Improve logic for <arm_acle.h> include (GH #568)

13 files changed, 103 insertions, 65 deletions

diff --git a/TestScripts/setenv-android-old.sh b/TestScripts/setenv-android-old.sh
index 3f7cac7b..d93b568d 100755
--- a/TestScripts/setenv-android-old.sh
+++ b/TestScripts/setenv-android-old.sh
@@ -108,21 +108,21 @@ case "$THE_ARCH" in
 	TOOLCHAIN_NAME="arm-linux-androideabi"
 	AOSP_ABI="armeabi"
 	AOSP_ARCH="arch-arm"
-	AOSP_FLAGS="-march=armv5te -mtune=xscale -mthumb -msoft-float -funwind-tables -fexceptions -frtti"
+	AOSP_FLAGS="-march=armv5te -mtune=xscale -mthumb -msoft-float -DCRYPTOPP_DISABLE_ASM -funwind-tables -fexceptions -frtti"
 	;;
   armv7a|armv7-a|armeabi-v7a)
 	TOOLCHAIN_ARCH="arm-linux-androideabi"
 	TOOLCHAIN_NAME="arm-linux-androideabi"
 	AOSP_ABI="armeabi-v7a"
 	AOSP_ARCH="arch-arm"
-	AOSP_FLAGS="-march=armv7-a -mthumb -mfpu=vfpv3-d16 -mfloat-abi=softfp -Wl,--fix-cortex-a8 -funwind-tables -fexceptions -frtti"
+	AOSP_FLAGS="-march=armv7-a -mthumb -mfpu=vfpv3-d16 -mfloat-abi=softfp -DCRYPTOPP_DISABLE_ASM -Wl,--fix-cortex-a8 -funwind-tables -fexceptions -frtti"
 	;;
   hard|armv7a-hard|armeabi-v7a-hard)
 	TOOLCHAIN_ARCH="arm-linux-androideabi"
 	TOOLCHAIN_NAME="arm-linux-androideabi"
 	AOSP_ABI="armeabi-v7a"
 	AOSP_ARCH="arch-arm"
-	AOSP_FLAGS="-mhard-float -D_NDK_MATH_NO_SOFTFP=1 -march=armv7-a -mfpu=vfpv3-d16 -mfloat-abi=softfp -Wl,--fix-cortex-a8 -funwind-tables -fexceptions -frtti -Wl,--no-warn-mismatch -Wl,-lm_hard"
+	AOSP_FLAGS="-mhard-float -D_NDK_MATH_NO_SOFTFP=1 -march=armv7-a -mfpu=vfpv3-d16 -DCRYPTOPP_DISABLE_ASM -mfloat-abi=softfp -Wl,--fix-cortex-a8 -funwind-tables -fexceptions -frtti -Wl,--no-warn-mismatch -Wl,-lm_hard"
 	;;
   neon|armv7a-neon)
 	TOOLCHAIN_ARCH="arm-linux-androideabi"
diff --git a/aria-simd.cpp b/aria-simd.cpp
index af90a42a..f855be90 100644
--- a/aria-simd.cpp
+++ b/aria-simd.cpp
@@ -10,16 +10,19 @@
 #include "config.h"

 #include "misc.h"

 

-#if !(defined(__ARM_NEON) || defined(_MSC_VER))

-# undef CRYPTOPP_ARM_NEON_AVAILABLE

+#if (CRYPTOPP_SSSE3_AVAILABLE)

+# include <tmmintrin.h>

 #endif

 

 #if (CRYPTOPP_ARM_NEON_AVAILABLE)

 # include <arm_neon.h>

 #endif

 

-#if (CRYPTOPP_SSSE3_AVAILABLE)

-# include <tmmintrin.h>

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

+#if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

+# include <arm_acle.h>

 #endif

 

 // Clang __m128i casts, http://bugs.llvm.org/show_bug.cgi?id=20670

diff --git a/blake2-simd.cpp b/blake2-simd.cpp
index d3bae574..76e7c153 100644
--- a/blake2-simd.cpp
+++ b/blake2-simd.cpp
@@ -22,10 +22,6 @@
 # undef CRYPTOPP_ARM_NEON_AVAILABLE

