Litecoin: More straightforward SSE2 detection

- Move CPU capabilities detection from init.cpp to scrypt.cpp
- scrypt_1024_1_1_256_sp is defined as a macro: no need to check for SSE2
  every time it is used
- Test correctness of both scrypt-generic and scrypt-sse2

src/init.cpp

@@ -22,18 +22,6 @@
 #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;
 
@@ -512,23 +500,6 @@ 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");
@@ -698,6 +669,10 @@ bool AppInit2(boost::thread_group& threadGroup)
 
     int64 nStart;
 
+#if defined(USE_SSE2)
+    scrypt_detect_sse2();
+#endif
+
     // ********************************************************* Step 5: verify wallet database integrity
 
     if (!fDisableWallet) {
@@ -846,10 +821,6 @@ 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.cpp

@@ -4596,20 +4596,7 @@ void static LitecoinMiner(CWallet *pwallet)
             char scratchpad[SCRYPT_SCRATCHPAD_SIZE];
             loop
             {
-#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(BEGIN(pblock->nVersion), BEGIN(thash), scratchpad);
-#else
-                // Detect SSE2: 32bit x86 Linux or Windows
                 scrypt_1024_1_1_256_sp(BEGIN(pblock->nVersion), BEGIN(thash), scratchpad);
-#endif
-#else
-                // Generic scrypt
-                scrypt_1024_1_1_256_sp_generic(BEGIN(pblock->nVersion), BEGIN(thash), scratchpad);
-#endif
 
                 if (thash <= hashTarget)
                 {

src/scrypt.cpp

@@ -34,6 +34,16 @@
 #include <string.h>
 #include <openssl/sha.h>
 
+#if defined(USE_SSE2) && !defined(USE_SSE2_ALWAYS)
+#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
+
 static inline uint32_t be32dec(const void *pp)
 {
 	const uint8_t *p = (uint8_t const *)pp;
@@ -202,26 +212,26 @@ static inline void xor_salsa8(uint32_t B[16], const uint32_t Bx[16])
 		/* Operate on columns. */
 		x04 ^= ROTL(x00 + x12,  7);  x09 ^= ROTL(x05 + x01,  7);
 		x14 ^= ROTL(x10 + x06,  7);  x03 ^= ROTL(x15 + x11,  7);
-		
+
 		x08 ^= ROTL(x04 + x00,  9);  x13 ^= ROTL(x09 + x05,  9);
 		x02 ^= ROTL(x14 + x10,  9);  x07 ^= ROTL(x03 + x15,  9);
-		
+
 		x12 ^= ROTL(x08 + x04, 13);  x01 ^= ROTL(x13 + x09, 13);
 		x06 ^= ROTL(x02 + x14, 13);  x11 ^= ROTL(x07 + x03, 13);
-		
+
 		x00 ^= ROTL(x12 + x08, 18);  x05 ^= ROTL(x01 + x13, 18);
 		x10 ^= ROTL(x06 + x02, 18);  x15 ^= ROTL(x11 + x07, 18);
-		
+
 		/* Operate on rows. */
 		x01 ^= ROTL(x00 + x03,  7);  x06 ^= ROTL(x05 + x04,  7);
 		x11 ^= ROTL(x10 + x09,  7);  x12 ^= ROTL(x15 + x14,  7);
-		
+
 		x02 ^= ROTL(x01 + x00,  9);  x07 ^= ROTL(x06 + x05,  9);
 		x08 ^= ROTL(x11 + x10,  9);  x13 ^= ROTL(x12 + x15,  9);
-		
+
 		x03 ^= ROTL(x02 + x01, 13);  x04 ^= ROTL(x07 + x06, 13);
 		x09 ^= ROTL(x08 + x11, 13);  x14 ^= ROTL(x13 + x12, 13);
-		
+
 		x00 ^= ROTL(x03 + x02, 18);  x05 ^= ROTL(x04 + x07, 18);
 		x10 ^= ROTL(x09 + x08, 18);  x15 ^= ROTL(x14 + x13, 18);
 	}
@@ -251,7 +261,7 @@ void scrypt_1024_1_1_256_sp_generic(const char *input, char *output, char *scrat
 	uint32_t i, j, k;
 
 	V = (uint32_t *)(((uintptr_t)(scratchpad) + 63) & ~ (uintptr_t)(63));
-	
+
 	PBKDF2_SHA256((const uint8_t *)input, 80, (const uint8_t *)input, 80, 1, B, 128);
 
 	for (k = 0; k < 32; k++)
@@ -277,47 +287,43 @@ void scrypt_1024_1_1_256_sp_generic(const char *input, char *output, char *scrat
 }
 
