Assembly coded android_memset16 and android_memset32

Change-Id: Ife2dd406e1dcb962e5e97788c515ac96f5c52e44

5 files changed, 1324 insertions, 0 deletions

diff --git a/libcutils/arch-x86/android_memset16.S b/libcutils/arch-x86/android_memset16.S
new file mode 100644
index 0000000..b1f09cb
--- /dev/null
+++ b/libcutils/arch-x86/android_memset16.S
@@ -0,0 +1,32 @@
+/*
+ * Copyright (C) 2010 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(USE_SSE2)
+
+# include "cache_wrapper.S"
+# undef __i686
+# define USE_AS_ANDROID
+# define sse2_memset16_atom android_memset16
+# include "sse2-memset16-atom.S"
+
+#else
+
+# include "memset16.S"
+
+#endif
diff --git a/libcutils/arch-x86/android_memset32.S b/libcutils/arch-x86/android_memset32.S
new file mode 100644
index 0000000..1fb2ffe
--- /dev/null
+++ b/libcutils/arch-x86/android_memset32.S
@@ -0,0 +1,33 @@
+/*
+ * Copyright (C) 2010 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(USE_SSE2)
+
+# include "cache_wrapper.S"
+# undef __i686
+# define USE_AS_ANDROID
+# define sse2_memset32_atom android_memset32
+# include "sse2-memset32-atom.S"
+
+#else
+
+# include "memset32.S"
+
+#endif
+
diff --git a/libcutils/arch-x86/cache_wrapper.S b/libcutils/arch-x86/cache_wrapper.S
new file mode 100644
index 0000000..508fdd3
--- /dev/null
+++ b/libcutils/arch-x86/cache_wrapper.S
@@ -0,0 +1,24 @@
+/*
+ * Copyright (C) 2010 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
+ */
+
+/* Values are optimized for Atom */
+#define SHARED_CACHE_SIZE       (512*1024)            /* Atom L2 Cache */
+#define DATA_CACHE_SIZE         (24*1024)             /* Atom L1 Data Cache */
+#define SHARED_CACHE_SIZE_HALF  (SHARED_CACHE_SIZE / 2)
+#define DATA_CACHE_SIZE_HALF    (DATA_CACHE_SIZE / 2)
diff --git a/libcutils/arch-x86/sse2-memset16-atom.S b/libcutils/arch-x86/sse2-memset16-atom.S
new file mode 100644
index 0000000..cafec82
--- /dev/null
+++ b/libcutils/arch-x86/sse2-memset16-atom.S
@@ -0,0 +1,722 @@
+/*
+ * Copyright (C) 2010 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
+ */
+
+#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 cfi_rel_offset
+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+#endif
+
+#ifndef cfi_restore
+# define cfi_restore(reg)		.cfi_restore reg
+#endif
+
+#ifndef cfi_adjust_cfa_offset
+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#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 CFI_PUSH(REG)						\
+  cfi_adjust_cfa_offset (4);					\
+  cfi_rel_offset (REG, 0)
+
+#define CFI_POP(REG)						\
+  cfi_adjust_cfa_offset (-4);					\
+  cfi_restore (REG)
+
+#define PUSH(REG)	pushl REG; CFI_PUSH (REG)
+#define POP(REG)	popl REG; CFI_POP (REG)
+
+#ifdef USE_AS_BZERO16
+# define DEST		PARMS
+# define LEN		DEST+4
+#else
