12 files changed, 267 insertions, 0 deletions
diff --git a/src/Crypto/Crypto.vcxproj b/src/Crypto/Crypto.vcxproj
index c17bd607..cad50c06 100644
--- a/src/Crypto/Crypto.vcxproj
+++ b/src/Crypto/Crypto.vcxproj
@@ -238,40 +238,44 @@
     <ClCompile Include="chacha-xmm.c" />
     <ClCompile Include="chacha256.c" />
     <ClCompile Include="chachaRng.c" />
     <ClCompile Include="cpu.c" />
     <ClCompile Include="jitterentropy-base.c">
       <Optimization Condition="'$(Configuration)|$(Platform)'=='Release|x64'">Disabled</Optimization>
       <Optimization Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">Disabled</Optimization>
     </ClCompile>
     <ClCompile Include="kuznyechik.c" />
     <ClCompile Include="kuznyechik_simd.c">
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|ARM64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">true</ExcludedFromBuild>
     </ClCompile>
     <ClCompile Include="rdrand.c">
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|ARM64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">true</ExcludedFromBuild>
     </ClCompile>
     <ClCompile Include="SerpentFast.c" />
     <ClCompile Include="SerpentFast_simd.cpp" />
     <ClCompile Include="Sha2.c" />
+    <ClCompile Include="sha256_armv8.c">
+      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
+      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">true</ExcludedFromBuild>
+    </ClCompile>
     <ClCompile Include="Sha2Intel.c" />
     <ClCompile Include="Streebog.c" />
     <ClCompile Include="t1ha2.c" />
     <ClCompile Include="t1ha2_selfcheck.c" />
     <ClCompile Include="t1ha_selfcheck.c" />
     <ClCompile Include="Twofish.c" />
     <ClCompile Include="Whirlpool.c" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Aes.h" />
     <ClInclude Include="Aes_hw_cpu.h" />
     <ClInclude Include="Aesopt.h" />
     <ClInclude Include="Aestab.h" />
     <ClInclude Include="Camellia.h" />
     <ClInclude Include="chacha256.h" />
     <ClInclude Include="chachaRng.h" />
     <ClInclude Include="chacha_u1.h" />
     <ClInclude Include="chacha_u4.h" />
     <ClInclude Include="config.h" />
     <ClInclude Include="cpu.h" />
diff --git a/src/Crypto/Crypto.vcxproj.filters b/src/Crypto/Crypto.vcxproj.filters
index f2b1b54d..099c3ce4 100644
--- a/src/Crypto/Crypto.vcxproj.filters
+++ b/src/Crypto/Crypto.vcxproj.filters
@@ -79,40 +79,43 @@
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="blake2s_SSE2.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="blake2s_SSE41.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="blake2s_SSSE3.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="Sha2Intel.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="Aescrypt.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="Aes_hw_armv8.c">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="sha256_armv8.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Aes.h">
       <Filter>Header Files</Filter>
     </ClInclude>
     <ClInclude Include="Aes_hw_cpu.h">
       <Filter>Header Files</Filter>
     </ClInclude>
     <ClInclude Include="Aesopt.h">
       <Filter>Header Files</Filter>
     </ClInclude>
     <ClInclude Include="Aestab.h">
       <Filter>Header Files</Filter>
     </ClInclude>
     <ClInclude Include="Camellia.h">
       <Filter>Header Files</Filter>
     </ClInclude>
     <ClInclude Include="config.h">
       <Filter>Header Files</Filter>
     </ClInclude>
diff --git a/src/Crypto/Sha2.c b/src/Crypto/Sha2.c
index 5ae9cae2..27e61c3d 100644
--- a/src/Crypto/Sha2.c
+++ b/src/Crypto/Sha2.c
@@ -298,40 +298,44 @@ void sha512(unsigned char * result, const unsigned char* source, uint_64t source
 #ifndef NO_OPTIMIZED_VERSIONS
 
 #if defined(__cplusplus)
 extern "C"
 {
 #endif
 
 #if CRYPTOPP_BOOL_X64
 	void sha256_sse4(void *input_data, uint_32t digest[8], uint_64t num_blks);
 	void sha256_rorx(void *input_data, uint_32t digest[8], uint_64t num_blks);
 	void sha256_avx(void *input_data, uint_32t digest[8], uint_64t num_blks);
 #if CRYPTOPP_SHANI_AVAILABLE
 	void sha256_intel(void *input_data, uint_32t digest[8], uint_64t num_blks);
 #endif
 #endif
 
 #if CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32
 	void VC_CDECL sha256_compress_nayuki(uint_32t state[8], const uint_8t block[64]);
 #endif
 
+#if CRYPTOPP_ARM_SHA2_AVAILABLE
+	void sha256_compress_digest_armv8(const void* input_data, uint_32t digest[8], uint_64t num_blks);
+#endif
+
 #if defined(__cplusplus)
 }
 #endif
 
