Litecoin: Runtime detection of SSE2 32bit x86 for Scrypt

* x86_64 and Intel MacOS X always uses scrypt-sse2, non-x86 uses scrypt-generic.
* x86 (not Mac) detects cpuid features and chooses sse2 or generic during runtime.

How to Build with SSE2 Support
==============================
* make litecoind with USE_SSE2=1
* qmake with USE_SSE2=1

bitcoin-qt.pro

@@ -316,6 +316,11 @@ DEFINES += BITCOIN_QT_TEST
   macx: CONFIG -= app_bundle
 }
 
+contains(USE_SSE2, 1) {
+SOURCES += src/scrypt-sse2.cpp
+DEFINES += USE_SSE2
+}
+
 # Todo: Remove this line when switching to Qt5, as that option was removed
 CODECFORTR = UTF-8
 

contrib/gitian-descriptors/gitian-win32.yml

@@ -51,7 +51,7 @@ script: |
   export LD_PRELOAD=/usr/lib/faketime/libfaketime.so.1
   export FAKETIME=$REFERENCE_DATETIME
   export TZ=UTC
-  $HOME/qt/src/bin/qmake -spec unsupported/win32-g++-cross MINIUPNPC_LIB_PATH=$HOME/build/miniupnpc MINIUPNPC_INCLUDE_PATH=$HOME/build/ BDB_LIB_PATH=$HOME/build/db-4.8.30.NC/build_unix BDB_INCLUDE_PATH=$HOME/build/db-4.8.30.NC/build_unix BOOST_LIB_PATH=$HOME/build/boost_1_50_0/stage/lib BOOST_INCLUDE_PATH=$HOME/build/boost_1_50_0 BOOST_LIB_SUFFIX=-mt-s BOOST_THREAD_LIB_SUFFIX=_win32-mt-s OPENSSL_LIB_PATH=$HOME/build/openssl-1.0.1c OPENSSL_INCLUDE_PATH=$HOME/build/openssl-1.0.1c/include QRENCODE_LIB_PATH=$HOME/build/qrencode-3.2.0/.libs QRENCODE_INCLUDE_PATH=$HOME/build/qrencode-3.2.0 USE_QRCODE=1 INCLUDEPATH=$HOME/build DEFINES=BOOST_THREAD_USE_LIB BITCOIN_NEED_QT_PLUGINS=1 QMAKE_LRELEASE=lrelease QMAKE_CXXFLAGS=-frandom-seed=litecoin USE_BUILD_INFO=1
+  $HOME/qt/src/bin/qmake -spec unsupported/win32-g++-cross MINIUPNPC_LIB_PATH=$HOME/build/miniupnpc MINIUPNPC_INCLUDE_PATH=$HOME/build/ BDB_LIB_PATH=$HOME/build/db-4.8.30.NC/build_unix BDB_INCLUDE_PATH=$HOME/build/db-4.8.30.NC/build_unix BOOST_LIB_PATH=$HOME/build/boost_1_50_0/stage/lib BOOST_INCLUDE_PATH=$HOME/build/boost_1_50_0 BOOST_LIB_SUFFIX=-mt-s BOOST_THREAD_LIB_SUFFIX=_win32-mt-s OPENSSL_LIB_PATH=$HOME/build/openssl-1.0.1c OPENSSL_INCLUDE_PATH=$HOME/build/openssl-1.0.1c/include QRENCODE_LIB_PATH=$HOME/build/qrencode-3.2.0/.libs QRENCODE_INCLUDE_PATH=$HOME/build/qrencode-3.2.0 USE_QRCODE=1 INCLUDEPATH=$HOME/build DEFINES=BOOST_THREAD_USE_LIB BITCOIN_NEED_QT_PLUGINS=1 QMAKE_LRELEASE=lrelease QMAKE_CXXFLAGS=-frandom-seed=litecoin USE_BUILD_INFO=1 USE_SSE2=1
   make $MAKEOPTS
   i586-mingw32msvc-strip release/litecoin-qt.exe
   cp release/litecoin-qt.exe $OUTDIR/
@@ -60,7 +60,7 @@ script: |
   export LD_PRELOAD=/usr/lib/faketime/libfaketime.so.1
   export FAKETIME=$REFERENCE_DATETIME
   export TZ=UTC
-  make -f makefile.linux-mingw $MAKEOPTS DEPSDIR=$HOME/build litecoind.exe USE_UPNP=0 DEBUGFLAGS="-frandom-seed=litecoin"
+  make -f makefile.linux-mingw $MAKEOPTS DEPSDIR=$HOME/build litecoind.exe USE_UPNP=0 DEBUGFLAGS="-frandom-seed=litecoin" USE_SSE2=1
   i586-mingw32msvc-strip litecoind.exe
   mkdir $OUTDIR/daemon
   cp litecoind.exe $OUTDIR/daemon

