aboutsummaryrefslogtreecommitdiffhomepage
path: root/Eigen/src/Core/arch/NEON/PacketMath.h
diff options
context:
space:
mode:
authorGravatar Benoit Jacob <benoitjacob@google.com>2017-03-15 06:57:25 -0400
committerGravatar Benoit Jacob <benoitjacob@google.com>2017-03-15 06:57:25 -0400
commit61160a21d23880749bce3b0d630d9880f70af6e5 (patch)
tree9d3b2d770f2b183156e460f10db96bcd5768c5dc /Eigen/src/Core/arch/NEON/PacketMath.h
parentf0f35911181cc7e2089a3319d966dcbd3596461b (diff)
ARM prefetch fixes: Implement prefetch on ARM64. Do not clobber cc on ARM32.
Diffstat (limited to 'Eigen/src/Core/arch/NEON/PacketMath.h')
-rw-r--r--Eigen/src/Core/arch/NEON/PacketMath.h13
1 files changed, 8 insertions, 5 deletions
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 84a56bdcc..aede4a6d5 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -51,14 +51,17 @@ typedef uint32x4_t Packet4ui;
#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
const Packet4i p4i_##NAME = pset1<Packet4i>(X)
-// arm64 does have the pld instruction. If available, let's trust the __builtin_prefetch built-in function
-// which available on LLVM and GCC (at least)
-#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
+#if EIGEN_ARCH_ARM64
+ // __builtin_prefetch tends to do nothing on ARM64 compilers because the
+ // prefetch instructions there are too detailed for __builtin_prefetch to map
+ // meaningfully to them.
+ #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__("prfm pldl1keep, [%[addr]]\n" ::[addr] "r"(ADDR) : );
+#elif EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
#define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR);
#elif defined __pld
#define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR)
-#elif !EIGEN_ARCH_ARM64
- #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
+#elif EIGEN_ARCH_ARM32
+ #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ("pld [%[addr]]\n" :: [addr] "r" (ADDR) : );
#else
// by default no explicit prefetching
#define EIGEN_ARM_PREFETCH(ADDR)