diff --git a/control/NodeControlClient.cpp b/control/NodeControlClient.cpp
index 6fd7b12ed..e5ea2c20e 100644
--- a/control/NodeControlClient.cpp
+++ b/control/NodeControlClient.cpp
@@ -33,6 +33,8 @@
 #include "IpcListener.hpp"
 #include "NodeControlService.hpp"
 
+#include "stddef.h"
+
 #ifdef __WINDOWS__
 #include <WinSock2.h>
 #include <Windows.h>
diff --git a/main.cpp b/main.cpp
index 438144c62..c18f726a6 100644
--- a/main.cpp
+++ b/main.cpp
@@ -573,8 +573,8 @@ static void printHelp(const char *cn,FILE *out)
 	fprintf(out,"Available switches:"ZT_EOL_S);
 	fprintf(out,"  -h                - Display this help"ZT_EOL_S);
 	fprintf(out,"  -v                - Show version"ZT_EOL_S);
-	fprintf(out,"  -p<port>          - Port for UDP (default: 9993)"ZT_EOL_S);
-	fprintf(out,"  -t<port>          - Port for TCP (default: disabled)"ZT_EOL_S);
+	fprintf(out,"  -p <port>         - Port for UDP (default: 9993)"ZT_EOL_S);
+	fprintf(out,"  -t <port>         - Port for TCP (default: disabled)"ZT_EOL_S);
 	//fprintf(out,"  -T<path>          - Override root topology, do not authenticate or update"ZT_EOL_S);
 #ifdef __UNIX_LIKE__
 	fprintf(out,"  -d                - Fork and run as daemon (Unix-ish OSes)"ZT_EOL_S);
diff --git a/make-freebsd.mk b/make-freebsd.mk
index 69f094abb..b50efa813 100644
--- a/make-freebsd.mk
+++ b/make-freebsd.mk
@@ -1,5 +1,5 @@
-CC=cc
-CXX=c++
+CC?=cc
+CXX?=c++
 
 INCLUDES=
 DEFS=
@@ -9,27 +9,6 @@ include objects.mk
 OBJS+=osnet/BSDEthernetTapFactory.o osnet/BSDEthernetTap.o osnet/BSDRoutingTable.o
 TESTNET_OBJS=testnet/SimNet.o testnet/SimNetSocketManager.o testnet/TestEthernetTap.o testnet/TestEthernetTapFactory.o testnet/TestRoutingTable.o
 
-# Enable SSE-optimized Salsa20 on x86 and x86_64 machines
-MACHINE=$(shell uname -m)
-ifeq ($(MACHINE),x86_64)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),amd64)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),i686)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),i586)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),i386)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),x86)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-
 # "make official" is a shortcut for this
 ifeq ($(ZT_OFFICIAL_RELEASE),1)
 	ZT_AUTO_UPDATE=1
@@ -42,19 +21,21 @@ endif
 # "make debug" is a shortcut for this
 ifeq ($(ZT_DEBUG),1)
 #	DEFS+=-DZT_TRACE -DZT_LOG_STDOUT 
-	CFLAGS=-Wall -g -pthread $(INCLUDES) $(DEFS)
-	LDFLAGS=
+	CFLAGS+=-Wall -g -pthread $(INCLUDES) $(DEFS)
+	LDFLAGS+=
 	STRIP=echo
 	# The following line enables optimization for the crypto code, since
 	# C25519 in particular is almost UNUSABLE in heavy testing without it.
-ext/lz4/lz4.o node/Salsa20.o node/SHA512.o node/C25519.o node/Poly1305.o: CFLAGS = -Wall -O2 -g -pthread $(INCLUDES) $(DEFS)
+ext/lz4/lz4.o node/Salsa20.o node/SHA512.o node/C25519.o node/Poly1305.o: CFLAGS?=-O2 CFLAGS+=-Wall -g -pthread $(INCLUDES) $(DEFS)
 else
-	CFLAGS=-Wall -O3 -fPIE -fvisibility=hidden -fstack-protector -pthread $(INCLUDES) -DNDEBUG $(DEFS)
-	LDFLAGS=-pie -Wl,-z,relro,-z,now
+	CFLAGS?=-O3 -fstack-protector
+	CFLAGS+=-Wall -fPIE -fvisibility=hidden -pthread $(INCLUDES) -DNDEBUG $(DEFS)
+	LDFLAGS+=-pie -Wl,-z,relro,-z,now
 	STRIP=strip --strip-all
 endif
 