+# define DEST		PARMS
+# define CHR		DEST+4
+# define LEN		CHR+4
+#endif
+
+#if 1
+# define SETRTNVAL
+#else
+# define SETRTNVAL	movl DEST(%esp), %eax
+#endif
+
+#ifdef SHARED
+# define ENTRANCE	PUSH (%ebx);
+# define RETURN_END	POP (%ebx); ret
+# define RETURN		RETURN_END; CFI_PUSH (%ebx)
+# define PARMS		8		/* Preserve EBX.  */
+# define JMPTBL(I, B)	I - B
+
+/* Load an entry in a jump table into EBX and branch to it.  TABLE is a
+   jump table with relative offsets.   */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE)				\
+    /* We first load PC into EBX.  */				\
+    call	__i686.get_pc_thunk.bx;				\
+    /* Get the address of the jump table.  */			\
+    add		$(TABLE - .), %ebx;				\
+    /* Get the entry and convert the relative offset to the	\
+       absolute address.  */					\
+    add		(%ebx,%ecx,4), %ebx;				\
+    /* We loaded the jump table and adjuested EDX. Go.  */	\
+    jmp		*%ebx
+
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	ALIGN (4)
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+#else
+# define ENTRANCE
+# define RETURN_END	ret
+# define RETURN		RETURN_END
+# define PARMS		4
+# define JMPTBL(I, B)	I
+
+/* Branch to an entry in a jump table.  TABLE is a jump table with
+   absolute offsets.  */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE)				\
+    jmp		*TABLE(,%ecx,4)
+#endif
+
+	.section .text.sse2,"ax",@progbits
+	ALIGN (4)
+ENTRY (sse2_memset16_atom)
+	ENTRANCE
+
+	movl	LEN(%esp), %ecx
+#ifdef USE_AS_ANDROID
+	shr	$1, %ecx
+#endif
+#ifdef USE_AS_BZERO16
+	xor	%eax, %eax
+#else
+	movzwl	CHR(%esp), %eax
+	mov	%eax, %edx
+	shl	$16, %eax
+	or	%edx, %eax
+#endif
+	movl	DEST(%esp), %edx
+	cmp	$32, %ecx
+	jae	L(32wordsormore)
+
+L(write_less32words):
+	lea	(%edx, %ecx, 2), %edx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_less32words))
+
+
+	.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	%eax, -56(%edx)
+	movl	%eax, -52(%edx)
+L(write_24words):
+	movl	%eax, -48(%edx)
+	movl	%eax, -44(%edx)
+L(write_20words):
+	movl	%eax, -40(%edx)
+	movl	%eax, -36(%edx)
+L(write_16words):
+	movl	%eax, -32(%edx)
+	movl	%eax, -28(%edx)
+L(write_12words):
+	movl	%eax, -24(%edx)
+	movl	%eax, -20(%edx)
+L(write_8words):
+	movl	%eax, -16(%edx)
+	movl	%eax, -12(%edx)
+L(write_4words):
+	movl	%eax, -8(%edx)
+	movl	%eax, -4(%edx)
+L(write_0words):
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_29words):
+	movl	%eax, -58(%edx)
+	movl	%eax, -54(%edx)
+L(write_25words):
+	movl	%eax, -50(%edx)
+	movl	%eax, -46(%edx)
+L(write_21words):
+	movl	%eax, -42(%edx)
+	movl	%eax, -38(%edx)
+L(write_17words):
+	movl	%eax, -34(%edx)
+	movl	%eax, -30(%edx)
+L(write_13words):
+	movl	%eax, -26(%edx)
+	movl	%eax, -22(%edx)
+L(write_9words):
+	movl	%eax, -18(%edx)
