diff --git a/src/Makefile b/src/Makefile
index e56dc622..90592ddf 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -5,7 +5,7 @@ CXXFLAGS=-g -fPIC
 
 LIB_OBJS=args.o obj.o mem.o core.o instruction.o enc.o util.o lex.yy.o
 
-all: harptool libharplib.so libharplib.a libqsim-harp.so
+all: harptool libharplib.so libharplib.a #libqsim-harp.so
 
 # Use -static so we don't have to install the library in order to just run
 # Harptool.
@@ -19,19 +19,27 @@ libharplib.a: $(LIB_OBJS)
 	ar rcs $@ $(LIB_OBJS)
 
 args.o : args.cpp include/args.h
-enc.o : enc.cpp include/types.h include/util.h include/enc.h include/archdef.h include/instruction.h
-harptool.o : harptool.cpp include/types.h include/core.h include/enc.h include/instruction.h include/mem.h include/obj.h include/archdef.h include/args.h include/help.h
+enc.o : enc.cpp include/types.h include/util.h include/enc.h include/archdef.h\
+                include/instruction.h
+harptool.o : harptool.cpp include/types.h include/core.h include/enc.h \
+             include/instruction.h include/mem.h include/obj.h \
+             include/archdef.h include/args.h include/help.h
 instruction.o : instruction.cpp include/instruction.h include/obj.h
-obj.o : obj.cpp include/types.h include/obj.h include/util.h include/asm-tokens.h
+obj.o : obj.cpp include/types.h include/obj.h include/util.h \
+        include/asm-tokens.h
 util.o : util.cpp include/types.h include/util.h
 mem.o : mem.cpp include/types.h include/util.h include/mem.h
-core.o : core.cpp include/types.h include/util.h include/mem.h include/archdef.h
+core.o : core.cpp include/types.h include/util.h include/mem.h \
+         include/debug.h include/archdef.h
 
 QSIM_CXXFLAGS=-DEMU_INSTRUMENTATION
 
-libqsim-harp.so: args.cpp enc.cpp instruction.cpp obj.cpp util.cpp mem.cpp core.cpp qsim-harp.cpp lex.yy.o include/qsim-harp.h include/types.h include/core.h include/util.h include/enc.h include/archdef.h include/instruction.h include/asm-tokens.h include/mem.h
-	$(CXX) $(CXXFLAGS) $(QSIM_CXXFLAGS) -shared -o $@ args.cpp enc.cpp instruction.cpp obj.cpp util.cpp mem.cpp core.cpp qsim-harp.cpp lex.yy.o
-
+libqsim-harp.so: args.cpp enc.cpp instruction.cpp obj.cpp util.cpp mem.cpp \
+                 core.cpp qsim-harp.cpp lex.yy.o include/qsim-harp.h \
+                 include/types.h include/core.h include/util.h include/enc.h \
+                 include/archdef.h include/instruction.h include/asm-tokens.h \
+                 include/mem.h
+	$(CXX) $(CXXFLAGS) $(QSIM_CXXFLAGS) -shared -o $@ $^
 
 lex.yy.cc: scanner.lex
 	flex scanner.lex
diff --git a/src/core.cpp b/src/core.cpp
index a80f6eda..4746b1c3 100644
--- a/src/core.cpp
+++ b/src/core.cpp
@@ -10,6 +10,7 @@
 #include "include/mem.h"
 #include "include/enc.h"
 #include "include/core.h"
+#include "include/debug.h"
 
 #ifdef EMU_INSTRUMENTATION
 #include "include/qsim-harp.h"
@@ -36,18 +37,13 @@ Core::Core(const ArchDef &a, Decoder &d, MemoryUnit &mem, Word id) :
   /* Build the register file. */
   Word regNum(0);
   for (Word j = 0; j < a.getNThds(); ++j) {
-    std::cout << "Pushing back a new register vector.\n";
     reg.push_back(vector<Reg<Word> >(0));
     for (Word i = 0; i < a.getNRegs(); ++i) {
-      std::cout << "Pushing back a new register in thread " << j << "\n";
       reg[j].push_back(Reg<Word>(id, regNum++));
     }
 
     pred.push_back(vector<Reg<bool> >(0));
-    std::cout << "getnpregs returns " << a.getNPRegs() << '\n';
     for (Word i = 0; i < a.getNPRegs(); ++i) {
-      std::cout << "Pushing back predicate reg " << i << ", thread " << j 
-                << "\n";
       pred[j].push_back(Reg<bool>(id, regNum++));
     }
   }
@@ -62,7 +58,7 @@ void Core::step() {
 
   if (activeThreads == 0) return;
 
