diff --git a/core/SoftFloat/softfloat.c b/core/SoftFloat/softfloat.c
index e4d3aab..7c45b61 100644
--- a/core/SoftFloat/softfloat.c
+++ b/core/SoftFloat/softfloat.c
@@ -631,10 +631,6 @@ static floatx80
     int8 roundingMode;
     flag roundNearestEven, increment, isTiny;
     int64 roundIncrement, roundMask, roundBits;
-
-    // [shoebill]
-    // printf("roundAndPackFloatx80: prec=%d sign=%d exp=%u sig0=0x%016llx sig1=0x%016llx\n",
-           // roundingPrecision, zSign, zExp, zSig0, zSig1);
     
     roundingMode = float_rounding_mode;
     roundNearestEven = ( roundingMode == float_round_nearest_even );
@@ -3191,9 +3187,7 @@ int32 floatx80_to_int32( floatx80 a )
     aExp = extractFloatx80Exp( a );
     aSign = extractFloatx80Sign( a );
     if ( ( aExp == 0x7FFF ) && (bits64) ( aSig<<1 ) ) aSign = 0;
-    // [shoebill] there is no way 0x4037 is the right constant to use here.
-    // 1.0 has exp=0x3fff, 0x4037 - 0x3fff == 56.
-    // (aSig >> 56) shifts out the mantissa bits, plus 8 more.
+    
     shiftCount = 0x4037 - aExp;
     if ( shiftCount <= 0 ) shiftCount = 1;
     shift64RightJamming( aSig, shiftCount, &aSig );
diff --git a/core/core_api.c b/core/core_api.c
index 4233853..d3a4d9c 100644
--- a/core/core_api.c
+++ b/core/core_api.c
@@ -115,15 +115,15 @@ void *_cpu_thread (void *arg)
     pthread_mutex_lock(&shoe.cpu_thread_lock);
     
     while (1) {
-        if (shoe.cpu_thread_notifications) {
+        if sunlikely(shoe.cpu_thread_notifications) {
             
             // If there's an interrupt pending
-            if (shoe.cpu_thread_notifications & 0xff) {
+            if slikely(shoe.cpu_thread_notifications & 0xff) {
                 // process_pending_interrupt() may clear SHOEBILL_STATE_STOPPED
                 process_pending_interrupt();
             }
             
-            if (shoe.cpu_thread_notifications & SHOEBILL_STATE_RETURN) {
+            if sunlikely(shoe.cpu_thread_notifications & SHOEBILL_STATE_RETURN) {
                 pthread_mutex_unlock(&shoe.cpu_thread_lock);
                 return NULL;
             }
diff --git a/core/cpu.c b/core/cpu.c
index 240df22..bffcbb7 100644
--- a/core/cpu.c
+++ b/core/cpu.c
@@ -55,9 +55,9 @@ static const _cc_func evaluate_cc[16] = {
 };
 
 
-#define nextword() ({const uint16_t w=lget(shoe.pc,2); if (shoe.abort) {return;}; shoe.pc+=2; w;})
-#define nextlong() ({const uint32_t L=lget(shoe.pc,4); if (shoe.abort) {return;}; shoe.pc+=4; L;})
-#define verify_supervisor() {if (!sr_s()) {throw_privilege_violation(); return;}}
+#define nextword() ({const uint16_t w=lget(shoe.pc,2); if sunlikely(shoe.abort) {return;}; shoe.pc+=2; w;})
+#define nextlong() ({const uint32_t L=lget(shoe.pc,4); if sunlikely(shoe.abort) {return;}; shoe.pc+=4; L;})
+#define verify_supervisor() {if sunlikely(!sr_s()) {throw_privilege_violation(); return;}}
 
 
 static void inst_callm(void) {
@@ -466,9 +466,9 @@ static void inst_abcd (void) {
         // FIXME: these addresses aren't predecremented (check whether a7 is incremented 2 bytes)
         assert(!"acbd is broken");
         packed_x = lget(shoe.a[x], 1);
-        if (shoe.abort) return ;
+        if sunlikely(shoe.abort) return ;
         packed_y = lget(shoe.a[y], 1);
-        if (shoe.abort) return ;
+        if sunlikely(shoe.abort) return ;
     }
     else {
         packed_x = shoe.d[x] & 0xff;
@@ -494,7 +494,7 @@ static void inst_abcd (void) {
     
     if (m) {
         lset(shoe.a[x]-1, 1, packed_sum);
-        if (shoe.abort) return ;
+        if sunlikely(shoe.abort) return ;
         
         shoe.a[x]--;
         if (x != y)
@@ -580,10 +580,10 @@ static void inst_stop (void) {
 
 static void inst_rtr (void) {
     const uint16_t ccr = lget(shoe.a[7], 2);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     const uint32_t pc = lget(shoe.a[7]+2, 4);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     shoe.a[7] += 6;
     
@@ -595,7 +595,7 @@ static void inst_rtr (void) {
 static void inst_rtd (void) {
     const int16_t disp = nextword();
     const uint32_t new_pc = lget(shoe.a[7], 4);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     shoe.pc = new_pc;
     shoe.a[7] += 4;
@@ -606,13 +606,13 @@ static void inst_rte (void) {
     verify_supervisor();
     
     const uint16_t sr = lget(shoe.a[7], 2);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     const uint32_t pc = lget(shoe.a[7]+2, 4);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     const uint16_t format_word = lget(shoe.a[7]+6, 2);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     // slog("rte: sr=0x%04x pc=0x%08x format=0x%04x, post-pop a7=0x%08x\n", sr, pc, format_word, shoe.a[7]+8);
     
