diff --git a/CMakeLists.txt b/CMakeLists.txt
new file mode 100644
index 0000000..99bba72
--- /dev/null
+++ b/CMakeLists.txt
@@ -0,0 +1,7 @@
+cmake_minimum_required(VERSION 3.3.0)
+project (linecount)
+
+add_library (linecount SHARED linecount.cpp linecount.h)
+add_library (linecount_static STATIC linecount.cpp linecount.h)
+add_executable (lc main.cpp)
+target_link_libraries (lc linecount_static)
diff --git a/create_testfiles.sh b/create_testfiles.sh
new file mode 100755
index 0000000..6b31d9f
--- /dev/null
+++ b/create_testfiles.sh
@@ -0,0 +1,5 @@
+#!/bin/bash
+
+dd if=/dev/urandom of=test1.txt bs=1000000 count=10
+dd if=/dev/urandom of=test2.txt bs=1000000 count=100
+dd if=/dev/urandom of=test3.txt bs=1000000 count=1000
diff --git a/linecount.cpp b/linecount.cpp
index dff6dd9..d2de1c7 100644
--- a/linecount.cpp
+++ b/linecount.cpp
@@ -8,7 +8,9 @@
 #define LCCLOSEFILE(handle) CloseHandle(handle)
 #define LCINVALIDHANDLE INVALID_HANDLE_VALUE
 #define LCSETREALLASTERROR(err, errstr) { setLastError((err), (errstr)); }
+#define MAP_FAILED NULL;
 typedef long long LCFILEOFFSET;
+
 #elif defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
 
 // POSIX
@@ -16,23 +18,28 @@ typedef long long LCFILEOFFSET;
 #include<sys/stat.h>
 #include<sys/fcntl.h>
 #include<sys/mman.h>
-#include<pthread.h>
-#if (defined (__APPLE__) && defined (__MACH__)
+#if (defined (__APPLE__) && defined (__MACH__))
 #include <sys/sysctl.h>
 typedef off_t LCFILEOFFSET;
-#define MMAP mmap
+#define MMAP ::mmap
+#define FSTAT ::fstat
+#define STAT ::stat
 #elif defined(__linux__)
 typedef off64_t LCFILEOFFSET;
-#define MMAP mmap64
+#define MMAP ::mmap64
+#define FSTAT ::fstat64
+#define STAT ::stat64
 #else
 typedef off_t LCFILEOFFSET;
-#define MMAP mmap
+#define MMAP ::mmap
+#define FSTAT ::fstat
+#define STAT ::stat
 #endif
 
-#define LCOPENFILE(name) open(name, O_RDONLY)
-#define LCCLOSEFILE(handle) (close(handle) != -1)
+#define LCOPENFILE(name) ::open(name, O_RDONLY)
+#define LCCLOSEFILE(handle) (::close(handle) != -1)
 #define LCINVALIDHANDLE -1
-#define LCSETREALLASTERROR(err, errstr) { int __err = errno; setLastError(__err, strerror(__err)); }
+#define LCSETREALLASTERROR(err, errstr) { int __err = errno; setLastError(__err, ::strerror(__err)); }
 
 #endif
 
@@ -51,37 +58,36 @@ CLineCount::CLineCount(PARAMETERS *parameters)
 
 	// Set line count parameter defaults
 	int cpucount;
+	int allocationgranularity;
 #ifdef _WIN32
 	SYSTEM_INFO sysinfo;
 	GetSystemInfo(&sysinfo);
 	cpucount = sysinfo.dwNumberOfProcessors;
-#elif defined(__linux__)
-	cpucount = sysconf(_SC_NPROCESSORS_ONLN);
-#elif (defined (__APPLE__) && defined (__MACH__))
-	intsize_t count_len = sizeof(cpucount);
-	sysctlbyname("hw.logicalcpu", &cpucount, &count_len, NULL, 0);
+	allocationgranularity = sysinfo.dwAllocationGranularity;
+//#elif defined(__linux__)
 #else
-	cpucount = 1;
+	cpucount = sysconf(_SC_NPROCESSORS_ONLN);
+	allocationgranularity = sysconf(_SC_PAGESIZE);
+//#elif (defined (__APPLE__) && defined (__MACH__))
+//	mmsize_t count_len = sizeof(cpucount);
+//	sysctlbyname("hw.logicalcpu", &cpucount, &count_len, NULL, 0);
+//#else
+//	cpucount = 1;
 #endif
-	m_parameters.workercount = cpucount;
-	m_parameters.windowsize = (1024 * 1024 * 8) / cpucount;
+	m_parameters.threadcount = cpucount;
+	m_parameters.buffersize = (1024 * 1024);
 	