-CXXFLAGS=$(CFLAGS) -fno-rtti
+CXXFLAGS?=-fno-rtti
+CXXFLAGS=$(CFLAGS)
 
 all:	one
 
diff --git a/make-linux.mk b/make-linux.mk
index 2033e35ce..9719f36a2 100644
--- a/make-linux.mk
+++ b/make-linux.mk
@@ -1,6 +1,6 @@
 # Pick clang or gcc, with preference for clang
-CC=$(shell which clang gcc cc 2>/dev/null | head -n 1)
-CXX=$(shell which clang++ g++ c++ 2>/dev/null | head -n 1)
+CC?=$(shell which clang gcc cc 2>/dev/null | head -n 1)
+CXX?=$(shell which clang++ g++ c++ 2>/dev/null | head -n 1)
 
 INCLUDES=
 DEFS=
@@ -10,27 +10,6 @@ include objects.mk
 OBJS+=osnet/LinuxRoutingTable.o osnet/LinuxEthernetTap.o osnet/LinuxEthernetTapFactory.o
 TESTNET_OBJS=testnet/SimNet.o testnet/SimNetSocketManager.o testnet/TestEthernetTap.o testnet/TestEthernetTapFactory.o testnet/TestRoutingTable.o
 
-# Enable SSE-optimized Salsa20 on x86 and x86_64 machines
-MACHINE=$(shell uname -m)
-ifeq ($(MACHINE),x86_64)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),amd64)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),i686)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),i586)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),i386)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-ifeq ($(MACHINE),x86)
-	DEFS+=-DZT_SALSA20_SSE 
-endif
-
 # "make official" is a shortcut for this
 ifeq ($(ZT_OFFICIAL_RELEASE),1)
 	ZT_AUTO_UPDATE=1
@@ -43,15 +22,16 @@ endif
 # "make debug" is a shortcut for this
 ifeq ($(ZT_DEBUG),1)
 #	DEFS+=-DZT_TRACE -DZT_LOG_STDOUT 
-	CFLAGS=-Wall -g -pthread $(INCLUDES) $(DEFS)
-	LDFLAGS=
+	CFLAGS+=-Wall -g -pthread $(INCLUDES) $(DEFS)
+	LDFLAGS+=
 	STRIP=echo
 	# The following line enables optimization for the crypto code, since
 	# C25519 in particular is almost UNUSABLE in heavy testing without it.
-ext/lz4/lz4.o node/Salsa20.o node/SHA512.o node/C25519.o node/Poly1305.o: CFLAGS = -Wall -O2 -g -pthread $(INCLUDES) $(DEFS)
+ext/lz4/lz4.o node/Salsa20.o node/SHA512.o node/C25519.o node/Poly1305.o: CFLAGS?=-O2 CFLAGS+=-Wall -g -pthread $(INCLUDES) $(DEFS)
 else
-	CFLAGS=-Wall -O3 -fPIE -fvisibility=hidden -fstack-protector -pthread $(INCLUDES) -DNDEBUG $(DEFS)
-	LDFLAGS=-pie -Wl,-z,relro,-z,now
+	CFLAGS?=-O3 -fstack-protector
+	CFLAGS+=-Wall -fPIE -fvisibility=hidden -pthread $(INCLUDES) -DNDEBUG $(DEFS)
+	LDFLAGS+=-pie -Wl,-z,relro,-z,now
 	STRIP=strip --strip-all
 endif
 
