Add CRYPTOPP_BUGGY_SIMD_LOAD_AND_STORE (GH #782)

I also suspect this has something to do with GH #768 and several others

config.h

@@ -291,12 +291,26 @@ const lword LWORD_MAX = W64LIT(0xffffffffffffffff);
 	#define CRYPTOPP_GCC_DIAGNOSTIC_AVAILABLE 1
 #endif
 
-// Some Clang cannot handle mixed asm with positional arguments, where the
+// Some Clang and SunCC cannot handle mixed asm with positional arguments,
-// body is Intel style with no prefix and the templates are AT&T style.
+// where the body is Intel style with no prefix and the templates are
-// Define this is the Makefile misdetects the configuration.
+// AT&T style. Define this is the Makefile misdetects the configuration.
 // Also see https://bugs.llvm.org/show_bug.cgi?id=39895 .
 // #define CRYPTOPP_DISABLE_MIXED_ASM 1
 
+// Several compilers discard SIMD code that loads unaligned data. For example,
+// Power7 can load unaligned data using vec_vsx_ld but GCC and Clang require
+// 16-byte aligned arrays when using builtin function. Define this to disable
+// Crypto++ code that uses problematic builtins or intrinsics.
+// #define CRYPTOPP_BUGGY_SIMD_LOAD_AND_STORE 1
+
+// This list will probably grow over time as more compilers are identified.
+#if defined(CRYPTOPP_BUGGY_SIMD_LOAD_AND_STORE)
+# define CRYPTOPP_DISABLE_LEA_SIMD 1
+# define CRYPTOPP_DISABLE_SIMON_SIMD 1
+# define CRYPTOPP_DISABLE_SPECK_SIMD 1
+# define CRYPTOPP_DISABLE_SM4_SIMD 1
+#endif
+
 // define hword, word, and dword. these are used for multiprecision integer arithmetic
 // Intel compiler won't have _umul128 until version 10.0. See http://softwarecommunity.intel.com/isn/Community/en-US/forums/thread/30231625.aspx
 #if (defined(_MSC_VER) && (!defined(__INTEL_COMPILER) || __INTEL_COMPILER >= 1000) && (defined(_M_X64) || defined(_M_IA64))) || (defined(__DECCXX) && defined(__alpha__)) || (defined(__INTEL_COMPILER) && defined(__x86_64__)) || (defined(__SUNPRO_CC) && defined(__x86_64__))

lea.h

@@ -16,7 +16,9 @@
 #include "algparam.h"
 
 #if (CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_ARM32 || CRYPTOPP_BOOL_ARMV8)
-# define CRYPTOPP_LEA_ADVANCED_PROCESS_BLOCKS 1
+# ifndef CRYPTOPP_DISABLE_LEA_SIMD
+#  define CRYPTOPP_LEA_ADVANCED_PROCESS_BLOCKS 1
+# endif
 #endif
 
 // Yet another SunStudio/SunCC workaround. Failed self tests

simon.h

@@ -20,13 +20,17 @@
 #if CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X86 || \
     CRYPTOPP_BOOL_ARM32 || CRYPTOPP_BOOL_ARMV8 || \
     CRYPTOPP_BOOL_PPC32 || CRYPTOPP_BOOL_PPC64
-# define CRYPTOPP_SIMON64_ADVANCED_PROCESS_BLOCKS 1
+# ifndef CRYPTOPP_DISABLE_SIMON_SIMD
+#  define CRYPTOPP_SIMON64_ADVANCED_PROCESS_BLOCKS 1
+# endif
 #endif
 
 #if CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X86 || \
     CRYPTOPP_BOOL_ARM32 || CRYPTOPP_BOOL_ARMV8 || \
     CRYPTOPP_BOOL_PPC32 || CRYPTOPP_BOOL_PPC64
-# define CRYPTOPP_SIMON128_ADVANCED_PROCESS_BLOCKS 1
+# ifndef CRYPTOPP_DISABLE_SIMON_SIMD
+#  define CRYPTOPP_SIMON128_ADVANCED_PROCESS_BLOCKS 1
+# endif
 #endif
 
 // Yet another SunStudio/SunCC workaround. Failed self tests

sm4.h

@@ -19,7 +19,9 @@
 #include "secblock.h"
 
 #if (CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X86)
-# define CRYPTOPP_SM4_ADVANCED_PROCESS_BLOCKS 1
+# ifndef CRYPTOPP_DISABLE_SM4_SIMD
+#  define CRYPTOPP_SM4_ADVANCED_PROCESS_BLOCKS 1
+# endif
 #endif
 
 NAMESPACE_BEGIN(CryptoPP)

speck.h

@@ -20,13 +20,17 @@
 #if CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X86 || \
     CRYPTOPP_BOOL_ARM32 || CRYPTOPP_BOOL_ARMV8 || \
     CRYPTOPP_BOOL_PPC32 || CRYPTOPP_BOOL_PPC64
-# define CRYPTOPP_SPECK64_ADVANCED_PROCESS_BLOCKS 1
+# ifndef CRYPTOPP_DISABLE_SPECK_SIMD
+#  define CRYPTOPP_SPECK64_ADVANCED_PROCESS_BLOCKS 1
+# endif
 #endif
 
 #if CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X86 || \
     CRYPTOPP_BOOL_ARM32 || CRYPTOPP_BOOL_ARMV8 || \
     CRYPTOPP_BOOL_PPC32 || CRYPTOPP_BOOL_PPC64
-# define CRYPTOPP_SPECK128_ADVANCED_PROCESS_BLOCKS 1
+# ifndef CRYPTOPP_DISABLE_SPECK_SIMD
+#  define CRYPTOPP_SPECK128_ADVANCED_PROCESS_BLOCKS 1
+# endif
 #endif
 
 // Yet another SunStudio/SunCC workaround. Failed self tests