 #if defined(USE_SSE2)
-#if defined(_M_X64) || defined(__x86_64__) || defined(_M_AMD64) || (defined(MAC_OSX) && defined(__i386__))
+// By default, set to generic scrypt function. This will prevent crash in case when scrypt_detect_sse2() wasn't called
-/* Always SSE2 */
+void (*scrypt_1024_1_1_256_sp_detected)(const char *input, char *output, char *scratchpad) = &scrypt_1024_1_1_256_sp_generic;
-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)
+void scrypt_detect_sse2()
 {
+#if defined(USE_SSE2_ALWAYS)
+    printf("scrypt: using scrypt-sse2 as built.\n");
+#else // USE_SSE2_ALWAYS
+    // 32bit x86 Linux or Windows, detect cpuid features
+    unsigned int cpuid_edx=0;
+#if defined(_MSC_VER)
+    // MSVC
+    int x86cpuid[4];
+    __cpuid(x86cpuid, 1);
+    cpuid_edx = (unsigned int)buffer[3];
+#else // _MSC_VER
+    // Linux or i686-w64-mingw32 (gcc-4.6.3)
+    unsigned int eax, ebx, ecx;
+    __get_cpuid(1, &eax, &ebx, &ecx, &cpuid_edx);
+#endif // _MSC_VER
+
     if (cpuid_edx & 1<<26)
     {
-        scrypt_1024_1_1_256_sp = &scrypt_1024_1_1_256_sp_sse2;
+        scrypt_1024_1_1_256_sp_detected = &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;
+        scrypt_1024_1_1_256_sp_detected = &scrypt_1024_1_1_256_sp_generic;
         printf("scrypt: using scrypt-generic, SSE2 unavailable.\n");
     }
+#endif // USE_SSE2_ALWAYS
 }
 #endif
-#endif
 
 void scrypt_1024_1_1_256(const char *input, char *output)
 {
 	char scratchpad[SCRYPT_SCRATCHPAD_SIZE];
-#if defined(USE_SSE2)
+    scrypt_1024_1_1_256_sp(input, output, scratchpad);
-        // 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
 }

src/scrypt.h

@@ -2,15 +2,25 @@
 #define SCRYPT_H
 #include <stdlib.h>
 #include <stdint.h>
+
 static const int SCRYPT_SCRATCHPAD_SIZE = 131072 + 63;
 
 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);
+#if defined(_M_X64) || defined(__x86_64__) || defined(_M_AMD64) || (defined(MAC_OSX) && defined(__i386__))
+#define USE_SSE2_ALWAYS 1
+#define scrypt_1024_1_1_256_sp(input, output, scratchpad) scrypt_1024_1_1_256_sp_sse2((input), (output), (scratchpad))
+#else
+#define scrypt_1024_1_1_256_sp(input, output, scratchpad) scrypt_1024_1_1_256_sp_detected((input), (output), (scratchpad))
+#endif
+
+void scrypt_detect_sse2();
 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);
+extern void (*scrypt_1024_1_1_256_sp_detected)(const char *input, char *output, char *scratchpad);
+#else
+#define scrypt_1024_1_1_256_sp(input, output, scratchpad) scrypt_1024_1_1_256_sp_generic((input), (output), (scratchpad))
 #endif
 
 void

src/test/scrypt_tests.cpp

@@ -11,6 +11,9 @@ BOOST_AUTO_TEST_CASE(scrypt_hashtest)
     #define HASHCOUNT 5
     const char* inputhex[HASHCOUNT] = { "020000004c1271c211717198227392b029a64a7971931d351b387bb80db027f270411e398a07046f7d4a08dd815412a8712f874a7ebf0507e3878bd24e20a3b73fd750a667d2f451eac7471b00de6659", "0200000011503ee6a855e900c00cfdd98f5f55fffeaee9b6bf55bea9b852d9de2ce35828e204eef76acfd36949ae56d1fbe81c1ac9c0209e6331ad56414f9072506a77f8c6faf551eac7471b00389d01", "02000000a72c8a177f523946f42f22c3e86b8023221b4105e8007e59e81f6beb013e29aaf635295cb9ac966213fb56e046dc71df5b3f7f67ceaeab24038e743f883aff1aaafaf551eac7471b0166249b", "010000007824bc3a8a1b4628485eee3024abd8626721f7f870f8ad4d2f33a27155167f6a4009d1285049603888fe85a84b6c803a53305a8d497965a5e896e1a00568359589faf551eac7471b0065434e", "0200000050bfd4e4a307a8cb6ef4aef69abc5c0f2d579648bd80d7733e1ccc3fbc90ed664a7f74006cb11bde87785f229ecd366c2d4e44432832580e0608c579e4cb76f383f7f551eac7471b00c36982" };
     const char* expected[HASHCOUNT] = { "00000000002bef4107f882f6115e0b01f348d21195dacd3582aa2dabd7985806" , "00000000003a0d11bdd5eb634e08b7feddcfbbf228ed35d250daf19f1c88fc94", "00000000000b40f895f288e13244728a6c2d9d59d8aff29c65f8dd5114a8ca81", "00000000003007005891cd4923031e99d8e8d72f6e8e7edc6a86181897e105fe", "000000000018f0b426a4afc7130ccb47fa02af730d345b4fe7c7724d3800ec8c" };
+#if defined(USE_SSE2)
+    scrypt_detect_sse2();
+#endif
     uint256 scrypthash;
     std::vector<unsigned char> inputbytes;
     char scratchpad[SCRYPT_SCRATCHPAD_SIZE];
@@ -19,6 +22,7 @@ BOOST_AUTO_TEST_CASE(scrypt_hashtest)
 #if defined(USE_SSE2)
         // Test SSE2 scrypt
         scrypt_1024_1_1_256_sp_sse2((const char*)&inputbytes[0], BEGIN(scrypthash), scratchpad);
+        BOOST_CHECK_EQUAL(scrypthash.ToString().c_str(), expected[i]);
 #endif
         // Test generic scrypt
         scrypt_1024_1_1_256_sp_generic((const char*)&inputbytes[0], BEGIN(scrypthash), scratchpad);