contrib/gitian-descriptors/gitian.yml

@@ -46,10 +46,10 @@ script: |
   cp $OUTDIR/src/doc/README.md $OUTDIR
   cp $OUTDIR/src/COPYING $OUTDIR
   cd src
-  make -f makefile.unix STATIC=1 OPENSSL_INCLUDE_PATH="$INSTDIR/include" OPENSSL_LIB_PATH="$INSTDIR/lib" $MAKEOPTS litecoind USE_UPNP=0 DEBUGFLAGS=
+  make -f makefile.unix STATIC=1 OPENSSL_INCLUDE_PATH="$INSTDIR/include" OPENSSL_LIB_PATH="$INSTDIR/lib" $MAKEOPTS litecoind USE_UPNP=0 DEBUGFLAGS= USE_SSE2=1
   mkdir -p $OUTDIR/bin/$GBUILD_BITS
   install -s litecoind $OUTDIR/bin/$GBUILD_BITS
   cd ..
-  qmake INCLUDEPATH="$INSTDIR/include" LIBS="-L$INSTDIR/lib" RELEASE=1 USE_QRCODE=1
+  qmake INCLUDEPATH="$INSTDIR/include" LIBS="-L$INSTDIR/lib" RELEASE=1 USE_QRCODE=1 USE_SSE2=1
   make $MAKEOPTS
   install litecoin-qt $OUTDIR/bin/$GBUILD_BITS

src/init.cpp

@@ -22,6 +22,18 @@
 #include <signal.h>
 #endif
 
+#if defined(USE_SSE2)
+#if !defined(MAC_OSX) && (defined(_M_IX86) || defined(__i386__) || defined(__i386))
+#ifdef _MSC_VER
+// MSVC 64bit is unable to use inline asm
+#include <intrin.h>
+#else
+// GCC Linux or i686-w64-mingw32
+#include <cpuid.h>
+#endif
+#endif
+#endif
+
 using namespace std;
 using namespace boost;
 
@@ -494,6 +506,23 @@ bool AppInit2(boost::thread_group& threadGroup)
     sigaction(SIGHUP, &sa_hup, NULL);
 #endif
 
+#if defined(USE_SSE2)
+    unsigned int cpuid_edx=0;
+#if !defined(MAC_OSX) && (defined(_M_IX86) || defined(__i386__) || defined(__i386))
+    // 32bit x86 Linux or Windows, detect cpuid features
+#if defined(_MSC_VER)
+    // MSVC
+    int x86cpuid[4];
+    __cpuid(x86cpuid, 1);
+    cpuid_edx = (unsigned int)buffer[3];
+#else
+    // Linux or i686-w64-mingw32 (gcc-4.6.3)
+    unsigned int eax, ebx, ecx;
+    __get_cpuid(1, &eax, &ebx, &ecx, &cpuid_edx);
+#endif
+#endif
+#endif
+
     // ********************************************************* Step 2: parameter interactions
 
     fTestNet = GetBoolArg("-testnet");
@@ -808,6 +837,10 @@ bool AppInit2(boost::thread_group& threadGroup)
 
     // ********************************************************* Step 7: load block chain
 
+#if defined(USE_SSE2)
+    scrypt_detect_sse2(cpuid_edx);
+#endif
+
     fReindex = GetBoolArg("-reindex");
 
     // Upgrading to 0.8; hard-link the old blknnnn.dat files into /blocks/

src/main.h

@@ -2234,6 +2234,9 @@ struct CBlockTemplate
     std::vector<int64_t> vTxSigOps;
 };
 
+#if defined(_M_IX86) || defined(__i386__) || defined(__i386) || defined(_M_X64) || defined(__x86_64__) || defined(_M_AMD64)
+extern unsigned int cpuid_edx;
+#endif
 
 
 

src/makefile.linux-mingw

@@ -94,6 +94,12 @@ OBJS= \
     obj/leveldb.o \
     obj/txdb.o
 
+ifdef USE_SSE2
+DEFS += -DUSE_SSE2
+OBJS_SSE2= obj/scrypt-sse2.o
+OBJS += $(OBJS_SSE2)
+endif
+
 all: litecoind.exe
 
