Add ARM SHA512 asm implementation from Cryptogams (GH #841)

Filelist.txt

@@ -318,6 +318,8 @@ sha1_armv4.h
 sha1_armv4.S
 sha256_armv4.h
 sha256_armv4.S
+sha512_armv4.h
+sha512_armv4.S
 sha3.cpp
 sha3.h
 shacal2.cpp

GNUmakefile

@@ -1059,7 +1059,7 @@ endif
 ifeq ($(IS_ARM32),1)
   CRYPTOGAMS_ARCH_FLAG = -march=armv7-a
   CRYPTOGAMS_ARCH_FLAG += -Wa,--noexecstack
-  SRCS += aes_armv4.S sha1_armv4.S sha256_armv4.S
+  SRCS += aes_armv4.S sha1_armv4.S sha256_armv4.S sha512_armv4.S
 endif
 
 # List cryptlib.cpp first, then cpu.cpp, then integer.cpp to tame C++ static initialization problems.
@@ -1513,6 +1513,10 @@ sha1_armv4.o : sha1_armv4.S
 sha256_armv4.o : sha256_armv4.S
 	$(CC) $(strip $(CXXFLAGS) $(CRYPTOGAMS_ARCH_FLAG) -c) $<
 
+# Cryptogams ARM asm implementation.
+sha512_armv4.o : sha512_armv4.S
+	$(CC) $(strip $(CXXFLAGS) $(CRYPTOGAMS_ARCH_FLAG) -c) $<
+
 sha3_simd.o : sha3_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(SHA3_FLAG) -c) $<
 

GNUmakefile-cross

@@ -594,7 +594,7 @@ ifeq ($(IS_ARM32),1)
   ifneq ($(IS_IOS),1)
     CRYPTOGAMS_ARCH_FLAG = -march=armv7-a
     CRYPTOGAMS_ARCH_FLAG += -Wa,--noexecstack
-    SRCS += aes_armv4.S sha1_armv4.S sha256_armv4.S
+    SRCS += aes_armv4.S sha1_armv4.S sha256_armv4.S sha512_armv4.S
   endif
 endif
 
@@ -873,6 +873,10 @@ sha1_armv4.o : sha1_armv4.S
 sha256_armv4.o : sha256_armv4.S
 	$(CC) $(strip $(CXXFLAGS) $(CRYPTOGAMS_ARCH_FLAG) -c) $<
 
+# Cryptogams ARM asm implementation.
+sha512_armv4.o : sha512_armv4.S
+	$(CC) $(strip $(CXXFLAGS) $(CRYPTOGAMS_ARCH_FLAG) -c) $<
+
 # SSE4.2/SHA-NI or ARMv8a available
 shacal2_simd.o : shacal2_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(SHA_FLAG) -c) $<

config_asm.h

@@ -366,6 +366,7 @@
 #  define CRYPTOGAMS_ARM_AES      1
 #  define CRYPTOGAMS_ARM_SHA1     1
 #  define CRYPTOGAMS_ARM_SHA256   1
+#  define CRYPTOGAMS_ARM_SHA512   1
 # endif
 #endif
 

sha.cpp

@@ -83,6 +83,11 @@ extern void SHA256_HashMultipleBlocks_POWER8(word32 *state, const word32 *data,
 extern void SHA512_HashMultipleBlocks_POWER8(word64 *state, const word64 *data, size_t length, ByteOrder order);
 #endif
 
+#if (CRYPTOGAMS_ARM_SHA512)
+extern "C" unsigned int CRYPTOGAMS_armcaps;
+extern "C" int sha512_block_data_order(word64* state, const word64 *data, size_t blocks);
+#endif
+
 // We add extern to export table to sha_simd.cpp, but it
 //  cleared http://github.com/weidai11/cryptopp/issues/502
 extern const word32 SHA256_K[64];
@@ -997,6 +1002,12 @@ std::string SHA512_AlgorithmProvider()
     if (HasSSE2())
         return "SSE2";
 #endif
+#if CRYPTOGAMS_ARM_SHA512
+    if (HasNEON())
+        return "NEON";
+    if (HasARMv7())
+        return "ARMv7";
+#endif
 #if (CRYPTOPP_POWER8_SHA_AVAILABLE)
     if (HasSHA512())
         return "Power8";
@@ -1303,6 +1314,23 @@ void SHA512::Transform(word64 *state, const word64 *data)
         return;
     }
 #endif
+#if CRYPTOGAMS_ARM_SHA512
+    if (HasARMv7())
+    {
+        // The Cryptogams code uses a global variable named CRYPTOGAMS_armcaps
+        // for capabilities like ARMv7 and NEON. Storage is allocated in the
+        // module. We still need to set CRYPTOGAMS_armcaps accordingly.
+        // The Cryptogams code defines NEON as 1<<0; see ARMV7_NEON.
+        static const unsigned int unused = CRYPTOGAMS_armcaps = HasNEON() ? (1<<0) : 0;
+        CRYPTOPP_UNUSED(unused);
+
+        word64 dataBuf[16];
+        ByteReverse(dataBuf, data, SHA512::BLOCKSIZE);
+        sha512_block_data_order(state, dataBuf, 1);
+
+        return;
+    }
+#endif
 #if CRYPTOPP_POWER8_SHA_AVAILABLE
     if (HasSHA512())
     {

sha512_armv4.h

@@ -0,0 +1,21 @@
+/* Header file for use with Cryptogam's ARMv4 SHA512.  */
+/* Also see http://www.openssl.org/~appro/cryptogams/  */
+/* https://wiki.openssl.org/index.php/Cryptogams_SHA.  */
+
+#ifndef CRYPTOGAMS_SHA512_ARMV4_H
+#define CRYPTOGAMS_SHA512_ARMV4_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void sha512_block_data_order(void *state, const void *data, size_t blocks);
+
+/* Cryptogams arm caps */
+#define ARMV7_NEON (1<<0)
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  /* CRYPTOGAMS_SHA512_ARMV4_H */