@@ -819,9 +819,9 @@ static void inst_addx (void) {
         const uint32_t predec_x = shoe.a[x] - sz;
         
         const uint32_t S = lget(predec_y, 4);
-        if (shoe.abort) return ;
+        if sunlikely(shoe.abort) return ;
         const uint32_t D = lget(predec_x, 4);
-        if (shoe.abort) return ;
+        if sunlikely(shoe.abort) return ;
         
         // Sm = mib(S + extend_bit, sz); // FIXME: sure you want to do this?
         Sm = mib(S, sz);
@@ -832,7 +832,7 @@ static void inst_addx (void) {
         
         const uint32_t chopped_R = chop(R, sz);
         lset(chopped_R, sz, shoe.dat);
-        if (shoe.abort) return ;
+        if sunlikely(shoe.abort) return ;
         
         shoe.a[y] = predec_y;
         shoe.a[x] = predec_x;
@@ -926,10 +926,10 @@ static void inst_cmpm (void) {
     const uint32_t ay = shoe.a[y] + ((x==y) ? sz : 0); // if x==y, then we pop the stack twice
     
     const uint32_t dst = lget(ax, sz);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     const uint32_t src = lget(ay, sz);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     // *I believe* that if x==y, then that register will only be incremented *once*
     // WRONG!
@@ -1202,7 +1202,7 @@ static void inst_move (void) {
     // So in this case, we need to switch back to the original a7 (set_sr(shoe.orig_sr)) before modifying it
     // (This is hacky)
     
-    if (shoe.abort) {
+    if sunlikely(shoe.abort) {
         if (m == 4 || m == 3) {
             const uint16_t new_sr = shoe.sr;
             const uint8_t delta = ((r==7) && (sz==1)) ? 2 : sz;
@@ -1357,12 +1357,12 @@ static void inst_moves (void) {
         assert(! ((a && (M&7)==r) && (((M>>3) == 3) || ((M>>3) == 4))) );
         
         lset_fc(addr, sz, data, fc);
-        if (shoe.abort)
+        if sunlikely(shoe.abort)
             return ;
     }
     else {
         uint32_t data = lget_fc(addr, sz, fc);
-        if (shoe.abort)
+        if sunlikely(shoe.abort)
             return ;
         
         // if destination is address register, data is sign extended to 32 bits
@@ -1758,7 +1758,7 @@ static void inst_pea (void) {
     call_ea_addr(M);
     // push it onto the stack
     lset(shoe.a[7]-4, 4, shoe.dat);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     // decrement the stack pointer if lset didn't abort
     shoe.a[7] -= 4;
 }
@@ -1775,10 +1775,10 @@ static void inst_subx (void) {
         const uint32_t predecrement_y = shoe.a[y]-predecrement_sz;
 
         const uint32_t src = lget(predecrement_x, sz);
-        if (shoe.abort) return ;
+        if sunlikely(shoe.abort) return ;
         
         const uint32_t dst = lget(predecrement_y, sz);
-        if (shoe.abort) return ;
+        if sunlikely(shoe.abort) return ;
         
         const uint32_t result = dst - src - (sr_x()?1:0);
         
@@ -1868,7 +1868,7 @@ static void inst_bsr (void) {
     }
     
     lset(shoe.a[7]-4, 4, shoe.pc);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     shoe.a[7] -= 4;
     shoe.pc = new_pc;
 }
@@ -1942,7 +1942,7 @@ static void inst_jsr (void) {
     // my quadra 800 doesn't object when a7 is odd... 
     // so, no extra error checking needed
     lset(shoe.a[7]-4, 4, shoe.pc);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     // slog("jsr: writing pc (0x%08x) to *0x%08x (phys=0x%08x)\n", shoe.pc, shoe.a[7]-4, shoe.physical_addr);
     
@@ -1997,7 +1997,7 @@ static void inst_link_word (void) {
     
     // push the contents of the address register onto the stack
     lset(shoe.a[7]-4, 4, shoe.a[r]);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     shoe.a[7] -= 4;
     
     // load the updated stack pointer into the address register
@@ -2011,7 +2011,7 @@ static void inst_unlk (void) {
     ~decompose(shoe.op, 0100 1110 0101 1 rrr);
     
     const uint32_t pop = lget(shoe.a[r], 4);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     // loads the stack pointer from the address register
     shoe.a[7] = shoe.a[r]+4;
@@ -2022,7 +2022,7 @@ static void inst_unlk (void) {
 
 static void inst_rts (void) {
     const uint32_t pop = lget(shoe.a[7], 4);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     
     shoe.a[7] += 4;
     shoe.pc = pop;
@@ -2059,7 +2059,7 @@ static void inst_link_long (void) {
     
     // push the contents of the address register onto the stack
     lset(shoe.a[7]-4, 4, shoe.a[r]);
-    if (shoe.abort) return ;
+    if sunlikely(shoe.abort) return ;
     shoe.a[7] -= 4;
     