 #endif

 

-#if !(defined(__ARM_NEON) || defined(_MSC_VER))

-# undef CRYPTOPP_ARM_NEON_AVAILABLE

-#endif

-

 #if (CRYPTOPP_SSE41_AVAILABLE)

 # include <emmintrin.h>

 # include <tmmintrin.h>

@@ -36,6 +32,13 @@
 # include <arm_neon.h>

 #endif

 

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

+#if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

+# include <arm_acle.h>

+#endif

+

 // Clang __m128i casts, http://bugs.llvm.org/show_bug.cgi?id=20670

 #define M128_CAST(x) ((__m128i *)(void *)(x))

 #define CONST_M128_CAST(x) ((const __m128i *)(const void *)(x))

diff --git a/config.h b/config.h
index 9ad864f8..8de736df 100644
--- a/config.h
+++ b/config.h
@@ -563,12 +563,20 @@ NAMESPACE_END
 // Requires ARMv7 and ACLE 1.0. Testing shows ARMv7 is really ARMv7a under most toolchains.

 // Android still uses ARMv5 and ARMv6 so we have to be conservative when enabling NEON.

 #if !defined(CRYPTOPP_ARM_NEON_AVAILABLE) && !defined(CRYPTOPP_DISABLE_ASM)

-# if defined(__ARM_NEON) || defined(__ARM_NEON_FP) || defined(__ARM_FEATURE_NEON) || defined(__ARM_FEATURE_ASIMD) || \

+# if defined(__ARM_NEON) || defined(__ARM_NEON_FP) || defined(__ARM_FEATURE_NEON) || \

 	(__ARM_ARCH >= 7) || (CRYPTOPP_MSC_VERSION >= 1700)

 #  define CRYPTOPP_ARM_NEON_AVAILABLE 1

 # endif

 #endif

 

+// ARMv8 and ASIMD, which is NEON. It is part of ARMv8 core.

+// TODO: Add MSC_VER and ARM-64 platform define when available

+#if !defined(CRYPTOPP_ARM_ASIMD_AVAILABLE) && !defined(CRYPTOPP_DISABLE_ASM)

+# if defined(__aarch32__) || defined(__aarch64__) || (CRYPTOPP_MSC_VERSION >= 1910)

+#  define CRYPTOPP_ARM_ASIMD_AVAILABLE 1

+# endif

+#endif

+

 // Requires ARMv8 and ACLE 2.0. GCC requires 4.8 and above.

 // LLVM Clang requires 3.5. Apple Clang is unknown at the moment.

 // Microsoft plans to support ARM-64, but its not clear how to detect it.

@@ -598,26 +606,33 @@ NAMESPACE_END
 // Microsoft plans to support ARM-64, but its not clear how to detect it.

 // TODO: Add Android ARMv8 support for AES and SHA

 // TODO: Add MSC_VER and ARM-64 platform define when available

-#if !defined(CRYPTOPP_ARM_CRYPTO_AVAILABLE) && !defined(CRYPTOPP_DISABLE_ASM) && !defined(__ANDROID__)

+#if !defined(CRYPTOPP_ARM_AES_AVAILABLE) && !defined(CRYPTOPP_DISABLE_ASM) && !defined(__ANDROID__)

 # if defined(__ARM_FEATURE_CRYPTO) || (CRYPTOPP_MSC_VERSION >= 1910) || \

 	defined(__aarch32__) || defined(__aarch64__)

 #  define CRYPTOPP_ARM_AES_AVAILABLE 1

+# endif

+#endif

+

+// Requires ARMv8 and ACLE 2.0. GCC requires 4.8 and above.

+// LLVM Clang requires 3.5. Apple Clang is unknown at the moment.

+// Microsoft plans to support ARM-64, but its not clear how to detect it.

