diff --git a/GNUmakefile b/GNUmakefile
index d57f2bb0..4981b73c 100755
--- a/GNUmakefile
+++ b/GNUmakefile
@@ -541,7 +541,7 @@ ifeq ($(IS_ARMV8),1)
     ifeq ($(strip $(HAVE_OPT)),0)
       SHA3_FLAG = -march=armv8.4-a+crypto
     endif
-	
+
   # ASIMD_FLAG
   endif
 
diff --git a/GNUmakefile-cross b/GNUmakefile-cross
index c5404c6c..15c16e22 100755
--- a/GNUmakefile-cross
+++ b/GNUmakefile-cross
@@ -3,6 +3,14 @@
 
 SHELL = /bin/sh
 
+# If needed
+TMPDIR ?= /tmp
+# Used for ARMv7 and NEON.
+FP_ABI ?= hard
+# Used for feature tests
+TOUT ?= a.out
+TOUT := $(strip $(TOUT))
+
 # Default CXXFLAGS if none were provided
 CXXFLAGS ?= -DNDEBUG -g2 -O3 -fPIC -pipe
 
@@ -22,32 +30,11 @@ INSTALL = install
 INSTALL_PROGRAM = $(INSTALL)
 INSTALL_DATA = $(INSTALL) -m 644
 
-# Attempt to determine host machine, fallback to "this" machine.
-#   The host machine is the one the package runs on. Most people
-#   call this the "target", but not Autotools.
-HOSTX := $(shell $(CXX) $(CXXFLAGS) -dumpmachine 2>/dev/null | cut -f 1 -d '-')
-ifeq ($(HOSTX),)
-  HOSTX := $(shell uname -m 2>/dev/null)
-endif
-
-IS_i686 := $(shell echo "$HOSTX" | $(EGREP) -v 64 | $(EGREP) -i -c 'i.86')
-IS_x86_64 := $(shell echo "$HOSTX" | $(EGREP) -i -c 'x86_64|amd64')
-IS_ARM32 := $(shell echo "$(HOSTX)" | $(EGREP) -i -c -E 'arm|armhf|arm7l|eabihf')
-IS_ARMv8 := $(shell echo "$HOSTX" | $(EGREP) -i -c 'aarch32|aarch64')
-
-GCC_COMPILER := $(shell $(CXX) --version 2>/dev/null | $(EGREP) -v -E '(llvm|clang)' | $(EGREP) -i -c -E '(gcc|g\+\+)')
-CLANG_COMPILER := $(shell $(CXX) --version 2>/dev/null | $(EGREP) -i -c -E '(llvm|clang)')
-
 IS_IOS ?= 0
 IS_ANDROID ?= 0
 IS_ARM_EMBEDDED ?= 0
 IS_NEON ?= 0
 
-# Fixup ARM
-ifeq ($(IS_ARMv8),1)
-  IS_ARM32 := 0
-endif
-
 # Can be used by Android and Embeeded cross-compiles. Disable by default because
 #   Android and embedded users typically don't run this configuration.
 HAS_SOLIB_VERSION ?= 0
@@ -83,14 +70,14 @@ endif
 
 # We honor ARFLAGS, but the "v" option used by default causes a noisy make
 ifeq ($(ARFLAGS),rv)
-ARFLAGS = r
+  ARFLAGS = r
 endif
 
 # Sadly, we can't actually use GCC_PRAGMA_AWARE because of GCC bug 53431.
 # Its a shame because GCC has so much to offer by the way of analysis.
 # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53431
 ifneq ($(CLANG_COMPILER),0)
-CXXFLAGS += -Wall
+  CXXFLAGS += -Wall
 endif
 
 ###########################################################
@@ -147,6 +134,355 @@ ifeq ($(IS_ARM_EMBEDDED),1)
   CXXFLAGS += $(ARM_EMBEDDED_FLAGS) --sysroot=$(ARM_EMBEDDED_SYSROOT)
 endif
 
