gh-123748: Add conditional compilation rules to allow HACL SIMD256 to be compiled on universal

Makefile.pre.in

@@ -1387,9 +1387,15 @@ Modules/_hacl/Hacl_Hash_Blake2b.o: $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2b.c $
 Modules/_hacl/Hacl_Hash_Blake2s_Simd128.o: $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2s_Simd128.c $(LIBHACL_BLAKE2_HEADERS)
 	$(CC) -c $(LIBHACL_CFLAGS) $(LIBHACL_SIMD128_FLAGS) -DHACL_CAN_COMPILE_VEC128 -o $@ $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2s_Simd128.c
 
+Modules/_hacl/Hacl_Hash_Blake2s_Simd128_universal2.o: $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2s_Simd128_universal2.c $(LIBHACL_BLAKE2_HEADERS)
+	$(CC) -c $(LIBHACL_CFLAGS) $(LIBHACL_SIMD128_FLAGS) -DHACL_CAN_COMPILE_VEC128 -o $@ $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2s_Simd128_universal2.c
+
 Modules/_hacl/Hacl_Hash_Blake2b_Simd256.o: $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2b_Simd256.c $(LIBHACL_BLAKE2_HEADERS)
 	$(CC) -c $(LIBHACL_CFLAGS) $(LIBHACL_SIMD256_FLAGS) -DHACL_CAN_COMPILE_VEC256 -o $@ $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2b_Simd256.c
 
+Modules/_hacl/Hacl_Hash_Blake2b_Simd256_universal2.o: $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2b_Simd256_universal2.c $(LIBHACL_BLAKE2_HEADERS)
+	$(CC) -c $(LIBHACL_CFLAGS) $(LIBHACL_SIMD256_FLAGS) -DHACL_CAN_COMPILE_VEC256 -o $@ $(srcdir)/Modules/_hacl/Hacl_Hash_Blake2b_Simd256_universal2.c
+
 Modules/_hacl/Lib_Memzero0.o: $(srcdir)/Modules/_hacl/Lib_Memzero0.c $(LIBHACL_BLAKE2_HEADERS)
 	$(CC) -c $(LIBHACL_CFLAGS) -o $@ $(srcdir)/Modules/_hacl/Lib_Memzero0.c
 

Modules/_hacl/Hacl_Hash_Blake2b_Simd256_universal2.c

@@ -0,0 +1,15 @@
+// This file isn't part of a standard HACL source tree.
+//
+// It is required for compatibility with universal2 macOS builds. The code in
+// Hacl_Hash_Blake2b_Simd256.c *will* compile on macOS x86_64, but *won't*
+// compile on ARM64. However, because universal2 builds are compiled in a
+// single pass, autoconf detects that the required compiler features *are*
+// available, and tries to compile this file, which then fails because of the
+// lack of support on ARM64.
+//
+// To compensate for this, autoconf will include *this* file instead of
+// Hacl_Hash_Blake2b_Simd256.c when compiling for universal. This allows the
+// underlying source code of HACL to remain unmodified.
+#if !(defined(__APPLE__) && defined(__arm64__))
+#include "Hacl_Hash_Blake2b_Simd256.c"
+#endif

Modules/_hacl/Hacl_Hash_Blake2s_Simd128_universal2.c

@@ -0,0 +1,14 @@
+// This file isn't part of a standard HACL source tree.
+//
+// It is required for compatibility with universal2 macOS builds. The code in
+// Hacl_Hash_Blake2s_Simd128.c will compile on macOS ARM64, but performance
+// isn't great, so it's disabled. However, because universal2 builds are
+// compiled in a single pass, autoconf detects that the required compiler
+// features *are* available, and tries to include this file.
+//
+// To compensate for this, autoconf will include *this* file instead of
+// Hacl_Hash_Blake2s_Simd128.c when compiling for universal. This allows the
+// underlying source code of HACL to remain unmodified.
+#if !(defined(__APPLE__) && defined(__arm64__))
+#include "Hacl_Hash_Blake2s_Simd128.c"
+#endif

