diff --git a/tools/bifrost-assemble.py b/tools/bifrost-assemble.py
index cfdc607d0430706c7b80958eb5d2f4c2373e7e0c..469c676d0fd7b4fa665cb8426fe2422cd04c7f07 100755
--- a/tools/bifrost-assemble.py
+++ b/tools/bifrost-assemble.py
@@ -956,18 +956,19 @@ class ImmediateSlot:
             return ConstantSrc(token.read_type == ImmediateToken.ReadType.HIGH64)
 
     def encode_contents(self):
-        return sum(Bits(length=t.bitlen, uint=t.value) for t in self.contents)
+        return sum(Bits(length=t.bitlen, uint=t.value) for t in
+                   reversed(self.contents))
 
     def encode_const_field(self):
-        return bitstring.pack('uint:1=0, uint:3, bits:4',
-                              self.IDX_MAP[self.idx],
-                              self.encode_contents()[60:64])
+        encoded = BitArray(8)
+        encoded[1:4] = self.IDX_MAP[self.idx]
+        encoded[4:8] = self.encode_contents()[60:64]
+        return encoded
 
     def __repr__(self):
         return "<ImmediateSlot #%d at 0x%x; contents=%s>" % (
             self.idx, id(self), self.contents)
 
-
 class Uniform:
     def __init__(self, idx):
         self.idx = idx
@@ -1110,6 +1111,24 @@ class Clause:
 
     MAX_CONSTS_ALLOWED = 5 # FIXME: maybe make this 6 later?
 
+    QUADWORD_FORMATS = {
+        # (instructions, constants): pos
+        (1,0): 0x0,
+        (2,0): 0x1,
+        (4,0): 0x2,
+        (3,1): 0x3,
+        (5,1): 0x4,
+        (4,2): 0x5,
+        (7,0): 0x6,
+        (6,1): 0x7,
+        (5,3): 0x8,
+        (8,1): 0x9,
+        (7,2): 0xa,
+        (6,3): 0xb,
+        (8,3): 0xc,
+        (7,4): 0xd
+    }
+
     def _header_flag(func):
         old_func = func
         def func(self, value):
@@ -1426,6 +1445,11 @@ class Clause:
             # immediate zero slot.
             if inst.has_pending_immediates():
                 pending = inst.reg_file.const_port
+                # FIXME: it looks like the compiler makes some interesting
+                # decisions regarding where it puts dummy clauses, for some
+                # reason consts 8/9 are swapped with consts 10/11 when the
+                # last instruction is the one reusing the same constant twice
+                # (and thus only requiring that half the slot be filled)
                 if all(t.value == 0 for t in pending):
                     inst.resolve_immediates(self, ImmediateZeroSlot)
                 elif pending.bitlen == 64:
@@ -1451,6 +1475,7 @@ class Clause:
                         first_inst.reg_file.add_reg_write(write.idx, stage)
         del self.__pending_writes
 
+        # Resolve any pending state in our instructions
         for inst in self.instructions:
             inst.reg_file.assign_reg_ports()
 
@@ -1468,10 +1493,8 @@ class Clause:
                 inst.resolve_immediates(self, slot)
                 self.immediate_slots.append(slot)
 
-    # TODO: Make sure we return whether or not the last format we used
-    # contained an inline immediate
     @staticmethod
-    def _next_immediate_to_encode(cls, immediates):
+    def _next_immediate_to_encode(immediates):
         if immediates:
             return immediates.pop(0)
         else:
@@ -1524,7 +1547,6 @@ class Clause:
         # Format 3
         quadword = BitArray(128)
         i3 = instructions.pop(0) if instructions else None
-        used_fmt_3_2 = False
 
         if not i3:
             # Format 3.1
@@ -1538,9 +1560,8 @@ class Clause:
             quadword[15:45] = i2[3:33]
             quadword[60:120] = next_immediate[0:60]
             quadword[120:128] = tag
-        elif instructions:
+        elif len(instructions) == 1:
             # Format 3.2
-            used_fmt_3_2 = True
             next_immediate = self._next_immediate_to_encode(immediates)
             tag = 0b10
 
@@ -1552,7 +1573,9 @@ class Clause:
             quadword[125:128] = i2[0:3]
         else:
             # Format 3.3
-            if immediates:
+            if instructions:
+                tag = 0b00000001
+            elif immediates:
                 tag = 0b00000101
             else:
                 tag = 0b01000101
@@ -1565,14 +1588,14 @@ class Clause:
 
         encoded.append(quadword)
         if not instructions:
-            return encoded, packed_immediate
+            return encoded
 
         # Format 4
         quadword = BitArray(128)
         i4 = instructions.pop(0)
         i5 = instructions.pop(0) if instructions else None
 
-        if used_fmt_3_2:
+        if not i5:
             # Format 4.1
             if immediates:
                 tag = 0b00010
@@ -1657,11 +1680,40 @@ class Clause:
         return encoded
 
     def encode(self):
-        encoded = BitString()
         immediates = [s.encode_contents() for s in self.immediate_slots]
 
-        # TODO: FINISH IT.
-        return self._encode_quadwords(immediates)
+        encoded = BitStream()
+        instructions = self._encode_quadwords(immediates)
+        encoded.append(instructions)
+
+        # We may have managed to pack one of the constants into the previous
+        # quadword
+        if len(immediates) < len(self.immediate_slots):
+            immediate_count = 1
+        else:
+            immediate_count = 0
+
+        instruction_count = len(self.instructions)
+        while immediates:
+            quadword = BitArray(128)
+            if len(immediates) > 2:
+                # More constants follow
+                tag = 0b0011
+            else:
+                tag = 0b0111
+
+            quadword[60:120] = immediates.pop(0)[0:60]
+            if immediates:
+                quadword[0:60] = immediates.pop(0)[0:60]
+
+            quadword[120:124] = tag
+            quadword[124:128] = self.QUADWORD_FORMATS[instruction_count,
+                                                      immediate_count]
+
+            encoded.append(quadword)
+            immediate_count += 2
+
+        return encoded
 
     def dump(self):
         print('Clause #%d' % self.idx)