+	movl	%eax, -14(%edx)
+L(write_5words):
+	movl	%eax, -10(%edx)
+	movl	%eax, -6(%edx)
+L(write_1words):
+	mov	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_30words):
+	movl	%eax, -60(%edx)
+	movl	%eax, -56(%edx)
+L(write_26words):
+	movl	%eax, -52(%edx)
+	movl	%eax, -48(%edx)
+L(write_22words):
+	movl	%eax, -44(%edx)
+	movl	%eax, -40(%edx)
+L(write_18words):
+	movl	%eax, -36(%edx)
+	movl	%eax, -32(%edx)
+L(write_14words):
+	movl	%eax, -28(%edx)
+	movl	%eax, -24(%edx)
+L(write_10words):
+	movl	%eax, -20(%edx)
+	movl	%eax, -16(%edx)
+L(write_6words):
+	movl	%eax, -12(%edx)
+	movl	%eax, -8(%edx)
+L(write_2words):
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_31words):
+	movl	%eax, -62(%edx)
+	movl	%eax, -58(%edx)
+L(write_27words):
+	movl	%eax, -54(%edx)
+	movl	%eax, -50(%edx)
+L(write_23words):
+	movl	%eax, -46(%edx)
+	movl	%eax, -42(%edx)
+L(write_19words):
+	movl	%eax, -38(%edx)
+	movl	%eax, -34(%edx)
+L(write_15words):
+	movl	%eax, -30(%edx)
+	movl	%eax, -26(%edx)
+L(write_11words):
+	movl	%eax, -22(%edx)
+	movl	%eax, -18(%edx)
+L(write_7words):
+	movl	%eax, -14(%edx)
+	movl	%eax, -10(%edx)
+L(write_3words):
+	movl	%eax, -6(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+
+L(32wordsormore):
+	shl	$1, %ecx
+	test	$0x01, %edx
+	jz	L(aligned2bytes)
+	mov	%eax, (%edx)
+	mov	%eax, -4(%edx, %ecx)
+	sub	$2, %ecx
+	add	$1, %edx
+	rol	$8, %eax
+L(aligned2bytes):
+#ifdef USE_AS_BZERO16
+	pxor	%xmm0, %xmm0
+#else
+	movd	%eax, %xmm0
+	pshufd	$0, %xmm0, %xmm0
+#endif
+	testl	$0xf, %edx
+	jz	L(aligned_16)
+/* ECX > 32 and EDX is not 16 byte aligned.  */
+L(not_aligned_16):
+	movdqu	%xmm0, (%edx)
+	movl	%edx, %eax
+	and	$-16, %edx
+	add	$16, %edx
+	sub	%edx, %eax
+	add	%eax, %ecx
+	movd	%xmm0, %eax
+
+	ALIGN (4)
+L(aligned_16):
+	cmp	$128, %ecx
+	jae	L(128bytesormore)
+
+L(aligned_16_less128bytes):
+	add	%ecx, %edx
+	shr	$1, %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	ALIGN (4)
+L(128bytesormore):
+#ifdef SHARED_CACHE_SIZE
+	PUSH (%ebx)
+	mov	$SHARED_CACHE_SIZE, %ebx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	mov	__x86_shared_cache_size@GOTOFF(%ebx), %ebx
+# else
+	PUSH (%ebx)
+	mov	__x86_shared_cache_size, %ebx
+# endif
+#endif
+	cmp	%ebx, %ecx
+	jae	L(128bytesormore_nt_start)
+
+	
+#ifdef DATA_CACHE_SIZE
+	POP (%ebx)
+# define RESTORE_EBX_STATE CFI_PUSH (%ebx)
+	cmp	$DATA_CACHE_SIZE, %ecx
+#else
+# ifdef SHARED
+#  define RESTORE_EBX_STATE
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmp	__x86_data_cache_size@GOTOFF(%ebx), %ecx
+# else
+	POP (%ebx)
+#  define RESTORE_EBX_STATE CFI_PUSH (%ebx)
+	cmp	__x86_data_cache_size, %ecx
+# endif
+#endif
+
+	jae	L(128bytes_L2_normal)
+	subl	$128, %ecx
+L(128bytesormore_normal):
