vm

as/k-as.py

@@ -308,14 +308,12 @@ def parse_preproc(line):
                         sys.exit(1)
 
                 written = b_data.write(ord(c).to_bytes(1, byteorder='little', signed=False))
-                assert(written == 1)
                 real_len += 1
                 pdata += 1
 
             # align
             for i in range(8 - len(s) % 8):
                 written = b_data.write(int(0).to_bytes(1, byteorder='little', signed=False))
-                assert(written == 1)
                 pdata += 1
 
             pdefs[label + "_len"] = str(real_len)
@@ -364,7 +362,7 @@ def get_cond_mask(cond, line):
         leave()
         sys.exit(1)
 
-    return (mask | pconds[cond]) << 10
+    return (mask | pconds[cond])
 
 #-------------------------------------------------------------------------------
 
@@ -419,9 +417,11 @@ def parse_instr(line):
 
     size = 4
 
-    # Word 2 (rep|cond|ft1|ft2)
+    # Word 2 (rep|ft3|ft2|ft1)
     w2 = 0
 
+    cond = None
+
     if len(instr) > 2 and '.' in instr:
         instr, suf = instr.split('.', 1)
         
@@ -440,20 +440,26 @@ def parse_instr(line):
             w2 |= 0x8000 # REP
         
         if len(suf) > 0:
-            w2 |= get_cond_mask(suf, line)
+            instrs.write("%%cond ")
+            cond = "%%imm16 {}".format(get_cond_mask(suf, line))
 
     instr_name = instr
     instr_args = ''
 
     if params == None or len(params) == 0:
         instrs.write("{} ".format(instr_name))
+
+        if cond != None:
+            size += 2
+            instrs.write("{} ".format(cond))
+
         instrs.write("%%imm16 {}".format(w2))
 
         return size
 
     tok = params.split(',')
     
-    # FT1 and FT2
+    # FTn
     fts = ''
 
     #
@@ -680,29 +686,43 @@ def parse_instr(line):
         fellthrough = False
 
     #
-    # Compute FT1 and FT2
+    # Compute FTn
     #
-    if ' ' in fts:
-        assert(len(fts.split()) == 2)
+    l = len(fts.split())
+
+    if l == 3:
+        ft1, ft2, ft3 = fts.split()
+
+        w2 |= get_fts_mask(ft3, line) << 10
+        w2 |= get_fts_mask(ft2, line) << 5
+        w2 |= get_fts_mask(ft1, line)
+
+    elif l == 2:
         ft1, ft2 = fts.split()
         
-        w2 |= get_fts_mask(ft1, line) << 5
-        w2 |= get_fts_mask(ft2, line)
+        w2 |= get_fts_mask(ft2, line) << 5
+        w2 |= get_fts_mask(ft1, line)
 
     else:
-        assert(len(fts) > 0)
-        w2 |= get_fts_mask(fts, line) << 5
+        assert(l == 1)
+        w2 |= get_fts_mask(fts, line)
 
+    if cond == None:
         instrs.write("{} %%imm16 {}{}".format(instr_name, w2, instr_args))
+    else:
+        size += 2
+        instrs.write("{} {} %%imm16 {}{}".format(instr_name, cond, w2, instr_args))
+
     return size
 
 #-------------------------------------------------------------------------------
 
 special_syms = {
+    "%%cond",
     "%%imm16",
     "%%imm32",
     "%%imm64",
-    "%%signed"
+    "%%signed",
 }
 
 def gentext():
@@ -712,6 +732,8 @@ def gentext():
 
     instrs.seek(0)
 
+    cond_mask = 0
+
     for _, line in enumerate(instrs):
         tok = line.strip().split()
 
@@ -742,7 +764,8 @@ def gentext():
                 continue
 
             if word in pinstrs:
-                idx = pinstrs.index(word)
+                idx = pinstrs.index(word) | cond_mask
+                cond_mask = 0
                 b_text.write(idx.to_bytes(2, byteorder='little', signed=False))
                 continue
 