-  //cout << "in step pc=0x" << hex << pc << '\n';
+  D(3, "in step pc=0x" << hex << pc);
 
   /* Fetch and decode. */
   if (wordSize < sizeof(pc)) pc &= ((1ll<<(wordSize*8))-1);
@@ -72,13 +68,14 @@ void Core::step() {
     /* Todo: speed this up for the byte encoder? */
     try {
       fetchMore = false;
-      fetchBuffer.resize(fetchPos + wordSize);
+      unsigned fetchSize(wordSize - (pc+fetchPos)%wordSize);
+      fetchBuffer.resize(fetchPos + fetchSize);
       Word fetched = mem.fetch(pc + fetchPos, supervisorMode);
-      writeWord(fetchBuffer, fetchPos, wordSize, fetched);
+      writeWord(fetchBuffer, fetchPos, fetchSize, fetched);
       decPos = 0;
       inst = iDec.decode(fetchBuffer, decPos);
     } catch (OutOfBytes o) {
-      //cout << "Caught OutOfBytes. Fetching more.\n";
+      D(3, "Caught OutOfBytes. Fetching more.");
       fetchMore = true;
     } catch (MemoryUnit::PageFault pf) {
       fetchPos = 0;
@@ -87,9 +84,8 @@ void Core::step() {
       interrupt(pf.notFound?1:2);
     }
   } while (fetchMore);
-  //cout << "Fetched at 0x" << hex << pc << '\n';
-  //cout << "0x" << hex << pc << ": " << *inst << '\n';
-  //cout << "sizeof(core)=" << dec << sizeof(*this) << '\n';
+  D(3, "Fetched at 0x" << hex << pc);
+  D(3, "0x" << hex << pc << ": " << *inst);
 
 #ifdef EMU_INSTRUMENTATION
   { Addr pcPhys(mem.virtToPhys(pc));
diff --git a/src/enc.cpp b/src/enc.cpp
index 4b44fc32..cc959e4f 100644
--- a/src/enc.cpp
+++ b/src/enc.cpp
@@ -8,6 +8,8 @@
 #include <iomanip>
 #include <vector>
 
+
+
 #include "include/types.h"
 #include "include/util.h"
 #include "include/enc.h"
diff --git a/src/include/debug.h b/src/include/debug.h
new file mode 100644
index 00000000..9f83f763
--- /dev/null
+++ b/src/include/debug.h
@@ -0,0 +1,23 @@
+/*******************************************************************************
+ HARPtools by Chad D. Kersey, Spring 2013
+*******************************************************************************/
+#ifndef __DEBUG_H
+#define __DEBUG_H
+
+#ifdef USE_DEBUG
+#include <iostream>
+
+#define D(lvl, x) do { \
+  using namespace std; \
+  if ((lvl) <= USE_DEBUG) { \
+    cout << "DEBUG " << __FILE__ << ':' << dec << __LINE__ << ": " \
+         << x << endl; \
+  } \
+} while(0)
+#else
+
+#define D(lvl, x) do {} while(0)
+
+#endif
+
+#endif
diff --git a/src/include/mem.h b/src/include/mem.h
index 34a2733b..7b93cb95 100644
--- a/src/include/mem.h
+++ b/src/include/mem.h
@@ -4,6 +4,7 @@
 #ifndef __MEM_H
 #define __MEM_H
 
+#include <ostream>
 #include <vector>
 #include <queue>
 #include <map>
diff --git a/src/include/obj.h b/src/include/obj.h
index 8db3fd8b..7dc19403 100644
--- a/src/include/obj.h
+++ b/src/include/obj.h
@@ -80,7 +80,9 @@ namespace Harp {
 
       return;
     noFit:
-      std::cout << "Attempt to bind a symbol to an address it cannot reach.\n";
+      std::cout << "Attempt to bind a " << bits << "-bit "
+                << (relative?"":"non-") << "relative symbol to an address"
+                                           " it cannot reach.\n";
       exit(1);
     }
 
diff --git a/src/instruction.cpp b/src/instruction.cpp
index 092baf84..41cc9e67 100644
--- a/src/instruction.cpp
+++ b/src/instruction.cpp
@@ -8,6 +8,7 @@
 #include "include/obj.h"
 #include "include/core.h"
 #include "include/harpfloat.h"
+#include "include/debug.h"
 
