diff --git a/as/k-as.py b/as/k-as.py
index 8612a48..8c40abb 100755
--- a/as/k-as.py
+++ b/as/k-as.py
@@ -437,6 +437,14 @@ pref2len = {
     "q" : 8,
 }
 
+i_aliases = {
+    "j_1" : "jmp_1",
+    "m_2" : "mov_2",
+    "b_3" : "bch_3",
+    "inc_2" : "add_2",
+    "dec_2" : "sub_2",
+}
+
 def parse_instr(line):
     if line == None or len(line) == 0:
         return 0
@@ -480,22 +488,20 @@ def parse_instr(line):
 
     if params == None or len(params) == 0:
         if b2 == 0:
-            instrs.write("{}".format(instr_name))
+            instrs.write("{}_0".format(instr_name))
 
         else:
             size += 1
-            instrs.write("%%suff {} %%imm8 {}".format(instr_name, b2))
+            instrs.write("%%suff {}_0 %%imm8 {}".format(instr_name, b2))
 
         return size
 
     tok = params.split(',')
 
-    # 'call' special case... temporary
-    if instr_name == 'call':
-        if len(tok) == 2:
-            instr_name = 'xcall2'
-        elif len(tok) == 3:
-            instr_name = 'xcall3'
+    instr_name += "_{}".format(len(tok))
+
+    if instr_name in i_aliases:
+        instr_name = i_aliases[instr_name]
 
     #
     # Parse operands
diff --git a/ka/crt/fmt/doprnt.k b/ka/crt/fmt/doprnt.k
index 95ea965..d1467e1 100644
--- a/ka/crt/fmt/doprnt.k
+++ b/ka/crt/fmt/doprnt.k
@@ -9,12 +9,8 @@ doprnt:
     push rbp
     mov rbp, rsp
 
-    push r12
-    push r13
-    push r14
-    push r15
-    push r16
-    push r17
+    push r12, r13, r14
+    push r15, r16, r17
 
     mov r12, ax2        ; fmt
     mov r14, ax3        ; va_list
@@ -145,13 +141,9 @@ doprnt:
 
 .epilogue:
     mov rax, r16
-    
-    pop r17
-    pop r16
-    pop r15
-    pop r14
-    pop r13
-    pop r12
+
+    pop r17, r16, r15
+    pop r14, r13, r12
 
     leave
     ret
diff --git a/ka/crt/fmt/ltostr.k b/ka/crt/fmt/ltostr.k
index 8517144..f1b9fdb 100644
--- a/ka/crt/fmt/ltostr.k
+++ b/ka/crt/fmt/ltostr.k
@@ -53,9 +53,10 @@ ltostr:
 
 .next:
     mov b[ax0], rdx
+
     inc ax0, 1
 
-    div ax1, ax1, ax2
+    div ax1, ax2
     jmp .conv
 
     ; add minus flag, null-terminate and reverse
diff --git a/ka/crt/fmt/strtol.k b/ka/crt/fmt/strtol.k
index 487bb8d..66f09dd 100644
--- a/ka/crt/fmt/strtol.k
+++ b/ka/crt/fmt/strtol.k
@@ -137,7 +137,7 @@ strtoq:
     ; too large for base?
     b.ae rcx, ax1, .done
 
-    mul rax, rax, ax1
+    mul rax, ax1
     inc rax, rcx
     jmp .main_loop
 
diff --git a/ka/crt/lib/time.k b/ka/crt/lib/time.k
index 9c6ac2c..78b7bba 100644
--- a/ka/crt/lib/time.k
+++ b/ka/crt/lib/time.k
@@ -54,12 +54,12 @@ GetTimeUTC:
 
     ; minutes
     div rcx, ax0, 60
-    rem rcx, rcx, 60
+    rem rcx, 60
     mov b[rdx+1], rcx
 
     ; hours
     div rcx, ax0, 3600
-    rem rcx, rcx, 24
+    rem rcx, 24
     mov b[rdx+2], rcx
 
     ; month days
diff --git a/ka/sys/dumprf.k b/ka/sys/dumprf.k
index af68f66..c9d8e1b 100644
--- a/ka/sys/dumprf.k
+++ b/ka/sys/dumprf.k
@@ -5,8 +5,9 @@ dumprf:
     push rbp
     mov rbp, rsp
 
-    push r12
+    push r12, r13
     mov r12, ax0
+    mov r13, rsp
 
     call RFS.LoadReg, r12, $cr2
     push rax
@@ -113,7 +114,8 @@ dumprf:
 
     call printf, .dmp6
 
-    pop r12
+    mov rsp, r13
+    pop r13, r12
 
     leave
     ret
diff --git a/ka/sys/intr/common.k b/ka/sys/intr/common.k
index 252e355..c8304bf 100644
--- a/ka/sys/intr/common.k
+++ b/ka/sys/intr/common.k
@@ -1,31 +1,6 @@
 ; The OS/K Team licenses this file to you under the MIT license.
 ; See the LICENSE file in the project root for more information.
 