 #endif
 
 CRYPTOPP_ALIGN_DATA(16) static const uint_32t SHA256_K[64] CRYPTOPP_SECTION_ALIGN16 = {
 		0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5, 0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
 		0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3, 0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
 		0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc, 0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
 		0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7, 0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
 		0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13, 0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
 		0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3, 0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
 		0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5, 0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
 		0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208, 0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
 	};
 
 #if (defined(CRYPTOPP_X86_ASM_AVAILABLE) || defined(CRYPTOPP_X32_ASM_AVAILABLE))
 
 #ifdef _MSC_VER
@@ -740,40 +744,47 @@ void AvxSha256Transform(sha256_ctx* ctx, void* mp, uint_64t num_blks)
 	if (num_blks > 1)
 		sha256_avx(mp, ctx->hash, num_blks);
 	else
 		sha256_sse4(mp, ctx->hash, num_blks);
 }
 
 void SSE4Sha256Transform(sha256_ctx* ctx, void* mp, uint_64t num_blks)
 {
 	sha256_sse4(mp, ctx->hash, num_blks);
 }
 
 #endif
 
 #if (defined(CRYPTOPP_X86_ASM_AVAILABLE) || defined(CRYPTOPP_X32_ASM_AVAILABLE))
 void SSE2Sha256Transform(sha256_ctx* ctx, void* mp, uint_64t num_blks)
 {
 	X86_SHA256_HashBlocks(ctx->hash, (const uint_32t*)mp, (size_t)(num_blks * 64));
 }
 #endif
 
+#if CRYPTOPP_ARM_SHA2_AVAILABLE
+void ArmSha256Transform(sha256_ctx* ctx, void* mp, uint_64t num_blks)
+{
+	sha256_compress_digest_armv8(mp, ctx->hash, num_blks);
+}
+#endif
+
 #if CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32
 void Sha256AsmTransform(sha256_ctx* ctx, void* mp, uint_64t num_blks)
 {
 	uint_64t i;
 	for (i = 0; i < num_blks; i++)
 		sha256_compress_nayuki(ctx->hash, (uint_8t*)mp + i * 64);
 }
 #endif
 
 #endif
 
 void sha256_begin(sha256_ctx* ctx)
 {
 	ctx->hash[0] = 0x6a09e667;
 	ctx->hash[1] = 0xbb67ae85;
 	ctx->hash[2] = 0x3c6ef372;
 	ctx->hash[3] = 0xa54ff53a;
 	ctx->hash[4] = 0x510e527f;
 	ctx->hash[5] = 0x9b05688c;
 	ctx->hash[6] = 0x1f83d9ab;
@@ -788,40 +799,46 @@ void sha256_begin(sha256_ctx* ctx)
 #if CRYPTOPP_SHANI_AVAILABLE
 		if (HasSHA256())
 			sha256transfunc = IntelSha256Transform;
 		else
 #endif
 		if (g_isIntel && HasSAVX2() && HasSBMI2())
 			sha256transfunc = Avx2Sha256Transform;
 		else if (g_isIntel && HasSAVX())
 				sha256transfunc = AvxSha256Transform;
 		else if (HasSSE41())
 				sha256transfunc = SSE4Sha256Transform;
 		else
 #endif
 
 #if (defined(CRYPTOPP_X86_ASM_AVAILABLE) || defined(CRYPTOPP_X32_ASM_AVAILABLE))
 		if (HasSSE2 ())
 			sha256transfunc = SSE2Sha256Transform;
 		else
 #endif
 
+#if CRYPTOPP_ARM_SHA2_AVAILABLE
+		if (HasSHA256())
+			sha256transfunc = ArmSha256Transform;
+		else
+#endif
+
 #if CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32
 			sha256transfunc = Sha256AsmTransform;
 #else
 			sha256transfunc = StdSha256Transform;
 #endif
 #else
 		sha256transfunc = StdSha256Transform;
 #endif
 	}
 }
 
 void sha256_end(unsigned char * result, sha256_ctx* ctx)
 {
 	int i;
 	uint_64t mlen, pos = ctx->count[0];
 	uint_8t* m = (uint_8t*) ctx->wbuf;
 	m[pos++] = 0x80;
 	if (pos > 56)
 	{
 		memset(m + pos, 0, (size_t) (64 - pos));
diff --git a/src/Crypto/config.h b/src/Crypto/config.h
index dd8e3f06..f485c07e 100644
--- a/src/Crypto/config.h
+++ b/src/Crypto/config.h
@@ -223,40 +223,53 @@
       (CRYPTOPP_GCC_VERSION >= 40800) || (CRYPTOPP_LLVM_CLANG_VERSION >= 30300) || \
       (CRYPTOPP_APPLE_CLANG_VERSION >= 40000) || (CRYPTOPP_MSC_VERSION >= 1916)
 #   define CRYPTOPP_ARM_NEON_AVAILABLE 1
 #   define CRYPTOPP_ARM_ASIMD_AVAILABLE 1
 #  endif  // Compilers
 # endif  // Platforms
 #endif
 
