fix compile on Sun CC

8 files changed, 42 insertions, 28 deletions

diff --git a/GNUmakefile b/GNUmakefile
index f8390024..23d13436 100644
--- a/GNUmakefile
+++ b/GNUmakefile
@@ -48,10 +48,8 @@ endif
 ifeq ($(UNAME),)	# for DJGPP, where uname doesn't exist
 CXXFLAGS += -mbnu210
 else
-ifneq ($(CXX),CC)	# don't use -pipe with CC (Solaris native C++ compiler)
 CXXFLAGS += -pipe
 endif
-endif
 
 ifeq ($(UNAME),Linux)
 LDFLAGS += -pthread
@@ -72,6 +70,14 @@ endif
 
 ifeq ($(UNAME),SunOS)
 LDLIBS += -lnsl -lsocket
+ifeq ($(CXX),CC)	# override flags for CC (Solaris native C++ compiler)
+CXXFLAGS = -DNDEBUG -O -g -native
+LDFLAGS =
+ifeq ($(ISX86),1)
+# SSE2 intrinsics should work in Sun Studio 12
+# CXXFLAGS += -xarch=sse2 -D__SSE2__
+endif
+endif
 endif
 
 SRCS = $(wildcard *.cpp)
diff --git a/config.h b/config.h
index 82fd4951..05d77fff 100644
--- a/config.h
+++ b/config.h
@@ -186,8 +186,10 @@ NAMESPACE_END
 #ifndef CRYPTOPP_ALIGN_DATA
 	#if defined(CRYPTOPP_MSVC6PP_OR_LATER)
 		#define CRYPTOPP_ALIGN_DATA(x) __declspec(align(x))
-	#elif defined(__GNUC__)
+	#elif defined(__GNUC__) || __SUNPRO_CC > 0x580
 		#define CRYPTOPP_ALIGN_DATA(x) __attribute__((aligned(x)))
+	#else
+		#define CRYPTOPP_ALIGN_DATA(x)
 	#endif
 #endif
 
@@ -323,7 +325,8 @@ NAMESPACE_END
 	#define CRYPTOPP_BOOL_X64 0
 #endif
 
-#if defined(_M_IX86) || defined(__i386__)
+// see http://predef.sourceforge.net/prearch.html
+#if defined(_M_IX86) || defined(__i386__) || defined(__i386) || defined(_X86_) || defined(__I86__) || defined(__INTEL__)
 	#define CRYPTOPP_BOOL_X86 1
 #else
 	#define CRYPTOPP_BOOL_X86 0
diff --git a/cpu.cpp b/cpu.cpp
index 643a0ffc..a4922504 100755
--- a/cpu.cpp
+++ b/cpu.cpp
@@ -76,7 +76,7 @@ static void SigIllHandlerSSE2(int)
 }
 #endif
 
-#elif _MSC_VER >= 1400
+#elif _MSC_VER >= 1400 && CRYPTOPP_BOOL_X64
 
 bool CpuId(word32 input, word32 *output)
 {
diff --git a/cpu.h b/cpu.h
index d8a5a1f7..6a212345 100755
--- a/cpu.h
+++ b/cpu.h
@@ -5,7 +5,7 @@
 
 NAMESPACE_BEGIN(CryptoPP)
 
-#if defined(CRYPTOPP_X86_ASM_AVAILABLE) || _MSC_VER >= 1400
+#if defined(CRYPTOPP_X86_ASM_AVAILABLE) || (_MSC_VER >= 1400 && CRYPTOPP_BOOL_X64)
 
 #define CRYPTOPP_CPUID_AVAILABLE
 
@@ -17,7 +17,10 @@ void DetectX86Features();
 
 bool CpuId(word32 input, word32 *output);
 
-#if !CRYPTOPP_BOOL_X64
+#if CRYPTOPP_BOOL_X64
+inline bool HasSSE2()	{return true;}
+inline bool HasMMX()	{return true;}
+#else
 
 inline bool HasSSE2()
 {
@@ -63,22 +66,20 @@ inline int GetCacheLineSize()
 	return CRYPTOPP_L1_CACHE_LINE_SIZE;
 }
 
-#endif		// #ifdef CRYPTOPP_X86_ASM_AVAILABLE || _MSC_VER >= 1400
-
-#if CRYPTOPP_BOOL_X64
-
-inline bool HasSSE2()
-{
-	return true;
-}
-
-inline bool HasMMX()
-{
-	return true;
-}
+inline bool HasSSSE3()	{return false;}
+inline bool IsP4()		{return false;}
 
+// assume MMX and SSE2 if intrinsics are enabled
+#if CRYPTOPP_BOOL_SSE2_INTRINSICS_AVAILABLE || CRYPTOPP_BOOL_X64
+inline bool HasSSE2()	{return true;}
+inline bool HasMMX()	{return true;}
+#else
+inline bool HasSSE2()	{return false;}
+inline bool HasMMX()	{return false;}
 #endif
 