Modules/blake2module.c

@@ -41,6 +41,16 @@
 
 #include <stdbool.h>
 
+// SIMD256 can't be compiled on macOS ARM64, and performance of SIMD128 isn't
+// great; but when compiling a universal2 binary, autoconf will set
+// HACL_CAN_COMPILE_SIMD128 and HACL_CAN_COMPILE_SIMD256 because they *can* be
+// compiled on x86_64. If we're on macOS ARM64, disable these preprocessor
+// symbols.
+#if defined(__APPLE__) && defined(__arm64__)
+#  undef HACL_CAN_COMPILE_SIMD128
+#  undef HACL_CAN_COMPILE_SIMD256
+#endif
+
 // ECX
 #define ECX_SSE3 (1 << 0)
 #define ECX_SSSE3 (1 << 9)

configure.ac

@@ -7805,17 +7805,42 @@ AC_SUBST([LIBHACL_CFLAGS])
 dnl This can be extended here to detect e.g. Power8, which HACL* should also support.
 AX_CHECK_COMPILE_FLAG([-msse -msse2 -msse3 -msse4.1 -msse4.2],[
   [LIBHACL_SIMD128_FLAGS="-msse -msse2 -msse3 -msse4.1 -msse4.2"]
-  [LIBHACL_SIMD128_OBJS="Modules/_hacl/Hacl_Hash_Blake2s_Simd128.o"]
+
   AC_DEFINE([HACL_CAN_COMPILE_SIMD128], [1], [HACL* library can compile SIMD128 implementations])
+
+  # macOS universal2 builds *support* the -msse etc flags because they're
+  # available on x86_64. However, performance of the HACL SIMD128 implementation
+  # isn't great, so it's disabled on ARM64.
+  AC_MSG_CHECKING([for HACL* SIMD128 implementation])
+  if test "$UNIVERSAL_ARCHS" == "universal2"; then
+    [LIBHACL_SIMD128_OBJS="Modules/_hacl/Hacl_Hash_Blake2s_Simd128_universal2.o"]
+    AC_MSG_RESULT([universal2])
+  else
+    [LIBHACL_SIMD128_OBJS="Modules/_hacl/Hacl_Hash_Blake2s_Simd128.o"]
+    AC_MSG_RESULT([standard])
+  fi
+
 ], [], [-Werror])
 
 AC_SUBST([LIBHACL_SIMD128_FLAGS])
 AC_SUBST([LIBHACL_SIMD128_OBJS])
 
 AX_CHECK_COMPILE_FLAG([-mavx2],[
   [LIBHACL_SIMD256_FLAGS="-mavx2"]
-  [LIBHACL_SIMD256_OBJS="Modules/_hacl/Hacl_Hash_Blake2b_Simd256.o"]
   AC_DEFINE([HACL_CAN_COMPILE_SIMD256], [1], [HACL* library can compile SIMD256 implementations])
+
+  # macOS universal2 builds *support* the -mavx2 compiler flag because it's
+  # available on x86_64; but the HACL SIMD256 build then fails because the
+  # implementation requires symbols that aren't available on ARM64. Use a
+  # wrapped implementation if we're building for universal2.
+  AC_MSG_CHECKING([for HACL* SIMD256 implementation])
+  if test "$UNIVERSAL_ARCHS" == "universal2"; then
+    [LIBHACL_SIMD256_OBJS="Modules/_hacl/Hacl_Hash_Blake2b_Simd256_universal2.o"]
+    AC_MSG_RESULT([universal2])
+  else
+    [LIBHACL_SIMD256_OBJS="Modules/_hacl/Hacl_Hash_Blake2b_Simd256.o"]
+    AC_MSG_RESULT([standard])
+  fi
 ], [], [-Werror])
 
 AC_SUBST([LIBHACL_SIMD256_FLAGS])