 #ifdef EMU_INSTRUMENTATION
 #include "include/qsim-harp.h"
@@ -63,7 +64,7 @@ Instruction::InstTableEntry Instruction::instTable[] = {
   {"andp",     false, false, false, false, AC_3PREG,    ITYPE_INTBASIC},
   {"orp",      false, false, false, false, AC_3PREG,    ITYPE_INTBASIC},
   {"xorp",     false, false, false, false, AC_3PREG,    ITYPE_INTBASIC},
-  {"notp",     false, false, false, false, AC_3PREG,    ITYPE_INTBASIC},
+  {"notp",     false, false, false, false, AC_2PREG,    ITYPE_INTBASIC},
   {"isneg",    false, false, false, false, AC_PREG_REG, ITYPE_INTBASIC},
   {"iszero",   false, false, false, false, AC_PREG_REG, ITYPE_INTBASIC},
   {"halt",     false, false, false, true,  AC_NONE,     ITYPE_NULL    },
@@ -106,6 +107,8 @@ ostream &Harp::operator<<(ostream& os, Instruction &inst) {
 }
 
 void Instruction::executeOn(Core &c) {
+  D(1, "Begin instruction execute.");
+
   /* If I try to execute a privileged instruction in user mode, throw an
      exception 3. */
   if (instTable[op].privileged && !c.supervisorMode) {
@@ -262,5 +265,7 @@ void Instruction::executeOn(Core &c) {
     if (instTable[op].controlFlow) break;
   }
 
+  D(3, "End instruction execute.");
+
   c.activeThreads = nextActiveThreads;
 }
diff --git a/src/mem.cpp b/src/mem.cpp
index 59caef77..732e7402 100644
--- a/src/mem.cpp
+++ b/src/mem.cpp
@@ -9,6 +9,7 @@
 #include <stdlib.h>
 #include <pthread.h>
 
+#include "include/debug.h"
 #include "include/types.h"
 #include "include/util.h"
 #include "include/mem.h"
@@ -39,12 +40,14 @@ void RomMemDevice::write(Addr, Word) {
 }
 
 Word RamMemDevice::read(Addr addr) {
-  Word w = readWord(contents, addr, wordSize);
+  D(2, "RAM read, addr=0x" << hex << addr);
+  Word w = readWord(contents, addr, wordSize - addr%wordSize);
   return w;
 }
 
 void RamMemDevice::write(Addr addr, Word w) {
-  writeWord(contents, addr, wordSize, w);
+  D(2, "RAM write, addr=0x" << hex << addr);
+  writeWord(contents, addr, wordSize - addr%wordSize, w);
 }
 
 MemDevice &MemoryUnit::ADecoder::doLookup(Addr a, Size &bit) {
@@ -104,8 +107,12 @@ MemoryUnit::TLBEntry MemoryUnit::tlbLookup(Addr vAddr, Word flagMask) {
   if ((i = tlb.find(vAddr/pageSize)) != tlb.end()) {
     TLBEntry &t = i->second;
     if (t.flags & flagMask) return t;
-    else throw PageFault(vAddr, false);
+    else {
+      D(2, "Page fault on addr 0x" << hex << vAddr << "(bad flags)");
+      throw PageFault(vAddr, false);
+    }
   } else {
+    D(2, "Page fault on addr 0x" << hex << vAddr << "(not in TLB)");
     throw PageFault(vAddr, true);
   }
 }
@@ -139,7 +146,7 @@ void MemoryUnit::write(Addr vAddr, Word w, bool sup) {
 }
 
 void MemoryUnit::tlbAdd(Addr virt, Addr phys, Word flags) {
-  cout << "tlbAdd(0x" << hex << virt << ", 0x" << phys << ", 0x" << flags << ")\n";
+  D(1, "tlbAdd(0x" << hex << virt << ", 0x" << phys << ", 0x" << flags << ')');
   tlb[virt/pageSize] = TLBEntry(phys/pageSize, flags);
 }
 
diff --git a/src/test/Makefile b/src/test/Makefile
index 1fac654d..e7cf5831 100644
--- a/src/test/Makefile
+++ b/src/test/Makefile
@@ -1,38 +1,44 @@
 HARPLD = ../harptool -L
 HARPAS = ../harptool -A
 HARPEM = ../harptool -E
-4WARCH = 6w16/2/2
+4BARCH = 4b16/16/2
 