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	lea	128(%edx), %edx
+	jb	L(128bytesless_normal)
+
+
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	lea	128(%edx), %edx
+	jae	L(128bytesormore_normal)
+
+L(128bytesless_normal):
+	lea	128(%ecx), %ecx
+	add	%ecx, %edx
+	shr	$1, %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	ALIGN (4)
+L(128bytes_L2_normal):
+	prefetcht0	0x380(%edx)
+	prefetcht0	0x3c0(%edx)
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movaps	%xmm0, 0x10(%edx)
+	movaps	%xmm0, 0x20(%edx)
+	movaps	%xmm0, 0x30(%edx)
+	movaps	%xmm0, 0x40(%edx)
+	movaps	%xmm0, 0x50(%edx)
+	movaps	%xmm0, 0x60(%edx)
+	movaps	%xmm0, 0x70(%edx)
+	add	$128, %edx
+	cmp	$128, %ecx 	
+	jae	L(128bytes_L2_normal)
+
+L(128bytesless_L2_normal):
+	add	%ecx, %edx
+	shr	$1, %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	RESTORE_EBX_STATE
+L(128bytesormore_nt_start):
+	sub	%ebx, %ecx
+	mov	%ebx, %eax
+	and	$0x7f, %eax
+	add	%eax, %ecx
+	movd	%xmm0, %eax
+	ALIGN (4)
+L(128bytesormore_shared_cache_loop):
+	prefetcht0	0x3c0(%edx)
+	prefetcht0	0x380(%edx)
+	sub	$0x80, %ebx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	add	$0x80, %edx
+	cmp	$0x80, %ebx
+	jae	L(128bytesormore_shared_cache_loop)
+	cmp	$0x80, %ecx
+	jb	L(shared_cache_loop_end)
+	ALIGN (4)
+L(128bytesormore_nt):
+	sub	$0x80, %ecx
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm0, 0x10(%edx)
+	movntdq	%xmm0, 0x20(%edx)
+	movntdq	%xmm0, 0x30(%edx)
+	movntdq	%xmm0, 0x40(%edx)
+	movntdq	%xmm0, 0x50(%edx)
+	movntdq	%xmm0, 0x60(%edx)
+	movntdq	%xmm0, 0x70(%edx)
+	add	$0x80, %edx
+	cmp	$0x80, %ecx
+	jae	L(128bytesormore_nt)
+	sfence
+L(shared_cache_loop_end):
+#if defined DATA_CACHE_SIZE || !defined SHARED
+	POP (%ebx)
+#endif
+	add	%ecx, %edx
+	shr	$1, %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+
+	.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(%edx)
+L(aligned_16_96bytes):
+	movdqa	%xmm0, -96(%edx)
+L(aligned_16_80bytes):
+	movdqa	%xmm0, -80(%edx)
+L(aligned_16_64bytes):
+	movdqa	%xmm0, -64(%edx)
+L(aligned_16_48bytes):
+	movdqa	%xmm0, -48(%edx)
+L(aligned_16_32bytes):
+	movdqa	%xmm0, -32(%edx)
+L(aligned_16_16bytes):
+	movdqa	%xmm0, -16(%edx)
+L(aligned_16_0bytes):
+	SETRTNVAL
+	RETURN
+
+
+	ALIGN (4)
+L(aligned_16_114bytes):
+	movdqa	%xmm0, -114(%edx)
+L(aligned_16_98bytes):
+	movdqa	%xmm0, -98(%edx)
+L(aligned_16_82bytes):
+	movdqa	%xmm0, -82(%edx)
+L(aligned_16_66bytes):
+	movdqa	%xmm0, -66(%edx)
+L(aligned_16_50bytes):
+	movdqa	%xmm0, -50(%edx)
+L(aligned_16_34bytes):
+	movdqa	%xmm0, -34(%edx)
+L(aligned_16_18bytes):
+	movdqa	%xmm0, -18(%edx)
+L(aligned_16_2bytes):
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_116bytes):
+	movdqa	%xmm0, -116(%edx)
+L(aligned_16_100bytes):
+	movdqa	%xmm0, -100(%edx)