+###########################################################
+#####              Compiler and Platform              #####
+###########################################################
+
+# Wait until CXXFLAGS have been set by setenv scripts
+
+GCC_COMPILER := $(shell $(CXX) --version 2>/dev/null | $(EGREP) -v -E '(llvm|clang)' | $(EGREP) -i -c -E '(gcc|g\+\+)')
+CLANG_COMPILER := $(shell $(CXX) --version 2>/dev/null | $(EGREP) -i -c -E '(llvm|clang)')
+
+HOSTX := $(shell $(CXX) $(CXXFLAGS) -dumpmachine 2>/dev/null | cut -f 1 -d '-')
+ifeq ($(HOSTX),)
+  HOSTX := $(shell uname -m 2>/dev/null)
+endif
+
+ifeq ($(CLANG_COMPILER),1)
+  IS_X86 := $(shell echo $(CXXFLAGS) | $(EGREP) -v 64 | $(EGREP) -i -c -E 'i.86')
+  IS_X64 := $(shell echo $(CXXFLAGS) | $(EGREP) -i -c -E 'x86_64|amd64')
+  IS_ARM32 := $(shell echo $(CXXFLAGS) | $(EGREP) -v 64 | $(EGREP) -i -c -E 'arm|armhf|arm7l|eabihf')
+  IS_ARMV8 := $(shell echo $(CXXFLAGS) | $(EGREP) -i -c -E 'aarch32|aarch64|arm64|armv8')
+else
+  IS_X86 := $(shell echo $(HOSTX) | $(EGREP) -v 64 | $(EGREP) -i -c -E 'i.86')
+  IS_X64 := $(shell echo $(HOSTX) | $(EGREP) -i -c -E 'x86_64|amd64')
+  IS_ARM32 := $(shell echo $(HOSTX) | $(EGREP) -v 64 | $(EGREP) -i -c -E 'arm|armhf|arm7l|eabihf')
+  IS_ARMV8 := $(shell echo $(HOSTX) | $(EGREP) -i -c -E 'aarch32|aarch64|arm64|armv8')
+endif
+
+$(info Here's what we think is happening... IS_X86: $(IS_X86), IS_X64: $(IS_X64), IS_ARM32: $(IS_ARM32), IS_ARMV8: $(IS_ARMV8))
+
+###########################################################
+#####                  Test Program                   #####
+###########################################################
+
+# Hack to skip CPU feature tests for some recipes
+DETECT_FEATURES ?= 1
+ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),-DCRYPTOPP_DISABLE_ASM)
+  DETECT_FEATURES := 0
+else ifeq ($(findstring clean,$(MAKECMDGOALS)),clean)
+  DETECT_FEATURES := 0
+else ifeq ($(findstring distclean,$(MAKECMDGOALS)),distclean)
+  DETECT_FEATURES := 0
+else ifeq ($(findstring distclean,$(MAKECMDGOALS)),trim)
+  DETECT_FEATURES := 0
+endif
+
+# Strip out -Wall, -Wextra and friends for feature testing
+ifeq ($(DETECT_FEATURES),1)
+  TCXXFLAGS := $(filter-out -Wall -Wextra -Werror -Wunused -Wconversion, $(CXXFLAGS))
+  ifneq ($(strip $(TCXXFLAGS)),)
+    $(info Using testing flags: $(TCXXFLAGS))
+  endif
+endif
+
+# For the previous messages
+$(info )
+
+###########################################################
+#####               X86/X32/X64 Options               #####
+###########################################################
