Fix ARM A-32 and Aarch64 feature tests (PR #737)

GNUmakefile

@@ -104,6 +104,16 @@ ifeq ($(IS_AIX),1)
   endif
 endif
 
+# Hack to skip CPU feature tests for some recipes
+DETECT_FEATURES ?= 1
+ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),-DCRYPTOPP_DISABLE_ASM)
+  DETECT_FEATURES := 0
+else ifeq ($(findstring clean,$(MAKECMDGOALS)),clean)
+  DETECT_FEATURES := 0
+else ifeq ($(findstring distclean,$(MAKECMDGOALS)),distclean)
+  DETECT_FEATURES := 0
+endif
+
 ###########################################################
 #####                General Variables                #####
 ###########################################################
@@ -190,7 +200,7 @@ ifneq ($(IS_X86)$(IS_X64),00)
 
 # Begin GCC and compatibles
 ifneq ($(GCC_COMPILER)$(CLANG_COMPILER)$(INTEL_COMPILER),000)
-ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),)
+ifeq ($(DETECT_FEATURES),1)
 
   # Tell MacPorts and Homebrew GCC to use Clang integrated assembler
   #   http://github.com/weidai11/cryptopp/issues/190
@@ -301,7 +311,7 @@ ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),)
   # https://github.com/weidai11/cryptopp/issues/738
   UNUSED := $(shell rm -f a.out && rm -rf a.out.dSYM/)
 
-# CRYPTOPP_DISABLE_ASM
+# DETECT_FEATURES
 endif
 
 # End GCC and compatibles
@@ -309,7 +319,7 @@ endif
 
 # Begin SunCC
 ifeq ($(SUN_COMPILER),1)
-ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),)
+ifeq ($(DETECT_FEATURES),1)
 
   TPROG = TestPrograms/test_sse2.cxx
   TOPT = -xarch=sse2
@@ -417,7 +427,7 @@ ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),)
   # https://github.com/weidai11/cryptopp/issues/738
   UNUSED := $(shell rm -f a.out && rm -rf a.out.dSYM/)
 
-# CRYPTOPP_DISABLE_ASM
+# DETECT_FEATURES
 endif
 
 # End SunCC
@@ -451,11 +461,11 @@ endif
 ###########################################################
 
 ifneq ($(IS_ARM32)$(IS_ARMV8)$(IS_NEON),000)
-ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),)
+ifeq ($(DETECT_FEATURES),1)
 
 ifeq ($(IS_ARM32)$(IS_NEON),11)
 
-  TPROG = TestPrograms/test_neon.cxx
+  TPROG = TestPrograms/test_arm_neon.cxx
   TOPT = -march=armv7-a -mfloat-abi=$(FP_ABI) -mfpu=neon
   HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
   ifeq ($(HAVE_OPT),0)
@@ -488,16 +498,18 @@ endif
 
 ifeq ($(IS_ARMV8),1)
 
-  TPROG = TestPrograms/test_asimd.cxx
+  TPROG = TestPrograms/test_arm_acle.cxx
   TOPT = -march=armv8-a
   HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
-  ifneq ($(HAVE_OPT),0)
+  ifeq ($(HAVE_OPT),0)
+	ACLE_FLAG += -DCRYPTOPP_ARM_ACLE_AVAILABLE=1
+  else
 	CXXFLAGS += -DCRYPTOPP_ARM_ACLE_AVAILABLE=0
   endif
 
-  TPROG = TestPrograms/test_asimd.cxx
+  TPROG = TestPrograms/test_arm_asimd.cxx
   TOPT = -march=armv8-a
-  HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
+  HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
   ifeq ($(HAVE_OPT),0)
     ASIMD_FLAG = -march=armv8-a
     ARIA_FLAG = -march=armv8-a
@@ -518,26 +530,46 @@ ifeq ($(IS_ARMV8),1)
   endif
 
   ifneq ($(ASIMD_FLAG),)
-    TPROG = TestPrograms/test_crc.cxx
+    TPROG = TestPrograms/test_arm_crc.cxx
-    TOPT = -march=armv8-a+crc
+    TOPT = -march=armv8.1-a+crc
-    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
     ifeq ($(HAVE_OPT),0)
-      CRC_FLAG = -march=armv8-a+crc
+      CRC_FLAG = -march=armv8.1-a+crc
+    else
+      CXXFLAGS += -DCRYPTOPP_ARM_CRC32_AVAILABLE=0
     endif
 
