diff --git a/libc/Android.mk b/libc/Android.mk
index 9881d59..2628507 100644
--- a/libc/Android.mk
+++ b/libc/Android.mk
@@ -471,6 +471,14 @@
   ifeq ($(ARCH_ARM_HAVE_TLS_REGISTER),true)
     libc_common_cflags += -DHAVE_ARM_TLS_REGISTER
   endif
+  #
+  # Define HAVE_32_BYTE_CACHE_LINES to indicate to C
+  # library it should use to 32-byte version of memcpy, and not
+  # the 64-byte version.
+  #
+  ifeq ($(ARCH_ARM_HAVE_32_BYTE_CACHE_LINES),true)
+    libc_common_cflags += -DHAVE_32_BYTE_CACHE_LINE
+  endif
 else # !arm
   ifeq ($(TARGET_ARCH),x86)
     libc_crt_target_cflags :=
diff --git a/libc/arch-arm/bionic/memcpy.S b/libc/arch-arm/bionic/memcpy.S
index 438fa00..815b5f6 100644
--- a/libc/arch-arm/bionic/memcpy.S
+++ b/libc/arch-arm/bionic/memcpy.S
@@ -34,23 +34,28 @@
         .text
         .fpu    neon
 
+#ifdef HAVE_32_BYTE_CACHE_LINE
+/* a prefetch distance of 2 cache-lines */
+#define CACHE_LINE_SIZE     32
+#define PREFETCH_DISTANCE   (CACHE_LINE_SIZE*2)
+#else
 /* a prefetch distance of 4 cache-lines works best experimentally */
 #define CACHE_LINE_SIZE     64
 #define PREFETCH_DISTANCE   (CACHE_LINE_SIZE*4)
+#endif
 
 ENTRY(memcpy)
         .save       {r0, lr}
-        stmfd       sp!, {r0, lr}
-
         /* start preloading as early as possible */
         pld         [r1, #(CACHE_LINE_SIZE*0)]
+        stmfd       sp!, {r0, lr}
         pld         [r1, #(CACHE_LINE_SIZE*1)]
 
         /* do we have at least 16-bytes to copy (needed for alignment below) */
         cmp         r2, #16
         blo         5f
 
-        /* align destination to half cache-line for the write-buffer */
+        /* align destination to cache-line for the write-buffer */
         rsb         r3, r0, #0
         ands        r3, r3, #0xF
         beq         0f
@@ -79,6 +84,26 @@
         pld         [r1, #(CACHE_LINE_SIZE*0)]
         pld         [r1, #(CACHE_LINE_SIZE*1)]
 
+#ifdef HAVE_32_BYTE_CACHE_LINE
+        /* make sure we have at least 32 bytes to copy */
+        subs        r2, r2, #32
+        blo         4f
+
+        /* preload all the cache lines we need.
+         * NOTE: the number of pld below depends on PREFETCH_DISTANCE,
+         * ideally would would increase the distance in the main loop to
+         * avoid the goofy code below. In practice this doesn't seem to make
+         * a big difference.
+         */
+        pld         [r1, #(PREFETCH_DISTANCE)]
+
+1:      /* The main loop copies 32 bytes at a time */
+        vld1.8      {d0  - d3},   [r1]!
+        pld         [r1, #(PREFETCH_DISTANCE)]
+        subs        r2, r2, #32
+        vst1.8      {d0  - d3},   [r0, :128]!
+        bhs         1b
+#else
         /* make sure we have at least 64 bytes to copy */
         subs        r2, r2, #64
         blo         2f
@@ -112,7 +137,7 @@
         subs        r2, r2, #32
         vst1.8      {d0 - d3},  [r0, :128]!
         bhs         3b
-
+#endif
 4:      /* less than 32 left */
         add         r2, r2, #32
         tst         r2, #0x10