-all: sieve.bin 2thread.bin sieve.4w.bin 2thread.4w.bin
+all: simple.bin sieve.bin 2thread.bin simple.4b.bin sieve.4b.bin 2thread.4b.bin
 
-run: sieve.out 2thread.out sieve.4w.out 2thread.4w.out
+run: simple.out sieve.out 2thread.out simple.4b.out sieve.4b.out 2thread.4b.out
 
-%.4w.out : %.4w.bin
-	$(HARPEM) -a $(4WARCH) -c $< > $@
+%.4b.out : %.4b.bin
+	$(HARPEM) -a $(4BARCH) -c $< > $@
 
 %.out : %.bin
 	$(HARPEM) -c $< > $@
 
-2thread.bin : lib.HOF 2thread.HOF boot.HOF
-	$(HARPLD) -o 2thread.bin boot.HOF 2thread.HOF lib.HOF
+2thread.bin : boot.HOF lib.HOF 2thread.HOF
+	$(HARPLD) -o 2thread.bin $^
 
-2thread.4w.bin : lib.4w.HOF 2thread.4w.HOF boot.4w.HOF
-	$(HARPLD) --arch $(4WARCH) -o 2thread.4w.bin boot.4w.HOF 2thread.4w.HOF 2thread.4w.HOF lib.4w.HOF
+2thread.4b.bin : boot.4b.HOF lib.4b.HOF 2thread.4b.HOF
+	$(HARPLD) --arch $(4BARCH) -o 2thread.4b.bin $^
 
-sieve.bin : lib.HOF sieve.HOF boot.HOF
-	$(HARPLD) -o sieve.bin boot.HOF sieve.HOF lib.HOF
+simple.bin : boot.HOF lib.HOF simple.HOF
+	$(HARPLD) -o $@ $^
 
-sieve.4w.bin : lib.4w.HOF sieve.4w.HOF boot.4w.HOF
-	$(HARPLD) --arch $(4WARCH) -o sieve.4w.bin boot.4w.HOF sieve.4w.HOF lib.4w.HOF
+sieve.bin : boot.HOF lib.HOF sieve.HOF
+	$(HARPLD) -o $@ $^
 
-%.4w.bin : %.4w.HOF
-	$(HARPLD) --arch $(4WARCH) -o $@ $<
+simple.4b.bin : boot.4b.HOF lib.4b.HOF simple.4b.HOF
+	$(HARPLD) --arch $(4BARCH) -o $@ $^
+
+sieve.4b.bin : boot.4b.HOF lib.4b.HOF sieve.4b.HOF
+	$(HARPLD) --arch $(4BARCH) -o $@ $^
+
+%.4b.bin : %.4b.HOF
+	$(HARPLD) --arch $(4BARCH) -o $@ $<
 
 %.bin : %.HOF
 	$(HARPLD) -o $@ $<
 
-%.4w.HOF : %.s
-	$(HARPAS) --arch $(4WARCH) -o $@ $<
+%.4b.HOF : %.s
+	$(HARPAS) --arch $(4BARCH) -o $@ $<
 
 %.HOF : %.s
 	$(HARPAS) -o $@ $<
diff --git a/src/test/sieve.s b/src/test/sieve.s
index 311251da..82c5dc9b 100644
--- a/src/test/sieve.s
+++ b/src/test/sieve.s
@@ -74,5 +74,5 @@ loop4:       ld   %r1, %r0, array;
 
 .global
 .word array 0 /* Basically, 0 and 1 are pre-cleared. */
-.word _0    0 /* I would love to not have to give this a name. */
-.word SIZE-2
+.word _0    0 /* Given a name, contents are zero. */
+.word 0x1ffe  /* Empty space of size SIZE-2 bytes. */