Port libcutils memset16/32 to x86_64.

Port libcutils memset16/32 assembly SSE2 optimizations to x86_64
architecture. Ensures the same performance on 64-bit arch.

Change-Id: I874a71a884c0d28a152933ddff9cb886c9a6e99e
Signed-off-by: Henrik Smiding <henrik.smiding@intel.com>

3 files changed, 967 insertions, 0 deletions

diff --git a/libcutils/arch-x86_64/android_memset16_SSE2-atom.S b/libcutils/arch-x86_64/android_memset16_SSE2-atom.S
new file mode 100644
index 0000000..48a10ed
--- /dev/null
+++ b/libcutils/arch-x86_64/android_memset16_SSE2-atom.S
@@ -0,0 +1,564 @@
+/*
+ * Copyright (C) 2014 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+/*
+ * Contributed by: Intel Corporation
+ */
+
+#include "cache.h"
+
+#ifndef L
+# define L(label)	.L##label
+#endif
+
+#ifndef ALIGN
+# define ALIGN(n)	.p2align n
+#endif
+
+#ifndef cfi_startproc
+# define cfi_startproc			.cfi_startproc
+#endif
+
+#ifndef cfi_endproc
+# define cfi_endproc			.cfi_endproc
+#endif
+
+#ifndef ENTRY
+# define ENTRY(name)			\
+	.type name,  @function; 	\
+	.globl name;			\
+	.p2align 4;			\
+name:					\
+	cfi_startproc
+#endif
+
+#ifndef END
+# define END(name)			\
+	cfi_endproc;			\
+	.size name, .-name
+#endif
+
+#define JMPTBL(I, B)	I - B
+
+/* Branch to an entry in a jump table.  TABLE is a jump table with
+   relative offsets.  INDEX is a register contains the index into the
+   jump table.  SCALE is the scale of INDEX.  */
+#define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \
+	lea    TABLE(%rip), %r11;						\
+	movslq (%r11, INDEX, SCALE), INDEX;				\
+	lea    (%r11, INDEX), INDEX;					\
+	jmp    *INDEX
+
+	.section .text.sse2,"ax",@progbits
+	ALIGN (4)
+ENTRY (android_memset16)	// Address in rdi
+	shr    $1, %rdx			// Count in rdx
+	movzwl %si, %ecx
+	/* Fill the whole ECX with pattern.  */
+	shl    $16, %esi
+	or     %esi, %ecx		// Pattern in ecx
+
+	cmp    $32, %rdx
+	jae    L(32wordsormore)
+
+L(write_less32words):
+	lea    (%rdi, %rdx, 2), %rdi
+	BRANCH_TO_JMPTBL_ENTRY (L(table_less32words), %rdx, 4)
+
+	.pushsection .rodata.sse2,"a",@progbits
+	ALIGN (2)
+L(table_less32words):
+	.int	JMPTBL (L(write_0words), L(table_less32words))
+	.int	JMPTBL (L(write_1words), L(table_less32words))
+	.int	JMPTBL (L(write_2words), L(table_less32words))
+	.int	JMPTBL (L(write_3words), L(table_less32words))
+	.int	JMPTBL (L(write_4words), L(table_less32words))
+	.int	JMPTBL (L(write_5words), L(table_less32words))
+	.int	JMPTBL (L(write_6words), L(table_less32words))
+	.int	JMPTBL (L(write_7words), L(table_less32words))
+	.int	JMPTBL (L(write_8words), L(table_less32words))
+	.int	JMPTBL (L(write_9words), L(table_less32words))
+	.int	JMPTBL (L(write_10words), L(table_less32words))
+	.int	JMPTBL (L(write_11words), L(table_less32words))
+	.int	JMPTBL (L(write_12words), L(table_less32words))
+	.int	JMPTBL (L(write_13words), L(table_less32words))
+	.int	JMPTBL (L(write_14words), L(table_less32words))
+	.int	JMPTBL (L(write_15words), L(table_less32words))
+	.int	JMPTBL (L(write_16words), L(table_less32words))
+	.int	JMPTBL (L(write_17words), L(table_less32words))
+	.int	JMPTBL (L(write_18words), L(table_less32words))
+	.int	JMPTBL (L(write_19words), L(table_less32words))
+	.int	JMPTBL (L(write_20words), L(table_less32words))
+	.int	JMPTBL (L(write_21words), L(table_less32words))