     // load the updated stack pointer into the address register
@@ -2093,7 +2093,7 @@ static void inst_movem (void) {
                     tmp = lget(shoe.dat, 4);
                 else 
                     tmp = (uint32_t)((int32_t)((int16_t)lget(shoe.dat, 2))); // sign-extend if short-mode
-                if (shoe.abort) goto abort;
+                if sunlikely(shoe.abort) goto abort;
                 shoe.d[i] = tmp;
                 shoe.dat += sz;
             }
@@ -2107,7 +2107,7 @@ static void inst_movem (void) {
                     tmp = lget(shoe.dat, 4);
                 else 
                     tmp = (uint32_t)((int32_t)((int16_t)lget(shoe.dat, 2))); // sign-extend if short-mode
-                if (shoe.abort) goto abort;
+                if sunlikely(shoe.abort) goto abort;
                 shoe.a[i] = tmp;
                 shoe.dat += sz;
             }
@@ -2152,7 +2152,7 @@ static void inst_movem (void) {
             // FIXME: determine what happens when the predecrementing address register is written
             if ((dfield >> i) & 1) {
                 lset(addr, sz, shoe.d[i]);
-                if (shoe.abort) 
+                if sunlikely(shoe.abort) 
                     goto abort; // FIXME: figure out how to abort cleanly
                 addr += sz;
             }
@@ -2175,7 +2175,7 @@ static void inst_movem (void) {
                 }
                 
                 lset(addr, sz, data);
-                if (shoe.abort) 
+                if sunlikely(shoe.abort) 
                     goto abort; // FIXME: figure out how to abort cleanly
                 addr += sz;
             }
@@ -2233,26 +2233,26 @@ static void inst_movep (void) {
     switch (m) {
         case 0: { // word, mem->reg
             uint16_t val = lget(addr, 1);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             val = (val << 8) | lget(addr + 2, 1);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             set_d(d, val, 2);
             break;
         }
         case 1: { // long, mem->reg
             uint32_t val = lget(addr, 1);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             val = (val << 8) | lget(addr + 2, 1);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             val = (val << 8) | lget(addr + 4, 1);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             val = (val << 8) | lget(addr + 6, 1);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             shoe.d[d] = val;
             break;
@@ -2260,26 +2260,26 @@ static void inst_movep (void) {
         case 2: { // word, reg->mem
             const uint16_t val = shoe.d[d];
             lset(addr + 0, 1, (val >> 8) & 0xff);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             lset(addr + 2, 1, (val >> 0) & 0xff);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             break;
         }
         case 3: { // long, reg->mem
             const uint32_t val = shoe.d[d];
             lset(addr + 0, 1, (val >> 24) & 0xff);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             lset(addr + 2, 1, (val >> 16) & 0xff);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             lset(addr + 4, 1, (val >>  8) & 0xff);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             lset(addr + 6, 1, (val >>  0) & 0xff);
-            if (shoe.abort) return;
+            if sunlikely(shoe.abort) return;
             
             break;
         }
@@ -2330,7 +2330,7 @@ void write_bitfield(const uint32_t width, const uint32_t offset, const uint32_t
             //slog("write_bitfield: byte_mask = 0x%02x field_mask = 0x%02x bit_offset=%u byte_offset=%u\n", byte_mask, field_mask, bit_offset, byte_offset);
             //slog("write_bitfield: changing byte at 0x%08x from 0x%02x to 0x%02x\n",
                    //first_byte_addr, old_byte, new_byte);
-            if (shoe.abort) return ;
+            if sunlikely(shoe.abort) return ;
         }
         else {
             uint32_t boff = bit_offset;
@@ -2340,7 +2340,7 @@ void write_bitfield(const uint32_t width, const uint32_t offset, const uint32_t
             uint32_t remaining_field = field<<(32-width); // left-aligned
             while (remaining_width > 0) {
                 const uint8_t byte = lget(addr, 1);
-                if (shoe.abort) return ;
+                if sunlikely(shoe.abort) return ;
                 
                 const uint8_t mask = ~~(((uint8_t)((0xff) << (8-curwidth))) >> boff);
                 const uint8_t field_chunk = remaining_field >> (32-curwidth);
@@ -2351,7 +2351,7 @@ void write_bitfield(const uint32_t width, const uint32_t offset, const uint32_t
                 lset(addr, 1, (byte & mask) | rotated_chunk);
                 //slog("write_bitfield: changing byte at 0x%08x from 0x%02x to 0x%02x\n",
                        //addr, byte, (byte & mask) | rotated_chunk);
-                if (shoe.abort) return ;
+                if sunlikely(shoe.abort) return ;
                 
                 addr++;
                 remaining_field <<= curwidth;
@@ -2399,10 +2399,10 @@ uint32_t extract_bitfield(const uint32_t width, const uint32_t offset, const uin
         
         field = bitchop(lget(first_byte_addr, 1), 8-bit_offset);
         //slog("debug: extract_bitfield: first byte field (low %u bits): 0x%02x\n", 8-bit_offset, field);
-        if (shoe.abort) return 0;
+        if sunlikely(shoe.abort) return 0;
         if (width > (8-bit_offset)) { // if the data isn't entirely contained in the first byte
             uint32_t last_long = lget(first_byte_addr+1, 4);
-            if (shoe.abort) return 0;
+            if sunlikely(shoe.abort) return 0;
             field = (field<<(width - (8-bit_offset))) | // first_byte, left shifted
             (last_long >> (32 - (width - (8-bit_offset)))); // last_long, right shifted
         }
@@ -2429,7 +2429,7 @@ static void inst_bfextu (void) {
     }
     
     const uint32_t field = extract_bitfield(width, offset, M, ea);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     
     shoe.d[r] = field;
     
