Add optimized SHA1 and SHA256 using ARMv8 Crypto Intrinsics.

ext/digest/arm_crypto.h

@@ -0,0 +1,88 @@
+/* arm_crypto.h - runtime detection of the ARMv8 Cryptographic Extension */
+#ifndef RB_DIGEST_ARM_CRYPTO_H
+#define RB_DIGEST_ARM_CRYPTO_H
+
+#if defined(HAVE_ARM_CRYPTO_EXT)
+
+#if defined(__APPLE__)
+# include <sys/sysctl.h>
+#elif defined(__linux__) || defined(__ANDROID__)
+# include <sys/auxv.h>
+/* Not every libc/kernel-header combination defines these (e.g. older
+ * <asm/hwcap.h> layouts), so fall back to the values Linux has used for
+ * aarch64 since they were introduced. */
+# ifndef HWCAP_SHA1
+#  define HWCAP_SHA1 (1 << 5)
+# endif
+# ifndef HWCAP_SHA2
+#  define HWCAP_SHA2 (1 << 6)
+# endif
+#endif
+
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && \
+    defined(__has_include) && __has_include(<stdatomic.h>) && !defined(_MSC_VER)
+# include <stdatomic.h>
+# define RB_DIGEST_ARM_ATOMIC_INT _Atomic int
+#else
+# define RB_DIGEST_ARM_ATOMIC_INT int
+#endif
+
+#if defined(__APPLE__)
+static inline int
+rb_digest_arm_sysctl_enabled(const char *name)
+{
+    int enabled = 0;
+    size_t size = sizeof(enabled);
+    if (sysctlbyname(name, &enabled, &size, NULL, 0) != 0) return 0;
+    return enabled != 0;
+}
+#endif
+
+static inline int
+rb_digest_have_arm_sha1(void)
+{
+    enum { UNKNOWN = -1, NO = 0, YES = 1 };
+    static RB_DIGEST_ARM_ATOMIC_INT cached = UNKNOWN;
+    int state = cached;
+
+    if (state == UNKNOWN) {
+#if defined(__APPLE__)
+        state = rb_digest_arm_sysctl_enabled("hw.optional.arm.FEAT_SHA1") ? YES : NO;
+#elif defined(__linux__) || defined(__ANDROID__)
+        state = (getauxval(AT_HWCAP) & HWCAP_SHA1) ? YES : NO;
+#else
+        state = NO;
+#endif
+        cached = state;
+    }
+    return state == YES;
+}
+
+static inline int
+rb_digest_have_arm_sha256(void)
+{
+    enum { UNKNOWN = -1, NO = 0, YES = 1 };
+    static RB_DIGEST_ARM_ATOMIC_INT cached = UNKNOWN;
+    int state = cached;
+
+    if (state == UNKNOWN) {
+#if defined(__APPLE__)
+        state = rb_digest_arm_sysctl_enabled("hw.optional.arm.FEAT_SHA256") ? YES : NO;
+#elif defined(__linux__) || defined(__ANDROID__)
+        state = (getauxval(AT_HWCAP) & HWCAP_SHA2) ? YES : NO;
+#else
+        state = NO;
+#endif
+        cached = state;
+    }
+    return state == YES;
+}
+
+#else /* !HAVE_ARM_CRYPTO_EXT */
+
+static inline int rb_digest_have_arm_sha1(void) { return 0; }
+static inline int rb_digest_have_arm_sha256(void) { return 0; }
+
+#endif /* HAVE_ARM_CRYPTO_EXT */
+
+#endif /* RB_DIGEST_ARM_CRYPTO_H */

ext/digest/digest_conf.rb

@@ -17,3 +17,21 @@ def digest_conf(name)
   $objs << "#{name}.#{$OBJEXT}"
   return
 end
+
+# Check whether this compiler on an aarch64/arm64 host can build +source+. Will
+# try with no additional cflags and also architecture specific flags when
+# compiling +source+.
+# 
+# Returns the flag that worked (nil if none was needed), or false if no
+# candidate compiled or the host isn't aarch64/arm64.
+def arm_crypto_flag(source)
+  return false unless RbConfig::CONFIG["host_cpu"] =~ /\A(aarch64|arm64)\z/i
+
+  [nil, "-march=armv8-a+crypto"].each do |flag|
+    label = "ARMv8 Crypto Extensions intrinsics" + (flag ? " (#{flag})" : "")
+    if checking_for(label) { try_compile(%{#include "#{$srcdir}/#{source}"\n}, flag) }
+      return flag
+    end
+  end
+  false
+end

ext/digest/sha1/extconf.rb

@@ -8,10 +8,42 @@
 
 $objs = [ "sha1init.#{$OBJEXT}" ]
 
-digest_conf("sha1")
+bundled = !digest_conf("sha1")
 
 have_header("sys/cdefs.h")
 
+# On aarch64/arm64, build the ARMv8 Crypto Extensions backend and let
+# SHA1_Transform() pick it at runtime when the CPU actually has the extension.
+# 
+# $CFLAGS can't carry the extra instruction-set flag for just this one file, 
+# so give it its own compile rule below when one is needed. This only applies
+# when CommonCrypto is not available.
+arm_cflags = {}
+if bundled && with_config("arm-crypto", true)
+  case flag = arm_crypto_flag("sha1-arm.c")
+  when false
+    # Not aarch64/arm64, or the toolchain can't build the intrinsics: fall
+    # back to the portable C implementation only.
+  else
+    $objs << "sha1-arm.#{$OBJEXT}"
+    $defs << "-DHAVE_ARM_CRYPTO_EXT"
+    arm_cflags["sha1-arm"] = flag if flag
+  end
+end
+
 $preload = %w[digest]
 
 create_makefile("digest/sha1")
+
+unless arm_cflags.empty?
+  File.open("Makefile", "a") do |mf|
+    mf.puts
+    mf.puts "# Per-file instruction-set flags for the ARMv8 Crypto Extensions backend."
+    arm_cflags.each do |obj, cflag|
+      target = "#{obj}.#{$OBJEXT}"
+      mf.puts "#{target}: $(srcdir)/#{obj}.c"
+      mf.puts "\t$(ECHO) compiling #{obj}.c"
+      mf.puts "\t$(Q) $(CC) $(INCFLAGS) $(CPPFLAGS) $(CFLAGS) #{cflag} $(COUTFLAG)$@ -c $(CSRCFLAG)$(srcdir)/#{obj}.c"
+    end
+  end
+end