@@ -763,13 +786,13 @@ def gentext():
                     lastimm = 4
                 elif word == "%%imm64":
                     lastimm = 8
-
+                elif word == "%%cond":
+                    cond_mask = (1 << 13)
                 elif word == "%%signed":
                     lastimm = 2
                     isSigned = True
                 else:
                     isSigned = False
-
                 continue
 
             if is_number(word):

ka/ABI

@@ -46,7 +46,7 @@ Between the 'enter' and the 'leave', the stack looks like this:
         |       .       |   rsp+8       |
         |       .       |   rsp+0       /
         |---------------|
-        .  (undefined)  .
+        .   (red zone)  .
         .               .
         .               .
 
@@ -131,9 +131,7 @@ The following registers can only be used by the supervisor:
 - Debugging registers for the supervisor:
     dr0-dr7
 
-The following registers cannot be referenced by machine code at all; they must
-be manipulated through specific instructions, and manipulating their values
-allow for controlling the machine in various ways:
+The following registers cannot be referenced by machine code at all:
     px0-px1, fc0-fc2, cr0-cr7
 
 #------------------------------------------------------------------------------#

ka/usr/cmd/dir.k

@@ -42,8 +42,8 @@ CMD.builtins.dir:
     j.b .print_file_name
 
     ; calculate where to put extension
-    lea r11, b[r10 - 1]
-    sub r11, .buf
+    ksub r11, r10, .buf
+    dec r11
 
 .ext_pad:
     ; print at least 11 non-space characters before extension

vm/in/FLAGS

@@ -27,19 +27,3 @@ cmc
 clc
 stc
 
-#
-# Load FLG register (LODF)
-#
-#       $1 = FLG
-#
-lodf rm
-
-#
-# Store FLG register - lower byte only (STOFB)
-#
-#       FLG[7:0] = $1[7:0]
-#
-# Note: FLG's lower byte contains CF, OF, ZF, SF, PF and DF
-#
-stofb rm
-

vm/in/INSTRS

@@ -13,4 +13,5 @@ include "FLAGS"
 include "INOUT"
 include "DEBUG"
 include "STRING"
+include "KINSTRS"
 

vm/in/JUMPS

@@ -6,7 +6,7 @@
 #---------------------------------------------------------------------------#
 
 #
-# Absolute jump (JMP) instruction
+# Jump (JMP) instruction
 #
 #       RIP = $1
 #
@@ -14,7 +14,7 @@ j ri
 jmp ri
 
 #
-# RCX-dependent absolute jump (LOOP) instruction
+# RCX-dependent jump (LOOP) instruction
 #
 #       IF (RCX > 0) THEN
 #           RCX = RCX - 1

vm/in/KINSTRS

@@ -0,0 +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.
+
+#---------------------------------------------------------------------------#
+# 3-operand instructions                                                    #
+#---------------------------------------------------------------------------#
+
+#
+# 2-sources logical & arithmetical operations
+#
+# Preserves CF and OF
+# Sets ZF and SF according to the result
+#
+kor rm rim rim
+kand rm rim rim
+kxor rm rim rim
+knor rm rim rim
+korn rm rim rim
+knand rm rim rim
+kandn rm rim rim
+kxnor rm rim rim
+kxorn rm rim rim
+kshl rm rim rim
+kshr rm rim rim
+kadd rm rim rim
+kadc rm rim rim
+kado rm rim rim
+ksub rm rim rim
+ksbb rm rim rim
+ksbo rm rim rim
+kmul rm rim rim
+kdiv rm rim rim
+krem rm rim rim
+

vm/in/arch_i.py

@@ -17,10 +17,11 @@ def getflag(s):
 
     return "__FLAG_ERROR__"
 
-def doprnt(fp, i, p1, p2):
+def doprnt(fp, i, p1, p2, p3):
     for c1 in p1:
         for c2 in p2:
-            fp.write("{} {} {}".format(i, c1, c2).strip())
+            for c3 in p3:
+                fp.write("{} {} {} {}".format(i, c1, c2, c3).strip())
                 fp.write('\n')
 
 def parse_1(fi, fp):
@@ -40,28 +41,34 @@ def parse_1(fi, fp):
             parse_1(fi2, fp)
             continue
 