+	.int	JMPTBL (L(write_22words), L(table_less32words))
+	.int	JMPTBL (L(write_23words), L(table_less32words))
+	.int	JMPTBL (L(write_24words), L(table_less32words))
+	.int	JMPTBL (L(write_25words), L(table_less32words))
+	.int	JMPTBL (L(write_26words), L(table_less32words))
+	.int	JMPTBL (L(write_27words), L(table_less32words))
+	.int	JMPTBL (L(write_28words), L(table_less32words))
+	.int	JMPTBL (L(write_29words), L(table_less32words))
+	.int	JMPTBL (L(write_30words), L(table_less32words))
+	.int	JMPTBL (L(write_31words), L(table_less32words))
+	.popsection
+
+	ALIGN (4)
+L(write_28words):
+	movl   %ecx, -56(%rdi)
+	movl   %ecx, -52(%rdi)
+L(write_24words):
+	movl   %ecx, -48(%rdi)
+	movl   %ecx, -44(%rdi)
+L(write_20words):
+	movl   %ecx, -40(%rdi)
+	movl   %ecx, -36(%rdi)
+L(write_16words):
+	movl   %ecx, -32(%rdi)
+	movl   %ecx, -28(%rdi)
+L(write_12words):
+	movl   %ecx, -24(%rdi)
+	movl   %ecx, -20(%rdi)
+L(write_8words):
+	movl   %ecx, -16(%rdi)
+	movl   %ecx, -12(%rdi)
+L(write_4words):
+	movl   %ecx, -8(%rdi)
+	movl   %ecx, -4(%rdi)
+L(write_0words):
+	ret
+
+	ALIGN (4)
+L(write_29words):
+	movl   %ecx, -58(%rdi)
+	movl   %ecx, -54(%rdi)
+L(write_25words):
+	movl   %ecx, -50(%rdi)
+	movl   %ecx, -46(%rdi)
+L(write_21words):
+	movl   %ecx, -42(%rdi)
+	movl   %ecx, -38(%rdi)
+L(write_17words):
+	movl   %ecx, -34(%rdi)
+	movl   %ecx, -30(%rdi)
+L(write_13words):
+	movl   %ecx, -26(%rdi)
+	movl   %ecx, -22(%rdi)
+L(write_9words):
+	movl   %ecx, -18(%rdi)
+	movl   %ecx, -14(%rdi)
+L(write_5words):
+	movl   %ecx, -10(%rdi)
+	movl   %ecx, -6(%rdi)
+L(write_1words):
+	mov	%cx, -2(%rdi)
+	ret
+
+	ALIGN (4)
+L(write_30words):
+	movl   %ecx, -60(%rdi)
+	movl   %ecx, -56(%rdi)
+L(write_26words):
+	movl   %ecx, -52(%rdi)
+	movl   %ecx, -48(%rdi)
+L(write_22words):
+	movl   %ecx, -44(%rdi)
+	movl   %ecx, -40(%rdi)
+L(write_18words):
+	movl   %ecx, -36(%rdi)
+	movl   %ecx, -32(%rdi)
+L(write_14words):
+	movl   %ecx, -28(%rdi)
+	movl   %ecx, -24(%rdi)
+L(write_10words):
+	movl   %ecx, -20(%rdi)
+	movl   %ecx, -16(%rdi)
+L(write_6words):
+	movl   %ecx, -12(%rdi)
+	movl   %ecx, -8(%rdi)
+L(write_2words):
+	movl   %ecx, -4(%rdi)
+	ret
+
+	ALIGN (4)
+L(write_31words):
+	movl   %ecx, -62(%rdi)
+	movl   %ecx, -58(%rdi)
+L(write_27words):
+	movl   %ecx, -54(%rdi)
+	movl   %ecx, -50(%rdi)
+L(write_23words):
+	movl   %ecx, -46(%rdi)
+	movl   %ecx, -42(%rdi)
+L(write_19words):
+	movl   %ecx, -38(%rdi)
+	movl   %ecx, -34(%rdi)
+L(write_15words):
+	movl   %ecx, -30(%rdi)
+	movl   %ecx, -26(%rdi)
+L(write_11words):
+	movl   %ecx, -22(%rdi)
+	movl   %ecx, -18(%rdi)
+L(write_7words):
+	movl   %ecx, -14(%rdi)
+	movl   %ecx, -10(%rdi)
+L(write_3words):
+	movl   %ecx, -6(%rdi)
+	movw   %cx, -2(%rdi)
+	ret
+
+	ALIGN (4)
+L(32wordsormore):
+	shl    $1, %rdx
+	test   $0x01, %edi
+	jz     L(aligned2bytes)
+	mov    %ecx, (%rdi)
+	mov    %ecx, -4(%rdi, %rdx)
+	sub    $2, %rdx
+	add    $1, %rdi
+	rol    $8, %ecx
+L(aligned2bytes):
+	/* Fill xmm0 with the pattern.  */
+	movd   %ecx, %xmm0
+	pshufd $0, %xmm0, %xmm0
+
+	testl  $0xf, %edi
+	jz     L(aligned_16)
+/* RDX > 32 and RDI is not 16 byte aligned.  */
+	movdqu %xmm0, (%rdi)
+	mov    %rdi, %rsi
+	and    $-16, %rdi
+	add    $16, %rdi
+	sub    %rdi, %rsi
+	add    %rsi, %rdx
+
+	ALIGN (4)
+L(aligned_16):
+	cmp    $128, %rdx
+	jge    L(128bytesormore)