@@ -2455,9 +2455,9 @@ static void inst_bfchg (void) {
     }
     
     const uint32_t field = extract_bitfield(width, offset, M, ea);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     write_bitfield(width, offset, M, ea, ~~field);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     
     set_sr_c(0);
     set_sr_v(0);
@@ -2481,7 +2481,7 @@ static void inst_bfexts (void) {
     }
     
     const uint32_t field = extract_bitfield(width, offset, M, ea);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     const uint32_t mib = (field >> (width-1))&1;
     
     const uint32_t maskA = (((uint32_t)0) - mib) << 1;
@@ -2511,9 +2511,9 @@ static void inst_bfclr (void) {
     }
     
     const uint32_t field = extract_bitfield(width, offset, M, ea);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     write_bitfield(width, offset, M, ea, 0);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     
     set_sr_c(0);
     set_sr_v(0);
@@ -2537,9 +2537,9 @@ static void inst_bfset (void) {
     }
     
     const uint32_t field = extract_bitfield(width, offset, M, ea);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     write_bitfield(width, offset, M, ea, 0xffffffff);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     
     set_sr_c(0);
     set_sr_v(0);
@@ -2563,7 +2563,7 @@ static void inst_bftst (void) {
     }
     
     const uint32_t field = extract_bitfield(width, offset, M, ea);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     
     set_sr_c(0);
     set_sr_v(0);
@@ -2587,7 +2587,7 @@ static void inst_bfffo (void) {
     }
     
     const uint32_t field = extract_bitfield(width, offset, M, ea);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     
     uint32_t i;
     for (i=1; (i<=width) && ((field>>(width-i))&1)==0; i++) ;
@@ -2618,7 +2618,7 @@ static void inst_bfins (void) {
     }
     
     write_bitfield(width, offset, M, ea, field);
-    if (shoe.abort) return;
+    if sunlikely(shoe.abort) return;
     
     set_sr_c(0);
     set_sr_v(0);
@@ -2752,20 +2752,20 @@ static void inst_ext (void) {
         case ~b(010): { // byte -> word
             uint16_t val = (int8_t)get_d(r, 1);
             set_d(r, val, 2);
-            set_sr_z(get_d(r, 2));
-            set_sr_n(mib(shoe.d[r], 2));
+            set_sr_z(!val);
+            set_sr_n(val >> 15);
             break;
         } case ~b(011): { // word -> long
             uint32_t val = (int16_t)get_d(r, 2);
             set_d(r, val, 4);
-            set_sr_z(get_d(r, 4));
-            set_sr_n(mib(shoe.d[r], 4));
+            set_sr_z(!val);
+            set_sr_n(val >> 31);
             break;
         } case ~b(111): { // byte -> long
             uint32_t val = (int8_t)get_d(r, 1);
             set_d(r, val, 4);
-            set_sr_z(get_d(r, 4));
-            set_sr_n(mib(shoe.d[r], 4));
+            set_sr_z(!val);
+            set_sr_n(val >> 31);
             break;
         }
     }
@@ -2980,14 +2980,16 @@ static void inst_trap (void) {
     set_sr_s(1);
     
     push_a7(vector_offset, 2);
-        if (shoe.abort) goto fail;
+    if sunlikely(shoe.abort) goto fail;
+    
     push_a7(shoe.pc, 4);
-        if (shoe.abort) goto fail;
+    if sunlikely(shoe.abort) goto fail;
+    
     push_a7(shoe.orig_sr, 2);
-        if (shoe.abort) goto fail;
+    if sunlikely(shoe.abort) goto fail;
     
     const uint32_t newpc = lget(shoe.vbr + vector_offset, 4);
-        if (shoe.abort) goto fail;
+    if sunlikely(shoe.abort) goto fail;
     
     shoe.pc = newpc;
     return ;
@@ -3007,7 +3009,7 @@ void cpu_step()
     shoe.orig_sr = shoe.sr;
     
     // Is this an odd address? Throw an address exception!
-    if (shoe.pc & 1) {
+    if sunlikely(shoe.pc & 1) {
         // throw_address_error(shoe.pc, 0);
         // I'm leaving this assert in here for now because it almost always indicates a bug in the emulator when it fires
         assert(!"odd PC address (probably a bug)");
@@ -3018,7 +3020,7 @@ void cpu_step()
     shoe.op = lget(shoe.pc, 2);
     
