
 #
-time r
+utime r r
 
 #
 # $1 = seconds since start of month
@@ -62,68 +47,11 @@ time r
 #
 ytime r r r
 
-#
-# Get timestamp in µseconds
-#
-utime r
-
-#
-# CPU Identification Number
-#
-# Does nothing (for now)
-#
-cpuid
-
-#---------------------------------------------------------------------------#
-# Clean-up misc. instructions                                               #
-#---------------------------------------------------------------------------#
-
 #
 # Clear all GPR registers except RBP/RSP
 #
 cls
 
-#
-# Clear base registers except RBP/RSP (CLR)
-#
-clr
-
-#
-# Clear argument registers (CLA)
-#
-cla
-
-#
-# Clear nonvolatile registers (CLN)
-#
-cln
-
-#---------------------------------------------------------------------------#
-# Flag manipulation instructions                                            #
-#---------------------------------------------------------------------------#
-
-#
-# Clear or set interrupt flag (CLI/STI)
-#
-# Throws:
-#       #SYS if not in supervisor mode
-#
-cli
-sti
-
-#
-# Clear or set direction flag (CLD/STD)
-#
-cld
-std
-
-#
-# Complement, clear or set carry flag (CMC/CLC/STC)
-#
-cmc
-clc
-stc
-
 #---------------------------------------------------------------------------#
 # Byte-wise / bit-wise manipulation instructions                            #
 #---------------------------------------------------------------------------#
@@ -144,7 +72,7 @@ dswap r r
 #
 # Send a character to standard output (PRN)
 #
-prn rim
+prn ri
 
 #
 # Print a string to standard output (PRN)
diff --git a/vm/in/MOV b/vm/in/MOV
index 17b4cd9..9d94547 100644
--- a/vm/in/MOV
+++ b/vm/in/MOV
@@ -52,6 +52,8 @@ b rm ri ri
 #       JMP(RIP)
 #
 call ri
+call i ri
+call i ri ri
 
 #
 # Return to caller (RET)
@@ -103,9 +105,9 @@ pop r
 #       $1 = ADDR($2)
 #
 # For instance:
-#       LEA RAX, [RBX + RCX + 4]
+#       LEA RAX, [RBX + RCX * 2 + 4]
 # will result in:
-#       RAX = RBX + RCX + 4
+#       RAX = RBX + RCX * 2 + 4
 #
 # Preserves all flags
 #
diff --git a/vm/in/STRING b/vm/in/STRING
index 4dd3f08..3fa33ff 100644
--- a/vm/in/STRING
+++ b/vm/in/STRING
@@ -20,24 +20,6 @@ stosw r ri
 stosl r ri
 stosq r ri
 
-#
-# Load value from string (LODSx)
-#
-#       $1 = [%2]
-#       IF (DF == 0) THEN
-#           %str = %str + sizeof(x)
-#       ELSE
-#           %str = %str - sizeof(x)
-#       FI
-#
-# Preserves CF, OF and SF
-# Sets ZF according to the loaded value
-#
-lodsb r r
-lodsw r r
-lodsl r r
-lodsq r r
-
 #
 # Scan string for a particular value (SCASx)
 #
@@ -65,61 +47,5 @@ scasw r ri
 scasl r ri
 scasq r ri
 
-#
-# Compare bytes in strings (CMPSx)
-#
-#       CMP([%1], [%2])
-#
-#       IF (DF == 0) THEN
-#           %1 = %1 + sizeof(x)
-#           %2 = %2 + sizeof(x)
-#       ELSE
-#           %1 = %1 - sizeof(x)
-#           %2 = %2 - sizeof(x)
-#       FI
-#
-# Sets CF, OF, ZF and SF according to the result of the comparison
-#
-# Moves past the compared values in any case!
-#
-cmpsb r r
-cmpsw r r
-cmpsl r r
-cmpsq r r
-
-#
-# Safe compare bytes in strings (CMPZSx)
-#
-# Behaves precisely like CMPSx, except in the following case:
-#   - If both [%1] and [%2] are zero, clears ZF (indicating NOT EQUAL)
-#
-# This prevents 'CMPZSx.REP.Z' from ignoring null-terminators when both strings
-# have the exact same content; this allows for short strcmp's
-#
-cmpzsb r r
-cmpzsw r r
-cmpzsl r r
-cmpzsq r r
-
-#
-# Move value from string to string (MOVSx)
-#
-#       [%1] = [%2]
-#       IF (DF == 0) THEN
-#           %1 = %1 + sizeof(x)
-#           %2 = %2 + sizeof(x)
-#       ELSE
-#           %1 = %1 - sizeof(x)
-#           %2 = %2 - sizeof(x)
-#       FI
-#
-# Preserves CF, OF and SF
-# Sets ZF according to the moved value
-#
-movsb r r
-movsw r r
-movsl r r
-movsq r r
-
 #---------------------------------------------------------------------------#
 