 	// Override defaults if specified
 	if (parameters)
 	{
-		if (parameters->windowsize != -1)
+		if (parameters->buffersize != -1)
 		{
-			m_parameters.windowsize = parameters->windowsize;
-#ifdef _WIN32
-			// Window size must be multiple of allocation granularity. Round up.
-			SYSTEM_INFO si;
-			GetSystemInfo(&si);
-			m_parameters.windowsize += (si.dwAllocationGranularity - (m_parameters.windowsize % si.dwAllocationGranularity)) % si.dwAllocationGranularity;
-#endif
+			m_parameters.buffersize = parameters->buffersize;
+			m_parameters.buffersize += (allocationgranularity - (m_parameters.buffersize % allocationgranularity)) % allocationgranularity;
 		}
-		if (parameters->workercount != -1)
+		if (parameters->threadcount != -1)
 		{
-			m_parameters.workercount = parameters->workercount;
+			m_parameters.threadcount = parameters->threadcount;
 		}
 	}
 	
@@ -197,14 +203,15 @@ void *threadProc(void *ctx)
 #endif
 {
 	LCTHREADCONTEXT *lctctx = (LCTHREADCONTEXT*)ctx;
-	return lctctx->m_this->countThread(lctctx->thread_number);
+	lctctx->m_this->countThread(lctctx->thread_number);
+	return NULL;
 }
 
 unsigned int CLineCount::countThread(int thread_number)
 {
-	LCFILEOFFSET windowsize = (LCFILEOFFSET)m_parameters.windowsize;
-	LCFILEOFFSET startoffset = windowsize * (LCFILEOFFSET)thread_number;
-	LCFILEOFFSET stride = windowsize * m_actual_thread_count;
+	LCFILEOFFSET buffersize = (LCFILEOFFSET)m_parameters.buffersize;
+	LCFILEOFFSET startoffset = buffersize * (LCFILEOFFSET)thread_number;
+	LCFILEOFFSET stride = buffersize * m_actual_thread_count;
 	LCFILEOFFSET curoffset = startoffset;
 	LCFILEOFFSET lastmapsize = 0;
 	LCLINECOUNT count = 0;
@@ -214,11 +221,12 @@ unsigned int CLineCount::countThread(int thread_number)
 	{
 		if (m_thread_fail)
 		{
+
 			return -1;
 		}
 
 		// Get best file mapping window size
-		size_t mapsize = (size_t)min((m_filesize - curoffset), windowsize);
+		size_t mapsize = (size_t)std::min((m_filesize - curoffset), buffersize);
 
 		// Map view of file
 #ifdef _WIN32
@@ -235,16 +243,17 @@ unsigned int CLineCount::countThread(int thread_number)
 #else
 		if (mem)
 		{
-			if(munmap(mem)!=0)
+			if(munmap(mem, lastmapsize) !=0)
 			{
 				LCSETREALLASTERROR(EINVAL, _T("memory unmap failed"));
 				m_thread_fail = true;
 				return -1;
 			}
 		}
-		mem = MMAP(NULL, mapsize, PROT_READ, MAP_FILE | MAP_SHARED, m_fd, curoffset);
+		mem = MMAP(NULL, mapsize, PROT_READ, MAP_FILE | MAP_SHARED, m_fh, curoffset);
+//		printf("%p %lld %lld\n",mem, mapsize, curoffset);
 #endif		
-		if (mem == NULL)
+		if (mem == MAP_FAILED)
 		{
 			LCSETREALLASTERROR(EINVAL, _T("memory map failed"));
 			m_thread_fail = true;
@@ -299,7 +308,7 @@ unsigned int CLineCount::countThread(int thread_number)
 #else
 	if (mem)
 	{
-		if (munmap(mem) != 0)
+		if (munmap(mem, lastmapsize) != 0)
 		{
 			LCSETREALLASTERROR(EINVAL, _T("memory unmap failed"));
 			m_thread_fail = true;
@@ -349,8 +358,8 @@ bool CLineCount::countLines(LCLINECOUNT & linecount)
 	}
 	m_filesize = li.QuadPart;
 #else
-	struct stat64 statbuf;
-	if(fstat64(m_fh,&statbuf)!=0)
+	struct STAT statbuf;
+	if(FSTAT(m_fh,&statbuf)!=0)
 	{
 		LCSETREALLASTERROR(EBADF, _T("unable to get file size"));
 		return false;
@@ -366,16 +375,18 @@ bool CLineCount::countLines(LCLINECOUNT & linecount)
 	}
 