+
+ifneq ($(IS_X86)$(IS_X64),00)
+ifeq ($(DETECT_FEATURES),1)
+
+  SSE2_FLAG = -msse2
+  SSE3_FLAG = -msse3
+  SSSE3_FLAG = -mssse3
+  SSE41_FLAG = -msse4.1
+  SSE42_FLAG = -msse4.2
+  CLMUL_FLAG = -mpclmul
+  AESNI_FLAG = -maes
+  AVX_FLAG = -mavx
+  AVX2_FLAG = -mavx2
+  SHANI_FLAG = -msha
+
+  TPROG = TestPrograms/test_x86_sse2.cxx
+  TOPT = $(SSE2_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    CHACHA_FLAG = $(SSE2_FLAG)
+  else
+    SSE2_FLAG =
+  endif
+
+  TPROG = TestPrograms/test_x86_ssse3.cxx
+  TOPT = $(SSSE3_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    ARIA_FLAG = $(SSSE3_FLAG)
+    CHAM_FLAG = $(SSSE3_FLAG)
+    LEA_FLAG = $(SSSE3_FLAG)
+    SIMECK_FLAG = $(SSSE3_FLAG)
+    SIMON64_FLAG = $(SSSE3_FLAG)
+    SIMON128_FLAG = $(SSSE3_FLAG)
+    SPECK64_FLAG = $(SSSE3_FLAG)
+    SPECK128_FLAG = $(SSSE3_FLAG)
+  else
+    SSSE3_FLAG =
+  endif
+
+  TPROG = TestPrograms/test_x86_sse41.cxx
+  TOPT = $(SSE41_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    BLAKE2B_FLAG = $(SSE41_FLAG)
+    BLAKE2S_FLAG = $(SSE41_FLAG)
+    SIMON64_FLAG = $(SSE41_FLAG)
+    SPECK64_FLAG = $(SSE41_FLAG)
+  else
+    SSE41_FLAG =
+  endif
+
+  TPROG = TestPrograms/test_x86_sse42.cxx
+  TOPT = $(SSE42_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    CRC_FLAG = $(SSE42_FLAG)
+  else
+    SSE42_FLAG =
+  endif
+
+  TPROG = TestPrograms/test_x86_clmul.cxx
+  TOPT = $(CLMUL_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    GCM_FLAG = $(SSSE3_FLAG) $(CLMUL_FLAG)
+  else
+    CLMUL_FLAG =
+  endif
+
+  TPROG = TestPrograms/test_x86_aes.cxx
+  TOPT = $(AESNI_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    AES_FLAG = $(SSE41_FLAG) $(AESNI_FLAG)
+    SM4_FLAG = $(SSSE3_FLAG) $(AESNI_FLAG)
+  else
+    AESNI_FLAG =
+  endif
+
+  TPROG = TestPrograms/test_x86_avx.cxx
+  TOPT = $(AVX_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    # XXX_FLAG = $(AVX_FLAG)
+  else
+    AVX_FLAG =
+  endif
+
+  TPROG = TestPrograms/test_x86_avx2.cxx
+  TOPT = $(AVX2_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    CHACHA_AVX2_FLAG = $(AVX2_FLAG)
+  else
+    AVX2_FLAG =
+  endif
+
+  TPROG = TestPrograms/test_x86_sha.cxx
+  TOPT = $(SHANI_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    SHA_FLAG = $(SSE42_FLAG) $(SHANI_FLAG)
+  else
+    SHANI_FLAG =
+  endif
+
+  ifeq ($(SSE2_FLAG),)
+    CXXFLAGS += -DCRYPTOPP_DISABLE_ASM
+  else ifeq ($(SSE3_FLAG),)
+    CXXFLAGS += -DCRYPTOPP_DISABLE_SSE3
+  else ifeq ($(SSSE3_FLAG),)
+    CXXFLAGS += -DCRYPTOPP_DISABLE_SSSE3