diff --git a/vm/in/SUPER b/vm/in/SUPER
index 519998a..4197855 100644
--- a/vm/in/SUPER
+++ b/vm/in/SUPER
@@ -58,6 +58,15 @@ trap ri
 #
 iret
 
+#
+# Clear or set interrupt flag (CLI/STI)
+#
+# Throws:
+#       #SYS if not in supervisor mode
+#
+cli
+sti
+
 #---------------------------------------------------------------------------#
 # Device control instructions                                               #
 #---------------------------------------------------------------------------#
@@ -70,7 +79,7 @@ iret
 # Throws:
 #       #SYS if not in supervisor mode
 #
-devctl ri ri
+devctl r r
 
 #
 # Call a device-defined function slot of device (IOCALL)
diff --git a/vm/in/alu.c b/vm/in/alu.c
index e1d2378..d68db41 100644
--- a/vm/in/alu.c
+++ b/vm/in/alu.c
@@ -4,26 +4,16 @@
 #include <in/instrs.h>
 
 #define _NEED_ARCH_I
-#include <in/arch_i.h>
+#include <ob/arch_i.h>
 
 //----------------------------------------------------------------------------//
 
-IMPL_START_2(not) { v1 = ~v2; } IMPL_OUT;
 IMPL_START_3(or) { v1 = v2 | v3; } IMPL_OUT;
 IMPL_START_3(and) { v1 = v2 & v3; } IMPL_OUT;
 IMPL_START_3(xor) { v1 = v2 ^ v3; } IMPL_OUT;
 
 //----------------------------------------------------------------------------//
 
-IMPL_START_3(orn) { v1 = v2 | ~v3; } IMPL_OUT;
-IMPL_START_3(nor) { v1 = ~(v2 | v3); } IMPL_OUT;
-IMPL_START_3(andn) { v1 = v2 & ~v3; } IMPL_OUT;
-IMPL_START_3(nand) { v1 = ~(v2 & v3); } IMPL_OUT;
-IMPL_START_3(xorn) { v1 = v2 ^ ~v3; } IMPL_OUT;
-IMPL_START_3(xnor) { v1 = ~(v2 ^ v3); } IMPL_OUT;
-
-//----------------------------------------------------------------------------//
-
 IMPL_START_3(shl) { v1 = v2 << v3; } IMPL_OUT;
 IMPL_START_3(shr) { v1 = v2 >> v3; } IMPL_OUT;
 IMPL_START_3(sal) { v1 = (ulong)((long)v2 << (long)v3); } IMPL_OUT;
@@ -31,81 +21,64 @@ IMPL_START_3(sar) { v1 = (ulong)((long)v2 >> (long)v3); } IMPL_OUT;
 
 //----------------------------------------------------------------------------//
 
-IMPL_START_2(sgn) { v1 = (!v2 ? 0: ((long)v2 < 0 ? (ulong)-1L : 1)); } IMPL_OUT;
-IMPL_START_2(neg) { v1 = ~v2 + 1; } IMPL_OUT;
-IMPL_START_1(inc) { v1++; } IMPL_OUT;
-IMPL_START_1(dec) { v1--; } IMPL_OUT;
-
-//----------------------------------------------------------------------------//
-
 IMPL_START_3(add) { v1 = v2 + v3; } IMPL_OUT;
-IMPL_START_3(addf) { COMPARE(v2, ~v3+1); v1 = v2 + v3; } IMPL_OUT;
-IMPL_START_3(addo) { COMPARE(v2, ~v3+1); v1 = v2 + v3; INTO(); } IMPL_OUT;
+IMPL_START_3(addf) { COMPARE_ADD(v2, v3); v1 = v2 + v3; } IMPL_OUT;
 