 DEFS += -I"$(CURDIR)/leveldb/include"
@@ -106,6 +112,9 @@ obj/build.h: FORCE
 version.cpp: obj/build.h
 DEFS += -DHAVE_BUILD_INFO
 
+obj/%-sse2.o: %-sse2.cpp
+	$(CXX) -c $(xCXXFLAGS) -msse2 -o $@ $<
+
 obj/%.o: %.cpp $(HEADERS)
 	$(CXX) -c $(xCXXFLAGS) -o $@ $<
 

src/makefile.mingw

@@ -102,6 +102,11 @@ OBJS= \
     obj/leveldb.o \
     obj/txdb.o
 
+ifdef USE_SSE2
+DEFS += -DUSE_SSE2
+OBJS_SSE2= obj/scrypt-sse2.o
+OBJS += $(OBJS_SSE2)
+endif
 
 all: litecoind.exe
 
@@ -117,6 +122,9 @@ DEFS += $(addprefix -I,$(CURDIR)/leveldb/helpers)
 leveldb/libleveldb.a:
 	cd leveldb && $(MAKE) CC=$(CC) CXX=$(CXX) OPT="$(CFLAGS)" TARGET_OS=NATIVE_WINDOWS libleveldb.a libmemenv.a && cd ..
 
+obj/%-sse2.o: %-sse2.cpp
+	$(CXX) -c $(CFLAGS) -msse2 -o $@ $<
+
 obj/%.o: %.cpp $(HEADERS)
 	$(CXX) -c $(CFLAGS) -o $@ $<
 

src/makefile.osx

@@ -105,6 +105,12 @@ OBJS= \
     obj/leveldb.o \
     obj/txdb.o
 
+ifdef USE_SSE2
+DEFS += -DUSE_SSE2
+OBJS_SSE2= obj/scrypt-sse2.o
+OBJS += $(OBJS_SSE2)
+endif
+
 ifndef USE_UPNP
 	override USE_UPNP = -
 endif
@@ -144,6 +150,13 @@ obj/build.h: FORCE
 version.cpp: obj/build.h
 DEFS += -DHAVE_BUILD_INFO
 
+obj/%-sse2.o: %-sse2.cpp
+	$(CXX) -c $(CFLAGS) -msse2 -MMD -MF $(@:%.o=%.d) -o $@ $<
+	@cp $(@:%.o=%.d) $(@:%.o=%.P); \
+	  sed -e 's/#.*//' -e 's/^[^:]*: *//' -e 's/ *\\$$//' \
+	      -e '/^$$/ d' -e 's/$$/ :/' < $(@:%.o=%.d) >> $(@:%.o=%.P); \
+	  rm -f $(@:%.o=%.d)
+
 obj/%.o: %.cpp
 	$(CXX) -c $(CFLAGS) -MMD -MF $(@:%.o=%.d) -o $@ $<
 	@cp $(@:%.o=%.d) $(@:%.o=%.P); \

src/makefile.unix

@@ -144,10 +144,10 @@ OBJS= \
     obj/leveldb.o \
     obj/txdb.o
 
-OBJS_SSE2= obj/scrypt-sse2.o
 
-ifdef SSE2
+ifdef USE_SSE2
 DEFS += -DUSE_SSE2
+OBJS_SSE2= obj/scrypt-sse2.o
 OBJS += $(OBJS_SSE2)
 endif
 

src/scrypt-sse2.cpp

@@ -26,7 +26,7 @@
  * This file was originally written by Colin Percival as part of the Tarsnap
  * online backup system.
  */
-#ifdef __SSE2__
+
 #include "scrypt.h"
 #include <stdlib.h>
 #include <stdint.h>
@@ -91,10 +91,9 @@ static inline void xor_salsa8_sse2(__m128i B[4], const __m128i Bx[4])
 	B[2] = _mm_add_epi32(B[2], X2);
 	B[3] = _mm_add_epi32(B[3], X3);
 }
-#endif
+
 void scrypt_1024_1_1_256_sp_sse2(const char *input, char *output, char *scratchpad)
 {
-#ifdef __SSE2__
 	uint8_t B[128];
 	union {
 		__m128i i128[8];
@@ -134,5 +133,4 @@ void scrypt_1024_1_1_256_sp_sse2(const char *input, char *output, char *scratchp
 	}
 