-TrapHandlers.prolog:
-    mov rbp, zero
-
-    ; rax = caller's cr2
-    call RFS.LoadReg, r14, $cr2
-
-    ; we don't preserve the r12 we got
-    mov r12, rax
-    mov rdx, zero
-
-    jmp rcx
-
-TrapHandlers.epilog:
-    ; TRAP return values: RAX-RDX
-
-    mov ax2, rax
-    call RFS.StoreReg, r14, $rax
-
-    mov ax2, rdx
-    call RFS.StoreReg, r14, $rdx
-
-    call IDT.DoneHandling, r13
-
-    iret
-
 ScreenOfDeath:
     push r12
     mov r12, ax0
@@ -51,8 +26,7 @@ ScreenOfDeath:
     push .scr2_usr
 
 .do_print:
-    push r14
-    push r12
+    push r14, r12
     call printf, .scr2
     inc rsp, 24
 
diff --git a/ka/sys/intr/trap0.k b/ka/sys/intr/trap0.k
index 1b34a41..3cd2cad 100644
--- a/ka/sys/intr/trap0.k
+++ b/ka/sys/intr/trap0.k
@@ -1,6 +1,34 @@
 ; The OS/K Team licenses this file to you under the MIT license.
 ; See the LICENSE file in the project root for more information.
 
+TrapHandlers.prolog:
+    mov rbp, zero
+
+    ; rax = caller's cr2
+    call RFS.LoadReg, r14, $cr2
+
+    ; we don't preserve the r12 we got
+    mov r12, rax
+    mov rdx, zero
+
+    jmp rcx
+
+TrapHandlers.epilog:
+    ; TRAP return values: RAX-RDX
+
+    mov ax2, rax
+    call RFS.StoreReg, r14, $rax
+
+    mov ax2, rdx
+    call RFS.StoreReg, r14, $rdx
+
+    call IDT.DoneHandling, r13
+
+    iret
+
+;
+; TRAP #0 handler
+;
 DefaultTrapHandler:
 
 .init:
diff --git a/ka/usr/cmd/dir.k b/ka/usr/cmd/dir.k
index 05dcae6..ee6feee 100644
--- a/ka/usr/cmd/dir.k
+++ b/ka/usr/cmd/dir.k
@@ -7,10 +7,8 @@ builtins.dir:
     push rbp
     mov rbp, rsp
 
-    push r12
-    push r13
-    push r14
-    push r15
+    push r12, r13
+    push r14, r15
 
     mov r12, zero       # no. of files found
     mov r13, zero       # no. of directories found
@@ -110,24 +108,19 @@ builtins.dir:
 .end:
     shr rax, r14, 10
     shr rdx, rax, 10
-    and rax, rax, 1023
-    and r14, r14, 1023
+    and rax, 1023
+    and r14, 1023
 
-    push r14
-    push rax
-    push rdx
+    push r14, rax, rdx
     call printf, .endstr0
     inc rsp, 24
 
-    push r13
-    push r12
+    push r13, r12
     call printf, .endstr1
     inc rsp, 16
 
-    pop r15
-    pop r14
-    pop r13
-    pop r12
+    pop r15, r14
+    pop r13, r12
 
     leave
     ret
diff --git a/ka/usr/cmd/main.k b/ka/usr/cmd/main.k
index 02463a9..3ca1335 100644
--- a/ka/usr/cmd/main.k
+++ b/ka/usr/cmd/main.k
@@ -211,12 +211,10 @@ main:
 .handle_DATE:
     call GetTimeUTC
 
-    push b[rax+3]
     mov rcx, b[rax+4]
     inc rcx, 1
-    push rcx
-    push w[rax+6]
 
+    push b[rax+3], rcx, w[rax+6]
     call printf, .datefmt
     inc rsp, 40
 
@@ -322,9 +320,7 @@ main:
 .handle_TIME:
     call GetTimeUTC
 
-    push b[rax]
-    push b[rax+1]
-    push b[rax+2]
+    push b[rax], b[rax+1], b[rax+2]
     call printf, .timefmt
     inc rsp, 24
 
@@ -383,8 +379,7 @@ main:
 .cnf_errmsg = ": command not found\n"
 
 .file_not_found:
-    push q[argv1pos]
-    push argv0
+    push q[argv1pos], argv0
     call printf, .fnf_errmsg
     inc rsp, 16
 
@@ -393,8 +388,7 @@ main:
 .fnf_errmsg = "%s: %s: file not found\n"
 
 .empty_file:
-    push q[argv1pos]
-    push argv0
+    push q[argv1pos], argv0
     call printf, .ef_errmsg
     inc rsp, 16
 
@@ -403,8 +397,7 @@ main:
 .ef_errmsg = "%s: %s: file was empty\n"
 
 .couldnt_open:
-    push q[argv1pos]
-    push argv0
+    push q[argv1pos], argv0
     call printf, .cno_errmsg
     inc rsp, 16
 
@@ -413,8 +406,7 @@ main:
 .cno_errmsg = "%s: %s: an error occured while opening file\n"
 
 .couldnt_remove:
-    push q[argv1pos]
-    push argv0
+    push q[argv1pos], argv0
     call printf, .cne_errmsg
     inc rsp, 16
 