+L(aligned_16_84bytes):
+	movdqa	%xmm0, -84(%edx)
+L(aligned_16_68bytes):
+	movdqa	%xmm0, -68(%edx)
+L(aligned_16_52bytes):
+	movdqa	%xmm0, -52(%edx)
+L(aligned_16_36bytes):
+	movdqa	%xmm0, -36(%edx)
+L(aligned_16_20bytes):
+	movdqa	%xmm0, -20(%edx)
+L(aligned_16_4bytes):
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+
+	ALIGN (4)
+L(aligned_16_118bytes):
+	movdqa	%xmm0, -118(%edx)
+L(aligned_16_102bytes):
+	movdqa	%xmm0, -102(%edx)
+L(aligned_16_86bytes):
+	movdqa	%xmm0, -86(%edx)
+L(aligned_16_70bytes):
+	movdqa	%xmm0, -70(%edx)
+L(aligned_16_54bytes):
+	movdqa	%xmm0, -54(%edx)
+L(aligned_16_38bytes):
+	movdqa	%xmm0, -38(%edx)
+L(aligned_16_22bytes):
+	movdqa	%xmm0, -22(%edx)
+L(aligned_16_6bytes):
+	movl	%eax, -6(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+
+	ALIGN (4)
+L(aligned_16_120bytes):
+	movdqa	%xmm0, -120(%edx)
+L(aligned_16_104bytes):
+	movdqa	%xmm0, -104(%edx)
+L(aligned_16_88bytes):
+	movdqa	%xmm0, -88(%edx)
+L(aligned_16_72bytes):
+	movdqa	%xmm0, -72(%edx)
+L(aligned_16_56bytes):
+	movdqa	%xmm0, -56(%edx)
+L(aligned_16_40bytes):
+	movdqa	%xmm0, -40(%edx)
+L(aligned_16_24bytes):
+	movdqa	%xmm0, -24(%edx)
+L(aligned_16_8bytes):
+	movq	%xmm0, -8(%edx)
+	SETRTNVAL
+	RETURN
+
+
+	ALIGN (4)
+L(aligned_16_122bytes):
+	movdqa	%xmm0, -122(%edx)
+L(aligned_16_106bytes):
+	movdqa	%xmm0, -106(%edx)
+L(aligned_16_90bytes):
+	movdqa	%xmm0, -90(%edx)
+L(aligned_16_74bytes):
+	movdqa	%xmm0, -74(%edx)
+L(aligned_16_58bytes):
+	movdqa	%xmm0, -58(%edx)
+L(aligned_16_42bytes):
+	movdqa	%xmm0, -42(%edx)
+L(aligned_16_26bytes):
+	movdqa	%xmm0, -26(%edx)
+L(aligned_16_10bytes):
+	movq	%xmm0, -10(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+
+	ALIGN (4)
+L(aligned_16_124bytes):
+	movdqa	%xmm0, -124(%edx)
+L(aligned_16_108bytes):
+	movdqa	%xmm0, -108(%edx)
+L(aligned_16_92bytes):
+	movdqa	%xmm0, -92(%edx)
+L(aligned_16_76bytes):
+	movdqa	%xmm0, -76(%edx)
+L(aligned_16_60bytes):
+	movdqa	%xmm0, -60(%edx)
+L(aligned_16_44bytes):
+	movdqa	%xmm0, -44(%edx)
+L(aligned_16_28bytes):
+	movdqa	%xmm0, -28(%edx)
+L(aligned_16_12bytes):
+	movq	%xmm0, -12(%edx)
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+
+	ALIGN (4)
+L(aligned_16_126bytes):
+	movdqa	%xmm0, -126(%edx)
+L(aligned_16_110bytes):
+	movdqa	%xmm0, -110(%edx)
+L(aligned_16_94bytes):
+	movdqa	%xmm0, -94(%edx)
+L(aligned_16_78bytes):
+	movdqa	%xmm0, -78(%edx)
+L(aligned_16_62bytes):
+	movdqa	%xmm0, -62(%edx)
+L(aligned_16_46bytes):
+	movdqa	%xmm0, -46(%edx)
+L(aligned_16_30bytes):
+	movdqa	%xmm0, -30(%edx)
+L(aligned_16_14bytes):
+	movq	%xmm0, -14(%edx)
+	movl	%eax, -6(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+END (sse2_memset16_atom)
diff --git a/libcutils/arch-x86/sse2-memset32-atom.S b/libcutils/arch-x86/sse2-memset32-atom.S