+  else ifeq ($(SSE41_FLAG),)
+    CXXFLAGS += -DCRYPTOPP_DISABLE_SSE4
+  else ifeq ($(SSE42_FLAG),)
+    CXXFLAGS += -DCRYPTOPP_DISABLE_SSE4
+  endif
+
+  # Unusual GCC/Clang on Macports. It assembles AES, but not CLMUL.
+  # test_x86_clmul.s:15: no such instruction: 'pclmulqdq $0, %xmm1,%xmm0'
+  ifeq ($(CLMUL_FLAG),)
+    CXXFLAGS += -DCRYPTOPP_DISABLE_CLMUL
+  endif
+  ifeq ($(AESNI_FLAG),)
+    CXXFLAGS += -DCRYPTOPP_DISABLE_AESNI
+  endif
+
+  ifneq ($(SSE42_FLAG),)
+    ifeq ($(AVX_FLAG),)
+      CXXFLAGS += -DCRYPTOPP_DISABLE_AVX
+    else ifeq ($(AVX2_FLAG),)
+      CXXFLAGS += -DCRYPTOPP_DISABLE_AVX2
+    else ifeq ($(SHANI_FLAG),)
+      CXXFLAGS += -DCRYPTOPP_DISABLE_SHANI
+    endif
+  endif
+
+# DETECT_FEATURES
+endif
+
+# IS_X86, IS_X32 and IS_X64
+endif
+
+###########################################################
+#####            ARM A-32, Aach64 and NEON            #####
+###########################################################
+
+ifneq ($(IS_ARM32)$(IS_ARMV8),00)
+ifeq ($(DETECT_FEATURES),1)
+
+ifeq ($(IS_ARM32),1)
+
+  ifneq ($(IS_IOS)$(IS_ANDROID),00)
+    NEON_FLAG =
+  else
+    NEON_FLAG = -march=armv7-a -mfloat-abi=$(FP_ABI) -mfpu=neon
+  endif
+
+  TPROG = TestPrograms/test_arm_neon.cxx
+  TOPT = $(NEON_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    ARIA_FLAG = $(NEON_FLAG)
+    AES_FLAG = $(NEON_FLAG)
+    CRC_FLAG = $(NEON_FLAG)
+    GCM_FLAG = $(NEON_FLAG)
+    BLAKE2B_FLAG = $(NEON_FLAG)
+    BLAKE2S_FLAG = $(NEON_FLAG)
+    CHACHA_FLAG = $(NEON_FLAG)
+    CHAM_FLAG = $(NEON_FLAG)
+    LEA_FLAG = $(NEON_FLAG)
+    SHA_FLAG = $(NEON_FLAG)
+    SIMECK_FLAG = $(NEON_FLAG)
+    SIMON64_FLAG = $(NEON_FLAG)
+    SIMON128_FLAG = $(NEON_FLAG)
+    SPECK64_FLAG = $(NEON_FLAG)
+    SPECK128_FLAG = $(NEON_FLAG)
+    SM4_FLAG = $(NEON_FLAG)
+  else
+    NEON_FLAG =
+    CXXFLAGS += -DCRYPTOPP_DISABLE_NEON
+  endif
+
+# IS_NEON
+endif
+
+ifeq ($(IS_ARMV8),1)
+
+  ifeq ($(IS_IOS),1)
+    ASIMD_FLAG =
+    CRC_FLAG =
+    AES_FLAG =
+    PMUL_FLAG =
+    SHA_FLAG =
+  else
+    ASIMD_FLAG = -march=armv8.1-a
+    CRC_FLAG = -march=armv8.1-a+crc
+    AES_FLAG = -march=armv8.1-a+crypto
+    PMULL_FLAG = -march=armv8.1-a+crypto
+    SHA_FLAG = -march=armv8.1-a+crypto
+  endif
+
+  TPROG = TestPrograms/test_arm_acle.cxx
+  TOPT = $(ASIMD_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(TCXXFLAGS) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    ACLE_FLAG += -DCRYPTOPP_ARM_ACLE_AVAILABLE=1
+  else
+    CXXFLAGS += -DCRYPTOPP_ARM_ACLE_AVAILABLE=0
+  endif
+
+  TPROG = TestPrograms/test_arm_asimd.cxx