@@ -423,8 +415,7 @@ main:
 .cne_errmsg = "%s: %s: an error occured while removing file\n"
 
 .couldnt_read:
-    push q[argv1pos]
-    push argv0
+    push q[argv1pos], argv0
     call printf, .cnr_errmsg
     inc rsp, 16
 
diff --git a/vm/in/ALU b/vm/in/ALU
index a095baa..f284dfd 100644
--- a/vm/in/ALU
+++ b/vm/in/ALU
@@ -12,6 +12,7 @@
 #
 # Preserves all flags
 #
+or 2
 or 3
 
 #
@@ -21,6 +22,7 @@ or 3
 #
 # Preserves all flags
 #
+and 2
 and 3
 
 #
@@ -30,6 +32,7 @@ and 3
 #
 # Preserves all flags
 #
+xor 2
 xor 3
 
 #
@@ -40,7 +43,9 @@ xor 3
 #
 # Preserves all flags
 #
+shl 2
 shl 3
+shr 2
 shr 3
 
 #
@@ -50,6 +55,7 @@ shr 3
 #
 # Preserves all flags
 #
+sar 2
 sar 3
 
 #---------------------------------------------------------------------------#
@@ -76,7 +82,7 @@ cmp 2
 # Sets OF is signed integer overflow occur, clears it otherwise
 # Sets ZF and SF according to the result
 #
-inc 2
+add 2
 add 3
 addf 3
 
@@ -89,7 +95,7 @@ addf 3
 # Sets OF is signed integer overflow occur, clears it otherwise
 # Sets ZF and SF according to the result
 #
-dec 2
+sub 2
 sub 3
 subf 3
 
@@ -103,7 +109,9 @@ subf 3
 # Sets OF is signed integer overflow occur, clears it otherwise
 # Sets ZF and SF according to the result
 #
+adcx 2
 adcx 3
+sbbx 2
 sbbx 3
 
 #
@@ -114,6 +122,7 @@ sbbx 3
 # Preserves ZF and SF
 # Sets CF and OF if HI($src1 * $src2) > 0, clears them otherwise
 #
+mul 2
 mul 3
 mulf 3
 
@@ -142,6 +151,7 @@ imulhi 3
 #
 # Preserves all flags
 #
+div 2
 div 3
 
 #
@@ -151,6 +161,7 @@ div 3
 #
 # Preserves all flags
 #
+idiv 2
 idiv 3
 
 #
@@ -160,5 +171,6 @@ idiv 3
 #
 # Preserves all flags
 #
+rem 2
 rem 3
 
diff --git a/vm/in/MEM b/vm/in/MEM
index 05779f8..481cefc 100644
--- a/vm/in/MEM
+++ b/vm/in/MEM
@@ -23,7 +23,7 @@ jmp 1
 loop 1
 
 #
-# Conditional absolute jumps (B)
+# Conditional absolute jumps (branches) (BCH)
 #
 #       COMPARE(SignExtend($1), $2)
 #
@@ -38,7 +38,7 @@ loop 1
 #
 # Suffixing B with the REP suffix results in undefined behavior
 #
-b 3
+bch 3
 
 
 #---------------------------------------------------------------------------#
@@ -52,8 +52,8 @@ b 3
 #       JMP(RIP)
 #
 call 1
-xcall2 2
-xcall3 3
+call 2
+call 3
 
 #
 # Return to caller (RET)
@@ -69,6 +69,7 @@ ret
 #       RBP = RSP
 #       RSP = RSP - $1
 #
+enter
 enter 1
 
 #
@@ -86,6 +87,8 @@ leave
 #       *RSP = $1
 #
 push 1
+push 2
+push 3
 
 #
 # POP value from stack
@@ -94,6 +97,8 @@ push 1
 #       RSP = RSP + 8
 #
 pop 1
+pop 2
+pop 3
 
 #---------------------------------------------------------------------------#
 # Movement instructions                                                     #
diff --git a/vm/in/MISC b/vm/in/MISC
index 7b0533c..443d14f 100644
--- a/vm/in/MISC
+++ b/vm/in/MISC
@@ -43,7 +43,7 @@ into
 #
 # Get timestamp in µseconds
 #
-utime 2
+utime 1
 