     // If there was an exception, then the pc changed. Restart execution from the beginning.
-    if (shoe.abort) {
+    if sunlikely(shoe.abort) {
         shoe.abort = 0;
         return ;
     }
diff --git a/core/mem.c b/core/mem.c
index bd1d0ef..af3e94c 100644
--- a/core/mem.c
+++ b/core/mem.c
@@ -34,17 +34,13 @@
 void _physical_get_ram (void)
 {
     uint64_t *addr;
-    if (shoe.physical_addr < shoe.physical_mem_size)
+    if slikely(shoe.physical_addr < shoe.physical_mem_size)
         addr = (uint64_t*)&shoe.physical_mem_base[shoe.physical_addr];
     else
         addr = (uint64_t*)&shoe.physical_mem_base[shoe.physical_addr % shoe.physical_mem_size];
     
     const uint8_t bits = (8 - shoe.physical_size) * 8;
     shoe.physical_dat = ntohll(*addr) >> bits;
-    
-    if ((shoe.physical_addr >= 256) && (shoe.physical_addr < 0x4000)) {
-        slog("LOMEM get: *0x%08x = 0x%x\n", shoe.physical_addr, (uint32_t)shoe.physical_dat);
-    }
 }
 
 void _physical_get_rom (void)
@@ -99,7 +95,7 @@ void _physical_get_io (void)
 void _physical_get_super_slot (void)
 {
     const uint32_t slot = shoe.physical_addr >> 28;
-    if (shoe.slots[slot].connected)
+    if slikely(shoe.slots[slot].connected)
         shoe.physical_dat = shoe.slots[slot].read_func(shoe.physical_addr,
                                                        shoe.physical_size,
                                                        slot);
@@ -111,7 +107,7 @@ void _physical_get_super_slot (void)
 void _physical_get_standard_slot (void)
 {
     const uint32_t slot = (shoe.physical_addr >> 24) & 0xf;
-    if (shoe.slots[slot].connected)
+    if slikely(shoe.slots[slot].connected)
         shoe.physical_dat = shoe.slots[slot].read_func(shoe.physical_addr,
                                                        shoe.physical_size,
                                                        slot);
@@ -342,7 +338,7 @@ static void translate_logical_addr()
     // TODO: Check limit here
     
     // If root descriptor is invalid, throw a bus error
-    if (rp_dt(rootp) == 0) {
+    if sunlikely(rp_dt(rootp) == 0) {
         throw_bus_error(shoe.logical_addr, shoe.logical_is_write);
         return ;
     }
@@ -377,7 +373,7 @@ static void translate_logical_addr()
         const uint8_t dt = desc_dt(desc, desc_size);
         
         // If this descriptor is invalid, throw a bus error
-        if (dt == 0) {
+        if sunlikely(dt == 0) {
             throw_bus_error(shoe.logical_addr, shoe.logical_is_write);
             return ;
         }
@@ -395,7 +391,7 @@ static void translate_logical_addr()
             get_desc(desc & 0xfffffff0, (4 << desc_size));
             
             // I think it's possible for an indirect descriptor to point to an invalid descriptor...
-            if (desc_dt(desc, desc_size) == 0) {
+            if sunlikely(desc_dt(desc, desc_size) == 0) {
                 throw_bus_error(shoe.logical_addr, shoe.logical_is_write);
                 return ;
             }
@@ -426,7 +422,7 @@ search_done:
     wp |= desc_wp(desc, desc_size); // or in the wp flag for this page descriptor
     
     // And finally throw a bus error
-    if (wp && shoe.logical_is_write) {
+    if sunlikely(wp && shoe.logical_is_write) {
         throw_bus_error(shoe.logical_addr, shoe.logical_is_write);
         return ;
     }
@@ -467,11 +463,11 @@ void logical_get (void)
 {
     
     // If address translation isn't enabled, this is a physical address
-    if (!tc_enable()) {
+    if sunlikely(!tc_enable()) {
         shoe.physical_addr = shoe.logical_addr;
         shoe.physical_size = shoe.logical_size;
         physical_get();
-        if (shoe.abort) {
+        if sunlikely(shoe.abort) {
             shoe.abort = 0;
             throw_long_bus_error(shoe.logical_addr, 0);
             return ;
@@ -491,21 +487,21 @@ void logical_get (void)
     shoe.logical_is_write = 0;
     
     // Common case: the read is contained entirely within a page
-    if (!((pageoffset + logical_size - 1) >> ps)) {
-        if (!check_pmmu_cache()) {
+    if slikely(!((pageoffset + logical_size - 1) >> ps)) {
+        if sunlikely(!check_pmmu_cache()) {
             translate_logical_addr();
-            if (shoe.abort)
+            if sunlikely(shoe.abort)
                 return ;
         }
         
-        if (shoe.physical_addr < shoe.physical_mem_size) {
+        if slikely(shoe.physical_addr < shoe.physical_mem_size) {
             // Fast path
             shoe.logical_dat = ntohll(*(uint64_t*)&shoe.physical_mem_base[shoe.physical_addr]) >> ((8-logical_size)*8);
         }
         else {
             shoe.physical_size = logical_size;
             physical_get();
-            if (shoe.abort) {
+            if sunlikely(shoe.abort) {
                 shoe.abort = 0;
                 throw_long_bus_error(logical_addr, 0);
                 return ;
@@ -521,9 +517,9 @@ void logical_get (void)
         
         shoe.logical_addr = addr_a;
         shoe.logical_size = size_a;
-        if (!check_pmmu_cache()) {
+        if sunlikely(!check_pmmu_cache()) {
             translate_logical_addr();
-            if (shoe.abort)
+            if sunlikely(shoe.abort)
                 return ;
         }
         