+// TODO: Add Android ARMv8 support for AES and SHA

+// TODO: Add MSC_VER and ARM-64 platform define when available

+#if !defined(CRYPTOPP_ARM_SHA_AVAILABLE) && !defined(CRYPTOPP_DISABLE_ASM) && !defined(__ANDROID__)

+# if defined(__ARM_FEATURE_CRYPTO) || (CRYPTOPP_MSC_VERSION >= 1910) || \

+	defined(__aarch32__) || defined(__aarch64__)

 #  define CRYPTOPP_ARM_SHA_AVAILABLE 1

-#  define CRYPTOPP_ARM_CRYPTO_AVAILABLE 1

 # endif

 #endif

 

-// Don't include <arm_acle.h> when using Apple Clang. Early Apple compilers

-//  fail to compile with <arm_acle.h> included. Later Apple compilers compile

-//  intrinsics without <arm_acle.h> included. Also avoid it with GCC 4.8,

-//  and avoid it on Android, too.

-#if defined(CRYPTOPP_ARM_NEON_AVAILABLE) || defined(CRYPTOPP_ARM_CRC32_AVAILABLE) || \

-	defined(CRYPTOPP_ARM_AES_AVAILABLE) || defined(CRYPTOPP_ARM_PMULL_AVAILABLE) || \

-	defined(CRYPTOPP_ARM_SHA_AVAILABLE)

-#  define CRYPTOPP_ARM_ACLE_AVAILABLE 1

+// Limit the <arm_acle.h> include.

+#if defined(__aarch32__) || defined(__aarch64__) || (__ARM_ARCH >= 8) || defined(__ARM_ACLE)

+# define CRYPTOPP_ARM_ACLE_AVAILABLE 1

 #endif

 

-#if (defined(__ANDROID__) || defined(ANDROID)) && !defined(__ARM_ACLE)

+// Man, this is borked. Apple Clang defines __ARM_ACLE but then fails

+// to compile with "fatal error: 'arm_acle.h' file not found"

+#if defined(__ANDROID__) || defined(ANDROID) || defined(__APPLE__)

 # undef CRYPTOPP_ARM_ACLE_AVAILABLE

 #endif

 

diff --git a/crc-simd.cpp b/crc-simd.cpp
index 2d55e008..ed709b28 100644
--- a/crc-simd.cpp
+++ b/crc-simd.cpp
@@ -14,10 +14,13 @@
 # include <nmmintrin.h>

 #endif

 

+// Use ARMv8 rather than NEON due to compiler inconsistencies

 #if (CRYPTOPP_ARM_CRC32_AVAILABLE)

 # include <arm_neon.h>

 #endif

 

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

 #if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

 # include <stdint.h>

 # include <arm_acle.h>

diff --git a/gcm-simd.cpp b/gcm-simd.cpp
index 88d260b0..32f04e40 100644
--- a/gcm-simd.cpp
+++ b/gcm-simd.cpp
@@ -39,11 +39,14 @@
 # include <wmmintrin.h>

 #endif

 

-#if (CRYPTOPP_ARM_NEON_AVAILABLE || CRYPTOPP_ARM_PMULL_AVAILABLE)

+#if (CRYPTOPP_ARM_NEON_AVAILABLE)

 # include <arm_neon.h>

 #endif

 

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

 #if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

 # include <arm_acle.h>

 #endif

 

diff --git a/neon-simd.cpp b/neon-simd.cpp
index 4c64b45a..5b2c8b05 100644
--- a/neon-simd.cpp
+++ b/neon-simd.cpp
@@ -10,15 +10,14 @@
 #include "config.h"

 #include "stdcpp.h"

 

-#if !(defined(__ARM_NEON) || defined(_MSC_VER))

-# undef CRYPTOPP_ARM_NEON_AVAILABLE

-#endif

-

 #if (CRYPTOPP_ARM_NEON_AVAILABLE)

 # include <arm_neon.h>

 #endif

 

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

 #if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

 # include <arm_acle.h>

 #endif

 

diff --git a/rijndael-simd.cpp b/rijndael-simd.cpp
index 032899e2..8a493606 100644
--- a/rijndael-simd.cpp
+++ b/rijndael-simd.cpp
@@ -25,12 +25,6 @@
 #include "misc.h"

 #include "adv-simd.h"

 

-// We set CRYPTOPP_ARM_AES_AVAILABLE based on compiler version.

-// If the crypto is not available, then we have to disable it here.