 // ARMv8 and AES. -march=armv8-a+crypto or above must be present
 // Requires GCC 4.8, Clang 3.3 or Visual Studio 2017
 #if !defined(CRYPTOPP_ARM_AES_AVAILABLE) && !defined(CRYPTOPP_DISABLE_ARM_AES)
 # if defined(__aarch32__) || defined(__aarch64__) || defined(_M_ARM64)
 #  if defined(__ARM_FEATURE_CRYPTO) || (CRYPTOPP_GCC_VERSION >= 40800) || \
       (CRYPTOPP_LLVM_CLANG_VERSION >= 30300) || (CRYPTOPP_APPLE_CLANG_VERSION >= 40300) || \
       (CRYPTOPP_MSC_VERSION >= 1916)
 #   define CRYPTOPP_ARM_AES_AVAILABLE 1
 #  endif  // Compilers
 # endif  // Platforms
 #endif
 
+// ARMv8 and SHA-1, SHA-256. -march=armv8-a+crypto or above must be present
+// Requires GCC 4.8, Clang 3.3 or Visual Studio 2017
+#if !defined(CRYPTOPP_ARM_SHA_AVAILABLE) && !defined(CRYPTOPP_DISABLE_ARM_SHA)
+# if defined(__aarch32__) || defined(__aarch64__) || defined(_M_ARM64)
+#  if defined(__ARM_FEATURE_CRYPTO) || (CRYPTOPP_GCC_VERSION >= 40800) || \
+      (CRYPTOPP_LLVM_CLANG_VERSION >= 30300) || (CRYPTOPP_APPLE_CLANG_VERSION >= 40300) || \
+      (CRYPTOPP_MSC_VERSION >= 1916)
+#   define CRYPTOPP_ARM_SHA1_AVAILABLE 1
+#   define CRYPTOPP_ARM_SHA2_AVAILABLE 1
+#  endif  // Compilers
+# endif  // Platforms
+#endif
+
 // Undo the ASM and Intrinsic related defines due to X32.
 #if CRYPTOPP_BOOL_X32
 # undef CRYPTOPP_BOOL_X64
 # undef CRYPTOPP_X64_ASM_AVAILABLE
 # undef CRYPTOPP_X64_MASM_AVAILABLE
 #endif
 
 #if !defined(CRYPTOPP_NO_UNALIGNED_DATA_ACCESS) && !defined(CRYPTOPP_ALLOW_UNALIGNED_DATA_ACCESS)
 #if (CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32 || defined(__powerpc__) || (__ARM_FEATURE_UNALIGNED >= 1))
     #define CRYPTOPP_ALLOW_UNALIGNED_DATA_ACCESS
 #endif
 #endif
 
 // this version of the macro is fastest on Pentium 3 and Pentium 4 with MSVC 6 SP5 w/ Processor Pack
 #define GETBYTE(x, y) (unsigned int)((unsigned char)((x)>>(8*(y))))
 // these may be faster on other CPUs/compilers
 // #define GETBYTE(x, y) (unsigned int)(((x)>>(8*(y)))&255)
 // #define GETBYTE(x, y) (((uint8 *)&(x))[y])
 
 #define CRYPTOPP_GET_BYTE_AS_BYTE(x, y) ((uint8)((x)>>(8*(y))))
diff --git a/src/Crypto/cpu.c b/src/Crypto/cpu.c
index 0f1ba54d..a5b5bb19 100644
--- a/src/Crypto/cpu.c
+++ b/src/Crypto/cpu.c
@@ -458,54 +458,83 @@ void DisableCPUExtendedFeatures ()
 	g_hasMMX = 0;
 	g_hasAVX = 0;
 	g_hasAVX2 = 0;
 	g_hasBMI2 = 0;
 	g_hasSSE42 = 0;
 	g_hasSSE41 = 0;
 	g_hasSSSE3 = 0;
 	g_hasAESNI = 0;
 	g_hasCLMUL = 0;
 	g_hasSHA256 = 0;
 }
 
 #endif
 
 #if CRYPTOPP_BOOL_ARMV8
 #if defined(__linux__) && defined(__aarch64__)
 #include <sys/auxv.h>
 #ifndef HWCAP_AES
 # define HWCAP_AES (1 << 3)
 #endif
+#ifndef HWCAP_SHA2
+# define HWCAP_SHA2 (1 << 6)
+#endif
 #endif
 
 volatile int g_hasAESARM = 0;
+volatile int g_hasSHA256ARM = 0;
 
 inline int CPU_QueryAES()
 {
 #if defined(CRYPTOPP_ARM_AES_AVAILABLE)
 #if defined(__linux__) && defined(__aarch64__)
 	if ((getauxval(AT_HWCAP) & HWCAP_AES) != 0)
 		return 1;
 #elif defined(__APPLE__) && defined(__aarch64__)
 	// Apple Sillcon (M1) and later
 	return 1;
 #elif defined(_WIN32) && defined(_M_ARM64)
 #ifdef TC_WINDOWS_DRIVER
 	if (ExIsProcessorFeaturePresent(PF_ARM_V8_CRYPTO_INSTRUCTIONS_AVAILABLE) != 0)
 		return 1;
 #else
 	if (IsProcessorFeaturePresent(PF_ARM_V8_CRYPTO_INSTRUCTIONS_AVAILABLE) != 0)
 		return 1;
 #endif
 #endif
 	return 0;
 #else
 	return 0;
 #endif
 }
 