+  TOPT = $(ASIMD_FLAG)
+  HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+  ifeq ($(strip $(HAVE_OPT)),0)
+    ARIA_FLAG = $(ASIMD_FLAG)
+    BLAKE2B_FLAG = $(ASIMD_FLAG)
+    BLAKE2S_FLAG = $(ASIMD_FLAG)
+    CHACHA_FLAG = $(ASIMD_FLAG)
+    CHAM_FLAG = $(ASIMD_FLAG)
+    LEA_FLAG = $(ASIMD_FLAG)
+    NEON_FLAG = $(ASIMD_FLAG)
+    SIMECK_FLAG = $(ASIMD_FLAG)
+    SIMON64_FLAG = $(ASIMD_FLAG)
+    SIMON128_FLAG = $(ASIMD_FLAG)
+    SPECK64_FLAG = $(ASIMD_FLAG)
+    SPECK128_FLAG = $(ASIMD_FLAG)
+    SM4_FLAG = $(ASIMD_FLAG)
+  else
+    ASIMD_FLAG =
+    CXXFLAGS += -DCRYPTOPP_DISABLE_ASM
+  endif
+
+    TPROG = TestPrograms/test_arm_crc.cxx
+    TOPT = $(CRC_FLAG)
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+    ifneq ($(strip $(HAVE_OPT)),0)
+      CRC_FLAG =
+      CXXFLAGS += -DCRYPTOPP_ARM_CRC32_AVAILABLE=0
+    endif
+
+    TPROG = TestPrograms/test_arm_aes.cxx
+    TOPT = $(AES_FLAG)
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+    ifneq ($(strip $(HAVE_OPT)),0)
+      AES_FLAG =
+      CXXFLAGS += -DCRYPTOPP_ARM_AES_AVAILABLE=0
+    endif
+
+    TPROG = TestPrograms/test_arm_pmull.cxx
+    TOPT = $(PMULL_FLAG)
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+    ifneq ($(strip $(HAVE_OPT)),0)
+      PMULL_FLAG =
+      CXXFLAGS += -DCRYPTOPP_ARM_PMULL_AVAILABLE=0
+    endif
+
+    TPROG = TestPrograms/test_arm_sha.cxx
+    TOPT = $(SHA_FLAG)
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+    ifneq ($(strip $(HAVE_OPT)),0)
+      SHA_FLAG =
+      CXXFLAGS += -DCRYPTOPP_ARM_SHA_AVAILABLE=0
+    endif
+
+    TPROG = TestPrograms/test_arm_sm3.cxx
+    TOPT = -march=armv8.4-a+crypto
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+    ifeq ($(strip $(HAVE_OPT)),0)
+      SM3_FLAG = -march=armv8.4-a+crypto
+      SM4_FLAG = -march=armv8.4-a+crypto
+    endif
+
+    TPROG = TestPrograms/test_arm_sha3.cxx
+    TOPT = -march=armv8.4-a+crypto
+    HAVE_OPT = $(shell $(CXX) $(CXXFLAGS) $(ACLE_FLAG) $(ZOPT) $(TOPT) $(TPROG) -o $(TOUT) 2>&1 | tr ' ' '\n' | wc -l)
+    ifeq ($(strip $(HAVE_OPT)),0)
+      SHA3_FLAG = -march=armv8.4-a+crypto
+      SHA512_FLAG = -march=armv8.4-a+crypto
+    endif
+
+# IS_ARMV8
+endif
+
+# DETECT_FEATURES
+endif
+
+# IS_ARM32, IS_ARMV8, IS_NEON
+endif
+
 ###########################################################
 #####                      Common                     #####
 ###########################################################
@@ -234,147 +570,23 @@ ifeq ($(findstring lean,$(MAKECMDGOALS)),lean)
   endif # MAKECMDGOALS
 endif # Dead code stripping
 