-    TPROG = TestPrograms/test_crypto_v81.cxx
+    TPROG = TestPrograms/test_arm_aes.cxx
-    TOPT = -march=armv8-a+crypto
+    TOPT = -march=armv8.1-a+crypto
-    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
     ifeq ($(HAVE_OPT),0)
-      AES_FLAG = -march=armv8-a+crypto
+      AES_FLAG = -march=armv8.1-a+crypto
-      GCM_FLAG = -march=armv8-a+crypto
+    else
-      SHA_FLAG = -march=armv8-a+crypto
+      CXXFLAGS += -DCRYPTOPP_ARM_AES_AVAILABLE=0
+    endif
+
+    TPROG = TestPrograms/test_arm_pmull.cxx
+    TOPT = -march=armv8.1-a+crypto
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
+    ifeq ($(HAVE_OPT),0)
+      GCM_FLAG = -march=armv8.1-a+crypto
+    else
+      CXXFLAGS += -DCRYPTOPP_ARM_PMULL_AVAILABLE=0
+    endif
+
+    TPROG = TestPrograms/test_arm_sha.cxx
+    TOPT = -march=armv8.1-a+crypto
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
+    ifeq ($(HAVE_OPT),0)
+      SHA_FLAG = -march=armv8.1-a+crypto
+    else
+      CXXFLAGS += -DCRYPTOPP_ARM_SHA_AVAILABLE=0
     endif
 
     ifneq ($(AES_FLAG),)
       TPROG = TestPrograms/test_crypto_v84.cxx
       TOPT = -march=armv8.4-a+crypto
-      HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
+      HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) 2>&1 | $(GREP) -i -c -E $(BAD_RESULT))
       ifeq ($(HAVE_OPT),0)
         SM3_FLAG = -march=armv8.4-a+crypto
         SM4_FLAG = -march=armv8.4-a+crypto
@@ -551,7 +583,7 @@ ifeq ($(IS_ARMV8),1)
 # IS_ARMV8
 endif
 
-# CRYPTOPP_DISABLE_ASM
+# DETECT_FEATURES
 endif
 
 # IS_ARM32, IS_ARMV8, IS_NEON
@@ -566,7 +598,7 @@ endif
 # front-end. XLC/LLVM only supplies POWER8 so we have to set the flags for
 # XLC/LLVM to POWER8. I've got a feeling LLVM is going to cause trouble.
 ifneq ($(IS_PPC32)$(IS_PPC64),00)
-ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),)
+ifeq ($(DETECT_FEATURES),1)
 
   # LLVM front-ends only provide POWER8 and need special options to
   # get XLC defines. The POWER8 really jambs us up for ppc_simd.cpp
@@ -722,7 +754,7 @@ ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),)
   # https://github.com/weidai11/cryptopp/issues/738
   UNUSED := $(shell rm -f a.out && rm -rf a.out.dSYM/)
 
-# CRYPTOPP_DISABLE_ASM
+# DETECT_FEATURES
 endif
 
 # IBM XL C/C++ compiler

TestPrograms/test_arm_acle.cxx

@@ -0,0 +1,6 @@
+#include <arm_acle.h>
+
+int main(int argc, char* argv[])
+{
+	return 0;
+}

TestPrograms/test_arm_aes.cxx

@@ -0,0 +1,19 @@
+#include <arm_neon.h>
+#include <stdint.h>
+
+// test_acle.h determines if this is available. Then,
+// -DCRYPTOPP_ARM_ACLE_AVAILABLE=0 is added to CXXFLAGS
+// if the ACLE header is not available.
+#if (CRYPTOPP_ARM_ACLE_AVAILABLE)
+#  include <arm_acle.h>
+#endif
+
+int main(int argc, char* argv[])
+{
+	uint8x16_t x;
+	x=vaeseq_u8(x,x);
+	x=vaesmcq_u8(x);
+	x=vaesdq_u8(x,x);
+	x=vaesimcq_u8(x);
+	return 0;
+}

TestPrograms/test_arm_asimd.cxx

@@ -0,0 +1,16 @@
+#include <arm_neon.h>
+#include <stdint.h>
+
+// test_acle.h determines if this is available. Then,
+// -DCRYPTOPP_ARM_ACLE_AVAILABLE=0 is added to CXXFLAGS
+// if the ACLE header is not available.
+#if (CRYPTOPP_ARM_ACLE_AVAILABLE)
+#  include <arm_acle.h>
+#endif
+
+int main(int argc, char* argv[])
+{
+	uint32x4_t x;
+	x=veorq_u32(x,x);
+	return 0;
+}

TestPrograms/test_arm_crc.cxx

@@ -1,11 +1,11 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-// This is how config.h includes it.
+// test_acle.h determines if this is available. Then,
-#if defined(__aarch32__) || defined(__aarch64__) || (__ARM_ARCH >= 8) || defined(__ARM_ACLE)
+// -DCRYPTOPP_ARM_ACLE_AVAILABLE=0 is added to CXXFLAGS
-# if !defined(__ANDROID__) && !defined(ANDROID) && !defined(__APPLE__)
+// if the ACLE header is not available.
+#if (CRYPTOPP_ARM_ACLE_AVAILABLE)
 #  include <arm_acle.h>
-# endif
 #endif
 
 int main(int argc, char* argv[])

TestPrograms/test_arm_neon.cxx

@@ -1,10 +1,6 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-#if (CRYPTOPP_ARM_ACLE_AVAILABLE)
-#  include <arm_acle.h>
-#endif
-
 int main(int argc, char* argv[])
 {
 	uint32x4_t x;