 	PBKDF2_SHA256((const uint8_t *)input, 80, B, 128, 1, (uint8_t *)output, 32);
-#endif
 }

src/scrypt.cpp

@@ -28,6 +28,7 @@
  */
 
 #include "scrypt.h"
+#include "util.h"
 #include <stdlib.h>
 #include <stdint.h>
 #include <string.h>
@@ -242,7 +243,7 @@ static inline void xor_salsa8(uint32_t B[16], const uint32_t Bx[16])
 	B[15] += x15;
 }
 
-void scrypt_1024_1_1_256_sp(const char *input, char *output, char *scratchpad)
+void scrypt_1024_1_1_256_sp_generic(const char *input, char *output, char *scratchpad)
 {
 	uint8_t B[128];
 	uint32_t X[32];
@@ -275,14 +276,48 @@ void scrypt_1024_1_1_256_sp(const char *input, char *output, char *scratchpad)
 	PBKDF2_SHA256((const uint8_t *)input, 80, B, 128, 1, (uint8_t *)output, 32);
 }
 
+#if defined(USE_SSE2)
+#if defined(_M_X64) || defined(__x86_64__) || defined(_M_AMD64) || (defined(MAC_OSX) && defined(__i386__))
+/* Always SSE2 */
+void scrypt_detect_sse2(unsigned int cpuid_edx)
+{
+    printf("scrypt: using scrypt-sse2 as built.\n");
+}
+#else
+/* Detect SSE2 */
+void (*scrypt_1024_1_1_256_sp)(const char *input, char *output, char *scratchpad);
+
+void scrypt_detect_sse2(unsigned int cpuid_edx)
+{
+    if (cpuid_edx & 1<<26)
+    {
+        scrypt_1024_1_1_256_sp = &scrypt_1024_1_1_256_sp_sse2;
+        printf("scrypt: using scrypt-sse2 as detected.\n");
+    }
+    else
+    {
+        scrypt_1024_1_1_256_sp = &scrypt_1024_1_1_256_sp_generic;
+        printf("scrypt: using scrypt-generic, SSE2 unavailable.\n");
+    }
+}
+#endif
+#endif
+
 void scrypt_1024_1_1_256(const char *input, char *output)
 {
 	char scratchpad[SCRYPT_SCRATCHPAD_SIZE];
-#ifdef USE_SSE2
-        // todo: runtime detection at startup and use function pointer
-        if(1)
-          scrypt_1024_1_1_256_sp_sse2(input, output, scratchpad);
-        else
+#if defined(USE_SSE2)
+        // Detection would work, but in cases where we KNOW it always has SSE2,
+        // it is faster to use directly than to use a function pointer or conditional.
+#if defined(_M_X64) || defined(__x86_64__) || defined(_M_AMD64) || (defined(MAC_OSX) && defined(__i386__))
+        // Always SSE2: x86_64 or Intel MacOS X
+        scrypt_1024_1_1_256_sp_sse2(input, output, scratchpad);
+#else
+        // Detect SSE2: 32bit x86 Linux or Windows
+        scrypt_1024_1_1_256_sp(input, output, scratchpad);
+#endif
+#else
+        // Generic scrypt
+        scrypt_1024_1_1_256_sp_generic(input, output, scratchpad);
 #endif
-          scrypt_1024_1_1_256_sp(input, output, scratchpad);
 }

src/scrypt.h

@@ -4,9 +4,14 @@
 #include <stdint.h>
 static const int SCRYPT_SCRATCHPAD_SIZE = 131072 + 63;
 
-void scrypt_1024_1_1_256_sp_sse2(const char *input, char *output, char *scratchpad);
-void scrypt_1024_1_1_256_sp(const char *input, char *output, char *scratchpad);
 void scrypt_1024_1_1_256(const char *input, char *output);
+void scrypt_1024_1_1_256_sp_generic(const char *input, char *output, char *scratchpad);
+
+#if defined(USE_SSE2)
+extern void scrypt_detect_sse2(unsigned int cpuid_edx);
+void scrypt_1024_1_1_256_sp_sse2(const char *input, char *output, char *scratchpad);
+extern void (*scrypt_1024_1_1_256_sp)(const char *input, char *output, char *scratchpad);
+#endif
 
 void
 PBKDF2_SHA256(const uint8_t *passwd, size_t passwdlen, const uint8_t *salt,
@@ -27,6 +32,4 @@ static inline void le32enc(void *pp, uint32_t x)
         p[2] = (x >> 16) & 0xff;
         p[3] = (x >> 24) & 0xff;
 }
-
-
 #endif