+
+L(aligned_16_less128bytes):
+	add    %rdx, %rdi
+	shr    $1, %rdx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes), %rdx, 4)
+
+	ALIGN (4)
+L(128bytesormore):
+	cmp    $SHARED_CACHE_SIZE, %rdx
+	jg     L(128bytesormore_nt)
+
+L(128bytesormore_normal):
+	sub    $128, %rdx
+	movdqa %xmm0, (%rdi)
+	movdqa %xmm0, 0x10(%rdi)
+	movdqa %xmm0, 0x20(%rdi)
+	movdqa %xmm0, 0x30(%rdi)
+	movdqa %xmm0, 0x40(%rdi)
+	movdqa %xmm0, 0x50(%rdi)
+	movdqa %xmm0, 0x60(%rdi)
+	movdqa %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jl     L(128bytesless_normal)
+
+	sub    $128, %rdx
+	movdqa %xmm0, (%rdi)
+	movdqa %xmm0, 0x10(%rdi)
+	movdqa %xmm0, 0x20(%rdi)
+	movdqa %xmm0, 0x30(%rdi)
+	movdqa %xmm0, 0x40(%rdi)
+	movdqa %xmm0, 0x50(%rdi)
+	movdqa %xmm0, 0x60(%rdi)
+	movdqa %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jl     L(128bytesless_normal)
+
+	sub    $128, %rdx
+	movdqa %xmm0, (%rdi)
+	movdqa %xmm0, 0x10(%rdi)
+	movdqa %xmm0, 0x20(%rdi)
+	movdqa %xmm0, 0x30(%rdi)
+	movdqa %xmm0, 0x40(%rdi)
+	movdqa %xmm0, 0x50(%rdi)
+	movdqa %xmm0, 0x60(%rdi)
+	movdqa %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jl     L(128bytesless_normal)
+
+	sub    $128, %rdx
+	movdqa %xmm0, (%rdi)
+	movdqa %xmm0, 0x10(%rdi)
+	movdqa %xmm0, 0x20(%rdi)
+	movdqa %xmm0, 0x30(%rdi)
+	movdqa %xmm0, 0x40(%rdi)
+	movdqa %xmm0, 0x50(%rdi)
+	movdqa %xmm0, 0x60(%rdi)
+	movdqa %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jge    L(128bytesormore_normal)
+
+L(128bytesless_normal):
+	add    %rdx, %rdi
+	shr    $1, %rdx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes), %rdx, 4)
+
+	ALIGN (4)
+L(128bytesormore_nt):
+	sub    $128, %rdx
+	movntdq %xmm0, (%rdi)
+	movntdq %xmm0, 0x10(%rdi)
+	movntdq %xmm0, 0x20(%rdi)
+	movntdq %xmm0, 0x30(%rdi)
+	movntdq %xmm0, 0x40(%rdi)
+	movntdq %xmm0, 0x50(%rdi)
+	movntdq %xmm0, 0x60(%rdi)
+	movntdq %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jge    L(128bytesormore_nt)
+
+	sfence
+	add    %rdx, %rdi
+	shr    $1, %rdx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes), %rdx, 4)
+
+	.pushsection .rodata.sse2,"a",@progbits
+	ALIGN (2)
+L(table_16_128bytes):
+	.int	JMPTBL (L(aligned_16_0bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_2bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_4bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_6bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_8bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_10bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_12bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_14bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_16bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_18bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_20bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_22bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_24bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_26bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_28bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_30bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_32bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_34bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_36bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_38bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_40bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_42bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_44bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_46bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_48bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_50bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_52bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_54bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_56bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_58bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_60bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_62bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_64bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_66bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_68bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_70bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_72bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_74bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_76bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_78bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_80bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_82bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_84bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_86bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_88bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_90bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_92bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_94bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_96bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_98bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_100bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_102bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_104bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_106bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_108bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_110bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_112bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_114bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_116bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_118bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_120bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_122bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_124bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_126bytes), L(table_16_128bytes))
+	.popsection
+
+	ALIGN (4)
+L(aligned_16_112bytes):
+	movdqa %xmm0, -112(%rdi)
+L(aligned_16_96bytes):
+	movdqa %xmm0, -96(%rdi)
+L(aligned_16_80bytes):
+	movdqa %xmm0, -80(%rdi)
+L(aligned_16_64bytes):
+	movdqa %xmm0, -64(%rdi)
+L(aligned_16_48bytes):
+	movdqa %xmm0, -48(%rdi)
+L(aligned_16_32bytes):
+	movdqa %xmm0, -32(%rdi)
+L(aligned_16_16bytes):
+	movdqa %xmm0, -16(%rdi)
+L(aligned_16_0bytes):
+	ret
+
+	ALIGN (4)
+L(aligned_16_114bytes):
+	movdqa %xmm0, -114(%rdi)
+L(aligned_16_98bytes):
+	movdqa %xmm0, -98(%rdi)
+L(aligned_16_82bytes):
+	movdqa %xmm0, -82(%rdi)
+L(aligned_16_66bytes):
+	movdqa %xmm0, -66(%rdi)
+L(aligned_16_50bytes):
+	movdqa %xmm0, -50(%rdi)
+L(aligned_16_34bytes):
+	movdqa %xmm0, -34(%rdi)
+L(aligned_16_18bytes):
+	movdqa %xmm0, -18(%rdi)
+L(aligned_16_2bytes):
+	movw   %cx, -2(%rdi)
+	ret
+
+	ALIGN (4)