+inline int CPU_QuerySHA2()
+{
+#if defined(CRYPTOPP_ARM_SHA2_AVAILABLE)
+#if defined(__linux__) && defined(__aarch64__)
+	if ((getauxval(AT_HWCAP) & HWCAP_SHA2) != 0)
+		return 1;
+#elif defined(__APPLE__) && defined(__aarch64__)
+	// Apple Sillcon (M1) and later
+	return 1;
+#elif defined(_WIN32) && defined(_M_ARM64)
+#ifdef TC_WINDOWS_DRIVER
+	if (ExIsProcessorFeaturePresent(PF_ARM_V8_CRYPTO_INSTRUCTIONS_AVAILABLE) != 0)
+		return 1;
+#else
+	if (IsProcessorFeaturePresent(PF_ARM_V8_CRYPTO_INSTRUCTIONS_AVAILABLE) != 0)
+		return 1;
+#endif
+#endif
+	return 0;
+#else
+	return 0;
+#endif
+}
+
 void DetectArmFeatures()
 {
 	g_hasAESARM  = CPU_QueryAES();
+	g_hasSHA256ARM = CPU_QuerySHA2();
 }
 
 #endif
 \ No newline at end of file
diff --git a/src/Crypto/cpu.h b/src/Crypto/cpu.h
index 761258f2..cb34ad1f 100644
--- a/src/Crypto/cpu.h
+++ b/src/Crypto/cpu.h
@@ -281,43 +281,45 @@ void DisableCPUExtendedFeatures ();
 #define HasRDSEED() g_hasRDSEED
 #define HasSHA256() g_hasSHA256
 #define IsCpuIntel() g_isIntel
 #define IsCpuAMD() g_isAMD
 #define GetCacheLineSize() g_cacheLineSize
 
 #if defined(__cplusplus)
 }
 #endif
 
 #elif CRYPTOPP_BOOL_ARMV8
 #if defined(__cplusplus)
 extern "C" {
 #endif
 
 #if !defined(CRYPTOPP_DISABLE_AESNI) && !defined(WOLFCRYPT_BACKEND)
 #define TC_AES_HW_CPU
 #endif
 
 extern volatile int g_hasAESARM;
+extern volatile int g_hasSHA256ARM;
 void DetectArmFeatures();
 