 //----------------------------------------------------------------------------//
 
-IMPL_START_2(cmp) { COMPARE(v1, v2); } IMPL_END;
+IMPL_START_2(cmp) { COMPARE_SUB(v1, v2); } IMPL_END;
 IMPL_START_3(sub) { v1 = v2 - v3; } IMPL_OUT;
-IMPL_START_3(subf){ COMPARE(v2, v3); v1 = v2 - v3; } IMPL_OUT;
-IMPL_START_3(subo){ COMPARE(v2, v3); v1 = v2 - v3; INTO(); } IMPL_OUT;
+IMPL_START_3(subf){ COMPARE_SUB(v2, v3); v1 = v2 - v3; } IMPL_OUT;
 
 //----------------------------------------------------------------------------//
 
-IMPL_START_3(adcx) { v1 = v2 + v3 + !!(R(RFX)&CF); } IMPL_OUT;
-IMPL_START_3(adox) { v1 = v2 + v3 + !!(R(RFX)&OF); } IMPL_OUT;
-IMPL_START_3(sbbx) { v1 = v2 - v3 - !!(R(RFX)&CF); } IMPL_OUT;
-IMPL_START_3(sbox) { v1 = v2 - v3 - !!(R(RFX)&OF); } IMPL_OUT;
+IMPL_START_3(adcx) { v3 += !!(R(RFX)&CF); COMPARE_ADD(v2, v3); v1 = v2 + v3; } IMPL_OUT;
+IMPL_START_3(sbbx) { v3 += !!(R(RFX)&CF); COMPARE_SUB(v2, v3); v1 = v2 - v3; } IMPL_OUT;
 
 //----------------------------------------------------------------------------//
 
+IMPL_START_3(mul) { v1 = v2 * v3; } IMPL_OUT;
 IMPL_START_3(rem) { v1 = v2 % v3; } IMPL_OUT;
 IMPL_START_3(div) {
     if (!v3)
         _except(ctx, E_DIV, "DIV by 0");
     v1 = v2 / v3;
 } IMPL_OUT;
+IMPL_START_3(idiv) {
+    if (!v3)
+        _except(ctx, E_DIV, "IDIV by 0");
+    v1 = (ulong)((long)v2/(long)v3);
+} IMPL_OUT;
 
 //----------------------------------------------------------------------------//
 
-//
-// www.codeproject.com/Tips/618570/UInt-Multiplication-Squaring
-//
-static void multiply(ulong u, ulong v, ulong *hi, ulong *lo)
+static void __unsigned_multiply128(ulong u, ulong v, ulong *hi, ulong *lo)
 {
-    ulong u1 = (u & 0xffffffff);
-    ulong v1 = (v & 0xffffffff);
-    ulong t = (u1 * v1);
-    ulong w3 = (t & 0xffffffff);
-    ulong k = (t >> 32);
-    
-    u >>= 32;
-    t = (u * v1) + k;
-    k = (t & 0xffffffff);
-    ulong w1 = (t >> 32);
-    
-    v >>= 32;
-    t = (u1 * v) + k;
-    k = (t >> 32);
-    
-    if (hi) *hi = (u * v) + w1 + k;
-    if (lo) *lo = (t << 32) + w3;
+    __uint128_t r = (__uint128_t)u * (__uint128_t)v;
+
+    *hi = r >> 64;
+    *lo = r;
 }
 
-IMPL_START_3(mul) { v1 = v2 * v3; } IMPL_OUT;
-IMPL_START_3(mulhi) { multiply(v2, v3, &v1, &v2); } IMPL_OUT_2;
+IMPL_START_3(mulhi) { __unsigned_multiply128(v2, v3, &v1, &v2); } IMPL_OUT_2;
 
 IMPL_START_3(mulf) {
     ulong tmp;
-    multiply(v2, v3, &tmp, &v1);
+    __unsigned_multiply128(v2, v3, &tmp, &v1);
     R(RFX) = v2 ? (R(RFX)|CF|OF) : R(RFX)&~(CF|OF);
 } IMPL_OUT;
 