 	// Figure out actual thread count
-	LCFILEOFFSET windowcount = (m_filesize + (m_parameters.windowsize - 1)) / m_parameters.windowsize;
-	if (windowcount < (LCFILEOFFSET) m_parameters.workercount)
+	LCFILEOFFSET windowcount = (m_filesize + (m_parameters.buffersize - 1)) / m_parameters.buffersize;
+	if (windowcount < (LCFILEOFFSET) m_parameters.threadcount)
 	{
 		m_actual_thread_count = (int)windowcount;
 	}
 	else
 	{
-		m_actual_thread_count = m_parameters.workercount;
+		m_actual_thread_count = m_parameters.threadcount;
 	}
 
+//	printf("act: %d\n",m_actual_thread_count);
+
 #ifdef _WIN32
 	// Prepare file mapping
 	m_filemapping = CreateFileMapping(m_fh, NULL, PAGE_READONLY, 0, 0, NULL);
@@ -405,7 +416,7 @@ bool CLineCount::countLines(LCLINECOUNT & linecount)
 #ifdef _WIN32
 			success = (WaitForSingleObject(m_threads[i], INFINITE) == WAIT_OBJECT_0);
 #else
-			success = pthread_join(m_threads[i]) == 0;
+			success = pthread_join(m_threads[i], NULL) == 0;
 #endif
 		}
 
@@ -429,7 +440,7 @@ bool CLineCount::countLines(LCLINECOUNT & linecount)
 		return false;
 	}
 
-	if (complete != m_parameters.workercount)
+	if (complete != m_actual_thread_count)
 	{
 		setLastError(ECHILD, _T("thread join failed"));
 		return false;
diff --git a/linecount.h b/linecount.h
index 4da2bb0..fbe1701 100644
--- a/linecount.h
+++ b/linecount.h
@@ -1,6 +1,9 @@
 #ifndef __INC_LINECOUNT_H
 #define __INC_LINECOUNT_H
 
+#define LINECOUNT_VERSION_MAJOR 1
+#define LINECOUNT_VERSION_MINOR 0
+
 ///////////////////////////////////////////// Headers
 
 ////////////// Platform independent
@@ -18,6 +21,7 @@
 #include<tchar.h>
 #elif defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) // POSIX
 #include<unistd.h>
+#include<pthread.h>
 #define _T(x) x
 #define TCHAR char
 #endif
@@ -45,7 +49,7 @@ BEGIN_LINECOUNT_NAMESPACE;
 	typedef std::string LCSTRING;
 	typedef int LCFILEHANDLE;
 	typedef errno_t LCERROR;
-	#if (defined (__APPLE__) && defined (__MACH__)
+	#if (defined (__APPLE__) && defined (__MACH__))
 		typedef off_t LCFILEOFFSET;
 		#define LCLINECOUNTFMT "%lld"
 	#elif defined(__linux__)
@@ -66,8 +70,8 @@ public:
 
 	struct PARAMETERS
 	{
-		size_t windowsize;
-		int workercount;
+		size_t buffersize;
+		int threadcount;
 	};
 
 private:
diff --git a/main.cpp b/main.cpp
index 097ccf0..7fa6f38 100644
--- a/main.cpp
+++ b/main.cpp
@@ -1,17 +1,46 @@
 #include"linecount.h"
+
 #ifdef _WIN32
+
 #include<tchar.h>
+
 #else
+
+#include<stdlib.h>
 #define _tprintf printf
 #define _ftprintf fprintf
 #define _tcscmp strcmp
+#define _tcslen strlen
 #define _ttoi atoi
 #define _tcstoui64 strtoull
 #define _T(x) x
+#define TCHAR char
+
 #endif
 
 using namespace LineCount;
 
+//////////////////////////////////////////////////////
+
+
+void help(const TCHAR *argv0)
+{
+	_ftprintf(stderr, _T("usage: %s [options] <file>\n"), argv0);
+	_ftprintf(stderr, _T("    -h  --help                         print this usage and exit\n"));
+	_ftprintf(stderr, _T("    -b  --buffersize <BUFFERSIZE>      size of buffer per-thread to use when reading (default is 1MB)\n"));   
+	_ftprintf(stderr, _T("    -t  --threadcount <THREADCOUNT>    number of threads to use (defaults to number of cpu cores)\n"));
+	_ftprintf(stderr, _T("    -v  --version                      print version information and exit\n"));
+}
+
+
+void version(void)
+{
+	_tprintf(_T("lc (linecount) %d.%2.2d\nCopyright (c) 2015 Christien Rioux\n"), LINECOUNT_VERSION_MAJOR, LINECOUNT_VERSION_MINOR);
+}
+
+//////////////////////////////////////////////////////
+
+
 #if defined(WIN32) && defined(_UNICODE)
 int wmain(int argc, TCHAR **argv)
 #else
@@ -23,27 +52,43 @@ int main(int argc, char **argv)
 	int posparam = 0;
 	