@@ -531,16 +527,16 @@ void logical_get (void)
         
         shoe.logical_addr = addr_b;
         shoe.logical_size = size_b;
-        if (!check_pmmu_cache()) {
+        if sunlikely(!check_pmmu_cache()) {
             translate_logical_addr();
-            if (shoe.abort)
+            if sunlikely(shoe.abort)
                 return ;
         }
         
         const uint32_t p_addr_b = shoe.physical_addr;
         shoe.physical_size = size_b;
         physical_get();
-        if (shoe.abort) {
+        if sunlikely(shoe.abort) {
             shoe.abort = 0;
             throw_long_bus_error(shoe.logical_addr, 0);
             return ;
@@ -550,7 +546,7 @@ void logical_get (void)
         shoe.physical_addr = p_addr_a;
         shoe.physical_size = size_a;
         physical_get();
-        if (shoe.abort) {
+        if sunlikely(shoe.abort) {
             shoe.abort = 0;
             throw_long_bus_error(shoe.logical_addr, 0);
             return ;
@@ -563,7 +559,7 @@ void logical_get (void)
 void logical_set (void)
 {
     // If address translation isn't enabled, this is a physical address
-    if (!tc_enable()) {
+    if sunlikely(!tc_enable()) {
         shoe.physical_addr = shoe.logical_addr;
         shoe.physical_size = shoe.logical_size;
         shoe.physical_dat = shoe.logical_dat;
@@ -583,11 +579,11 @@ void logical_set (void)
     shoe.logical_is_write = 1;
     
     // Common case: this write is contained entirely in one page
-    if (!((pageoffset + logical_size - 1) >> ps)) {
+    if slikely(!((pageoffset + logical_size - 1) >> ps)) {
         // Common case: the write is contained entirely within a page
-        if (!check_pmmu_cache()) {
+        if sunlikely(!check_pmmu_cache()) {
             translate_logical_addr();
-            if (shoe.abort)
+            if sunlikely(shoe.abort)
                 return ;
         }
         
@@ -605,18 +601,18 @@ void logical_set (void)
         
         shoe.logical_addr = addr_a;
         shoe.logical_size = size_a;
-        if (!check_pmmu_cache()) {
+        if sunlikely(!check_pmmu_cache()) {
             translate_logical_addr();
-            if (shoe.abort)
+            if sunlikely(shoe.abort)
                 return ;
         }
         const uint32_t p_addr_a = shoe.physical_addr;
         
         shoe.logical_addr = addr_b;
         shoe.logical_size = size_b;
-        if (!check_pmmu_cache()) {
+        if sunlikely(!check_pmmu_cache()) {
             translate_logical_addr();
-            if (shoe.abort)
+            if sunlikely(shoe.abort)
                 return ;
         }
         const uint32_t p_addr_b = shoe.physical_addr;
@@ -639,8 +635,8 @@ void logical_set (void)
 /* --- EA routines --- */
 #pragma mark EA routines
 
-#define nextword(pc) ({const uint16_t w=lget((pc),2);if (shoe.abort){return;}(pc)+=2; w;})
-#define nextlong(pc) ({const uint32_t L=lget((pc),4);if (shoe.abort){return;}(pc)+=4; L;})
+#define nextword(pc) ({const uint16_t w=lget((pc),2);if sunlikely(shoe.abort){return;}(pc)+=2; w;})
+#define nextlong(pc) ({const uint32_t L=lget((pc),4);if sunlikely(shoe.abort){return;}(pc)+=4; L;})
 
 // ea_decode_extended() - find the EA for those hiddeous 68020 addr modes
 static void ea_decode_extended()
@@ -760,7 +756,7 @@ static void ea_decode_extended()
             case ~b(1001): case ~b(1010): case ~b(1011): {
                 // Indirect preindexed
                 const uint32_t intermediate = lget(base_addr + base_disp + index_val, 4);
-                if (shoe.abort) return ;
+                if sunlikely(shoe.abort) return ;
                 shoe.extended_addr = intermediate + outer_disp;
                 shoe.extended_len = mypc - start_pc;
                 // slog("addr=0x%x len=%u\n", shoe.extended_addr, shoe.extended_len);
@@ -770,7 +766,7 @@ static void ea_decode_extended()
             case ~b(0101): case ~b(0110): case ~b(0111): {
                 // Indirect postindexed
                 const uint32_t intermediate = lget(base_addr + base_disp, 4);
-                if (shoe.abort) return ;
+                if sunlikely(shoe.abort) return ;
                 shoe.extended_addr = intermediate + index_val + outer_disp;
                 shoe.extended_len = mypc - start_pc;
                 return ;
@@ -847,7 +843,7 @@ void _ea_011_write (void)
     const uint8_t delta = ((reg==7) && (shoe.sz==1)) ? 2 : shoe.sz;
     
     lset(shoe.a[reg], shoe.sz, shoe.dat);
-    if (!shoe.abort)
+    if slikely(!shoe.abort)
         shoe.a[reg] += delta;
 }
 
@@ -870,7 +866,7 @@ void _ea_100_write (void)
     const uint8_t delta = ((reg==7) && (shoe.sz==1)) ? 2 : shoe.sz;
     
     lset(shoe.a[reg] - delta, shoe.sz, shoe.dat);
-    if (!shoe.abort)
+    if slikely(!shoe.abort)
         shoe.a[reg] -= delta;
 }
 