-# Pickup ARMv7 and NEON. Do it after Android, iOS and Embedded flags have been set.
-ifeq ($(IS_ARM32),1)
-  IS_ARMv7 := $(shell $(CXX) $(CXXFLAGS) -dM -E pch.cpp 2>&1 | $(EGREP) -i -c '__ARM_ARCH 7')
-  ifeq ($(IS_ARMv7),1)
-    IS_NEON := 1
-  else
-    IS_NEON := $(shell $(CXX) $(CXXFLAGS) -dM -E pch.cpp 2>&1 | $(EGREP) -i -c -E '\<__ARM_NEON\>')
-  endif
-endif
-
-# ARMv7-a
-ifeq ($(IS_ARMv7),1)
-  ifeq ($(findstring -march=armv7-a,$(CXXFLAGS)),)
-    NEON_FLAG = -march=armv7-a
-    GCM_FLAG = -march=armv7-a
-    ARIA_FLAG = -march=armv7-a
-    BLAKE2_FLAG = -march=armv7-a
-    CHACHA_FLAG = -march=armv7-a
-    LEA_FLAG = -march=armv7-a
-  endif
-endif
-
-# NEON
-ifeq ($(IS_NEON),1)
-  ifeq ($(findstring -mfpu=neon,$(CXXFLAGS)),)
-    NEON_FLAG += -mfpu=neon
-    GCM_FLAG += -mfpu=neon
-    ARIA_FLAG += -mfpu=neon
-    BLAKE2_FLAG += -mfpu=neon
-    CHACHA_FLAG += -mfpu=neon
-    LEA_FLAG += -mfpu=neon
-    SIMECK_FLAG += -mfpu=neon
-    SIMON64_FLAG += -mfpu=neon
-    SIMON128_FLAG += -mfpu=neon
-    SPECK64_FLAG += -mfpu=neon
-    SPECK128_FLAG += -mfpu=neon
-    ifeq ($(IS_ANDROID),1)
-      ifeq ($(findstring -mfloat-abi=softfp,$(CXXFLAGS)),)
-        NEON_FLAG += -mfloat-abi=softfp
-        GCM_FLAG += -mfloat-abi=softfp
-        ARIA_FLAG += -mfloat-abi=softfp
-        BLAKE2_FLAG += -mfloat-abi=softfp
-        CHACHA_FLAG += -mfloat-abi=softfp
-        LEA_FLAG += -mfloat-abi=softfp
-        SIMECK_FLAG += -mfloat-abi=softfp
-        SIMON64_FLAG += -mfloat-abi=softfp
-        SIMON128_FLAG += -mfloat-abi=softfp
-        SPECK64_FLAG += -mfloat-abi=softfp
-        SPECK128_FLAG += -mfloat-abi=softfp
-      endif
-    endif
-  endif
-endif
-
-# ARMv8-a
-ifneq ($(IS_ARMv8),0)
-  IS_NEON := $(shell $(CXX) $(CXXFLAGS) -march=armv8-a -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __ARM_NEON)
-  ifeq ($(IS_NEON),1)
-    ARIA_FLAG = -march=armv8-a
-    BLAKE2_FLAG = -march=armv8-a
-	CHACHA_FLAG = -march=armv8-a
-    LEA_FLAG = -march=armv8-a
-    NEON_FLAG = -march=armv8-a
-    SIMECK_FLAG = -march=armv8-a
-    SIMON64_FLAG = -march=armv8-a
-    SIMON128_FLAG = -march=armv8-a
-    SPECK64_FLAG = -march=armv8-a
-    SPECK128_FLAG = -march=armv8-a
-  endif
-  HAVE_CRC := $(shell $(CXX) $(CXXFLAGS) -march=armv8-a+crc -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __ARM_FEATURE_CRC32)
-  ifeq ($(HAVE_CRC),1)
-    CRC_FLAG = -march=armv8-a+crc
-  endif
-  HAVE_CRYPTO := $(shell $(CXX) $(CXXFLAGS) -march=armv8-a+crypto -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __ARM_FEATURE_CRYPTO)
-  ifeq ($(HAVE_CRYPTO),1)
-    AES_FLAG = -march=armv8-a+crypto
-    GCM_FLAG = -march=armv8-a+crypto
-    SHA_FLAG = -march=armv8-a+crypto
-  endif
-  HAVE_CRYPTO := $(shell $(CXX) $(CXXFLAGS) -march=armv8.4-a+crypto -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __ARM_FEATURE_CRYPTO)
-  ifeq ($(HAVE_CRYPTO),1)
-    SM4_FLAG = -march=armv8.4-a+crypto
-  endif
-endif
-
-# i686 and x86_64
-ifneq ($(IS_i686)$(IS_x86_64),00)
-# SSE2 is a core feature of x86_64
-ifeq ($(findstring -DCRYPTOPP_DISABLE_ASM,$(CXXFLAGS)),)
-  SSE_FLAG = -msse2
-  CHACHA_FLAG = -msse2
-endif
-  HAVE_SSSE3 = $(shell $(CXX) $(CXXFLAGS) -mssse3 -dM -E pch.cpp 2>&1 |  $(EGREP) -i -c __SSSE3__)
-  ifeq ($(HAVE_SSSE3),1)
-    ARIA_FLAG = -mssse3
-    CHAM_FLAG = -mssse3
-    LEA_FLAG = -mssse3
-    SSSE3_FLAG = -mssse3
-    SIMECK_FLAG = -mssse3
-    SIMON128_FLAG = -mssse3
-    SPECK128_FLAG = -mssse3
-  endif
-  HAVE_SSE4 = $(shell $(CXX) $(CXXFLAGS) -msse4.1 -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __SSE4_1__)
-  ifeq ($(HAVE_SSE4),1)
-    SIMON64_FLAG = -msse4.1
-    SPECK64_FLAG = -msse4.1
-  endif
-  HAVE_SSE4 = $(shell $(CXX) $(CXXFLAGS) -msse4.2 -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __SSE4_2__)
-  ifeq ($(HAVE_SSE4),1)
-    BLAKE2_FLAG = -msse4.2
-    CRC_FLAG = -msse4.2
-  endif
-  HAVE_CLMUL = $(shell $(CXX) $(CXXFLAGS) -mssse3 -mpclmul -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __PCLMUL__ )
-  ifeq ($(HAVE_CLMUL),1)
-    GCM_FLAG = -mssse3 -mpclmul
-  endif
-  HAVE_AES = $(shell $(CXX) $(CXXFLAGS) -msse4.1 -maes -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __AES__)
-  ifeq ($(HAVE_AES),1)
-    AES_FLAG = -msse4.1 -maes
-    SM4_FLAG = -mssse3 -maes
-  endif
-  HAVE_SHA = $(shell $(CXX) $(CXXFLAGS) -msse4.2 -msha -dM -E pch.cpp 2>&1 | $(EGREP) -i -c __SHA__)
-  ifeq ($(HAVE_SHA),1)
-    SHA_FLAG = -msse4.2 -msha
-  endif
-endif
-
 ###########################################################
 #####              Source and object files            #####
 ###########################################################
 