 #define HasAESNI() g_hasAESARM
+#define HasSHA256() g_hasSHA256ARM
 
 #if defined(__cplusplus)
 }
 #endif
 
 #else
 
 #define HasSSE2()	0
 #define HasISSE()	0
 
 #define HasMMX()	0
 #define HasSSE42() 0
 #define HasSSE41() 0
 #define HasSAVX() 0
 #define HasSAVX2() 0
 #define HasSBMI2() 0
 #define HasSSSE3() 0
 #define HasAESNI() 0
 #define HasCLMUL() 0
 #define IsP4() 0
diff --git a/src/Crypto/sha256_armv8.c b/src/Crypto/sha256_armv8.c
new file mode 100644
index 00000000..1599350a
--- /dev/null
+++ b/src/Crypto/sha256_armv8.c
@@ -0,0 +1,184 @@
+/*
+* SHA-256 using CPU instructions in ARMv8
+*
+* Contributed by Jeffrey Walton. Based on public domain code by
+* Johannes Schneiders, Skip Hovsmith and Barry O'Rourke.
+*
+* Further changes (C) 2020 Jack Lloyd
+*
+* Botan is released under the Simplified BSD License (see license.txt)
+*/
+
+/* Modified and adapted for VeraCrypt */
+
+#include "Common/Tcdefs.h"
+#if !defined(_UEFI)
+#include <memory.h>
+#include <stdlib.h>
+#endif
+#include "cpu.h"
+#include "misc.h"
+
+#if CRYPTOPP_ARM_SHA2_AVAILABLE
+
+#include <arm_neon.h>
+
+CRYPTOPP_ALIGN_DATA(64) static const uint32 K[] = {
+    0x428A2F98, 0x71374491, 0xB5C0FBCF, 0xE9B5DBA5, 0x3956C25B, 0x59F111F1, 0x923F82A4, 0xAB1C5ED5,
+    0xD807AA98, 0x12835B01, 0x243185BE, 0x550C7DC3, 0x72BE5D74, 0x80DEB1FE, 0x9BDC06A7, 0xC19BF174,
+    0xE49B69C1, 0xEFBE4786, 0x0FC19DC6, 0x240CA1CC, 0x2DE92C6F, 0x4A7484AA, 0x5CB0A9DC, 0x76F988DA,
+    0x983E5152, 0xA831C66D, 0xB00327C8, 0xBF597FC7, 0xC6E00BF3, 0xD5A79147, 0x06CA6351, 0x14292967,
+    0x27B70A85, 0x2E1B2138, 0x4D2C6DFC, 0x53380D13, 0x650A7354, 0x766A0ABB, 0x81C2C92E, 0x92722C85,
+    0xA2BFE8A1, 0xA81A664B, 0xC24B8B70, 0xC76C51A3, 0xD192E819, 0xD6990624, 0xF40E3585, 0x106AA070,
+    0x19A4C116, 0x1E376C08, 0x2748774C, 0x34B0BCB5, 0x391C0CB3, 0x4ED8AA4A, 0x5B9CCA4F, 0x682E6FF3,
+    0x748F82EE, 0x78A5636F, 0x84C87814, 0x8CC70208, 0x90BEFFFA, 0xA4506CEB, 0xBEF9A3F7, 0xC67178F2,
+};
+
+void sha256_compress_digest_armv8(void* input_data, uint32 digest[8], uint64 num_blks) {
+
+
+    // Load initial values
+    uint32x4_t STATE0 = vld1q_u32(&digest[0]);
+    uint32x4_t STATE1 = vld1q_u32(&digest[4]);
+
+    // Intermediate void* cast due to https://llvm.org/bugs/show_bug.cgi?id=20670
+    const uint32* input32 = (const uint32*)(const void*)input_data;
+
+    while (num_blks > 0) {
+        // Save current state
+        const uint32x4_t ABCD_SAVE = STATE0;
+        const uint32x4_t EFGH_SAVE = STATE1;
+
+        uint32x4_t MSG0 = vld1q_u32(input32 + 0);
+        uint32x4_t MSG1 = vld1q_u32(input32 + 4);
+        uint32x4_t MSG2 = vld1q_u32(input32 + 8);
+        uint32x4_t MSG3 = vld1q_u32(input32 + 12);
+
+        MSG0 = vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(MSG0)));
+        MSG1 = vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(MSG1)));
+        MSG2 = vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(MSG2)));
+        MSG3 = vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(MSG3)));
+
+        uint32x4_t MSG_K, TSTATE;
+
+        // Rounds 0-3
+        MSG_K = vaddq_u32(MSG0, vld1q_u32(&K[4 * 0]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG0 = vsha256su1q_u32(vsha256su0q_u32(MSG0, MSG1), MSG2, MSG3);
+
+        // Rounds 4-7
+        MSG_K = vaddq_u32(MSG1, vld1q_u32(&K[4 * 1]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG1 = vsha256su1q_u32(vsha256su0q_u32(MSG1, MSG2), MSG3, MSG0);
+
+        // Rounds 8-11
+        MSG_K = vaddq_u32(MSG2, vld1q_u32(&K[4 * 2]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG2 = vsha256su1q_u32(vsha256su0q_u32(MSG2, MSG3), MSG0, MSG1);
+
+        // Rounds 12-15
+        MSG_K = vaddq_u32(MSG3, vld1q_u32(&K[4 * 3]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG3 = vsha256su1q_u32(vsha256su0q_u32(MSG3, MSG0), MSG1, MSG2);
+
+        // Rounds 16-19
+        MSG_K = vaddq_u32(MSG0, vld1q_u32(&K[4 * 4]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG0 = vsha256su1q_u32(vsha256su0q_u32(MSG0, MSG1), MSG2, MSG3);
+
+        // Rounds 20-23
+        MSG_K = vaddq_u32(MSG1, vld1q_u32(&K[4 * 5]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG1 = vsha256su1q_u32(vsha256su0q_u32(MSG1, MSG2), MSG3, MSG0);
+
+        // Rounds 24-27
+        MSG_K = vaddq_u32(MSG2, vld1q_u32(&K[4 * 6]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG2 = vsha256su1q_u32(vsha256su0q_u32(MSG2, MSG3), MSG0, MSG1);
+
+        // Rounds 28-31
+        MSG_K = vaddq_u32(MSG3, vld1q_u32(&K[4 * 7]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG3 = vsha256su1q_u32(vsha256su0q_u32(MSG3, MSG0), MSG1, MSG2);
+
+        // Rounds 32-35
+        MSG_K = vaddq_u32(MSG0, vld1q_u32(&K[4 * 8]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG0 = vsha256su1q_u32(vsha256su0q_u32(MSG0, MSG1), MSG2, MSG3);
+
+        // Rounds 36-39
+        MSG_K = vaddq_u32(MSG1, vld1q_u32(&K[4 * 9]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG1 = vsha256su1q_u32(vsha256su0q_u32(MSG1, MSG2), MSG3, MSG0);
+
+        // Rounds 40-43
+        MSG_K = vaddq_u32(MSG2, vld1q_u32(&K[4 * 10]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG2 = vsha256su1q_u32(vsha256su0q_u32(MSG2, MSG3), MSG0, MSG1);
+
+        // Rounds 44-47
+        MSG_K = vaddq_u32(MSG3, vld1q_u32(&K[4 * 11]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+        MSG3 = vsha256su1q_u32(vsha256su0q_u32(MSG3, MSG0), MSG1, MSG2);
+