@@ -902,7 +898,7 @@ void _ea_101_addr (void)
 void _ea_110_read (void)
 {
     ea_decode_extended();
-    if (!shoe.abort)
+    if slikely(!shoe.abort)
         shoe.dat = lget(shoe.extended_addr, shoe.sz);
     shoe.uncommitted_ea_read_pc = shoe.pc + shoe.extended_len;
 }
@@ -913,16 +909,16 @@ void _ea_110_read_commit (void)
 void _ea_110_write (void)
 {
     ea_decode_extended();
-    if (!shoe.abort) {
+    if slikely(!shoe.abort) {
         lset(shoe.extended_addr, shoe.sz, shoe.dat);
-        if (!shoe.abort)
+        if slikely(!shoe.abort)
             shoe.pc += shoe.extended_len;
     }
 }
 void _ea_110_addr (void)
 {
     ea_decode_extended();
-    if (!shoe.abort) {
+    if slikely(!shoe.abort) {
         shoe.dat = shoe.extended_addr;
         shoe.pc += shoe.extended_len;
     }
@@ -998,7 +994,7 @@ void _ea_111_010_addr (void)
 void _ea_111_011_read (void)
 {
     ea_decode_extended();
-    if (!shoe.abort)
+    if slikely(!shoe.abort)
         shoe.dat = lget(shoe.extended_addr, shoe.sz);
     shoe.uncommitted_ea_read_pc = shoe.pc + shoe.extended_len;
 }
@@ -1009,7 +1005,7 @@ void _ea_111_011_read_commit (void)
 void _ea_111_011_addr (void)
 {
     ea_decode_extended();
-    if (!shoe.abort) {
+    if slikely(!shoe.abort) {
         shoe.dat = shoe.extended_addr;
         shoe.pc += shoe.extended_len;
     }
diff --git a/core/shoebill.h b/core/shoebill.h
index 30e02fe..5b6d79b 100644
--- a/core/shoebill.h
+++ b/core/shoebill.h
@@ -79,6 +79,9 @@
 #endif
 
 
+#define slikely(e) (__builtin_expect(!!(e), 1))
+#define sunlikely(e) (__builtin_expect(!!(e), 0))
+
 /*
  * core_api.c stuff
  */
@@ -876,7 +879,12 @@ extern const physical_get_ptr physical_get_jump_table[16];
 extern const physical_set_ptr physical_set_jump_table[16];
 
 #define physical_set() physical_set_jump_table[shoe.physical_addr >> 28]()
-#define pset(addr, s, val) {shoe.physical_addr=(addr); shoe.physical_size=(s); shoe.physical_dat=(val); physical_set();}
+#define pset(addr, s, val) do { \
+    shoe.physical_addr=(addr); \
+    shoe.physical_size=(s); \
+    shoe.physical_dat=(val); \
+    physical_set(); \
+} while (0)
 
 #define physical_get() physical_get_jump_table[shoe.physical_addr >> 28]()
 #define pget(addr, s) ({shoe.physical_addr=(addr); shoe.physical_size=(s); physical_get(); shoe.physical_dat;})
@@ -889,26 +897,17 @@ void logical_get (void);
     logical_get(); \
     shoe.logical_dat; \
 })
-
-#define lget(addr, s) ({ \
-    shoe.logical_addr=(addr); \
-    shoe.logical_size=(s); \
-    shoe.logical_fc = (sr_s() ? 5 : 1); \
-    logical_get(); \
-    shoe.logical_dat; \
-})
+#define lget(addr, s) lget_fc((addr), (s), (sr_s() ? 5 : 1))
 
 void logical_set (void);
-#define lset_fc(addr, s, val, fc) {\
+#define lset_fc(addr, s, val, fc) do { \
     shoe.logical_addr=(addr); \
     shoe.logical_size=(s); \
     shoe.logical_dat=(val); \
     shoe.logical_fc = (fc); \
     logical_set();\
-}
-#define lset(addr, s, val) { \
-    lset_fc((addr), (s), (val), sr_s() ? 5 : 1) \
-}
+} while (0)
+#define lset(addr, s, val) lset_fc((addr), (s), (val), sr_s() ? 5 : 1)
 
 typedef void (*_ea_func) (void);
 extern const _ea_func ea_read_jump_table[64];
@@ -922,10 +921,10 @@ extern const _ea_func ea_addr_jump_table[64];
 #define ea_addr() ea_addr_jump_table[shoe.mr]()
 
 
-#define call_ea_read(M, s) {shoe.mr=(M);shoe.sz=(s);ea_read();if (shoe.abort) return;}
-#define call_ea_write(M, s) {shoe.mr=(M);shoe.sz=(s);ea_write();if (shoe.abort) return;}
-#define call_ea_read_commit(M, s) {shoe.mr=(M);shoe.sz=(s);ea_read_commit();if (shoe.abort) return;}
-#define call_ea_addr(M) {shoe.mr=(M);ea_addr();if (shoe.abort) return;}
+#define call_ea_read(M, s) {shoe.mr=(M);shoe.sz=(s);ea_read();if sunlikely(shoe.abort) return;}
+#define call_ea_write(M, s) {shoe.mr=(M);shoe.sz=(s);ea_write();if sunlikely(shoe.abort) return;}
+#define call_ea_read_commit(M, s) {shoe.mr=(M);shoe.sz=(s);ea_read_commit();if sunlikely(shoe.abort) return;}
+#define call_ea_addr(M) {shoe.mr=(M);ea_addr();if sunlikely(shoe.abort) return;}
 