 # List cryptlib.cpp first, then cpu.cpp, then integer.cpp to tame C++ static initialization problems.
-SRCS := cryptlib.cpp cpu.cpp integer.cpp $(filter-out cryptlib.cpp cpu.cpp integer.cpp pch.cpp simple.cpp winpipes.cpp cryptlib_bds.cpp,$(sort $(wildcard *.cpp)))
+SRCS := cryptlib.cpp cpu.cpp integer.cpp $(filter-out cryptlib.cpp cpu.cpp integer.cpp pch.cpp simple.cpp cryptlib_bds.cpp,$(sort $(wildcard *.cpp)))
 # For Makefile.am; resource.h is Windows
 INCL := $(filter-out resource.h,$(sort $(wildcard *.h)))
 
 # Cryptogams AES for ARMv4 and above. We couple to ARMv7.
+# Avoid iOS. It cannot consume the assembly.
 ifeq ($(IS_ARM32),1)
-CRYPTOGAMS_AES_FLAG = -march=armv7-a
-CRYPTOGAMS_AES_FLAG += -Wa,--noexecstack
-SRCS += aes_armv4.S
+  ifneq ($(IS_IOS),1)
+    CRYPTOGAMS_AES_FLAG = -march=armv7-a
+    CRYPTOGAMS_AES_FLAG += -Wa,--noexecstack
+    SRCS += aes_armv4.S
+  endif
 endif
 