+        // Rounds 48-51
+        MSG_K = vaddq_u32(MSG0, vld1q_u32(&K[4 * 12]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+
+        // Rounds 52-55
+        MSG_K = vaddq_u32(MSG1, vld1q_u32(&K[4 * 13]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+
+        // Rounds 56-59
+        MSG_K = vaddq_u32(MSG2, vld1q_u32(&K[4 * 14]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+
+        // Rounds 60-63
+        MSG_K = vaddq_u32(MSG3, vld1q_u32(&K[4 * 15]));
+        TSTATE = vsha256hq_u32(STATE0, STATE1, MSG_K);
+        STATE1 = vsha256h2q_u32(STATE1, STATE0, MSG_K);
+        STATE0 = TSTATE;
+
+        // Add back to state
+        STATE0 = vaddq_u32(STATE0, ABCD_SAVE);
+        STATE1 = vaddq_u32(STATE1, EFGH_SAVE);
+
+        input32 += 64 / 4;
+        num_blks--;
+    }
+
+    // Save state
+    vst1q_u32(&digest[0], STATE0);
+    vst1q_u32(&digest[4], STATE1);
+}
+#endif
diff --git a/src/Driver/Driver.vcxproj b/src/Driver/Driver.vcxproj
index f6aacf9a..76ca90ed 100644
--- a/src/Driver/Driver.vcxproj
+++ b/src/Driver/Driver.vcxproj
@@ -255,40 +255,44 @@ copy $(OutDir)veracrypt.inf "$(SolutionDir)Debug\Setup Files\veracrypt.inf"</Com
     <ClCompile Include="..\Crypto\kuznyechik_simd.c">
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|ARM64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">false</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">false</ExcludedFromBuild>
     </ClCompile>
     <ClCompile Include="..\Crypto\rdrand.c">
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|ARM64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">false</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">false</ExcludedFromBuild>
     </ClCompile>
     <ClCompile Include="..\Crypto\SerpentFast.c" />
     <ClCompile Include="..\Crypto\SerpentFast_simd.cpp">
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|ARM64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">false</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">false</ExcludedFromBuild>
     </ClCompile>
     <ClCompile Include="..\Crypto\Sha2.c" />
+    <ClCompile Include="..\Crypto\sha256_armv8.c">
+      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
+      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">true</ExcludedFromBuild>
+    </ClCompile>
     <ClCompile Include="..\Crypto\Sha2Intel.c" />
     <ClCompile Include="..\Crypto\Streebog.c" />
     <ClCompile Include="..\Crypto\t1ha2.c" />
     <ClCompile Include="..\Crypto\t1ha2_selfcheck.c" />
     <ClCompile Include="..\Crypto\t1ha_selfcheck.c" />
     <ClCompile Include="..\Crypto\Twofish.c" />
     <ClCompile Include="..\Crypto\Whirlpool.c" />
     <ClCompile Include="..\Driver\DriveFilter.c" />
     <ClCompile Include="..\Driver\DumpFilter.c" />
     <ClCompile Include="..\Driver\EncryptedIoQueue.c" />
     <ClCompile Include="..\Driver\Ntdriver.c" />
     <ClCompile Include="..\Driver\Ntvol.c" />
     <ClCompile Include="..\Driver\VolumeFilter.c" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\Common\Tcdefs.h" />
     <ClInclude Include="..\Crypto\Aes.h" />
     <ClInclude Include="..\Crypto\Aesopt.h" />
     <ClInclude Include="..\Crypto\AesSmall.h" />
     <ClInclude Include="..\Crypto\Aestab.h" />
diff --git a/src/Driver/Driver.vcxproj.filters b/src/Driver/Driver.vcxproj.filters
index 6f43b0e8..3c427dad 100644
--- a/src/Driver/Driver.vcxproj.filters
+++ b/src/Driver/Driver.vcxproj.filters
@@ -151,40 +151,43 @@
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="..\Driver\DumpFilter.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="..\Driver\EncryptedIoQueue.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="..\Driver\Ntdriver.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="..\Driver\Ntvol.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="..\Driver\VolumeFilter.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="..\Crypto\Aes_hw_armv8.c">
       <Filter>Crypto\Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="..\Crypto\sha256_armv8.c">
+      <Filter>Crypto\Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\Common\Tcdefs.h">
       <Filter>Common</Filter>
     </ClInclude>
     <ClInclude Include="..\Crypto\Aes.h">
       <Filter>Crypto\Header Files</Filter>
     </ClInclude>
     <ClInclude Include="..\Crypto\Aesopt.h">
       <Filter>Crypto\Header Files</Filter>
     </ClInclude>
     <ClInclude Include="..\Crypto\AesSmall.h">
       <Filter>Crypto\Header Files</Filter>
     </ClInclude>
     <ClInclude Include="..\Crypto\Aestab.h">
       <Filter>Crypto\Header Files</Filter>
     </ClInclude>
     <ClInclude Include="..\Crypto\Aes_hw_cpu.h">
       <Filter>Crypto\Header Files</Filter>
     </ClInclude>
diff --git a/src/Release/Setup Files/veracrypt-arm64.sys b/src/Release/Setup Files/veracrypt-arm64.sys
index 8925f812..707fa0d6 100644
--- a/src/Release/Setup Files/veracrypt-arm64.sys
+++ b/src/Release/Setup Files/veracrypt-arm64.sys
diff --git a/src/Release/Setup Files/veracrypt-x64.sys b/src/Release/Setup Files/veracrypt-x64.sys
index 42e677f7..ae31c516 100644
--- a/src/Release/Setup Files/veracrypt-x64.sys
+++ b/src/Release/Setup Files/veracrypt-x64.sys
diff --git a/src/Volume/Volume.make b/src/Volume/Volume.make
index e38542bb..a4f62562 100644
--- a/src/Volume/Volume.make
+++ b/src/Volume/Volume.make
@@ -33,72 +33,74 @@ OBJS += VolumePassword.o
 OBJS += VolumePasswordCache.o
 