-        tok = line.split(' ')
+        tok = line.split()
 
         if len(tok) == 0:
             continue
 
-        i = tok[0].strip()
+        i = tok[0]
 
         if len(tok) == 1:
-            doprnt(fp, i, ' ', ' ')
+            doprnt(fp, i, ' ', ' ', ' ')
             continue
 
         if len(tok) == 2:
-            p = tok[1].strip()
-            doprnt(fp, i, p, ' ')
+            p = tok[1]
+            doprnt(fp, i, p, ' ', ' ')
             continue
 
-        assert(len(tok) == 3)
+        if len(tok) == 3:
+            p1 = tok[1]
+            p2 = tok[2]
+            doprnt(fp, i, p1, p2, ' ')
+            continue
 
-        p1 = tok[1].strip()
-        p2 = tok[2].strip()
+        assert(len(tok) == 4)
+        p1 = tok[1]
+        p2 = tok[2]
+        p3 = tok[3]
 
-        doprnt(fp, i, p1, p2)
+        doprnt(fp, i, p1, p2, p3)
 
     fi.close()
 
@@ -84,30 +91,40 @@ def parse_2(fp):
             name = tok[0]
             p1 = "NOPRM"
             p2 = "NOPRM"
+            p3 = "NOPRM"
 
         elif len(tok) == 2:
-            name = "{}_{}".format(tok[0], tok[1].strip())
+            name = "{}_{}".format(tok[0], tok[1])
             p1 = getflag(tok[1])
             p2 = "NOPRM"
+            p3 = "NOPRM"
 
         elif len(tok) == 3:
-            name = "{}_{}_{}".format(tok[0], tok[1], tok[2].strip())
+            name = "{}_{}_{}".format(tok[0], tok[1], tok[2])
             p1 = getflag(tok[1])
             p2 = getflag(tok[2])
+            p3 = "NOPRM"
+
+        elif len(tok) == 4:
+            name = "{}_{}_{}_{}".format(tok[0], tok[1], tok[2], tok[3])
+            p1 = getflag(tok[1])
+            p2 = getflag(tok[2])
+            p3 = getflag(tok[3])
 
         else:
             name = "__TOK_ERROR__"
             p1 = "__TOK_ERROR__"
             p2 = "__TOK_ERROR__"
+            p3 = "__TOK_ERROR__"
 
         ls.write("{}{}\n".format(deprecated, name))
 
         hd.write("#ifdef _NEED_ARCH_I\n")
-        hd.write('{{ "{}{}", "{}{}", {}, {}, i_{} }},\n'\
-                    .format(deprecated, tok[0], deprecated, name, p1, p2, tok[0]))
+        hd.write('{{ "{}{}", "{}{}", {}, {}, {}, i_{} }},\n'\
+                    .format(deprecated, tok[0], deprecated, name, p1, p2, p3, tok[0]))
         hd.write("#else\n")
         hd.write("#define I_{} {}\n".format(name.upper(), count))
-        hd.write("extern bool i_{}(ctx_t *, acc_t *, acc_t *, ulong *, ulong *);\n"
+        hd.write("extern bool i_{}(ctx_t *, acc_t *, acc_t *, acc_t *, ulong *, ulong *, ulong *);\n"
                     .format(tok[0]))
         hd.write("#endif\n\n")
 

vm/in/flags.c

@@ -55,17 +55,3 @@ IMPL_END;
 
 //----------------------------------------------------------------------------//
 
-IMPL_START_1(lodf)
-{
-    v1 = flg;
-}
-IMPL_OUT;
-
-IMPL_START_1(stofb)
-{
-    flg = (flg & ~0xFF) | (v1 & 0xFF);
-}
-IMPL_END;
-
-//----------------------------------------------------------------------------//
-

vm/in/instrs.h

@@ -31,20 +31,31 @@
 //----------------------------------------------------------------------------//
 
