Merge pull request #2773 from patricklodder/1.14.6-fix-scrypt-sse2

Fix and enable scrypt sse2 integration

.github/workflows/ci.yml

@@ -170,7 +170,7 @@ jobs:
             run-bench: true
             run-tests: true
             dep-opts: "AVX2=1"
-            config-opts: "--with-intel-avx2 --enable-gui=qt5 --enable-zmq --enable-glibc-back-compat --enable-reduce-exports"
+            config-opts: "--enable-scrypt-sse2 --with-intel-avx2 --enable-gui=qt5 --enable-zmq --enable-glibc-back-compat --enable-reduce-exports"
             goal: install
 
     runs-on: ${{ matrix.os }}

configure.ac

@@ -177,6 +177,12 @@ AC_ARG_ENABLE([zmq],
   [use_zmq=$enableval],
   [use_zmq=yes])
 
+AC_ARG_ENABLE([scrypt-sse2],
+  [AS_HELP_STRING([--enable-scrypt-sse2],
+  [Build with scrypt sse2 implementation (default is no)])],
+  [use_scrypt_sse2=$enableval],
+  [use_scrypt_sse2=no])
+
 AC_ARG_WITH([intel-avx2],
   [AS_HELP_STRING([--with-intel-avx2],
   [Build with intel avx2 (default is no)])],
@@ -725,6 +731,11 @@ BOOST_LIBS="$BOOST_LDFLAGS $BOOST_SYSTEM_LIB $BOOST_FILESYSTEM_LIB $BOOST_PROGRA
 
 fi
 
+# Configure Scrypt SSE2
+if test x$use_scrypt_sse2 = xyes; then
+  AC_DEFINE(USE_SSE2, 1, [Define this symbol if SSE2 works])
+fi
+
 if test x$intel_avx2 = xyes; then
   case $host in
      x86_64-*-linux*)
@@ -1028,6 +1039,7 @@ AM_CONDITIONAL([USE_LCOV],[test x$use_lcov = xyes])
 AM_CONDITIONAL([GLIBC_BACK_COMPAT],[test x$use_glibc_compat = xyes])
 AM_CONDITIONAL([HARDEN],[test x$use_hardening = xyes])
 AM_CONDITIONAL([WORDS_BIGENDIAN],[test x$ac_cv_c_bigendian = xyes])
+AM_CONDITIONAL([USE_SCRYPT_SSE2], [test x$use_scrypt_sse2 = xyes])
 
 AC_DEFINE(CLIENT_VERSION_MAJOR, _CLIENT_VERSION_MAJOR, [Major version])
 AC_DEFINE(CLIENT_VERSION_MINOR, _CLIENT_VERSION_MINOR, [Minor version])

src/Makefile.am

@@ -267,6 +267,11 @@ crypto_libdogecoin_crypto_a_SOURCES = \
   crypto/sha512.cpp \
   crypto/sha512.h
 
+# only include SSE2 scrypt sources if USE_SCRYPT_SSE2 is defined
+if USE_SCRYPT_SSE2
+crypto_libdogecoin_crypto_a_SOURCES += crypto/scrypt-sse2.cpp
+endif
+
 # consensus: shared between all executables that validate any consensus rules.
 libdogecoin_consensus_a_CPPFLAGS = $(AM_CPPFLAGS) $(BITCOIN_INCLUDES)
 libdogecoin_consensus_a_CXXFLAGS = $(AM_CXXFLAGS) $(PIE_FLAGS)

src/crypto/scrypt.cpp

@@ -222,10 +222,11 @@ void scrypt_1024_1_1_256_sp_generic(const char *input, char *output, char *scrat
 // By default, set to generic scrypt function. This will prevent crash in case when scrypt_detect_sse2() wasn't called
 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()
+bool scrypt_detect_sse2()
 {
+    bool fUsingSSE2;
 #if defined(USE_SSE2_ALWAYS)
-    printf("scrypt: using scrypt-sse2 as built.\n");
+    fUsingSSE2 = true;
 #else // USE_SSE2_ALWAYS
     // 32bit x86 Linux or Windows, detect cpuid features
     unsigned int cpuid_edx=0;
@@ -243,14 +244,16 @@ void scrypt_detect_sse2()
     if (cpuid_edx & 1<<26)
     {
         scrypt_1024_1_1_256_sp_detected = &scrypt_1024_1_1_256_sp_sse2;
-        printf("scrypt: using scrypt-sse2 as detected.\n");
+        fUsingSSE2 = true;
     }
     else
     {
         scrypt_1024_1_1_256_sp_detected = &scrypt_1024_1_1_256_sp_generic;
-        printf("scrypt: using scrypt-generic, SSE2 unavailable.\n");
+        fUsingSSE2 = false;
     }
 #endif // USE_SSE2_ALWAYS
+
+    return fUsingSSE2;
 }
 #endif
 

src/crypto/scrypt.h

@@ -1,8 +1,12 @@
-#ifndef SCRYPT_H
-#define SCRYPT_H
+#ifndef BITCOIN_CRYPTO_SCRYPT_H
+#define BITCOIN_CRYPTO_SCRYPT_H
 #include <stdlib.h>
 #include <stdint.h>
 
+#if defined(HAVE_CONFIG_H)
+#include "bitcoin-config.h" // for USE_SSE2
+#endif
+
 static const int SCRYPT_SCRATCHPAD_SIZE = 131072 + 63;
 
 void scrypt_1024_1_1_256(const char *input, char *output);
@@ -16,7 +20,7 @@ void scrypt_1024_1_1_256_sp_generic(const char *input, char *output, char *scrat
 #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();
+bool 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_detected)(const char *input, char *output, char *scratchpad);
 #else
@@ -44,4 +48,4 @@ static inline void le32enc(void *pp, uint32_t x)
         p[3] = (x >> 24) & 0xff;
 }
 #endif
-#endif
+#endif // BITCOIN_CRYPTO_SCRYPT_H

src/init.cpp

@@ -17,6 +17,7 @@
 #include "checkpoints.h"
 #include "compat/sanity.h"
 #include "consensus/validation.h"
+#include "crypto/scrypt.h" // for scrypt_detect_sse2
 #include "httpserver.h"
 #include "httprpc.h"
 #include "key.h"
@@ -1231,7 +1232,11 @@ bool AppInitMain(boost::thread_group& threadGroup, CScheduler& scheduler)
     int64_t nStart;
 
 #if defined(USE_SSE2)
-    scrypt_detect_sse2();
+    if (scrypt_detect_sse2()) {
+        LogPrintf("scrypt: using SSE2 implementation\n");
+    } else {
+        LogPrintf("scrypt: using generic implementation\n");
+    }
 #endif
 
     // ********************************************************* Step 5: verify wallet database integrity