@@ -60,7 +40,8 @@ endif
 #LDFLAGS=
 #STRIP=echo
 
-CXXFLAGS=$(CFLAGS) -fno-rtti
+CXXFLAGS?=-fno-rtti
+CXXFLAGS+=$(CFLAGS)
 
 all:	one
 
diff --git a/make-mac.mk b/make-mac.mk
index 26ca8fe0b..c9e05c5a4 100644
--- a/make-mac.mk
+++ b/make-mac.mk
@@ -1,5 +1,5 @@
-CC=clang
-CXX=clang++
+CC?=clang
+CXX?=clang++
 
 INCLUDES=
 DEFS=
@@ -25,22 +25,21 @@ ifeq ($(ZT_AUTO_UPDATE),1)
 	DEFS+=-DZT_AUTO_UPDATE 
 endif
 
-# Enable SSE-optimized Salsa20 -- all Intel macs support SSE2
-DEFS+=-DZT_SALSA20_SSE
-
 ifeq ($(ZT_DEBUG),1)
 #	DEFS+=-DZT_TRACE -DZT_LOG_STDOUT 
-	CFLAGS=-Wall -g -pthread $(INCLUDES) $(DEFS)
+	CFLAGS+=-Wall -g -pthread $(INCLUDES) $(DEFS)
 	STRIP=echo
 	# The following line enables optimization for the crypto code, since
 	# C25519 in particular is almost UNUSABLE in heavy testing without it.
-ext/lz4/lz4.o node/Salsa20.o node/SHA512.o node/C25519.o node/Poly1305.o: CFLAGS = -Wall -O2 -g -pthread $(INCLUDES) $(DEFS)
+ext/lz4/lz4.o node/Salsa20.o node/SHA512.o node/C25519.o node/Poly1305.o: CFLAGS?=-O2 CFLAGS+=-Wall -g -pthread $(INCLUDES) $(DEFS)
 else
-	CFLAGS=-arch i386 -arch x86_64 -Wall -O3 -flto -fPIE -fvectorize -fstack-protector -pthread -mmacosx-version-min=10.6 -DNDEBUG -Wno-unused-private-field $(INCLUDES) $(DEFS)
+	CFLAGS?=-O3 -fstack-protector
+	CFLAGS+=-arch i386 -arch x86_64 -Wall -flto -fPIE -fvectorize -pthread -mmacosx-version-min=10.6 -DNDEBUG -Wno-unused-private-field $(INCLUDES) $(DEFS)
 	STRIP=strip
 endif
 
-CXXFLAGS=$(CFLAGS) -fno-rtti
+CXXFLAGS?=-fno-rtti
+CXXFLAGS=$(CFLAGS)
 
 all: one
 
diff --git a/node/Array.hpp b/node/Array.hpp
index ce179c16c..a1779909a 100644
--- a/node/Array.hpp
+++ b/node/Array.hpp
@@ -31,12 +31,14 @@
 #include <string>
 #include <algorithm>
 
+#include "stddef.h"
+
 namespace ZeroTier {
 
 /**
  * Static array -- a simple thing that's belonged in STL since the time of the dinosaurs
  */
-template<typename T,std::size_t S>
+template<typename T,size_t S>
 class Array
 {
 public:
@@ -44,19 +46,19 @@ public:
 
 	Array(const Array &a)
 	{
-		for(std::size_t i=0;i<S;++i)
+		for(size_t i=0;i<S;++i)
 			data[i] = a.data[i];
 	}
 
 	Array(const T *ptr)
 	{
-		for(std::size_t i=0;i<S;++i)
+		for(size_t i=0;i<S;++i)
 			data[i] = ptr[i];
 	}
 
 	inline Array &operator=(const Array &a)
 	{
-		for(std::size_t i=0;i<S;++i)
+		for(size_t i=0;i<S;++i)
 			data[i] = a.data[i];
 		return *this;
 	}
@@ -68,8 +70,8 @@ public:
 	typedef const T& const_reference;
 	typedef T* iterator;
 	typedef const T* const_iterator;
-	typedef std::size_t size_type;
-	typedef std::ptrdiff_t difference_type;
+	typedef size_t size_type;
+	typedef ptrdiff_t difference_type;
 	typedef std::reverse_iterator<iterator> reverse_iterator;
 	typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
 