new file mode 100644
index 0000000..4a52484
--- /dev/null
+++ b/libcutils/arch-x86/sse2-memset32-atom.S
@@ -0,0 +1,513 @@
+/*
+ * Copyright (C) 2010 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
+ */
+
+#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 cfi_rel_offset
+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+#endif
+
+#ifndef cfi_restore
+# define cfi_restore(reg)		.cfi_restore reg
+#endif
+
+#ifndef cfi_adjust_cfa_offset
+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#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 CFI_PUSH(REG)						\
+  cfi_adjust_cfa_offset (4);					\
+  cfi_rel_offset (REG, 0)
+
+#define CFI_POP(REG)						\
+  cfi_adjust_cfa_offset (-4);					\
+  cfi_restore (REG)
+
+#define PUSH(REG)	pushl REG; CFI_PUSH (REG)
+#define POP(REG)	popl REG; CFI_POP (REG)
+
+#ifdef USE_AS_BZERO32
+# define DEST		PARMS
+# define LEN		DEST+4
+#else
+# define DEST		PARMS
+# define DWDS		DEST+4
+# define LEN		DWDS+4
+#endif
+
+#ifdef USE_AS_WMEMSET32
+# define SETRTNVAL	movl DEST(%esp), %eax
+#else
+# define SETRTNVAL
+#endif
+
+#ifdef SHARED
+# define ENTRANCE	PUSH (%ebx);
+# define RETURN_END	POP (%ebx); ret
+# define RETURN		RETURN_END; CFI_PUSH (%ebx)
+# define PARMS		8		/* Preserve EBX.  */
+# define JMPTBL(I, B)	I - B
+
+/* Load an entry in a jump table into EBX and branch to it.  TABLE is a
+   jump table with relative offsets.   */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE)				\
+    /* We first load PC into EBX.  */				\
+    call	__i686.get_pc_thunk.bx;				\
+    /* Get the address of the jump table.  */			\
+    add		$(TABLE - .), %ebx;				\
+    /* Get the entry and convert the relative offset to the	\
+       absolute address.  */					\
+    add		(%ebx,%ecx,4), %ebx;				\
+    /* We loaded the jump table and adjuested EDX. Go.  */	\
+    jmp		*%ebx
+
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	ALIGN (4)
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+#else
+# define ENTRANCE
+# define RETURN_END	ret
+# define RETURN		RETURN_END
+# define PARMS		4
+# define JMPTBL(I, B)	I
+
+/* Branch to an entry in a jump table.  TABLE is a jump table with
+   absolute offsets.  */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE)				\
+    jmp		*TABLE(,%ecx,4)
+#endif
+
+	.section .text.sse2,"ax",@progbits
+	ALIGN (4)
+ENTRY (sse2_memset32_atom)
+	ENTRANCE
+
+	movl	LEN(%esp), %ecx
+#ifdef USE_AS_ANDROID
+	shr     $2, %ecx
+#endif
+#ifdef USE_AS_BZERO32
+	xor	%eax, %eax
+#else
+	mov	DWDS(%esp), %eax
+	mov	%eax, %edx
+#endif
+	movl	DEST(%esp), %edx
+	cmp	$16, %ecx
+	jae	L(16dbwordsormore)
+
+L(write_less16dbwords):
+	lea	(%edx, %ecx, 4), %edx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_less16dbwords))
+