 #
 # $1 = seconds since start of month
diff --git a/vm/in/alu.c b/vm/in/alu.c
index 35adff5..c697e63 100644
--- a/vm/in/alu.c
+++ b/vm/in/alu.c
@@ -8,30 +8,46 @@
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(or) { SRCP(p2); SRCP(p3); *r1 = p2->val | p3->val; return 1; }
-IMPL_START(and) { SRCP(p2); SRCP(p3); *r1 = p2->val & p3->val; return 1; }
-IMPL_START(xor) { SRCP(p2); SRCP(p3); *r1 = p2->val ^ p3->val; return 1; }
+IMPL_START(or, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val | p2->val; return 1; }
+IMPL_START(and, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val & p2->val; return 1; }
+IMPL_START(xor, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val ^ p2->val; return 1; }
+
+IMPL_START(or, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val | p3->val; return 1; }
+IMPL_START(and, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val & p3->val; return 1; }
+IMPL_START(xor, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val ^ p3->val; return 1; }
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(shl) { SRCP(p2); SRCP(p3); *r1 = p2->val << p3->val; return 1; }
-IMPL_START(shr) { SRCP(p2); SRCP(p3); *r1 = p2->val >> p3->val; return 1; }
-IMPL_START(sar) { SRCP(p2); SRCP(p3); *r1 = (ulong)((long)p2->val >> (long)p3->val); return 1; }
+IMPL_START(shl, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val << p2->val; return 1; }
+IMPL_START(shr, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val >> p2->val; return 1; }
+IMPL_START(sar, 2) { SRCP(p1); SRCP(p2); *r1 = (ulong)((long)p1->val >> (long)p2->val); return 1; }
+
+IMPL_START(shl, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val << p3->val; return 1; }
+IMPL_START(shr, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val >> p3->val; return 1; }
+IMPL_START(sar, 3) { SRCP(p2); SRCP(p3); *r1 = (ulong)((long)p2->val >> (long)p3->val); return 1; }
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(inc) { SRCP(p1); SRCP(p2); *r1 = p1->val + p2->val; return 1; }
-IMPL_START(add) { SRCP(p2); SRCP(p3); *r1 = p2->val + p3->val; return 1; }
-IMPL_START(addf) { SRCP(p2); SRCP(p3); COMPARE_ADD(p2->val, p3->val); *r1 = p2->val + p3->val; return 1; }
+IMPL_START(add, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val + p2->val; return 1; }
+IMPL_START(add, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val + p3->val; return 1; }
+IMPL_START(addf, 3) { SRCP(p2); SRCP(p3); COMPARE_ADD(p2->val, p3->val); *r1 = p2->val + p3->val; return 1; }
 
-IMPL_START(dec) { SRCP(p1); SRCP(p2); *r1 = p1->val - p2->val; return 1; }
-IMPL_START(sub) { SRCP(p2); SRCP(p3); *r1 = p2->val - p3->val; return 1; }
-IMPL_START(subf) { SRCP(p2); SRCP(p3); COMPARE_SUB(p2->val, p3->val); *r1 = p2->val - p3->val; return 1; }
+IMPL_START(sub, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val - p2->val; return 1; }
+IMPL_START(sub, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val - p3->val; return 1; }
+IMPL_START(subf, 3) { SRCP(p2); SRCP(p3); COMPARE_SUB(p2->val, p3->val); *r1 = p2->val - p3->val; return 1; }
 
-IMPL_START(adcx)
-{
-    SRCP(p2);
-    SRCP(p3);
+IMPL_START(adcx, 2) {
+    SRCP(p1); SRCP(p2);
+
+    p2->val += !!(R(RFX)&CF);
+    COMPARE_ADD(p1->val, p2->val);
+    *r1 = p1->val + p2->val;
+
+    return 1;
+}
+
+IMPL_START(adcx, 3) {
+    SRCP(p2); SRCP(p3);
 
     p3->val += !!(R(RFX)&CF);
     COMPARE_ADD(p2->val, p3->val);
@@ -40,10 +56,18 @@ IMPL_START(adcx)
     return 1;
 }
 
-IMPL_START(sbbx)
-{
-    SRCP(p2);
-    SRCP(p3);
+IMPL_START(sbbx, 2) {
+    SRCP(p1); SRCP(p2);
+
+    p2->val += !!(R(RFX)&CF); 
+    COMPARE_SUB(p1->val, p2->val);
+    *r1 = p1->val - p2->val;
+
+    return 1;
+}
+
+IMPL_START(sbbx, 3) {
+    SRCP(p2); SRCP(p3);
 
     p3->val += !!(R(RFX)&CF); 
     COMPARE_SUB(p2->val, p3->val);
@@ -54,29 +78,43 @@ IMPL_START(sbbx)
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(mul) { SRCP(p2); SRCP(p3); *r1 = p2->val * p3->val; return 1; }
-IMPL_START(rem) { SRCP(p2); SRCP(p3); *r1 = p2->val % p3->val; return 1; }
+IMPL_START(mul, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val * p2->val; return 1; }
+IMPL_START(mul, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val * p3->val; return 1; }
 
-IMPL_START(div)
-{
-    SRCP(p2);
-    SRCP(p3);
+IMPL_START(rem, 2) { SRCP(p1); SRCP(p2); *r1 = p1->val % p2->val; return 1; }
+IMPL_START(rem, 3) { SRCP(p2); SRCP(p3); *r1 = p2->val % p3->val; return 1; }
 
-    if (!p3->val)
-        _except(E_DIV, "DIV by 0");
+IMPL_START(div, 2) {
+    SRCP(p1); SRCP(p2);
+
+    if (!p2->val) _except(E_DIV, "DIV by 0");
+    *r1 = p1->val / p2->val;
+
+    return 1;
+}
+
+IMPL_START(div, 3) {
+    SRCP(p2); SRCP(p3);
+
+    if (!p3->val) _except(E_DIV, "DIV by 0");
     *r1 = p2->val / p3->val;
 
     return 1;
 }
 
-IMPL_START(idiv)
-{
-    SRCP(p2);
-    SRCP(p3);
+IMPL_START(idiv, 2) {
+    SRCP(p1); SRCP(p2);
 
