diff --git a/blake2.cpp b/blake2.cpp
index 0cb858b1..7e86a98f 100644
--- a/blake2.cpp
+++ b/blake2.cpp
@@ -22,9 +22,14 @@ NAMESPACE_BEGIN(CryptoPP)
 # undef CRYPTOPP_BOOL_SSE2_INTRINSICS_AVAILABLE
 #endif
 
-// Testing shows Sun CC needs 12.4 for _mm_set_epi64x
-#if (__SUNPRO_CC <= 0x5130)
+// SunCC needs 12.4 for _mm_set_epi64x, _mm_blend_epi16, _mm_shuffle_epi16, etc
+#if defined(__SUNPRO_CC) && (__SUNPRO_CC < 0x5130)
 # undef CRYPTOPP_BOOL_SSE2_INTRINSICS_AVAILABLE
+# undef CRYPTOPP_BOOL_SSE4_INTRINSICS_AVAILABLE
+#elif (__SUNPRO_CC >= 0x5130)
+# include <emmintrin.h>    // _mm_set_epi64x
+# include <smmintrin.h>    // _mm_blend_epi16
+# include <tmmintrin.h>    // _mm_shuffle_epi16
 #endif
 
 // Visual Studio needs VS2008 (1500); no dependency on _mm_set_epi64x()
diff --git a/crc.cpp b/crc.cpp
index 738d1238..3cdcc9a9 100644
--- a/crc.cpp
+++ b/crc.cpp
@@ -13,6 +13,11 @@ NAMESPACE_BEGIN(CryptoPP)
 # undef CRYPTOPP_BOOL_SSE4_INTRINSICS_AVAILABLE
 #endif
 
+// SunCC 12.4 and above
+#if defined(CRYPTOPP_BOOL_SSE4_INTRINSICS_AVAILABLE) && (__SUNPRO_CC >= 0x5130)
+# include <nmmintrin.h>
+#endif
+
 /* Table of CRC-32's of all single byte values (made by makecrc.c) */
 const word32 CRC32::m_tab[] = {
 #ifdef IS_LITTLE_ENDIAN
diff --git a/cryptest.sh b/cryptest.sh
index dd02a697..427c598e 100755
--- a/cryptest.sh
+++ b/cryptest.sh
@@ -22,6 +22,8 @@
 
 # The fastest results (in running time) will most likely use:
 #     HAVE_VALGRIND=0 WANT_BENCHMARKS=0 ./cryptest.sh
+# Using 'fast' is shorthand for it:
+#     ./cryptest.sh fast
 
 ############################################
 # Set to suite your taste
@@ -343,6 +345,12 @@ else
 	fi
 fi
 
+# Fixup... SunCC appears to botch the code generation
+if [[ ("$SUN_COMPILER" -ne "0" )]];then
+	HAVE_O5=0
+	OPT_O5=
+fi
+
 # Hit or miss, mostly hit
 HAVE_OS=0
 OPT_OS=
@@ -500,7 +508,7 @@ HAVE_X86_AES=0
 HAVE_X86_RDRAND=0
 HAVE_X86_RDSEED=0
 HAVE_X86_PCLMUL=0
-if [[ (("$IS_X86" -ne "0") || ("$IS_X64" -ne "0")) ]]; then
+if [[ (("$IS_X86" -ne "0") || ("$IS_X64" -ne "0")) && ("$SUN_COMPILER" -eq "0") ]]; then
 	rm -f "$TMP/adhoc.exe" > /dev/null 2>&1
 	"$CXX" -DCRYPTOPP_ADHOC_MAIN -maes adhoc.cpp -o "$TMP/adhoc.exe" > /dev/null 2>&1
 	if [[ "$?" -eq "0" ]]; then
@@ -529,8 +537,9 @@ fi
 # ld-gold linker testing
 if [[ (-z "$HAVE_LDGOLD") ]]; then
 	HAVE_LDGOLD=0
-	LD_GOLD=$(which ld.gold 2>/dev/null | "$GREP" -v "no ld.gold" | head -1)
-	if [[ (! -z "$LD_GOLD") ]]; then
+	LD_GOLD=$(which ld.gold 2>&1 | "$GREP" -v "no ld.gold" | head -1)
+	ELF_FILE=$(which file 2>&1 | "$GREP" -v "no file" | head -1)
+	if [[ (! -z "$LD_GOLD") && (! -z "$ELF_FILE") ]]; then
 		HAVE_LDGOLD=$(file "$LD_GOLD" | cut -d":" -f 2 | "$EGREP" -i -c "elf")
 	fi
 fi
diff --git a/rdrand.asm b/rdrand.asm
index 73f5a964..9fcf6e0a 100644
--- a/rdrand.asm
+++ b/rdrand.asm
@@ -108,7 +108,7 @@ Call_RDRAND_EAX:
 			;; RDRAND is not available prior to VS2012. Just emit
 			;;   the byte codes using DB. This is `rdrand eax`.
 	DB		0Fh, 0C7h, 0F0h
-	
+
 			;; If CF=1, the number returned by RDRAND is valid.
 			;; If CF=0, a random number was not available.
 	jc 		RDRAND_succeeded
@@ -126,13 +126,13 @@ RDRAND_succeeded:
 