-IMPL_START_3(mulo) {
-    ulong tmp;
-    multiply(v2, v3, &tmp, &v1);
-    R(RFX) = v2 ? (R(RFX)|CF|OF) : R(RFX)&~(CF|OF);
-    INTO();
-} IMPL_OUT;
-
+//----------------------------------------------------------------------------//
+
+static void __signed_multiply128(ulong u, ulong v, ulong *hi, ulong *lo)
+{
+    __int128_t r = (__int128_t)(long)u * (__int128_t)(long)v;
+
+    *hi = r >> 64;
+    *lo = r;
+}
+
+IMPL_START_3(imulhi) { __signed_multiply128(v2, v3, &v1, &v2); } IMPL_OUT_2;
+
 //----------------------------------------------------------------------------//
 
diff --git a/vm/in/arch_i.py b/vm/in/arch_i.py
index 5888a4d..15f6126 100644
--- a/vm/in/arch_i.py
+++ b/vm/in/arch_i.py
@@ -131,8 +131,8 @@ def parse_2(fp):
         count = count + 1
 
 fi = open("INSTRS")
-hd = open("arch_i.h", "w")
-ls = open("instrs.lst", "w")
+hd = open("../ob/arch_i.h", "w")
+ls = open("../ob/instrs.lst", "w")
 
 count = 0
 
diff --git a/vm/in/instrs.h b/vm/in/instrs.h
index 3fff8d2..c52e564 100644
--- a/vm/in/instrs.h
+++ b/vm/in/instrs.h
@@ -2,7 +2,6 @@
 // See the LICENSE file in the project root for more information.
 
 #include <pc/arch.h>
-#include <in/arch_i.h>
 
 //----------------------------------------------------------------------------//
 
@@ -108,18 +107,29 @@ IMPL_START_2(name)                                                          \
 
 //----------------------------------------------------------------------------//
 
-#define COMPARE(v1, v2)                                                     \
-    ulong _u1 = (ulong)v1, _u2 = (ulong)v2;                                 \
-    long _s1 = (long)v1, _s2 = (long)v2;                                    \
+#define COMPARE_ADD(v1, v2)                                                 \
+    ulong _u1 = (ulong)v1, _u2 = (ulong)v2, _u3 = (_u1 + _u2);              \
+                                                                            \
+    if (_u1 + _u2 < _u2) R(RFX) |= CF;                                      \
+    else R(RFX) &= ~CF;                                                     \
+                                                                            \
+    if (((_u1 ^ _u3) & (_u2 ^ _u3)) >> 63)                                  \
+        R(RFX) |= OF;                                                       \
+    else R(RFX) &= ~OF;                                                     \
+                                                                            \
+    SET_ZSF(_u3);
+
+#define COMPARE_SUB(v1, v2)                                                 \
+    ulong _u1 = (ulong)v1, _u2 = (ulong)v2, _u3 = (_u1 - _u2);              \
                                                                             \
     if (_u1 < _u2) R(RFX) |= CF;                                            \
     else R(RFX) &= ~CF;                                                     \
                                                                             \
-    if ( ((_s1 < 0) && (_s1 > LONG_MAX + _s2))                              \
-      || ((_s2 > 0) && (_s1 < LONG_MIN + _s2)) )                            \
+    if (((_u1 ^ _u3) & (_u1 ^ _u2)) >> 63)                                  \
         R(RFX) |= OF;                                                       \
     else R(RFX) &= ~OF;                                                     \
-    SET_ZSF(_u1 - _u2);
+                                                                            \
+    SET_ZSF(_u3);
 
 //----------------------------------------------------------------------------//
 
diff --git a/vm/in/misc.c b/vm/in/misc.c
index 905fd6c..62dc4ba 100644
--- a/vm/in/misc.c
+++ b/vm/in/misc.c
@@ -9,11 +9,8 @@
 
 //----------------------------------------------------------------------------//
 
-IMPL_START_0(nop) {} IMPL_END;
-IMPL_START_0(cpuid) { rax = rdx = 0; } IMPL_END;
-
+IMPL_START_0(into) { INTO(); } IMPL_END;
 IMPL_START_0(pause) { usleep(5000); } IMPL_END;
-IMPL_START_0(xpause) { CHK_SUPERV(); usleep(25000); } IMPL_END;
 
 //----------------------------------------------------------------------------//
 
@@ -46,8 +43,6 @@ IMPL_END;
 
 //----------------------------------------------------------------------------//
 
-IMPL_START_X_NOIN(time) { v1 = time(NULL); } IMPL_OUT;
-
 IMPL_START_X_NOIN(ytime)
 {
     time_t t = time(NULL);
@@ -77,24 +72,6 @@ IMPL_START_0(cls) {
     for (int i = R10; i <= NX7; i++) R(i) = 0;
 } IMPL_END;
 