 #define IMPL_START_0(name)                                                  \
-bool i_##name(ctx_t *ctx, acc_t *p1, acc_t *p2, ulong *r1, ulong *r2)       \
+bool i_##name(ctx_t *ctx, acc_t *p1, acc_t *p2, acc_t *p3,                  \
+                          ulong *r1, ulong *r2, ulong *r3)                  \
 {
 
 #define IMPL_START_1(name)                                                  \
-bool i_##name(ctx_t *ctx, acc_t *p1, acc_t *p2, ulong *r1, ulong *r2)       \
+bool i_##name(ctx_t *ctx, acc_t *p1, acc_t *p2, acc_t *p3,                  \
+                          ulong *r1, ulong *r2, ulong *r3)                  \
 {                                                                           \
     DECV(v1, p1);
 
 #define IMPL_START_2(name)                                                  \
-bool i_##name(ctx_t *ctx, acc_t *p1, acc_t *p2, ulong *r1, ulong *r2)       \
+bool i_##name(ctx_t *ctx, acc_t *p1, acc_t *p2, acc_t *p3,                  \
+                          ulong *r1, ulong *r2, ulong *r3)                  \
 {                                                                           \
     DECV(v1, p1);                                                           \
     DECV(v2, p2);
 
+#define IMPL_START_3(name)                                                  \
+bool i_##name(ctx_t *ctx, acc_t *p1, acc_t *p2, acc_t *p3,                  \
+                          ulong *r1, ulong *r2, ulong *r3)                  \
+{                                                                           \
+    DECV(v1, p1);                                                           \
+    DECV(v2, p2);                                                           \
+    DECV(v3, p3);
+
 #define IMPL_OUT_ZSF                                                        \
     SET_ZSF(v1);                                                            \
     IMPL_OUT
@@ -60,6 +71,13 @@ bool i_##name(ctx_t *ctx, acc_t *p1, acc_t *p2, ulong *r1, ulong *r2)       \
     return 2;                                                               \
 }
 
+#define IMPL_OUT_3                                                          \
+    *r1 = v1;                                                               \
+    *r2 = v2;                                                               \
+    *r3 = v3;                                                               \
+    return 3;                                                               \
+}
+
 #define IMPL_END                                                            \
     return 0;                                                               \
 }

vm/in/kinstrs.c

@@ -0,0 +1,40 @@
+// The OS/K Team licenses this file to you under the MIT license.
+// See the LICENSE file in the project root for more information.
+
+#include <in/instrs.h>
+
+//----------------------------------------------------------------------------//
+
+#define K_IMPL_ALU(name, expr)                                              \
+IMPL_START_3(k##name)                                                       \
+{                                                                           \
+    v1 = expr;                                                              \
+}                                                                           \
+IMPL_OUT_ZSF
+
+K_IMPL_ALU(or, v2 | v3);
+K_IMPL_ALU(and, v2 & v3);
+K_IMPL_ALU(xor, v2 ^ v3);
+
+K_IMPL_ALU(nor, ~(v2 | v3));
+K_IMPL_ALU(orn, ~v2 | v3);
+K_IMPL_ALU(nand, ~(v2 & v3));
+K_IMPL_ALU(andn, ~v2 & v3);
+K_IMPL_ALU(xnor, ~(v2 ^ v3));
+K_IMPL_ALU(xorn, ~v2 ^ v3);
+
+K_IMPL_ALU(shl, v2 << v3);
+K_IMPL_ALU(shr, v2 >> v3);
+
+K_IMPL_ALU(add, v2 + v3);
+K_IMPL_ALU(adc, v2 + v3 + !!(flg&CF));
+K_IMPL_ALU(ado, v2 + v3 + !!(flg&OF));
+K_IMPL_ALU(sub, v2 - v3);
+K_IMPL_ALU(sbb, v2 - v3 - !!(flg&CF));
+K_IMPL_ALU(sbo, v2 - v3 - !!(flg&OF));
+K_IMPL_ALU(mul, v2 * v3);
+K_IMPL_ALU(div, v2 / v3);
+K_IMPL_ALU(rem, v2 % v3);
+
+//----------------------------------------------------------------------------//
+

vm/pc/DECD

@@ -5,14 +5,15 @@ Instruction encoding:
 