 	cmp		bsize, MWSIZE
 	jb		Partial_Machine_Word
-	
+
 Full_Machine_Word:
 
 	mov		DWORD PTR [buffer], eax
 	add		buffer, MWSIZE		;; No need for Intel Core 2 slow workarounds, like
 	sub		bsize, MWSIZE		;;   `lea buffer,[buffer+MWSIZE]` for faster adds
-	
+
 			;; Continue
 	jmp		GenerateBlock_Top
 
@@ -146,9 +146,9 @@ Partial_Machine_Word:
 	mov		WORD PTR [buffer], ax
 	shr		eax, 16
 	add 	buffer, 2
-	
+
 Bit_1_Not_Set:
-	
+
 			;; Test bit 0 to see if size is at least 1
 	test	bsize, 1
 	jz  	GenerateBlock_Success
@@ -164,14 +164,14 @@ GenerateBlock_PreRet:
 
 			;; Test for success (was the request completely fulfilled?)
 	cmp 	bsize, 0
-	je 		GenerateBlock_Success	
-	
+	je 		GenerateBlock_Success
+
 GenerateBlock_Failure:
 
 	xor		eax, eax
 	mov		al, RDRAND_FAILURE
 	ret
-	
+
 GenerateBlock_Success:
 
 	xor		eax, eax
@@ -241,7 +241,7 @@ RDRAND_succeeded:
 
 	cmp		bsize, MWSIZE
 	jb		Partial_Machine_Word
-	
+
 Full_Machine_Word:
 
 	mov		QWORD PTR [buffer], rax
@@ -284,19 +284,19 @@ Bit_0_Not_Set:
 
 			;; We've hit all the bits
 	jmp		GenerateBlock_Success
-		
+
 GenerateBlock_PreRet:
 
 			;; Test for success (was the request completely fulfilled?)
 	cmp 	bsize, 0
 	je 		GenerateBlock_Success
-	
+
 GenerateBlock_Failure:
 
 	xor		rax, rax
 	mov		al, RDRAND_FAILURE
 	ret
-	
+
 GenerateBlock_Success:
 
 	xor		rax, rax
@@ -352,7 +352,7 @@ Call_RDSEED_EAX:
 			;; RDSEED is not available prior to VS2012. Just emit
 			;;   the byte codes using DB. This is `rdseed eax`.
 	DB		0Fh, 0C7h, 0F8h
-	
+
 			;; If CF=1, the number returned by RDSEED is valid.
 			;; If CF=0, a random number was not available.
 	jc 		RDSEED_succeeded
@@ -370,13 +370,13 @@ RDSEED_succeeded:
 
 	cmp		bsize, MWSIZE
 	jb		Partial_Machine_Word
-	
+
 Full_Machine_Word:
 
 	mov		DWORD PTR [buffer], eax
 	add		buffer, MWSIZE		;; No need for Intel Core 2 slow workarounds, like
 	sub		bsize, MWSIZE		;;   `lea buffer,[buffer+MWSIZE]` for faster adds
-	
+
 			;; Continue
 	jmp		GenerateBlock_Top
 