@@ -83,11 +85,11 @@ public:
 	inline const_reverse_iterator rbegin() const throw() { return const_reverse_iterator(begin()); }
 	inline const_reverse_iterator rend() const throw() { return const_reverse_iterator(end()); }
 
-	inline std::size_t size() const throw() { return S; }
-	inline std::size_t max_size() const throw() { return S; }
+	inline size_t size() const throw() { return S; }
+	inline size_t max_size() const throw() { return S; }
 
-	inline reference operator[](const std::size_t n) throw() { return data[n]; }
-	inline const_reference operator[](const std::size_t n) const throw() { return data[n]; }
+	inline reference operator[](const size_t n) throw() { return data[n]; }
+	inline const_reference operator[](const size_t n) const throw() { return data[n]; }
 
 	inline reference front() throw() { return data[0]; }
 	inline const_reference front() const throw() { return data[0]; }
diff --git a/node/InetAddress.cpp b/node/InetAddress.cpp
index 832acf04c..5e36be0d5 100644
--- a/node/InetAddress.cpp
+++ b/node/InetAddress.cpp
@@ -137,7 +137,7 @@ std::string InetAddress::toString() const
 
 void InetAddress::fromString(const std::string &ipSlashPort)
 {
-	std::size_t slashAt = ipSlashPort.find('/');
+	size_t slashAt = ipSlashPort.find('/');
 	if ((slashAt == std::string::npos)||(slashAt >= ipSlashPort.length()))
 		set(ipSlashPort,0);
 	else {
diff --git a/node/Peer.hpp b/node/Peer.hpp
index 78506fd23..fe5362e3f 100644
--- a/node/Peer.hpp
+++ b/node/Peer.hpp
@@ -400,13 +400,4 @@ private:
 
 } // namespace ZeroTier
 
-// Add a swap() for shared ptr's to peers to speed up peer sorts
-namespace std {
-	template<>
-	inline void swap(ZeroTier::SharedPtr<ZeroTier::Peer> &a,ZeroTier::SharedPtr<ZeroTier::Peer> &b)
-	{
-		a.swap(b);
-	}
-}
-
 #endif
diff --git a/node/Salsa20.hpp b/node/Salsa20.hpp
index 0c1f3be46..6f51a160b 100644
--- a/node/Salsa20.hpp
+++ b/node/Salsa20.hpp
@@ -11,6 +11,11 @@
 
 #include "Constants.hpp"
 
+/* Enable SSE-optimized Salsa20 on SSE2 machines */
+#ifdef __SSE2__
+#define ZT_SALSA20_SSE
+#endif
+
 #ifdef ZT_SALSA20_SSE
 #include <emmintrin.h>
 #endif // ZT_SALSA20_SSE
diff --git a/windows/ZeroTierOne/ZeroTierOne.vcxproj b/windows/ZeroTierOne/ZeroTierOne.vcxproj
index a651b5f73..49392b54d 100644
--- a/windows/ZeroTierOne/ZeroTierOne.vcxproj
+++ b/windows/ZeroTierOne/ZeroTierOne.vcxproj
@@ -270,7 +270,7 @@
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <SDLCheck>true</SDLCheck>
       <AdditionalIncludeDirectories>$(SolutionDir)\ext\bin\libcrypto\include</AdditionalIncludeDirectories>
-      <PreprocessorDefinitions>ZT_OFFICIAL_RELEASE;ZT_AUTO_UPDATE;ZT_SALSA20_SSE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>ZT_OFFICIAL_RELEASE;ZT_AUTO_UPDATE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
       <StringPooling>true</StringPooling>
@@ -294,7 +294,7 @@
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <SDLCheck>true</SDLCheck>
       <AdditionalIncludeDirectories>$(SolutionDir)\ext\bin\libcrypto\include</AdditionalIncludeDirectories>
-      <PreprocessorDefinitions>ZT_OFFICIAL_RELEASE;ZT_AUTO_UPDATE;ZT_SALSA20_SSE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>ZT_OFFICIAL_RELEASE;ZT_AUTO_UPDATE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <EnableEnhancedInstructionSet>NotSet</EnableEnhancedInstructionSet>
       <StringPooling>true</StringPooling>