 15             0 15             0
 xxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxx
-││└────────────┘ │└───┘└───┘└───┘
-││    INSTR.     │ CND  FT1  FT2
-││               │
-LM               R
+│││└───────────┘ │└───┘└───┘└───┘
+│││   INSTR.     │ FT3  FT2  FT1
+│││              │
+LMC              R
 
 M       No more words
-L       LOCK prefix
-R       REP prefix
+C       COND suffix
+L       LOCK suffix
+R       REP suffix
 FT1     First operand format
 FT2     Second operand format
 

vm/pc/decode.c

@@ -7,7 +7,7 @@
 // Read the "DECD" file before reading this code
 //
 
-static void check_param_type(ctx_t *ctx, instr_t *in, uint prm, uchar fmt)
+static void check_param_type(ctx_t *ctx, instr_t *in, uint prm, uchar fmt, int which)
 {
     bool ok;
 
@@ -22,8 +22,8 @@ static void check_param_type(ctx_t *ctx, instr_t *in, uint prm, uchar fmt)
 
     if (!ok)
         _except(ctx, E_ILL,
-            "FT1 or FT2 not matching %s's expected parameter types: "
-            "fmt=0x%x prm=0x%x", in->full, fmt, prm);
+            "FT%d not matching %s's expected parameter types: "
+            "fmt=0x%x prm=0x%x", which, in->full, fmt, prm);
 }
 
 //
@@ -146,27 +146,21 @@ void extract_param(ctx_t *ctx, acc_t *p, uchar fmt)
 
 void decode(ctx_t *ctx)
 {
-    char *illmsg;
-
     instr_t *in;
 
     acc_t p1 = { 0 };
     acc_t p2 = { 0 };
+    acc_t p3 = { 0 };
 
     bool rep = 0;
     uint cond = 0;
     bool lock, nomore;
 
     ushort w1, w2;
-    uchar f1 = 0, f2 = 0;
+    uchar f1 = 0, f2 = 0, f3 = 0;
 
     rpc = rip;
 
-/*
-    sym_t *sym = find_sym_by_addr(pc);
-    if (sym)
-        trace("0x%lX: %s:\n", pc, sym->name);
-*/
     //
     // Process the first word of the instruction
     //
@@ -174,10 +168,15 @@ void decode(ctx_t *ctx)
 
     // Extract first word flags
 
-    lock = !!(w1 & PREF_LOCK);
-    nomore = !!(w1 & PREF_NOMORE);
+    lock = !!(w1 & SUFF_LOCK);
+    nomore = !!(w1 & SUFF_NOMORE);
 
-    w1 &= ~(PREF_LOCK|PREF_NOMORE);
+    if (w1 & SUFF_COND)
+    {
+        cond = ctx->get(ctx);
+    }
+
+    w1 &= ~(SUFF_LOCK|SUFF_NOMORE|SUFF_COND);
 
     // Find instruction
 
@@ -197,12 +196,12 @@ void decode(ctx_t *ctx)
         w2 = ctx->get(ctx);
 
         // REP and COND
-        rep = !!(w2 & PREF_REP);
-        cond = (w2 & BITS_COND) >> COND_SHIFT;
+        rep = !!(w2 & SUFF_REP);
 
-        // F1 and F2
-        f1 = (w2 >> F1_SHIFT) & Fx_MASK;
-        f2 = w2 & Fx_MASK;
+        // Fn
+        f3 = (w2 >> F3_SHIFT) & Fx_MASK;
+        f2 = (w2 >> F2_SHIFT) & Fx_MASK;
+        f1 = w2 & Fx_MASK;
     }
 