 # List cryptlib.cpp first, then cpu.cpp, then integer.cpp to tame C++ static initialization problems.
@@ -564,32 +776,44 @@ ifeq ($(wildcard GNUmakefile.deps),GNUmakefile.deps)
 -include GNUmakefile.deps
 endif # Dependencies
 
+# Cryptogams ARM asm implementation.
+aes_armv4.o : aes_armv4.S
+	$(CXX) $(strip $(CXXFLAGS) -fpermissive $(CRYPTOGAMS_AES_FLAG) -c) $<
+
 cpu-features.o: cpu-features.h cpu-features.c
 	$(CXX) $(strip $(CXXFLAGS) -fpermissive -c) cpu-features.c
 
-# Cryptogams ARM asm implementation.
-aes_armv4.o : aes_armv4.S
-	$(CXX) $(strip $(CXXFLAGS) $(CRYPTOGAMS_AES_FLAG) -c) $<
-
-# SSE4.2 or NEON available
+# SSSE3 or NEON available
 aria_simd.o : aria_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(ARIA_FLAG) -c) $<
 
-# SSE4.2 or ARMv8a available
-blake2_simd.o : blake2-simd.cpp
-	$(CXX) $(strip $(CXXFLAGS) $(BLAKE2_FLAG) -c) $<
+# SSE, NEON or POWER7 available
+blake2s_simd.o : blake2s_simd.cpp
+	$(CXX) $(strip $(CXXFLAGS) $(BLAKE2S_FLAG) -c) $<
+
+# SSE, NEON or POWER8 available
+blake2b_simd.o : blake2b_simd.cpp
+	$(CXX) $(strip $(CXXFLAGS) $(BLAKE2B_FLAG) -c) $<
 
 # SSE2 or NEON available
 chacha_simd.o : chacha_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(CHACHA_FLAG) -c) $<
 
+# AVX2 available
+chacha_avx.o : chacha_avx.cpp
+	$(CXX) $(strip $(CXXFLAGS) $(CHACHA_AVX2_FLAG) -c) $<
+
 # SSSE3 available
 cham_simd.o : cham_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(CHAM_FLAG) -c) $<
 
+# Power9 available
+darn.o : darn.cpp
+	$(CXX) $(strip $(CXXFLAGS) $(DARN_FLAG) -c) $<
+
 # SSE2 on i586
-cpu.o : cpu.cpp
-	$(CXX) $(strip $(CXXFLAGS) $(CPU_FLAG) -c) $<
+sse_simd.o : sse_simd.cpp
+	$(CXX) $(strip $(CXXFLAGS) $(SSE2_FLAG) -c) $<
 
 # SSE4.2 or ARMv8a available
 crc_simd.o : crc_simd.cpp
@@ -599,7 +823,7 @@ crc_simd.o : crc_simd.cpp
 gcm_simd.o : gcm_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(GCM_FLAG) -c) $<
 
-# SSSE3 or ARMv8a available
+# SSSE3 available
 lea_simd.o : lea_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(LEA_FLAG) -c) $<
 
@@ -639,6 +863,10 @@ speck64_simd.o : speck64_simd.cpp
 speck128_simd.o : speck128_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(SPECK128_FLAG) -c) $<
 
+# ARMv8.4 available
+sm3_simd.o : sm3_simd.cpp
+	$(CXX) $(strip $(CXXFLAGS) $(SM3_FLAG) -c) $<
+
 # AESNI available
 sm4_simd.o : sm4_simd.cpp
 	$(CXX) $(strip $(CXXFLAGS) $(SM4_FLAG) -c) $<
@@ -649,4 +877,3 @@ sm4_simd.o : sm4_simd.cpp
 .PHONY: dep deps depend
 dep deps depend GNUmakefile.deps:
 	$(CXX) $(strip $(CXXFLAGS) -DCRYPTOPP_DISABLE_ASM) -MM *.cpp > GNUmakefile.deps
-