-#if !(defined(__ARM_FEATURE_CRYPTO) || defined(_MSC_VER))

-# undef CRYPTOPP_ARM_AES_AVAILABLE

-#endif

-

 // We set CRYPTOPP_POWER8_CRYPTO_AVAILABLE based on compiler version.

 // If the crypto is not available, then we have to disable it here.

 #if !(defined(__CRYPTO) || defined(_ARCH_PWR8) || defined(_ARCH_PWR9))

@@ -43,11 +37,16 @@
 # include <wmmintrin.h>

 #endif

 

+// Use ARMv8 rather than NEON due to compiler inconsistencies

 #if (CRYPTOPP_ARM_AES_AVAILABLE)

 # include <arm_neon.h>

-# if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

-#  include <arm_acle.h>

-# endif

+#endif

+

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

+#if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

+# include <arm_acle.h>

 #endif

 

 #if defined(CRYPTOPP_POWER8_AES_AVAILABLE)

diff --git a/setenv-android.sh b/setenv-android.sh
index 04191133..3066eeb3 100755
--- a/setenv-android.sh
+++ b/setenv-android.sh
@@ -101,21 +101,21 @@ case "$THE_ARCH" in
 	TOOLCHAIN_NAME="arm-linux-androideabi"
 	AOSP_ABI="armeabi"
 	AOSP_ARCH="arch-arm"
-	AOSP_FLAGS="-march=armv5te -mtune=xscale -mthumb -msoft-float -funwind-tables -fexceptions -frtti"
+	AOSP_FLAGS="-march=armv5te -mtune=xscale -mthumb -msoft-float -DCRYPTOPP_DISABLE_ASM -funwind-tables -fexceptions -frtti"
 	;;
   armv7a|armv7-a|armeabi-v7a)
 	TOOLCHAIN_ARCH="arm-linux-androideabi"
 	TOOLCHAIN_NAME="arm-linux-androideabi"
 	AOSP_ABI="armeabi-v7a"
 	AOSP_ARCH="arch-arm"
-	AOSP_FLAGS="-march=armv7-a -mthumb -mfpu=vfpv3-d16 -mfloat-abi=softfp -Wl,--fix-cortex-a8 -funwind-tables -fexceptions -frtti"
+	AOSP_FLAGS="-march=armv7-a -mthumb -mfpu=vfpv3-d16 -mfloat-abi=softfp -DCRYPTOPP_DISABLE_ASM -Wl,--fix-cortex-a8 -funwind-tables -fexceptions -frtti"
 	;;
   hard|armv7a-hard|armeabi-v7a-hard)
 	TOOLCHAIN_ARCH="arm-linux-androideabi"
 	TOOLCHAIN_NAME="arm-linux-androideabi"
 	AOSP_ABI="armeabi-v7a"
 	AOSP_ARCH="arch-arm"
-	AOSP_FLAGS="-mhard-float -D_NDK_MATH_NO_SOFTFP=1 -march=armv7-a -mfpu=vfpv3-d16 -mfloat-abi=softfp -Wl,--fix-cortex-a8 -funwind-tables -fexceptions -frtti -Wl,--no-warn-mismatch -Wl,-lm_hard"
+	AOSP_FLAGS="-mhard-float -D_NDK_MATH_NO_SOFTFP=1 -march=armv7-a -mfpu=vfpv3-d16 -DCRYPTOPP_DISABLE_ASM -mfloat-abi=softfp -Wl,--fix-cortex-a8 -funwind-tables -fexceptions -frtti -Wl,--no-warn-mismatch -Wl,-lm_hard"
 	;;
   neon|armv7a-neon)
 	TOOLCHAIN_ARCH="arm-linux-androideabi"
diff --git a/sha-simd.cpp b/sha-simd.cpp
index 25805f46..e3a45d17 100644
--- a/sha-simd.cpp
+++ b/sha-simd.cpp
@@ -11,22 +11,21 @@
 #include "sha.h"

 #include "misc.h"

 

-// We set CRYPTOPP_ARM_SHA_AVAILABLE based on compiler version.

-// If the crypto is not available, then we have to disable it here.