+L(aligned_16_116bytes):
+	movdqa %xmm0, -116(%rdi)
+L(aligned_16_100bytes):
+	movdqa %xmm0, -100(%rdi)
+L(aligned_16_84bytes):
+	movdqa %xmm0, -84(%rdi)
+L(aligned_16_68bytes):
+	movdqa %xmm0, -68(%rdi)
+L(aligned_16_52bytes):
+	movdqa %xmm0, -52(%rdi)
+L(aligned_16_36bytes):
+	movdqa %xmm0, -36(%rdi)
+L(aligned_16_20bytes):
+	movdqa %xmm0, -20(%rdi)
+L(aligned_16_4bytes):
+	movl   %ecx, -4(%rdi)
+	ret
+
+	ALIGN (4)
+L(aligned_16_118bytes):
+	movdqa %xmm0, -118(%rdi)
+L(aligned_16_102bytes):
+	movdqa %xmm0, -102(%rdi)
+L(aligned_16_86bytes):
+	movdqa %xmm0, -86(%rdi)
+L(aligned_16_70bytes):
+	movdqa %xmm0, -70(%rdi)
+L(aligned_16_54bytes):
+	movdqa %xmm0, -54(%rdi)
+L(aligned_16_38bytes):
+	movdqa %xmm0, -38(%rdi)
+L(aligned_16_22bytes):
+	movdqa %xmm0, -22(%rdi)
+L(aligned_16_6bytes):
+	movl   %ecx, -6(%rdi)
+	movw   %cx, -2(%rdi)
+	ret
+
+	ALIGN (4)
+L(aligned_16_120bytes):
+	movdqa %xmm0, -120(%rdi)
+L(aligned_16_104bytes):
+	movdqa %xmm0, -104(%rdi)
+L(aligned_16_88bytes):
+	movdqa %xmm0, -88(%rdi)
+L(aligned_16_72bytes):
+	movdqa %xmm0, -72(%rdi)
+L(aligned_16_56bytes):
+	movdqa %xmm0, -56(%rdi)
+L(aligned_16_40bytes):
+	movdqa %xmm0, -40(%rdi)
+L(aligned_16_24bytes):
+	movdqa %xmm0, -24(%rdi)
+L(aligned_16_8bytes):
+	movq   %xmm0, -8(%rdi)
+	ret
+
+	ALIGN (4)
+L(aligned_16_122bytes):
+	movdqa %xmm0, -122(%rdi)
+L(aligned_16_106bytes):
+	movdqa %xmm0, -106(%rdi)
+L(aligned_16_90bytes):
+	movdqa %xmm0, -90(%rdi)
+L(aligned_16_74bytes):
+	movdqa %xmm0, -74(%rdi)
+L(aligned_16_58bytes):
+	movdqa %xmm0, -58(%rdi)
+L(aligned_16_42bytes):
+	movdqa %xmm0, -42(%rdi)
+L(aligned_16_26bytes):
+	movdqa %xmm0, -26(%rdi)
+L(aligned_16_10bytes):
+	movq   %xmm0, -10(%rdi)
+	movw   %cx, -2(%rdi)
+	ret
+
+	ALIGN (4)
+L(aligned_16_124bytes):
+	movdqa %xmm0, -124(%rdi)
+L(aligned_16_108bytes):
+	movdqa %xmm0, -108(%rdi)
+L(aligned_16_92bytes):
+	movdqa %xmm0, -92(%rdi)
+L(aligned_16_76bytes):
+	movdqa %xmm0, -76(%rdi)
+L(aligned_16_60bytes):
+	movdqa %xmm0, -60(%rdi)
+L(aligned_16_44bytes):
+	movdqa %xmm0, -44(%rdi)
+L(aligned_16_28bytes):
+	movdqa %xmm0, -28(%rdi)
+L(aligned_16_12bytes):
+	movq   %xmm0, -12(%rdi)
+	movl   %ecx, -4(%rdi)
+	ret
+
+	ALIGN (4)
+L(aligned_16_126bytes):
+	movdqa %xmm0, -126(%rdi)
+L(aligned_16_110bytes):
+	movdqa %xmm0, -110(%rdi)
+L(aligned_16_94bytes):
+	movdqa %xmm0, -94(%rdi)
+L(aligned_16_78bytes):
+	movdqa %xmm0, -78(%rdi)
+L(aligned_16_62bytes):
+	movdqa %xmm0, -62(%rdi)
+L(aligned_16_46bytes):
+	movdqa %xmm0, -46(%rdi)
+L(aligned_16_30bytes):
+	movdqa %xmm0, -30(%rdi)
+L(aligned_16_14bytes):
+	movq   %xmm0, -14(%rdi)
+	movl   %ecx, -6(%rdi)
+	movw   %cx, -2(%rdi)
+	ret
+
+END (android_memset16)
diff --git a/libcutils/arch-x86_64/android_memset32_SSE2-atom.S b/libcutils/arch-x86_64/android_memset32_SSE2-atom.S
new file mode 100644
index 0000000..4bdea8e
--- /dev/null
+++ b/libcutils/arch-x86_64/android_memset32_SSE2-atom.S
@@ -0,0 +1,372 @@
+/*
+ * Copyright (C) 2014 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+/*
+ * Contributed by: Intel Corporation
+ */
+
+#include "cache.h"
+
+#ifndef L
+# define L(label)	.L##label
+#endif
+
+#ifndef ALIGN
+# define ALIGN(n)	.p2align n
+#endif
+
+#ifndef cfi_startproc
+# define cfi_startproc			.cfi_startproc
+#endif
+
+#ifndef cfi_endproc
+# define cfi_endproc			.cfi_endproc
+#endif
+