-IMPL_START_0(clr) {
-    R(RAX) = R(RBX) = R(RCX) = R(RDX) = R(RDI) = R(RSI) = R(RFX) = 0;
-    for (int i = R10; i <= R15; i++) R(i) = 0;
-} IMPL_END;
-
-IMPL_START_0(cla) { for (int i = AX0; i <= AX7; i++) R(i) = 0; } IMPL_END;
-IMPL_START_0(cln) { for (int i = NX0; i <= NX7; i++) R(i) = 0; } IMPL_END;
-
-//----------------------------------------------------------------------------//
-
-IMPL_START_0(cli) { CHK_SUPERV(); R(CR0) &= ~IF; } IMPL_END;
-IMPL_START_0(sti) { CHK_SUPERV(); R(CR0) |= IF; } IMPL_END;
-IMPL_START_0(cld) { R(RFX) &= ~DF; } IMPL_END;
-IMPL_START_0(clc) { R(RFX) &= ~CF; } IMPL_END;
-IMPL_START_0(std) { R(RFX) |= DF; } IMPL_END;
-IMPL_START_0(stc) { R(RFX) |= CF; } IMPL_END;
-IMPL_START_0(cmc) { R(RFX) = (R(RFX)&CF ? R(RFX)&~CF : R(RFX)|CF); } IMPL_END;
-
 //----------------------------------------------------------------------------//
 
 IMPL_START_2(bswap)
@@ -151,7 +128,7 @@ IMPL_START_0(prns)
 {
     uchar ch = readmemzx(ctx, R(p1->reg), 1);
 
-    COMPARE(ch, 0);
+    COMPARE_SUB(ch, 0);
 
     if ((R(RFX) & ZF) == 0)
     {
diff --git a/vm/in/mov.c b/vm/in/mov.c
index cec4ee4..10b524b 100644
--- a/vm/in/mov.c
+++ b/vm/in/mov.c
@@ -15,7 +15,7 @@ IMPL_START_1(loop) {
 IMPL_END;
 
 IMPL_START_3(b) {
-    COMPARE(v1, v2);
+    COMPARE_SUB(v1, v2);
 
     if (eval_cond(ctx, ctx->cond))
         R(RIP) = v3;
@@ -51,11 +51,32 @@ IMPL_START_1(pop) {
     R(RSP) += 8;
 } IMPL_OUT;
 
+/*
 IMPL_START_1(call) {
     R(RSP) -= 8;
     writemem(ctx, R(RIP), R(RSP), 8);
     R(RIP) = v1;
 } IMPL_END;
+*/
+
+IMPL_START_1(call) {
+    R(RSP) -= 8;
+    writemem(ctx, R(RIP), R(RSP), 8);
+    R(RIP) = v1;
+
+    if (p2)
+    {
+        DECV(v2, p2);
+        ax0 = v2;
+
+        if (p3)
+        {
+            DECV(v3, p3);
+            ax1 = v3;
+        }
+    }
+
+} IMPL_END;
 
 IMPL_START_0(ret) {
     R(RIP) = readmem(ctx, R(RSP), 8); R(RSP) += 8;
diff --git a/vm/in/string.c b/vm/in/string.c
index 8bfe35a..edc850d 100644
--- a/vm/in/string.c
+++ b/vm/in/string.c
@@ -25,26 +25,12 @@ IMPL_START_0(stosq) { stos_impl(ctx, p1, p2, 8); } IMPL_END;
 
 //----------------------------------------------------------------------------//
 
-static void lods_impl(ctx_t *ctx, acc_t *p1, acc_t *p2, uint len)
-{
-    R(p1->reg) = readmem(ctx, R(p2->reg), len);
-    R(RFX) = (R(p1->reg) == 0 ? R(RFX)|ZF : R(RFX)&~ZF);
-    STR_MOVE(p2->reg, len);
-}
-
-IMPL_START_0(lodsb) { lods_impl(ctx, p1, p2, 1); } IMPL_END;
-IMPL_START_0(lodsw) { lods_impl(ctx, p1, p2, 2); } IMPL_END;
-IMPL_START_0(lodsl) { lods_impl(ctx, p1, p2, 4); } IMPL_END;
-IMPL_START_0(lodsq) { lods_impl(ctx, p1, p2, 8); } IMPL_END;
-
-//----------------------------------------------------------------------------//
-
 static void scas_impl(ctx_t *ctx, acc_t *p1, acc_t *p2, uint len)
 {
     DECV(v2, p2);
 
     ulong x = readmem(ctx, R(p1->reg), len);
-    COMPARE(x, v2);
+    COMPARE_SUB(x, v2);
 
     if (x == 0) {
         R(RFX) |= ZF;
@@ -62,45 +48,6 @@ IMPL_START_0(scasq) { scas_impl(ctx, p1, p2, 8); } IMPL_END;
 