-#if !(defined(__ARM_FEATURE_CRYPTO) || defined(_MSC_VER))

-# undef CRYPTOPP_ARM_SHA_AVAILABLE

-#endif

-

 #if (CRYPTOPP_SHANI_AVAILABLE)

 # include <nmmintrin.h>

 # include <immintrin.h>

 #endif

 

+// Use ARMv8 rather than NEON due to compiler inconsistencies

 #if (CRYPTOPP_ARM_SHA_AVAILABLE)

 # include <arm_neon.h>

-# if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

-#  include <arm_acle.h>

-# endif

+#endif

+

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

+#if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

+# include <arm_acle.h>

 #endif

 

 #if CRYPTOPP_POWER8_SHA_AVAILABLE

diff --git a/shacal2-simd.cpp b/shacal2-simd.cpp
index 93cb6e07..6fb7f2b0 100644
--- a/shacal2-simd.cpp
+++ b/shacal2-simd.cpp
@@ -17,21 +17,21 @@
 #include "sha.h"

 #include "misc.h"

 

-// Clang and GCC hoops...

-#if !(defined(__ARM_FEATURE_CRYPTO) || defined(_MSC_VER))

-# undef CRYPTOPP_ARM_SHA_AVAILABLE

-#endif

-

 #if (CRYPTOPP_SHANI_AVAILABLE)

 # include <nmmintrin.h>

 # include <immintrin.h>

 #endif

 

+// Use ARMv8 rather than NEON due to compiler inconsistencies

 #if (CRYPTOPP_ARM_SHA_AVAILABLE)

 # include <arm_neon.h>

-# if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

-#  include <arm_acle.h>

-# endif

+#endif

+

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

+#if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

+# include <arm_acle.h>

 #endif

 

 // Clang __m128i casts, http://bugs.llvm.org/show_bug.cgi?id=20670

diff --git a/simon-simd.cpp b/simon-simd.cpp
index 26b41c73..81a73204 100644
--- a/simon-simd.cpp
+++ b/simon-simd.cpp
@@ -18,10 +18,6 @@
 // #undef CRYPTOPP_SSE41_AVAILABLE

 // #undef CRYPTOPP_ARM_NEON_AVAILABLE

 

-#if (CRYPTOPP_ARM_NEON_AVAILABLE)

-# include <arm_neon.h>

-#endif

-

 #if (CRYPTOPP_SSSE3_AVAILABLE)

 # include <pmmintrin.h>

 # include <tmmintrin.h>

@@ -36,6 +32,17 @@
 # include <immintrin.h>

 #endif

 

+#if (CRYPTOPP_ARM_NEON_AVAILABLE)

+# include <arm_neon.h>

+#endif

+

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

+#if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

+# include <arm_acle.h>

+#endif

+

 // https://www.spinics.net/lists/gcchelp/msg47735.html and

 // https://www.spinics.net/lists/gcchelp/msg47749.html

 #if (CRYPTOPP_GCC_VERSION >= 40900)

diff --git a/speck-simd.cpp b/speck-simd.cpp
index b946460a..04db1030 100644
--- a/speck-simd.cpp
+++ b/speck-simd.cpp
@@ -27,10 +27,6 @@
 # define WORKAROUND_GCC_AARCH64_BUG 1

 #endif

 

-#if (CRYPTOPP_ARM_NEON_AVAILABLE)

-# include <arm_neon.h>

-#endif

-

 #if (CRYPTOPP_SSSE3_AVAILABLE)

 # include <pmmintrin.h>

 # include <tmmintrin.h>

@@ -45,6 +41,17 @@
 # include <immintrin.h>

 #endif

 

+#if (CRYPTOPP_ARM_NEON_AVAILABLE)

+# include <arm_neon.h>

+#endif

+

+// Can't use CRYPTOPP_ARM_XXX_AVAILABLE because too many

+// compilers don't follow ACLE conventions for the include.

+#if defined(CRYPTOPP_ARM_ACLE_AVAILABLE)

+# include <stdint.h>

+# include <arm_acle.h>

+#endif

+

 // https://www.spinics.net/lists/gcchelp/msg47735.html and

 // https://www.spinics.net/lists/gcchelp/msg47749.html

 #if (CRYPTOPP_GCC_VERSION >= 40900)