 ifeq "$(ENABLE_WOLFCRYPT)" "0"
 OBJS += EncryptionModeXTS.o
 else
 OBJS += EncryptionModeWolfCryptXTS.o
 endif
 
 ifeq "$(ENABLE_WOLFCRYPT)" "0"
 ifeq "$(PLATFORM)" "MacOSX"
 ifneq "$(COMPILE_ASM)" "false"
 	OBJSEX += ../Crypto/Aes_asm.oo
 	OBJS += ../Crypto/Aes_hw_cpu.o
 	OBJSEX += ../Crypto/Aes_hw_armv8.oo
 	OBJS += ../Crypto/Aescrypt.o
 	OBJSEX += ../Crypto/Twofish_asm.oo
 	OBJSEX += ../Crypto/Camellia_asm.oo
 	OBJSEX += ../Crypto/Camellia_aesni_asm.oo
 	OBJSEX += ../Crypto/sha256-nayuki.oo
 	OBJSEX += ../Crypto/sha512-nayuki.oo
+	OBJSEX += ../Crypto/sha256_armv8.oo
 	OBJSEX += ../Crypto/sha256_avx1.oo
 	OBJSEX += ../Crypto/sha256_avx2.oo
 	OBJSEX += ../Crypto/sha256_sse4.oo
 	OBJSEX += ../Crypto/sha512_avx1.oo
 	OBJSEX += ../Crypto/sha512_avx2.oo
 	OBJSEX += ../Crypto/sha512_sse4.oo
 endif
 else ifeq "$(CPU_ARCH)" "x86"
 	OBJS += ../Crypto/Aes_x86.o
 	ifeq "$(DISABLE_AESNI)" "0"
 		OBJS += ../Crypto/Aes_hw_cpu.o
 	endif
 	OBJS += ../Crypto/sha256-x86-nayuki.o
 	OBJS += ../Crypto/sha512-x86-nayuki.o
 else ifeq "$(CPU_ARCH)" "x64"
 	OBJS += ../Crypto/Aes_x64.o
 	ifeq "$(DISABLE_AESNI)" "0"
 		OBJS += ../Crypto/Aes_hw_cpu.o
 	endif
 	OBJS += ../Crypto/Twofish_x64.o
 	OBJS += ../Crypto/Camellia_x64.o
 	OBJS += ../Crypto/Camellia_aesni_x64.o
 	OBJS += ../Crypto/sha512-x64-nayuki.o
 	OBJS += ../Crypto/sha256_avx1_x64.o
 	OBJS += ../Crypto/sha256_avx2_x64.o
 	OBJS += ../Crypto/sha256_sse4_x64.o
 	OBJS += ../Crypto/sha512_avx1_x64.o
 	OBJS += ../Crypto/sha512_avx2_x64.o
 	OBJS += ../Crypto/sha512_sse4_x64.o
 else ifeq "$(CPU_ARCH)" "arm64"
 	OBJARMV8CRYPTO += ../Crypto/Aes_hw_armv8.oarmv8crypto
 	OBJS += ../Crypto/Aescrypt.o
+	OBJARMV8CRYPTO += ../Crypto/sha256_armv8.oarmv8crypto
 else
 	OBJS += ../Crypto/Aescrypt.o
 endif
 
 ifeq "$(GCC_GTEQ_430)" "1"
 	OBJSSSE41 += ../Crypto/blake2s_SSE41.osse41
 	OBJSSSSE3 += ../Crypto/blake2s_SSSE3.ossse3
 else
 	OBJS += ../Crypto/blake2s_SSE41.o
 	OBJS += ../Crypto/blake2s_SSSE3.o
 endif
 ifeq "$(GCC_GTEQ_500)" "1"
 	OBJSHANI += ../Crypto/Sha2Intel.oshani
 else
 	OBJS += ../Crypto/Sha2Intel.o
 endif
 else
 OBJS += ../Crypto/wolfCrypt.o
 endif
 