 //----------------------------------------------------------------------------//
 
-static void cmps_impl(ctx_t *ctx, acc_t *p1, acc_t *p2, uint len)
-{
-    ulong x1 = readmem(ctx, R(p1->reg), len);
-    ulong x2 = readmem(ctx, R(p2->reg), len);
-
-    COMPARE(x1, x2);
-
-    STR_MOVE(p1->reg, len);
-    STR_MOVE(p2->reg, len);
-}
-
-IMPL_START_0(cmpsb) { cmps_impl(ctx, p1, p2, 1); } IMPL_END;
-IMPL_START_0(cmpsw) { cmps_impl(ctx, p1, p2, 2); } IMPL_END;
-IMPL_START_0(cmpsl) { cmps_impl(ctx, p1, p2, 4); } IMPL_END;
-IMPL_START_0(cmpsq) { cmps_impl(ctx, p1, p2, 8); } IMPL_END;
-
-//----------------------------------------------------------------------------//
-
-static void cmpzs_impl(ctx_t *ctx, acc_t *p1, acc_t *p2, uint len)
-{
-    ulong x1 = readmem(ctx, R(p1->reg), len);
-    ulong x2 = readmem(ctx, R(p2->reg), len);
-
-    COMPARE(x1, x2);
-
-    if (!x1 && !x2)
-        R(RFX) &= ~ZF;
-
-    STR_MOVE(p1->reg, len);
-    STR_MOVE(p2->reg, len);
-}
-
-IMPL_START_0(cmpzsb) { cmpzs_impl(ctx, p1, p2, 1); } IMPL_END;
-IMPL_START_0(cmpzsw) { cmpzs_impl(ctx, p1, p2, 2); } IMPL_END;
-IMPL_START_0(cmpzsl) { cmpzs_impl(ctx, p1, p2, 4); } IMPL_END;
-IMPL_START_0(cmpzsq) { cmpzs_impl(ctx, p1, p2, 8); } IMPL_END;
-
-//----------------------------------------------------------------------------//
-
 static void movs_impl(ctx_t *ctx, acc_t *p1, acc_t *p2, uint len)
 {
     ulong x = readmem(ctx, R(p2->reg), len);
diff --git a/vm/in/super.c b/vm/in/super.c
index cb362d4..9ccfa34 100644
--- a/vm/in/super.c
+++ b/vm/in/super.c
@@ -42,7 +42,6 @@ IMPL_START_1(trap) {
     _except(ctx, v1 + 256, "TRAP instruction");
 } IMPL_END;
 
-IMPL_START_0(into) { INTO(); } IMPL_END;
 
 IMPL_START_0(iret) {
     if (ctx->dumpsw)
@@ -55,6 +54,9 @@ IMPL_START_0(iret) {
 }
 IMPL_END;
 
+IMPL_START_0(cli) { CHK_SUPERV(); R(CR0) &= ~IF; } IMPL_END;
+IMPL_START_0(sti) { CHK_SUPERV(); R(CR0) |= IF; } IMPL_END;
+
 //----------------------------------------------------------------------------//
 
 //
diff --git a/vm/pc/arch.h b/vm/pc/arch.h
index 87fd207..6ab9981 100644
--- a/vm/pc/arch.h
+++ b/vm/pc/arch.h
@@ -102,7 +102,7 @@ extern void do_hlt(ctx_t *ctx);
 #include <pc/regs.h>
 #include <pc/decode.h>
 #include <pc/except.h>
-#include <in/arch_i.h>
+#include <ob/arch_i.h>
 
 extern ctx_t main_ctx;
 extern reg_t arch_r[NREGS];
diff --git a/vm/pc/exec.c b/vm/pc/exec.c
index 093f079..01412f9 100644
--- a/vm/pc/exec.c
+++ b/vm/pc/exec.c
@@ -2,7 +2,6 @@
 // See the LICENSE file in the project root for more information.
 
 #include <pc/arch.h>
-#include <in/arch_i.h>
 
 #define rfx R(RFX)