-    if (!p3->val)
-        _except(E_DIV, "IDIV by 0");
+    if (!p3->val) _except(E_DIV, "IDIV by 0");
+    *r1 = (ulong)((long)p1->val/(long)p2->val);
 
+    return 1;
+}
+
+IMPL_START(idiv, 3) {
+    SRCP(p2); SRCP(p3);
+
+    if (!p3->val) _except(E_DIV, "IDIV by 0");
     *r1 = (ulong)((long)p2->val/(long)p3->val);
 
     return 1;
@@ -92,7 +130,7 @@ static void __unsigned_multiply128(ulong u, ulong v, ulong *hi, ulong *lo)
     *lo = r;
 }
 
-IMPL_START(mulhi)
+IMPL_START(mulhi, 3)
 {
     SRCP(p2);
     SRCP(p3);
@@ -101,7 +139,7 @@ IMPL_START(mulhi)
     return 2;
 }
 
-IMPL_START(mulf)
+IMPL_START(mulf, 3)
 {
     SRCP(p2);
     SRCP(p3);
@@ -123,7 +161,7 @@ static void __signed_multiply128(ulong u, ulong v, ulong *hi, ulong *lo)
     *lo = r;
 }
 
-IMPL_START(imulhi)
+IMPL_START(imulhi, 3)
 {
     SRCP(p2);
     SRCP(p3);
diff --git a/vm/in/arch_i.py b/vm/in/arch_i.py
index 232a40c..fa6f668 100644
--- a/vm/in/arch_i.py
+++ b/vm/in/arch_i.py
@@ -47,15 +47,15 @@ def parse_2(fp):
             assert(len(tok) == 2)
             n = tok[1]
 
-        ls.write("{}{}\n".format(deprecated, tok[0]))
+        ls.write("{}{}_{}\n".format(deprecated, tok[0], n))
 
         hd.write("#ifdef _NEED_ARCH_I\n")
-        hd.write('{{ "{}{}", {}, i_{} }},\n'\
-                    .format(deprecated, tok[0], n, tok[0]))
+        hd.write('{{ "{}{}", {}, i_{}_{} }},\n'\
+                    .format(deprecated, tok[0], n, tok[0], n))
         hd.write("#else\n")
-        hd.write("#define I_{} {}\n".format(tok[0].upper(), count))
-        hd.write("extern bool i_{}(acc_t *, acc_t *, acc_t *, ulong *, ulong *, ulong *);\n"
-                    .format(tok[0]))
+        hd.write("#define I_{}_{} {}\n".format(tok[0].upper(), n, count))
+        hd.write("extern bool i_{}_{}(acc_t *, acc_t *, acc_t *, ulong *, ulong *, ulong *);\n"
+                    .format(tok[0], n))
         hd.write("#endif\n\n")
 
         count = count + 1
diff --git a/vm/in/instrs.h b/vm/in/instrs.h
index d36e1f9..c1c925f 100644
--- a/vm/in/instrs.h
+++ b/vm/in/instrs.h
@@ -5,8 +5,8 @@
 
 //----------------------------------------------------------------------------//
 
-#define IMPL_START(name)                                                    \
-uint i_##name(acc_t *p1, acc_t *p2, acc_t *p3,                              \
+#define IMPL_START(name, n)                                                 \
+uint i_##name##_##n(acc_t *p1, acc_t *p2, acc_t *p3,                        \
                           ulong *r1, ulong *r2, ulong *r3)                  \
 
 #define XSRCP(v, p, x)                                                      \
diff --git a/vm/in/mem.c b/vm/in/mem.c
index db5d936..d9ebcc6 100644
--- a/vm/in/mem.c
+++ b/vm/in/mem.c
@@ -5,9 +5,9 @@
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(jmp) { R(RIP) = p1->val; return 0; }
+IMPL_START(jmp, 1) { R(RIP) = p1->val; return 0; }
 
-IMPL_START(loop) {
+IMPL_START(loop, 1) {
     if (R(RCX) > 0) {
         R(RCX)--;
         R(RIP) = p1->val;
@@ -15,7 +15,7 @@ IMPL_START(loop) {
     return 0;
 }
 
-IMPL_START(b)
+IMPL_START(bch, 3)
 {
     SRCP(p1);
     SRCP(p2);
@@ -28,7 +28,7 @@ IMPL_START(b)
     return 0;
 }
 
-IMPL_START(cmp)
+IMPL_START(cmp, 2)
 {
     SRCP(p1);
 
@@ -39,15 +39,15 @@ IMPL_START(cmp)
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(lea) { *r1 = p2->addr; return 1; }
-IMPL_START(mov) { XSRCP(*r1, p2, sx); return 1; }
-IMPL_START(movzx) { XSRCP(*r1, p2, zx); return 1; }
+IMPL_START(lea, 2) { *r1 = p2->addr; return 1; }
+IMPL_START(mov, 2) { XSRCP(*r1, p2, sx); return 1; }
+IMPL_START(movzx, 2) { XSRCP(*r1, p2, zx); return 1; }
 