 	CLineCount::PARAMETERS params;
-	params.windowsize = -1;
-	params.workercount = -1;
+	params.buffersize = -1;
+	params.threadcount = -1;
 
 	TCHAR *filename = NULL;
 
+	if(argc==1)
+	{
+		help(argv[0]);
+		exit(0);
+	}
+
 	while (arg < argc)
 	{
-		if (_tcscmp(argv[arg], _T("--windowsize")) == 0)
+		if (_tcscmp(argv[arg], _T("-h")) == 0 || _tcscmp(argv[arg], _T("--help")) == 0) 
+		{
+			help(argv[0]);
+			exit(0);
+		}
+		else if (_tcscmp(argv[arg], _T("-v")) == 0 || _tcscmp(argv[arg], _T("--version")) == 0) 
+		{
+			version();
+			exit(0);
+		}
+		else if (_tcscmp(argv[arg], _T("-b")) == 0 || _tcscmp(argv[arg], _T("--buffersize")) == 0)
 		{
 			arg++;
 			if (arg == argc)
 			{
-				_ftprintf(stderr, _T("missing argument to --windowsize"));
+				_ftprintf(stderr, _T("%s: missing argument to %s\n"), argv[0], argv[arg-1]);
 				return 1;
 			}
 
-			_TCHAR *wsstr = argv[arg];
+			TCHAR *wsstr = argv[arg];
 
 			// Check for size multipliers			
 			size_t multiplier = 1;
-			_TCHAR *lastchar = wsstr + (_tcslen(wsstr) - 1);
+			TCHAR *lastchar = wsstr + (_tcslen(wsstr) - 1);
 			if (*lastchar == _T('k') || *lastchar == _T('K'))
 			{
 				multiplier = 1024;
@@ -60,20 +105,25 @@ int main(int argc, char **argv)
 				lastchar = 0;
 			}
 
-			_TCHAR *endptr;
-			params.windowsize = ((size_t)_tcstoui64(argv[arg], &endptr, 10)) * multiplier;
+			TCHAR *endptr;
+			params.buffersize = ((size_t)_tcstoui64(argv[arg], &endptr, 10)) * multiplier;
 
 		}
-		else if (_tcscmp(argv[arg], _T("--workercount")) == 0)
+		else if (_tcscmp(argv[arg], _T("-t")) == 0 || _tcscmp(argv[arg], _T("--threadcount")) == 0)
 		{
 			arg++;
 			if (arg == argc)
 			{
-				_ftprintf(stderr, _T("missing argument to --workercount"));
+				_ftprintf(stderr, _T("%s: Missing argument to %s\n"), argv[0], argv[arg-1]);
 				return 1;
 			}
 
-			params.workercount = _ttoi(argv[arg]);
+			params.threadcount = _ttoi(argv[arg]);
+			if(params.threadcount<=0)
+			{
+				_ftprintf(stderr, _T("%s: Invalid thread count\n"), argv[0]);
+				return 1;
+			}
 		}
 		else
 		{
@@ -83,7 +133,7 @@ int main(int argc, char **argv)
 			}
 			else
 			{
-				_ftprintf(stderr, _T("too many arguments"));
+				_ftprintf(stderr, _T("%s: Too many arguments\n"), argv[0]);
 				return 1;
 			}
 			posparam++;
@@ -94,7 +144,7 @@ int main(int argc, char **argv)
 
 	if (posparam != 1)
 	{
-		_ftprintf(stderr, _T("missing required argument"));
+		_ftprintf(stderr, _T("%s: Missing required argument\n"), argv[0]);
 		return 1;
 	}
 	
@@ -106,7 +156,7 @@ int main(int argc, char **argv)
 		LCERROR err = lc.lastError();
 		LCSTRING errstr = lc.lastErrorString();
 
-		_ftprintf(stderr, _T("error (%d): %s\n"), err, errstr.c_str());
+		_ftprintf(stderr, _T("%s: Error %d (%s)\n"), argv[0], err, errstr.c_str());
 		return err;
 	}
 
@@ -117,7 +167,7 @@ int main(int argc, char **argv)
 		LCERROR err = lc.lastError();
 		LCSTRING errstr = lc.lastErrorString();
 
-		_ftprintf(stderr, _T("error (%d): %s\n"), err, errstr.c_str());
+		_ftprintf(stderr, _T("%s: Error %d: (%s)\n"), argv[0], err, errstr.c_str());
 		return err;
 	}