arch: add NEON cpu feature detection

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Dan Mick <dan.mick@inktank.com>

configure.ac

@@ -106,6 +106,9 @@ AC_DEFUN([AC_CHECK_CC_FLAG],
 AC_CHECK_CC_FLAG([-Wtype-limits], [WARN_TYPE_LIMITS])
 AC_CHECK_CC_FLAG([-Wignored-qualifiers], [WARN_IGNORED_QUALIFIERS])
 
+# Checks for architecture stuff
+AM_CONDITIONAL([ENABLE_FPU_NEON], [case $target_cpu in arm*) true;; *) false;; esac])
+
 # Checks for libraries.
 ACX_PTHREAD
 AC_CHECK_LIB([uuid], [uuid_parse], [true], AC_MSG_FAILURE([libuuid not found]))

src/Makefile.am

@@ -775,6 +775,10 @@ unittest_addrs_CXXFLAGS = ${AM_CXXFLAGS} ${UNITTEST_CXXFLAGS}
 unittest_addrs_LDADD = $(LIBGLOBAL_LDA) ${UNITTEST_LDADD}
 check_PROGRAMS += unittest_addrs
 
+unittest_arch_SOURCES = test/test_arch.c arch/intel.c arch/neon.c arch/probe.cc
+unittest_arch_CXXFLAGS = ${AM_CFLAGS}
+check_PROGRAMS += unittest_arch
+
 unittest_sharedptr_registry_SOURCES = test/common/test_sharedptr_registry.cc
 unittest_sharedptr_registry_CXXFLAGS = ${AM_CXXFLAGS} ${UNITTEST_CXXFLAGS}
 unittest_sharedptr_registry_LDADD = libcommon.la ${LIBGLOBAL_LDA} ${UNITTEST_LDADD}
@@ -1340,6 +1344,11 @@ AM_CXXFLAGS = \
 	-Wnon-virtual-dtor \
 	-Wno-invalid-offsetof
 
+if ENABLE_FPU_NEON
+AM_CFLAGS += -mfpu=neon
+AM_CXXFLAGS += -mfpu=neon
+endif
+
 if !CLANG
     AM_CXXFLAGS += -Wstrict-null-sentinel
 endif
@@ -1521,6 +1530,7 @@ libcommon_files = \
 	./ceph_ver.c \
 	arch/probe.cc \
 	arch/intel.c \
+	arch/neon.c \
 	auth/AuthAuthorizeHandler.cc \
 	auth/AuthClientHandler.cc \
 	auth/AuthSessionHandler.cc \
@@ -1772,6 +1782,7 @@ noinst_HEADERS = \
 	rados_sync.h \
 	arch/probe.h \
 	arch/intel.h \
+	arch/neon.h \
 	auth/cephx/CephxAuthorizeHandler.h\
 	auth/cephx/CephxKeyServer.h\
 	auth/cephx/CephxProtocol.h\

src/arch/neon.c

@@ -0,0 +1,51 @@
+#include "arch/probe.h"
+
+/* flags we export */
+int ceph_arch_neon = 0;
+
+#include <stdio.h>
+
+#if __linux__
+
+#include <elf.h>
+#include <link.h> // ElfW macro
+
+#if __arm__
+#include <asm/hwcap.h>
+#endif // __arm__
+
+static unsigned long get_auxval(unsigned long type)
+{
+	unsigned long result = 0;
+	FILE *f = fopen("/proc/self/auxv", "r");
+	if (f) {
+		ElfW(auxv_t) entry;
+		while (fread(&entry, sizeof(entry), 1, f)) {
+			if (entry.a_type == type) {
+				result = entry.a_un.a_val;
+				break;
+			}
+		}
+		fclose(f);
+	}
+	return result;
+}
+
+static unsigned long get_hwcap(void)
+{
+	return get_auxval(AT_HWCAP);
+}
+
+#endif // __linux__
+
+int ceph_arch_neon_probe(void)
+{
+#if __arm__ && __linux__
+	ceph_arch_neon = (get_hwcap() & HWCAP_NEON) == HWCAP_NEON;
+#else
+	if (0)
+		get_hwcap();  // make compiler shut up
+#endif
+	return 0;
+}
+

src/arch/neon.h

@@ -0,0 +1,16 @@
+#ifndef CEPH_ARCH_NEON_H
+#define CEPH_ARCH_NEON_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern int ceph_arch_neon;  /* true if we have ARM NEON abilities */
+
+extern int ceph_arch_neon_probe(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

src/arch/probe.cc

@@ -4,6 +4,7 @@
 #include "arch/probe.h"
 
 #include "arch/intel.h"
+#include "arch/neon.h"
 
 int ceph_arch_probe(void)
 {
@@ -11,6 +12,7 @@ int ceph_arch_probe(void)
     return 1;
 
   ceph_arch_intel_probe();
+  ceph_arch_neon_probe();
 
   ceph_arch_probed = 1;
   return 1;

src/test/test_arch.c

@@ -0,0 +1,19 @@
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+
+#include "arch/probe.h"
+#include "arch/intel.h"
+#include "arch/neon.h"
+
+int main(int argc, char **argv)
+{
+	ceph_arch_probe();
+	assert(ceph_arch_probed);
+
+	printf("ceph_arch_intel_sse42 = %d\n", ceph_arch_intel_sse42);
+	printf("ceph_arch_neon = %d\n", ceph_arch_neon);
+
+	return 0;
+}