     //
@@ -211,17 +210,18 @@ void decode(ctx_t *ctx)
 
     if (in->prm1 == NOPRM)
     {
-        if (f1 || f2)
+        if (f1 || f2 || f3)
         {
-            illmsg = "FT1 and/or FT2 filled for 0-param INSTR";
-            goto ill;
+            _except(ctx, E_ILL,
+                "FTn (%u,%u,%u) filled for 0-param INSTR '%s'",
+                    f1, f2, f3, in->full);
         }
 
-        exec_instr(ctx, in, NULL, NULL, lock, rep, cond);
+        exec_instr(ctx, in, NULL, NULL, NULL, lock, rep, cond);
         return;
     }
 
-    check_param_type(ctx, in, in->prm1, f1);
+    check_param_type(ctx, in, in->prm1, f1, 1);
     extract_param(ctx, &p1, f1);
 
     //
@@ -230,23 +230,40 @@ void decode(ctx_t *ctx)
 
     if (in->prm2 == NOPRM)
     {
-        if (f2)
+        if (f2 || f3)
         {
-            illmsg = "FT2 filled for 1-param INSTR";
-            goto ill;
+            _except(ctx, E_ILL,
+                "FT3/FT2 (%u,%u,%u) filled for 1-param INSTR '%s'",
+                    f1, f2, f3, in->full);
         }
 
-        exec_instr(ctx, in, &p1, NULL, lock, rep, cond);
+        exec_instr(ctx, in, &p1, NULL, NULL, lock, rep, cond);
         return;
     }
 
-    check_param_type(ctx, in, in->prm2, f2);
+    check_param_type(ctx, in, in->prm2, f2, 2);
     extract_param(ctx, &p2, f2);
 
-    exec_instr(ctx, in, &p1, &p2, lock, rep, cond);
-    return;
+    //
+    // Deal with operand 3
+    //
 
-ill:
-    _except(ctx, E_ILL, illmsg);
+    if (in->prm3 == NOPRM)
+    {
+        if (f3)
+        {
+            _except(ctx, E_ILL,
+                "FT3 (%u,%u,%u) filled for 2-param INSTR '%s'",
+                    f1, f2, f3, in->full);
+        }
+
+        exec_instr(ctx, in, &p1, &p2, NULL, lock, rep, cond);
+        return;
+    }
+
+    check_param_type(ctx, in, in->prm3, f3, 3);
+    extract_param(ctx, &p3, f3);
+
+    exec_instr(ctx, in, &p1, &p2, &p3, lock, rep, cond);
 }
 

vm/pc/decode.h

@@ -13,18 +13,14 @@ enum
 
 enum
 {
-    PREF_LOCK = (1 << 15),
-    PREF_NOMORE = (1 << 14),
-
-    PREF_REP = (1 << 15),
-
-    BITS_COND = (1 << 14) | (1 << 13) | (1 << 12)
-              | (1 << 11) | (1 << 10),
-
-    COND_SHIFT = 10,
+    SUFF_LOCK = (1 << 15),
+    SUFF_NOMORE = (1 << 14),
+    SUFF_COND = (1 << 13),
+    SUFF_REP = (1 << 15),
 
     Fx_MASK = 0x1F,
-    F1_SHIFT = 5,
+    F3_SHIFT = 10,
+    F2_SHIFT = 5,
 };
 
 enum
@@ -79,14 +75,17 @@ struct instr_t
 
     uint    prm1;
     uint    prm2;
+    uint    prm3;
 