@@ -133,40 +135,46 @@ OBJS += ../Common/SCardReader.o
 OBJS += ../Common/Token.o
 OBJS += ../Common/Crc.o
 OBJS += ../Common/TLVParser.o
 OBJS += ../Common/EMVCard.o
 OBJS += ../Common/EMVToken.o
 OBJS += ../Common/Endian.o
 OBJS += ../Common/GfMul.o
 OBJS += ../Common/SecurityToken.o
 
 VolumeLibrary: Volume.a
 
 ifeq "$(ENABLE_WOLFCRYPT)" "0"
 ifeq "$(PLATFORM)" "MacOSX"
 ifneq "$(COMPILE_ASM)" "false"
 ../Crypto/Aes_hw_armv8.oo: ../Crypto/Aes_hw_armv8.c
 	@echo Compiling $(<F)
 	$(CC) $(CFLAGS_ARM64) -c ../Crypto/Aes_hw_armv8.c -o ../Crypto/Aes_hw_armv8_arm64.o
 	$(CC) $(CFLAGS_X64) -c ../Crypto/Aes_hw_armv8.c -o ../Crypto/Aes_hw_armv8_x64.o
 	lipo -create ../Crypto/Aes_hw_armv8_arm64.o ../Crypto/Aes_hw_armv8_x64.o -output ../Crypto/Aes_hw_armv8.oo
 	rm -fr ../Crypto/Aes_hw_armv8_arm64.o ../Crypto/Aes_hw_armv8_x64.o
+../Crypto/sha256_armv8.oo: ../Crypto/sha256_armv8.c
+	@echo Compiling $(<F)
+	$(CC) $(CFLAGS_ARM64) -c ../Crypto/sha256_armv8.c -o ../Crypto/sha256_armv8_arm64.o
+	$(CC) $(CFLAGS_X64) -c ../Crypto/sha256_armv8.c -o ../Crypto/sha256_armv8_x64.o
+	lipo -create ../Crypto/sha256_armv8_arm64.o ../Crypto/sha256_armv8_x64.o -output ../Crypto/sha256_armv8.oo
+	rm -fr ../Crypto/sha256_armv8_arm64.o ../Crypto/sha256_armv8_x64.o
 ../Crypto/Aes_asm.oo: ../Crypto/Aes_x86.asm ../Crypto/Aes_x64.asm
 	@echo Assembling $(<F)
 	$(AS) $(ASFLAGS32) -o ../Crypto/Aes_x86.o ../Crypto/Aes_x86.asm
 	$(AS) $(ASFLAGS64) -o ../Crypto/Aes_x64.o ../Crypto/Aes_x64.asm
 	lipo -create ../Crypto/Aes_x86.o ../Crypto/Aes_x64.o -output ../Crypto/Aes_asm.oo
 	rm -fr ../Crypto/Aes_x86.o ../Crypto/Aes_x64.o
 ../Crypto/Twofish_asm.oo: ../Crypto/Twofish_x64.S
 	@echo Assembling $(<F)
 	$(AS) $(ASFLAGS64) -p gas -o ../Crypto/Twofish_asm.oo ../Crypto/Twofish_x64.S
 ../Crypto/Camellia_asm.oo: ../Crypto/Camellia_x64.S
 	@echo Assembling $(<F)
 	$(AS) $(ASFLAGS64) -p gas -o ../Crypto/Camellia_asm.oo ../Crypto/Camellia_x64.S
 ../Crypto/Camellia_aesni_asm.oo: ../Crypto/Camellia_aesni_x64.S
 	@echo Assembling $(<F)
 	$(AS) $(ASFLAGS64) -p gas -o ../Crypto/Camellia_aesni_asm.oo ../Crypto/Camellia_aesni_x64.S
 ../Crypto/sha256-nayuki.oo: ../Crypto/sha256-x86-nayuki.S
 	@echo Assembling $(<F)
 	$(AS) $(ASFLAGS32) -p gas -o ../Crypto/sha256-x86-nayuki.o ../Crypto/sha256-x86-nayuki.S
 	$(AS) $(ASFLAGS64) -p gas -o ../Crypto/sha256-x64-nayuki.o ../Crypto/sha256-x64-nayuki.S
 	lipo -create ../Crypto/sha256-x86-nayuki.o ../Crypto/sha256-x64-nayuki.o -output ../Crypto/sha256-nayuki.oo