-IMPL_START(movsxb) { SRCP(p2); *r1 = (ulong)(long)(char)p2->val; return 1; }
-IMPL_START(movsxw) { SRCP(p2); *r1 = (ulong)(long)(short)p2->val; return 1; }
-IMPL_START(movsxd) { SRCP(p2); *r1 = (ulong)(long)(int)p2->val; return 1; }
+IMPL_START(movsxb, 2) { SRCP(p2); *r1 = (ulong)(long)(char)p2->val; return 1; }
+IMPL_START(movsxw, 2) { SRCP(p2); *r1 = (ulong)(long)(short)p2->val; return 1; }
+IMPL_START(movsxd, 2) { SRCP(p2); *r1 = (ulong)(long)(int)p2->val; return 1; }
 
-IMPL_START(xchg)
+IMPL_START(xchg, 2)
 {
     SRCP(p1);
     SRCP(p2);
@@ -60,11 +60,7 @@ IMPL_START(xchg)
 
 //----------------------------------------------------------------------------//
 
-//
-// Stack manipulation instructions
-//
-
-IMPL_START(push)
+IMPL_START(push, 1)
 {
     XSRCP(p1->val, p1, zx);
 
@@ -74,7 +70,35 @@ IMPL_START(push)
     return 0;
 }
 
-IMPL_START(pop)
+IMPL_START(push, 2)
+{
+    XSRCP(p1->val, p1, zx);
+    XSRCP(p2->val, p2, zx);
+
+    R(RSP) -= 16;
+    writemem(p1->val, R(RSP) + 8, 8);
+    writemem(p2->val, R(RSP) + 0, 8);
+
+    return 0;
+}
+
+IMPL_START(push, 3)
+{
+    XSRCP(p1->val, p1, zx);
+    XSRCP(p2->val, p2, zx);
+    XSRCP(p3->val, p3, zx);
+
+    R(RSP) -= 24;
+    writemem(p1->val, R(RSP) + 16, 8);
+    writemem(p2->val, R(RSP) + 8, 8);
+    writemem(p3->val, R(RSP) + 0, 8);
+
+    return 0;
+}
+
+//----------------------------------------------------------------------------//
+
+IMPL_START(pop, 1)
 {
     *r1 = readmemzx(R(RSP), 8);
     R(RSP) += 8;
@@ -82,7 +106,28 @@ IMPL_START(pop)
     return 1;
 }
 
-IMPL_START(call)
+IMPL_START(pop, 2)
+{
+    *r1 = readmemzx(R(RSP) + 0, 8);
+    *r2 = readmemzx(R(RSP) + 8, 8);
+    R(RSP) += 16;
+    
+    return 2;
+}
+
+IMPL_START(pop, 3)
+{
+    *r1 = readmemzx(R(RSP) + 0, 8);
+    *r2 = readmemzx(R(RSP) + 8, 8);
+    *r3 = readmemzx(R(RSP) + 16, 8);
+    R(RSP) += 24;
+    
+    return 3;
+}
+
+//----------------------------------------------------------------------------//
+
+IMPL_START(call, 1)
 {
     SRCP(p1);
 
@@ -93,7 +138,7 @@ IMPL_START(call)
     return 0;
 }
 
-IMPL_START(xcall2)
+IMPL_START(call, 2)
 {
     SRCP(p1);
     SRCP(p2);
@@ -107,7 +152,7 @@ IMPL_START(xcall2)
     return 0;
 }
 
-IMPL_START(xcall3)
+IMPL_START(call, 3)
 {
     SRCP(p1);
     SRCP(p2);
@@ -123,7 +168,9 @@ IMPL_START(xcall3)
     return 0;
 }
 
-IMPL_START(ret)
+//----------------------------------------------------------------------------//
+
+IMPL_START(ret, 0)
 {
     R(RIP) = readmemzx(R(RSP), 8);
     R(RSP) += 8;
@@ -131,7 +178,16 @@ IMPL_START(ret)
     return 0;
 }
 
-IMPL_START(enter)
+IMPL_START(enter, 0)
+{
+    R(RSP) -= 8;
+    writemem(R(RBP), R(RSP), 8);
+    R(RBP) = R(RSP);
+
+    return 0;
+}
+
+IMPL_START(enter, 1)
 {
     writemem(R(RBP), R(RSP) - 8, 8);
     R(RBP) = R(RSP) - 8;
@@ -140,7 +196,7 @@ IMPL_START(enter)
     return 0;
 }
 
-IMPL_START(leave)
+IMPL_START(leave, 0)
 {
     R(RSP) = R(RBP) + 8;
     R(RBP) = readmemzx(R(RBP), 8);
diff --git a/vm/in/misc.c b/vm/in/misc.c
index a17ac6d..e28faf8 100644
--- a/vm/in/misc.c
+++ b/vm/in/misc.c
@@ -9,13 +9,13 @@
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(into) { INTO(); return 0; }
-IMPL_START(pause) { usleep(5000); return 0; }
-IMPL_START(udf) { _except(E_UDF, "UDF instruction"); }
+IMPL_START(into, 0) { INTO(); return 0; }
+IMPL_START(pause, 0) { usleep(5000); return 0; }
+IMPL_START(udf, 0) { _except(E_UDF, "UDF instruction"); }
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(break)
+IMPL_START(break, 0)
 {
 #ifndef NDEBUG
     trace("\nExecuting BREAK INSTR\n");
@@ -28,7 +28,7 @@ IMPL_START(break)
     return 0;
 }
 