+#ifndef ENTRY
+# define ENTRY(name)			\
+	.type name,  @function; 	\
+	.globl name;			\
+	.p2align 4;			\
+name:					\
+	cfi_startproc
+#endif
+
+#ifndef END
+# define END(name)			\
+	cfi_endproc;			\
+	.size name, .-name
+#endif
+
+#define JMPTBL(I, B)	I - B
+
+/* Branch to an entry in a jump table.  TABLE is a jump table with
+   relative offsets.  INDEX is a register contains the index into the
+   jump table.  SCALE is the scale of INDEX.  */
+#define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \
+	lea    TABLE(%rip), %r11;						\
+	movslq (%r11, INDEX, SCALE), INDEX;				\
+	lea    (%r11, INDEX), INDEX;					\
+	jmp    *INDEX
+
+	.section .text.sse2,"ax",@progbits
+	ALIGN (4)
+ENTRY (android_memset32)	// Address in rdi
+	shr    $2, %rdx			// Count in rdx
+	movl   %esi, %ecx		// Pattern in ecx
+
+	cmp    $16, %rdx
+	jae    L(16dbwordsormore)
+
+L(write_less16dbwords):
+	lea    (%rdi, %rdx, 4), %rdi
+	BRANCH_TO_JMPTBL_ENTRY (L(table_less16dbwords), %rdx, 4)
+
+	.pushsection .rodata.sse2,"a",@progbits
+	ALIGN (2)
+L(table_less16dbwords):
+	.int	JMPTBL (L(write_0dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_1dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_2dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_3dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_4dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_5dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_6dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_7dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_8dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_9dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_10dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_11dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_12dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_13dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_14dbwords), L(table_less16dbwords))
+	.int	JMPTBL (L(write_15dbwords), L(table_less16dbwords))
+	.popsection
+
+	ALIGN (4)
+L(write_15dbwords):
+	movl   %ecx, -60(%rdi)
+L(write_14dbwords):
+	movl   %ecx, -56(%rdi)
+L(write_13dbwords):
+	movl   %ecx, -52(%rdi)
+L(write_12dbwords):
+	movl   %ecx, -48(%rdi)
+L(write_11dbwords):
+	movl   %ecx, -44(%rdi)
+L(write_10dbwords):
+	movl   %ecx, -40(%rdi)
+L(write_9dbwords):
+	movl   %ecx, -36(%rdi)
+L(write_8dbwords):
+	movl   %ecx, -32(%rdi)
+L(write_7dbwords):
+	movl   %ecx, -28(%rdi)
+L(write_6dbwords):
+	movl   %ecx, -24(%rdi)
+L(write_5dbwords):
+	movl   %ecx, -20(%rdi)
+L(write_4dbwords):
+	movl   %ecx, -16(%rdi)
+L(write_3dbwords):
+	movl   %ecx, -12(%rdi)
+L(write_2dbwords):
+	movl   %ecx, -8(%rdi)
+L(write_1dbwords):
+	movl   %ecx, -4(%rdi)
+L(write_0dbwords):
+	ret
+
+	ALIGN (4)
+L(16dbwordsormore):
+	test   $3, %edi
+	jz     L(aligned4bytes)
+	mov    %ecx, (%rdi)
+	mov    %ecx, -4(%rdi, %rdx, 4)
+	sub    $1, %rdx
+	rol    $24, %ecx
+	add    $1, %rdi
+	test   $3, %edi
+	jz     L(aligned4bytes)
+	ror    $8, %ecx
+	add    $1, %rdi
+	test   $3, %edi
+	jz     L(aligned4bytes)
+	ror    $8, %ecx
+	add    $1, %rdi
+L(aligned4bytes):
+	shl    $2, %rdx
+
+	/* Fill xmm0 with the pattern.  */
+	movd   %ecx, %xmm0
+	pshufd $0, %xmm0, %xmm0
+
+	testl  $0xf, %edi
+	jz     L(aligned_16)
+/* RDX > 32 and RDI is not 16 byte aligned.  */
+	movdqu %xmm0, (%rdi)
+	mov    %rdi, %rsi
+	and    $-16, %rdi