TestPrograms/test_arm_pmull.cxx

@@ -0,0 +1,23 @@
+#include <arm_neon.h>
+#include <stdint.h>
+
+// test_acle.h determines if this is available. Then,
+// -DCRYPTOPP_ARM_ACLE_AVAILABLE=0 is added to CXXFLAGS
+// if the ACLE header is not available.
+#if (CRYPTOPP_ARM_ACLE_AVAILABLE)
+#  include <arm_acle.h>
+#endif
+
+int main(int argc, char* argv[])
+{
+	poly64_t   a1={0x9090909090909090}, b1={0xb0b0b0b0b0b0b0b0};
+	poly8x16_t a2={0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,
+	               0xa0,0xa0,0xa0,0xa0,0xa0,0xa0,0xa0,0xa0},
+	b2={0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,
+	    0xe0,0xe0,0xe0,0xe0,0xe0,0xe0,0xe0,0xe0};
+
+	poly128_t r1 = vmull_p64(a1, b1);
+	poly128_t r2 = vmull_high_p64((poly64x2_t)(a2), (poly64x2_t)(b2));
+
+	return 0;
+}

TestPrograms/test_arm_sha.cxx

@@ -0,0 +1,21 @@
+#include <arm_neon.h>
+#include <stdint.h>
+
+// test_acle.h determines if this is available. Then,
+// -DCRYPTOPP_ARM_ACLE_AVAILABLE=0 is added to CXXFLAGS
+// if the ACLE header is not available.
+#if (CRYPTOPP_ARM_ACLE_AVAILABLE)
+#  include <arm_acle.h>
+#endif
+
+int main(int argc, char* argv[])
+{
+	uint32x4_t y;
+	y=vsha1cq_u32(y,0,y);
+	y=vsha1mq_u32(y,1,y);
+	y=vsha1pq_u32(y,2,y);
+	y = vsha256hq_u32(y, y, y);
+	y = vsha256h2q_u32(y, y, y);
+	y = vsha256su1q_u32(y, y, y);
+	return 0;
+}

TestPrograms/test_asimd.cxx

@@ -1,16 +0,0 @@
-#include <arm_neon.h>
-#include <stdint.h>
-
-// This is how config.h includes it.
-#if defined(__aarch32__) || defined(__aarch64__) || (__ARM_ARCH >= 8) || defined(__ARM_ACLE)
-# if !defined(__ANDROID__) && !defined(ANDROID) && !defined(__APPLE__)
-#  include <arm_acle.h>
-# endif
-#endif
-
-int main(int argc, char* argv[])
-{
-	uint32x4_t x;
-	x=veorq_u32(x,x);
-	return 0;
-}

TestPrograms/test_crypto_v81.cxx

@@ -1,23 +0,0 @@
-#include <arm_neon.h>
-#include <stdint.h>
-
-// This is how config.h includes it.
-#if defined(__aarch32__) || defined(__aarch64__) || (__ARM_ARCH >= 8) || defined(__ARM_ACLE)
-# if !defined(__ANDROID__) && !defined(ANDROID) && !defined(__APPLE__)
-#  include <arm_acle.h>
-# endif
-#endif
-
-int main(int argc, char* argv[])
-{
-	uint8x16_t x;
-	x=vaeseq_u8(x,x);
-	x=vaesdq_u8(x,x);
-
-	uint32x4_t y;
-	y=vsha1cq_u32(y,0,y);
-	y=vsha1mq_u32(y,1,y);
-	y=vsha1pq_u32(y,2,y);
-
-	return 0;
-}

TestPrograms/test_crypto_v84.cxx

@@ -1,11 +1,11 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-// This is how config.h includes it.
+// test_acle.h determines if this is available. Then,
-#if defined(__aarch32__) || defined(__aarch64__) || (__ARM_ARCH >= 8) || defined(__ARM_ACLE)
+// -DCRYPTOPP_ARM_ACLE_AVAILABLE=0 is added to CXXFLAGS
-# if !defined(__ANDROID__) && !defined(ANDROID) && !defined(__APPLE__)
+// if the ACLE header is not available.
+#if (CRYPTOPP_ARM_ACLE_AVAILABLE)
 #  include <arm_acle.h>
-# endif
 #endif
 
 int main(int argc, char* argv[])

config.h

@@ -729,9 +729,11 @@ NAMESPACE_END
 #endif
 
 // Limit the <arm_acle.h> include.
-#if defined(__aarch32__) || defined(__aarch64__) || (__ARM_ARCH >= 8) || defined(__ARM_ACLE)
+#if !defined(CRYPTOPP_ARM_ACLE_AVAILABLE)
-# if !defined(__ANDROID__) && !defined(ANDROID) && !defined(__APPLE__)
+# if defined(__aarch32__) || defined(__aarch64__) || (__ARM_ARCH >= 8) || defined(__ARM_ACLE)
-#  define CRYPTOPP_ARM_ACLE_AVAILABLE 1
+#  if !defined(__ANDROID__) && !defined(ANDROID) && !defined(__APPLE__)
+#   define CRYPTOPP_ARM_ACLE_AVAILABLE 1
+#  endif
 # endif
 #endif