-IMPL_START(dump)
+IMPL_START(dump, 0)
 {
 #ifndef NDEBUG
     if (ctx->dumpsw)
@@ -44,7 +44,7 @@ IMPL_START(dump)
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(ytime)
+IMPL_START(ytime, 3)
 {
     time_t t = time(NULL);
     struct tm *tm = localtime(&t);
@@ -59,7 +59,7 @@ IMPL_START(ytime)
     return 3;
 }
 
-IMPL_START(utime)
+IMPL_START(utime, 1)
 {
     struct timeval time;
     gettimeofday(&time, NULL);
@@ -70,7 +70,7 @@ IMPL_START(utime)
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(cls)
+IMPL_START(cls, 0)
 {
     R(RFX) = 0;
     for (int i = RAX; i <= R20; i++) R(i) = 0;
@@ -80,7 +80,7 @@ IMPL_START(cls)
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(bswap)
+IMPL_START(bswap, 2)
 {
     SRCP(p2);
 
@@ -98,7 +98,7 @@ IMPL_START(bswap)
     return 1;
 }
 
-IMPL_START(wswap)
+IMPL_START(wswap, 2)
 {
     SRCP(p2);
 
@@ -112,7 +112,7 @@ IMPL_START(wswap)
     return 1;
 }
 
-IMPL_START(dswap)
+IMPL_START(dswap, 2)
 {
     SRCP(p2);
 
@@ -128,7 +128,7 @@ IMPL_START(dswap)
 #define PRN_CLEAR_MAGIC 0x8BF00001
 #define PRN_FLUSH_MAGIC 0x8BF00002
 
-IMPL_START(prn)
+IMPL_START(prn, 1)
 {
     SRCP(p1);
 
@@ -158,7 +158,7 @@ IMPL_START(prn)
     return 0;
 }
 
-IMPL_START(prns)
+IMPL_START(prns, 1)
 {
     if (p1->type != A_REG)
         _except(E_ILL, "PRNS given a non-REG operand");
@@ -180,7 +180,7 @@ IMPL_START(prns)
     return 0;
 }
 
-IMPL_START(scan)
+IMPL_START(scan, 1)
 {
     *r1 = console_scankeybuf();
 
diff --git a/vm/in/string.c b/vm/in/string.c
index 4d04e2f..878f19c 100644
--- a/vm/in/string.c
+++ b/vm/in/string.c
@@ -20,10 +20,10 @@ static void stos_impl(acc_t *p1, acc_t *p2, uint len)
     STR_MOVE(p1->reg, len);
 }
 
-IMPL_START(stosb) { stos_impl(p1, p2, 1); return 0; }
-IMPL_START(stosw) { stos_impl(p1, p2, 2); return 0; }
-IMPL_START(stosd) { stos_impl(p1, p2, 4); return 0; }
-IMPL_START(stosq) { stos_impl(p1, p2, 8); return 0; }
+IMPL_START(stosb, 2) { stos_impl(p1, p2, 1); return 0; }
+IMPL_START(stosw, 2) { stos_impl(p1, p2, 2); return 0; }
+IMPL_START(stosd, 2) { stos_impl(p1, p2, 4); return 0; }
+IMPL_START(stosq, 2) { stos_impl(p1, p2, 8); return 0; }
 
 //----------------------------------------------------------------------------//
 
@@ -44,10 +44,10 @@ static void scas_impl(acc_t *p1, acc_t *p2, uint len)
     }
 }
 
-IMPL_START(scasb) { scas_impl(p1, p2, 1); return 0; }
-IMPL_START(scasw) { scas_impl(p1, p2, 2); return 0; }
-IMPL_START(scasd) { scas_impl(p1, p2, 4); return 0; }
-IMPL_START(scasq) { scas_impl(p1, p2, 8); return 0; }
+IMPL_START(scasb, 2) { scas_impl(p1, p2, 1); return 0; }
+IMPL_START(scasw, 2) { scas_impl(p1, p2, 2); return 0; }
+IMPL_START(scasd, 2) { scas_impl(p1, p2, 4); return 0; }
+IMPL_START(scasq, 2) { scas_impl(p1, p2, 8); return 0; }
 
 //----------------------------------------------------------------------------//
 
diff --git a/vm/in/super.c b/vm/in/super.c
index 1730e0a..b760666 100644
--- a/vm/in/super.c
+++ b/vm/in/super.c
@@ -31,20 +31,20 @@ void do_hlt()
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(hlt) { CHK_SUPERV(); do_hlt(); return 0; }
-IMPL_START(stop) { CHK_SUPERV(); _except(E_SHT, "STOP INSTR"); }
-IMPL_START(crash) { CHK_SUPERV(); _except(1023, "CRASH instruction"); }
+IMPL_START(hlt, 0) { CHK_SUPERV(); do_hlt(); return 0; }
+IMPL_START(stop, 0) { CHK_SUPERV(); _except(E_SHT, "STOP INSTR"); }
+IMPL_START(crash, 0) { CHK_SUPERV(); _except(1023, "CRASH instruction"); }
 
