From 1f878636c85edaf4c99bdf599e54d2e4c46297b4 Mon Sep 17 00:00:00 2001
From: Ayan Shafqat <ayan.x.shafqat@gmail.com>
Date: Sun, 23 Aug 2020 14:27:45 -0400
Subject: [PATCH] Adding NEON detection during runtime

---
 libs/ardour/globals.cc |  4 ++--
 libs/pbd/fpu.cc        | 13 +++++++++++++
 libs/pbd/pbd/fpu.h     |  4 +++-
 3 files changed, 18 insertions(+), 3 deletions(-)

diff --git a/libs/ardour/globals.cc b/libs/ardour/globals.cc
index ae33ab32b3..8a7391f5f9 100644
--- a/libs/ardour/globals.cc
+++ b/libs/ardour/globals.cc
@@ -219,7 +219,7 @@ setup_hardware_optimization (bool try_optimization)
 
 #elif defined ARM_NEON_SUPPORT
 		/* Use NEON routines */
-		do {
+		if (fpu->has_neon ()) {
 			info << "Using ARM NEON optimized routines" << endmsg;
 
 			compute_peak          = arm_neon_compute_peak;
@@ -230,7 +230,7 @@ setup_hardware_optimization (bool try_optimization)
 			copy_vector           = arm_neon_copy_vector;
 
 			generic_mix_functions = false;
-		} while (0);
+		}
 
 #elif defined(__APPLE__) && defined(BUILD_VECLIB_OPTIMIZATIONS)
 
diff --git a/libs/pbd/fpu.cc b/libs/pbd/fpu.cc
index 206ce2499f..9ad7cc94e9 100644
--- a/libs/pbd/fpu.cc
+++ b/libs/pbd/fpu.cc
@@ -31,6 +31,12 @@
 #include <intrin.h>
 #endif
 
+#ifdef ARM_NEON_SUPPORT
+/* Needed for ARM NEON detection */
+#include <sys/auxv.h>
+#include <asm/hwcap.h>
+#endif
+
 #include "pbd/compose.h"
 #include "pbd/fpu.h"
 #include "pbd/error.h"
@@ -155,6 +161,13 @@ FPU::FPU ()
 		return;
 	}
 
+
+#ifdef ARM_NEON_SUPPORT
+	if (getauxval(AT_HWCAP) & HWCAP_NEON) {
+		_flags = Flags(_flags & HasNEON);
+	}
+#endif
+
 #if !( (defined __x86_64__) || (defined __i386__) || (defined _M_X64) || (defined _M_IX86) ) // !ARCH_X86
 	/* Non-Intel architecture, nothing to do here */
 	return;
diff --git a/libs/pbd/pbd/fpu.h b/libs/pbd/pbd/fpu.h
index 91992f77c0..b0979daa9e 100644
--- a/libs/pbd/pbd/fpu.h
+++ b/libs/pbd/pbd/fpu.h
@@ -31,7 +31,8 @@ class LIBPBD_API FPU {
 		HasDenormalsAreZero = 0x2,
 		HasSSE = 0x4,
 		HasSSE2 = 0x8,
-		HasAVX = 0x10
+		HasAVX = 0x10,
+		HasNEON = 0x20,
 	};
 
   public:
@@ -45,6 +46,7 @@ class LIBPBD_API FPU {
 	bool has_sse () const { return _flags & HasSSE; }
 	bool has_sse2 () const { return _flags & HasSSE2; }
 	bool has_avx () const { return _flags & HasAVX; }
+	bool has_neon () const { return _flags & HasNEON; }
 
   private:
 	Flags _flags;