+	add    $16, %rdi
+	sub    %rdi, %rsi
+	add    %rsi, %rdx
+
+	ALIGN (4)
+L(aligned_16):
+	cmp    $128, %rdx
+	jge    L(128bytesormore)
+
+L(aligned_16_less128bytes):
+	add    %rdx, %rdi
+	shr    $2, %rdx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes), %rdx, 4)
+
+	ALIGN (4)
+L(128bytesormore):
+	cmp    $SHARED_CACHE_SIZE, %rdx
+	jg     L(128bytesormore_nt)
+
+L(128bytesormore_normal):
+	sub    $128, %rdx
+	movdqa %xmm0, (%rdi)
+	movdqa %xmm0, 0x10(%rdi)
+	movdqa %xmm0, 0x20(%rdi)
+	movdqa %xmm0, 0x30(%rdi)
+	movdqa %xmm0, 0x40(%rdi)
+	movdqa %xmm0, 0x50(%rdi)
+	movdqa %xmm0, 0x60(%rdi)
+	movdqa %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jl     L(128bytesless_normal)
+
+	sub    $128, %rdx
+	movdqa %xmm0, (%rdi)
+	movdqa %xmm0, 0x10(%rdi)
+	movdqa %xmm0, 0x20(%rdi)
+	movdqa %xmm0, 0x30(%rdi)
+	movdqa %xmm0, 0x40(%rdi)
+	movdqa %xmm0, 0x50(%rdi)
+	movdqa %xmm0, 0x60(%rdi)
+	movdqa %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jl     L(128bytesless_normal)
+
+	sub    $128, %rdx
+	movdqa %xmm0, (%rdi)
+	movdqa %xmm0, 0x10(%rdi)
+	movdqa %xmm0, 0x20(%rdi)
+	movdqa %xmm0, 0x30(%rdi)
+	movdqa %xmm0, 0x40(%rdi)
+	movdqa %xmm0, 0x50(%rdi)
+	movdqa %xmm0, 0x60(%rdi)
+	movdqa %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jl     L(128bytesless_normal)
+
+	sub    $128, %rdx
+	movdqa %xmm0, (%rdi)
+	movdqa %xmm0, 0x10(%rdi)
+	movdqa %xmm0, 0x20(%rdi)
+	movdqa %xmm0, 0x30(%rdi)
+	movdqa %xmm0, 0x40(%rdi)
+	movdqa %xmm0, 0x50(%rdi)
+	movdqa %xmm0, 0x60(%rdi)
+	movdqa %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jge    L(128bytesormore_normal)
+
+L(128bytesless_normal):
+	add    %rdx, %rdi
+	shr    $2, %rdx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes), %rdx, 4)
+
+	ALIGN (4)
+L(128bytesormore_nt):
+	sub    $128, %rdx
+	movntdq %xmm0, (%rdi)
+	movntdq %xmm0, 0x10(%rdi)
+	movntdq %xmm0, 0x20(%rdi)
+	movntdq %xmm0, 0x30(%rdi)
+	movntdq %xmm0, 0x40(%rdi)
+	movntdq %xmm0, 0x50(%rdi)
+	movntdq %xmm0, 0x60(%rdi)
+	movntdq %xmm0, 0x70(%rdi)
+	lea    128(%rdi), %rdi
+	cmp    $128, %rdx
+	jge    L(128bytesormore_nt)
+
+	sfence
+	add    %rdx, %rdi
+	shr    $2, %rdx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes), %rdx, 4)
+
+	.pushsection .rodata.sse2,"a",@progbits
+	ALIGN (2)
+L(table_16_128bytes):
+	.int	JMPTBL (L(aligned_16_0bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_4bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_8bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_12bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_16bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_20bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_24bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_28bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_32bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_36bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_40bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_44bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_48bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_52bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_56bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_60bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_64bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_68bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