@@ -390,9 +390,9 @@ Partial_Machine_Word:
 	mov		WORD PTR [buffer], ax
 	shr		eax, 16
 	add 	buffer, 2
-	
+
 Bit_1_Not_Set:
-	
+
 			;; Test bit 0 to see if size is at least 1
 	test	bsize, 1
 	jz  	GenerateBlock_Success
@@ -408,14 +408,14 @@ GenerateBlock_PreRet:
 
 			;; Test for success (was the request completely fulfilled?)
 	cmp 	bsize, 0
-	je 		GenerateBlock_Success	
-	
+	je 		GenerateBlock_Success
+
 GenerateBlock_Failure:
 
 	xor		eax, eax
 	mov		al, RDSEED_FAILURE
 	ret
-	
+
 GenerateBlock_Success:
 
 	xor		eax, eax
@@ -485,7 +485,7 @@ RDSEED_succeeded:
 
 	cmp		bsize, MWSIZE
 	jb		Partial_Machine_Word
-	
+
 Full_Machine_Word:
 
 	mov		QWORD PTR [buffer], rax
@@ -528,19 +528,19 @@ Bit_0_Not_Set:
 
 			;; We've hit all the bits
 	jmp		GenerateBlock_Success
-		
+
 GenerateBlock_PreRet:
 
 			;; Test for success (was the request completely fulfilled?)
 	cmp 	bsize, 0
 	je 		GenerateBlock_Success
-	
+
 GenerateBlock_Failure:
 
 	xor		rax, rax
 	mov		al, RDSEED_FAILURE
 	ret
-	
+
 GenerateBlock_Success:
 
 	xor		rax, rax
diff --git a/rdrand.cpp b/rdrand.cpp
index 4edb1bd8..fec720c0 100644
--- a/rdrand.cpp
+++ b/rdrand.cpp
@@ -68,7 +68,7 @@
 
 #if defined(CRYPTOPP_CPUID_AVAILABLE)
 # define MSC_INTRIN_COMPILER ((CRYPTOPP_MSC_VERSION >= 1700) || (CRYPTOPP_LLVM_CLANG_VERSION >= 30200) || (_INTEL_COMPILER >= 1210))
-# define GCC_INTRIN_COMPILER ((CRYPTOPP_GCC_VERSION >= 40600) || (CRYPTOPP_LLVM_CLANG_VERSION >= 30200) || (_INTEL_COMPILER >= 1210))
+# define GCC_INTRIN_COMPILER ((CRYPTOPP_GCC_VERSION >= 40600) || (CRYPTOPP_LLVM_CLANG_VERSION >= 30200) || (_INTEL_COMPILER >= 1210)) || (__SUNPRO_CC >= 0x5130)
 #else
 # define MSC_INTRIN_COMPILER 0
 # define GCC_INTRIN_COMPILER 0
@@ -99,6 +99,17 @@
 #  elif CRYPTOPP_BOOL_RDSEED_ASM
 #    define GCC_RDSEED_ASM_AVAILABLE 1
 #  endif
+#elif defined(CRYPTOPP_CPUID_AVAILABLE) && (__SUNPRO_CC >= 0x5100)
+#  if GCC_INTRIN_COMPILER && defined(__RDRND__) && (__SUNPRO_CC >= 0x5130)
+#    define ALL_RDRAND_INTRIN_AVAILABLE 1
+#  elif CRYPTOPP_BOOL_RDRAND_ASM
+#    define GCC_RDRAND_ASM_AVAILABLE 1
+#  endif
+#  if GCC_INTRIN_COMPILER && defined(__RDSEED__) && (__SUNPRO_CC >= 0x5150)
+#    define ALL_RDSEED_INTRIN_AVAILABLE 1
+#  elif CRYPTOPP_BOOL_RDSEED_ASM
+#    define GCC_RDSEED_ASM_AVAILABLE 1
+#  endif
 #endif
 
 // Debug diagnostics
@@ -131,7 +142,7 @@
 /////////////////////////////////////////////////////////////////////
 
 #if (ALL_RDRAND_INTRIN_AVAILABLE || ALL_RDSEED_INTRIN_AVAILABLE)
-# include <immintrin.h> // rdrand, MSC, ICC, and GCC
+# include <immintrin.h> // rdrand, MSC, ICC, GCC, and SunCC
 # if defined(__GNUC__) && (CRYPTOPP_GCC_VERSION >= 40600)
 #  include <x86intrin.h> // rdseed for some compilers, like GCC
 # endif