 //----------------------------------------------------------------------------//
 
-IMPL_START(trap)
+IMPL_START(trap, 1)
 {
     if (p1->val > 255) _except(E_ILL, "TRAP number greater than 255");
     _except(p1->val + 256, "TRAP instruction");
 }
 
 
-IMPL_START(iret) {
+IMPL_START(iret, 0) {
     if (ctx->dumpsw)
         trace("\nReturning from exception #%ld\n\n", R(R13));
 
@@ -56,8 +56,8 @@ IMPL_START(iret) {
     return 0;
 }
 
-IMPL_START(cli) { CHK_SUPERV(); R(CR0) &= ~IF; return 0; }
-IMPL_START(sti) { CHK_SUPERV(); R(CR0) |= IF; return 0; }
+IMPL_START(cli, 0) { CHK_SUPERV(); R(CR0) &= ~IF; return 0; }
+IMPL_START(sti, 0) { CHK_SUPERV(); R(CR0) |= IF; return 0; }
 
 //----------------------------------------------------------------------------//
 
@@ -76,7 +76,7 @@ dev_t *devctl_common(ulong idx)
     return NULL;
 }
 
-IMPL_START(devctl)
+IMPL_START(devctl, 2)
 {
     CHK_SUPERV();
 
@@ -98,7 +98,7 @@ IMPL_START(devctl)
     return 0;
 }
 
-IMPL_START(iocall)
+IMPL_START(iocall, 2)
 {
     CHK_SUPERV();
 
diff --git a/vm/pc/decode.c b/vm/pc/decode.c
index af94461..0a3272f 100644
--- a/vm/pc/decode.c
+++ b/vm/pc/decode.c
@@ -199,6 +199,8 @@ void decode(void)
     ushort b1, b2, rep = 0, lock;
     acc_t p1 = { 0 }, p2 = { 0 }, p3 = { 0 };
 
+    //logerr("decodin'\n");
+
     ctx->cur_pc = R(RIP);
 
     // Address range check
diff --git a/vm/pc/exec.c b/vm/pc/exec.c
index 8f63221..d9d3517 100644
--- a/vm/pc/exec.c
+++ b/vm/pc/exec.c
@@ -66,7 +66,7 @@ void exec_instr(instr_t *in,
     // For REPs we evaluate the condition AFTER running the instruction,
     // in a do ... while(cond) fashion
     //
-    if (ctx->cond && in->func != i_b)     // 'B' instruction is special
+    if (ctx->cond && in->func != i_bch_3)     // 'B' instruction is special
         if (!rep && !eval_cond(ctx->cond))
             return;
 
diff --git a/vm/pc/main.c b/vm/pc/main.c
index effe1aa..70ab912 100644
--- a/vm/pc/main.c
+++ b/vm/pc/main.c
@@ -13,21 +13,10 @@ static ushort *fwprog;
 
 ctx_t main_ctx;
 
-void sigcommon(void)
-{
-    _except(E_BRK, "SIGNAL'ed");
-}
+void sigcommon(void) { _except(E_BRK, "SIGNAL'ed"); }
+void sigint(int _) { sigcommon(); }
 
-void sigint(int _)
-{
-    sigcommon();
-}
-
-void sigsegv(int _)
-{
-    logerr("Segmentation fault\n");
-    sigcommon();
-}
+void sigsegv(int _) { logerr("\nCaught: Segmentation fault\n\n"); dumpregs(); die(1); }
 
 jmp_buf exc_jmp_buf;
 
@@ -38,16 +27,12 @@ void main_loop(void)
 {
     setjmp(exc_jmp_buf);
 
-    //
     // Start decoding
-    //
     while (!dying) {
         // Execute one instruction
         decode();
-
         console_update();
     }
-
     die(0);
 }
 
@@ -55,13 +40,13 @@ int main(int argc, char **argv)
 {
     FILE *fwfile;
 
-    main_ctx.r = arch_r;
-    main_ctx.i = arch_i;
-
 #if 0 && !defined(NDEBUG)
     main_ctx.dumpsw = 1;
 #endif
 
+    main_ctx.r = arch_r;
+    main_ctx.i = arch_i;
+
     //
     // srand
     //
diff --git a/vm/pc/mem.c b/vm/pc/mem.c
index 4758ea0..2df44d3 100644
--- a/vm/pc/mem.c
+++ b/vm/pc/mem.c
@@ -162,8 +162,6 @@ void writemem(ulong val, ulong addr, uint len)
     GETREAL();
     CHK_RANGE();
 
-    // log("writemem: 0x%lX: 0x%lX (%d)\n", addr, val, len);
-
     switch (len) {
         case 1: writemem8(val, real, addr); break;
         case 2: writemem16(val, real, addr); break;