-    bool  (*func)(ctx_t *, acc_t *, acc_t *, ulong *, ulong *);
+    bool  (*func)(ctx_t *, acc_t *, acc_t *, acc_t *,
+                           ulong *, ulong *, ulong *);
 };
 
 void exec_instr(ctx_t *ctx,
                 instr_t *in,
                 acc_t *p1,
                 acc_t *p2,
+                acc_t *p3,
                 bool lock,
                 bool rep,
                 uint cond);
@@ -99,6 +98,7 @@ void dump_instr(ctx_t *ctx,
                 instr_t *in,
                 acc_t *p1,
                 acc_t *p2,
+                acc_t *p3,
                 bool lock,
                 bool rep,
                 uint cond);

vm/pc/dump.c

@@ -23,6 +23,7 @@ void dump_instr(ctx_t *ctx,
                 instr_t *in,
                 acc_t *p1,
                 acc_t *p2,
+                acc_t *p3,
                 bool lock,
                 bool rep,
                 uint cond)
@@ -62,12 +63,16 @@ void dump_instr(ctx_t *ctx,
         trace("\t");
 
     if (p1)
-    {
         dump_acc(ctx, p1);
+
     if (p2) {
         trace(", ");
         dump_acc(ctx, p2);
     }
+
+    if (p3) {
+        trace(", ");
+        dump_acc(ctx, p3);
     }
 
     trace("\n");

vm/pc/exec.c

@@ -48,12 +48,13 @@ void exec_instr(ctx_t *ctx,
                 instr_t *in,
                 acc_t *p1,
                 acc_t *p2, 
+                acc_t *p3, 
                 bool lock,
                 bool rep,
                 uint cond)
 {
     bool out;
-    ulong r1 = 0, r2 = 0;
+    ulong r1 = 0, r2 = 0, r3 = 0;
 
     // Global instruction counter
     ctx->ninstrs++;
@@ -75,14 +76,13 @@ void exec_instr(ctx_t *ctx,
         return;
     }
 
-
 #ifndef NDEBUG
-    dump_instr(ctx, in, p1, p2, lock, rep, cond);
+    dump_instr(ctx, in, p1, p2, p3, lock, rep, cond);
 #endif
 
 do_rep:
 
-    out = in->func(ctx, p1, p2, &r1, &r2);
+    out = in->func(ctx, p1, p2, p3, &r1, &r2, &r3);
 
     if (out)
     {
@@ -99,7 +99,7 @@ do_rep:
         }
     }
     
-    if (out == 2)
+    if (out >= 2)
     {
         if (p2->type == A_REG)
             R(p2->reg) = r2;
@@ -114,6 +114,21 @@ do_rep:
         }
     }
 
+    if (out >= 3)
+    {
+        if (p3->type == A_REG)
+            R(p3->reg) = r3;
+
+        else if (p3->type == A_IMM64)
+            _except(ctx, E_ACC, "Trying to output to an IMM64");
+
+        else
+        {
+            assert(ACC_IS_MEM(p3));
+            writemem(ctx, r3, p3->addr, p3->mlen);
+        }
+    }
+
     if (rep)
     {
         // RCX remains untouched when condition fails
@@ -130,7 +145,7 @@ do_rep:
 
 #ifndef NDEBUG
         // Show that we're REP'ing
-        dump_instr(ctx, in, p1, p2, lock, rep, cond);
+        dump_instr(ctx, in, p1, p2, p3, lock, rep, cond);
 #endif
 
         goto do_rep;

vm/pc/regs.c

@@ -6,11 +6,11 @@
 reg_t arch_r[] =
 {
     { "inv", RES },
-    { "flg", RES },
+    { "flg", GPR },
     { "rip", GPR },
     { "rpc", GPR },
-    { "px0", CTL },
-    { "px1", CTL },
+    { "px0", RES },
+    { "px1", RES },
     { "fc1", RES },
     { "fc2", RES },
     { "sa0", SYS },

vm/pc/regs.h

@@ -37,8 +37,6 @@ struct reg_t
     ulong   flags;
 };
 
-#define rx0 ctx->ninstrs
-
 enum
 {
     INV, FLG, RIP, RPC, PX0, PX1, FC1, FC2,
@@ -64,6 +62,8 @@ enum
     NREGS
 };
 
+#define rx0 ctx->ninstrs
+
 #define inv R(INV)
 #define rip R(RIP)
 #define rpc R(RPC)