+#endif		// #ifdef CRYPTOPP_X86_ASM_AVAILABLE || _MSC_VER >= 1400
+
 #if defined(__GNUC__)
 	// define these in two steps to allow arguments to be expanded
 	#define GNU_AS1(x) #x ";"
diff --git a/integer.cpp b/integer.cpp
index 8909ec94..64f3cea0 100644
--- a/integer.cpp
+++ b/integer.cpp
@@ -692,6 +692,7 @@ int CRYPTOPP_FASTCALL Baseline_Add(size_t N, word *C, const word *A, const word
 	assert (N%2 == 0);
 
 	Declare2Words(u);
+	AssignWord(u, 0);
 	for (size_t i=0; i<N; i+=2)
 	{
 		AddWithCarry(u, A[i], B[i]);
@@ -707,6 +708,7 @@ int CRYPTOPP_FASTCALL Baseline_Sub(size_t N, word *C, const word *A, const word
 	assert (N%2 == 0);
 
 	Declare2Words(u);
+	AssignWord(u, 0);
 	for (size_t i=0; i<N; i+=2)
 	{
 		SubtractWithBorrow(u, A[i], B[i]);
diff --git a/misc.h b/misc.h
index c898040f..f0f7bd20 100644
--- a/misc.h
+++ b/misc.h
@@ -24,7 +24,7 @@
 	#define CRYPTOPP_FAST_ROTATE(x) ((x) == 32)
 #elif defined(__GNUC__) && (CRYPTOPP_BOOL_X64 || CRYPTOPP_BOOL_X86)	// depend on GCC's peephole optimization to generate rotate instructions
 	#define CRYPTOPP_FAST_ROTATE(x) 1
-#elif
+#else
 	#define CRYPTOPP_FAST_ROTATE(x) 0
 #endif
 
@@ -795,19 +795,19 @@ inline void GetUserKey(ByteOrder order, T *out, size_t outlen, const byte *in, s
 }
 
 #ifndef CRYPTOPP_ALLOW_UNALIGNED_DATA_ACCESS
-inline byte UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, byte*)
+inline byte UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, const byte *)
 {
 	return block[0];
 }
 
-inline word16 UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, word16*)
+inline word16 UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, const word16 *)
 {
 	return (order == BIG_ENDIAN_ORDER)
 		? block[1] | (block[0] << 8)
 		: block[0] | (block[1] << 8);
 }
 
-inline word32 UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, word32*)
+inline word32 UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, const word32 *)
 {
 	return (order == BIG_ENDIAN_ORDER)
 		? word32(block[3]) | (word32(block[2]) << 8) | (word32(block[1]) << 16) | (word32(block[0]) << 24)
@@ -815,7 +815,7 @@ inline word32 UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, wo
 }
 
 #ifdef WORD64_AVAILABLE
-inline word64 UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, word64*)
+inline word64 UnalignedGetWordNonTemplate(ByteOrder order, const byte *block, const word64 *)
 {
 	return (order == BIG_ENDIAN_ORDER)
 		?
@@ -936,7 +936,7 @@ inline T GetWord(bool assumeAligned, ByteOrder order, const byte *block)
 {
 #ifndef CRYPTOPP_ALLOW_UNALIGNED_DATA_ACCESS
 	if (!assumeAligned)
-		return UnalignedGetWordNonTemplate(order, block);
+		return UnalignedGetWordNonTemplate(order, block, (T*)NULL);
 	assert(IsAligned<T>(block));
 #endif
 	return ConditionalByteReverse(order, *reinterpret_cast<const T *>(block));
@@ -953,7 +953,7 @@ inline void PutWord(bool assumeAligned, ByteOrder order, byte *block, T value, c
 {
 #ifndef CRYPTOPP_ALLOW_UNALIGNED_DATA_ACCESS
 	if (!assumeAligned)
-		return UnalignedGetWordNonTemplate(order, block, value, xorBlock);
+		return UnalignedPutWordNonTemplate(order, block, value, xorBlock);
 	assert(IsAligned<T>(block));
 	assert(IsAligned<T>(xorBlock));
 #endif
diff --git a/salsa.cpp b/salsa.cpp
index 40fffc43..4781218a 100755
--- a/salsa.cpp
+++ b/salsa.cpp
@@ -6,7 +6,9 @@
 #include "argnames.h"
 #include "cpu.h"
 
+#if CRYPTOPP_BOOL_SSE2_INTRINSICS_AVAILABLE
 #include <emmintrin.h>
+#endif
 
 NAMESPACE_BEGIN(CryptoPP)
 
diff --git a/strciphr.h b/strciphr.h
index e9bcf2e6..724bbbfd 100644
--- a/strciphr.h
+++ b/strciphr.h
@@ -83,7 +83,7 @@ template <typename WT, unsigned int W, unsigned int X = 1, class BASE = Additive
 struct CRYPTOPP_NO_VTABLE AdditiveCipherConcretePolicy : public BASE
 {
 	typedef WT WordType;
-	CRYPTOPP_CONSTANT(BYTES_PER_ITERATION = sizeof(WordType) * W);
+	CRYPTOPP_CONSTANT(BYTES_PER_ITERATION = sizeof(WordType) * W)
 
 #if !(CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X64)
 	unsigned int GetAlignment() const {return GetAlignmentOf<WordType>();}