_72bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_76bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_80bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_84bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_88bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_92bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_96bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_100bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_104bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_108bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_112bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_116bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_120bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_124bytes), L(table_16_128bytes))
+	.popsection
+
+	ALIGN (4)
+L(aligned_16_112bytes):
+	movdqa	%xmm0, -112(%rdi)
+L(aligned_16_96bytes):
+	movdqa	%xmm0, -96(%rdi)
+L(aligned_16_80bytes):
+	movdqa	%xmm0, -80(%rdi)
+L(aligned_16_64bytes):
+	movdqa	%xmm0, -64(%rdi)
+L(aligned_16_48bytes):
+	movdqa	%xmm0, -48(%rdi)
+L(aligned_16_32bytes):
+	movdqa	%xmm0, -32(%rdi)
+L(aligned_16_16bytes):
+	movdqa	%xmm0, -16(%rdi)
+L(aligned_16_0bytes):
+	ret
+
+	ALIGN (4)
+L(aligned_16_116bytes):
+	movdqa	%xmm0, -116(%rdi)
+L(aligned_16_100bytes):
+	movdqa	%xmm0, -100(%rdi)
+L(aligned_16_84bytes):
+	movdqa	%xmm0, -84(%rdi)
+L(aligned_16_68bytes):
+	movdqa	%xmm0, -68(%rdi)
+L(aligned_16_52bytes):
+	movdqa	%xmm0, -52(%rdi)
+L(aligned_16_36bytes):
+	movdqa	%xmm0, -36(%rdi)
+L(aligned_16_20bytes):
+	movdqa	%xmm0, -20(%rdi)
+L(aligned_16_4bytes):
+	movl	%ecx, -4(%rdi)
+	ret
+
+	ALIGN (4)
+L(aligned_16_120bytes):
+	movdqa	%xmm0, -120(%rdi)
+L(aligned_16_104bytes):
+	movdqa	%xmm0, -104(%rdi)
+L(aligned_16_88bytes):
+	movdqa	%xmm0, -88(%rdi)
+L(aligned_16_72bytes):
+	movdqa	%xmm0, -72(%rdi)
+L(aligned_16_56bytes):
+	movdqa	%xmm0, -56(%rdi)
+L(aligned_16_40bytes):
+	movdqa	%xmm0, -40(%rdi)
+L(aligned_16_24bytes):
+	movdqa	%xmm0, -24(%rdi)
+L(aligned_16_8bytes):
+	movq	%xmm0, -8(%rdi)
+	ret
+
+	ALIGN (4)
+L(aligned_16_124bytes):
+	movdqa	%xmm0, -124(%rdi)
+L(aligned_16_108bytes):
+	movdqa	%xmm0, -108(%rdi)
+L(aligned_16_92bytes):
+	movdqa	%xmm0, -92(%rdi)
+L(aligned_16_76bytes):
+	movdqa	%xmm0, -76(%rdi)
+L(aligned_16_60bytes):
+	movdqa	%xmm0, -60(%rdi)
+L(aligned_16_44bytes):
+	movdqa	%xmm0, -44(%rdi)
+L(aligned_16_28bytes):
+	movdqa	%xmm0, -28(%rdi)
+L(aligned_16_12bytes):
+	movq	%xmm0, -12(%rdi)
+	movl	%ecx, -4(%rdi)
+	ret
+
+END (android_memset32)
diff --git a/libcutils/arch-x86_64/cache.h b/libcutils/arch-x86_64/cache.h
new file mode 100644
index 0000000..ab5dd2f
--- /dev/null
+++ b/libcutils/arch-x86_64/cache.h
@@ -0,0 +1,31 @@
+/*
+ * Copyright (C) 2014 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+/*
+ * Contributed by: Intel Corporation
+ */
+
+#if defined(__slm__)
+/* Values are optimized for Silvermont */
+#define SHARED_CACHE_SIZE	(1024*1024)			/* Silvermont L2 Cache */
+#define DATA_CACHE_SIZE		(24*1024)			/* Silvermont L1 Data Cache */
+#else
+/* Values are optimized for Atom */
+#define SHARED_CACHE_SIZE	(512*1024)			/* Atom L2 Cache */
+#define DATA_CACHE_SIZE		(24*1024)			/* Atom L1 Data Cache */
+#endif
+
+#define SHARED_CACHE_SIZE_HALF	(SHARED_CACHE_SIZE / 2)
+#define DATA_CACHE_SIZE_HALF	(DATA_CACHE_SIZE / 2)