 #define push_a7(_dat, _sz) {shoe.a[7]-=(_sz);lset(shoe.a[7], (_sz), (_dat));}
 
diff --git a/core/video.c b/core/video.c
index 58ee7e8..6d7c71c 100644
--- a/core/video.c
+++ b/core/video.c
@@ -145,7 +145,7 @@ uint32_t nubus_video_read_func(const uint32_t rawaddr, const uint32_t size,
     const uint32_t addr = rawaddr & 0x00ffffff;
     
     // ROM and control registers
-    if ((addr >> 20) == 0xf) {
+    if sunlikely((addr >> 20) == 0xf) {
         
         slog("nubus_video_read_func: got a read to 0x%08x sz=%u\n", rawaddr, size);
         
@@ -172,7 +172,7 @@ uint32_t nubus_video_read_func(const uint32_t rawaddr, const uint32_t size,
     // Else, this is video ram
     
     uint32_t i, result = 0;
-    if (addr < (ctx->pixels * 4)) {
+    if slikely(addr < (ctx->pixels * 4)) {
         for (i=0; i<size; i++)
             result = (result << 8) | ((uint8_t*)ctx->direct_buf)[addr + i];
     }
@@ -234,7 +234,7 @@ void nubus_video_write_func(const uint32_t rawaddr, const uint32_t size,
     uint32_t i;
     
     // ROM and control registers
-    if ((addr >> 20) == 0xf) {
+    if sunlikely((addr >> 20) == 0xf) {
         
         slog("nubus_video_write_func: got a write to 0x%08x sz=%u data=0x%x\n", rawaddr, size, data);
         
@@ -340,7 +340,7 @@ void nubus_video_write_func(const uint32_t rawaddr, const uint32_t size,
     
     // Else, this is video ram
     
-    if (addr < (ctx->pixels * 4)) {
+    if slikely(addr < (ctx->pixels * 4)) {
         uint32_t mydata, myaddr;
         for (myaddr = addr + size, mydata = data; addr < myaddr; ) {
             ((uint8_t*)ctx->direct_buf)[--myaddr] = mydata & 0xff;
diff --git a/sdl-gui/lin_build.sh b/sdl-gui/lin_build.sh
index da2227c..f543072 100755
--- a/sdl-gui/lin_build.sh
+++ b/sdl-gui/lin_build.sh
@@ -8,7 +8,7 @@ for i in adb fpu mc68851 mem via floppy core_api cpu dis; do
 	files="$files $i.post.c"
 done
 
-for i in atrap_tab coff exception macii_symbols redblack scsi video filesystem alloc_pool toby_frame_buffer ethernet sound; do
+for i in SoftFloat/softfloat atrap_tab coff exception macii_symbols redblack scsi video filesystem alloc_pool toby_frame_buffer ethernet sound; do
 	files="$files ../core/$i.c"
 done
 
diff --git a/sdl-gui/osx_build.sh b/sdl-gui/osx_build.sh
index 16f14f3..eb13bab 100755
--- a/sdl-gui/osx_build.sh
+++ b/sdl-gui/osx_build.sh
@@ -8,7 +8,7 @@ for i in adb fpu mc68851 mem via floppy core_api cpu dis; do
 	files="$files $i.post.c"
 done
 
-for i in atrap_tab coff exception macii_symbols redblack scsi video filesystem alloc_pool toby_frame_buffer ethernet sound; do
+for i in SoftFloat/softfloat atrap_tab coff exception macii_symbols redblack scsi video filesystem alloc_pool toby_frame_buffer ethernet sound; do
 	files="$files ../core/$i.c"
 done
 
diff --git a/sdl-gui/win_build.bat b/sdl-gui/win_build.bat
index 4891afc..b25163d 100755
--- a/sdl-gui/win_build.bat
+++ b/sdl-gui/win_build.bat
@@ -6,4 +6,4 @@ gcc -O1 ..\core\decoder_gen.c -o decoder_gen
 decoder_gen inst .
 decoder_gen dis .
 
-gcc -O3 -flto -mno-ms-bitfields sdl.c adb.post.c fpu.post.c mc68851.post.c mem.post.c via.post.c floppy.post.c core_api.post.c cpu.post.c dis.post.c ..\core\atrap_tab.c ..\core\coff.c ..\core\exception.c ..\core\macii_symbols.c ..\core\redblack.c ..\core\scsi.c ..\core\video.c ..\core\filesystem.c ..\core\alloc_pool.c ..\core\toby_frame_buffer.c ..\core\ethernet.c ..\core\sound.c -lmingw32 -lopengl32 -lsdl2main -lsdl2 -o shoebill
\ No newline at end of file
+gcc -O3 -flto -mno-ms-bitfields sdl.c adb.post.c fpu.post.c mc68851.post.c mem.post.c via.post.c floppy.post.c core_api.post.c cpu.post.c dis.post.c ..\core\atrap_tab.c ..\core\coff.c ..\core\exception.c ..\core\macii_symbols.c ..\core\redblack.c ..\core\scsi.c ..\core\video.c ..\core\filesystem.c ..\core\alloc_pool.c ..\core\toby_frame_buffer.c ..\core\ethernet.c ..\core\sound.c ..\core\SoftFloat\softfloat.c -lmingw32 -lopengl32 -lsdl2main -lsdl2 -o shoebill