+	.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	%eax, -60(%edx)
+L(write_14dbwords):
+	movl	%eax, -56(%edx)
+L(write_13dbwords):
+	movl	%eax, -52(%edx)
+L(write_12dbwords):
+	movl	%eax, -48(%edx)
+L(write_11dbwords):
+	movl	%eax, -44(%edx)
+L(write_10dbwords):
+	movl	%eax, -40(%edx)
+L(write_9dbwords):
+	movl	%eax, -36(%edx)
+L(write_8dbwords):
+	movl	%eax, -32(%edx)
+L(write_7dbwords):
+	movl	%eax, -28(%edx)
+L(write_6dbwords):
+	movl	%eax, -24(%edx)
+L(write_5dbwords):
+	movl	%eax, -20(%edx)
+L(write_4dbwords):
+	movl	%eax, -16(%edx)
+L(write_3dbwords):
+	movl	%eax, -12(%edx)
+L(write_2dbwords):
+	movl	%eax, -8(%edx)
+L(write_1dbwords):
+	movl	%eax, -4(%edx)
+L(write_0dbwords):
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(16dbwordsormore):
+	test	$3, %edx
+	jz	L(aligned4bytes)
+	mov	%eax, (%edx)
+	mov	%eax, -4(%edx, %ecx, 4)
+	sub	$1, %ecx
+	rol	$24, %eax
+	add	$1, %edx
+	test	$3, %edx
+	jz	L(aligned4bytes)
+	ror	$8, %eax
+	add	$1, %edx
+	test	$3, %edx
+	jz	L(aligned4bytes)
+	ror	$8, %eax
+	add	$1, %edx
+L(aligned4bytes):
+	shl	$2, %ecx
+
+#ifdef USE_AS_BZERO32
+	pxor	%xmm0, %xmm0
+#else
+	movd	%eax, %xmm0
+	pshufd	$0, %xmm0, %xmm0
+#endif
+	testl	$0xf, %edx
+	jz	L(aligned_16)
+/* ECX > 32 and EDX is not 16 byte aligned.  */
+L(not_aligned_16):
+	movdqu	%xmm0, (%edx)
+	movl	%edx, %eax
+	and	$-16, %edx
+	add	$16, %edx
+	sub	%edx, %eax
+	add	%eax, %ecx
+	movd	%xmm0, %eax
+	ALIGN (4)
+L(aligned_16):
+	cmp	$128, %ecx
+	jae	L(128bytesormore)
+
+L(aligned_16_less128bytes):
+	add	%ecx, %edx
+	shr	$2, %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	ALIGN (4)
+L(128bytesormore):
+#ifdef SHARED_CACHE_SIZE
+	PUSH (%ebx)
+	mov	$SHARED_CACHE_SIZE, %ebx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	mov	__x86_shared_cache_size@GOTOFF(%ebx), %ebx
+# else
+	PUSH (%ebx)
+	mov	__x86_shared_cache_size, %ebx
+# endif
+#endif
+	cmp	%ebx, %ecx
+	jae	L(128bytesormore_nt_start)
+	
+#ifdef DATA_CACHE_SIZE
+	POP (%ebx)
+# define RESTORE_EBX_STATE CFI_PUSH (%ebx)
+	cmp	$DATA_CACHE_SIZE, %ecx
+#else
+# ifdef SHARED
+#  define RESTORE_EBX_STATE
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmp	__x86_data_cache_size@GOTOFF(%ebx), %ecx
+# else
+	POP (%ebx)
+#  define RESTORE_EBX_STATE CFI_PUSH (%ebx)
+	cmp	__x86_data_cache_size, %ecx
+# endif
+#endif
+
+	jae	L(128bytes_L2_normal)
+	subl	$128, %ecx
+L(128bytesormore_normal):
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	lea	128(%edx), %edx
+	jb	L(128bytesless_normal)
+
+
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	lea	128(%edx), %edx
+	jae	L(128bytesormore_normal)
+
+L(128bytesless_normal):
+	lea	128(%ecx), %ecx
+	add	%ecx, %edx
+	shr	$2, %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	ALIGN (4)
+L(128bytes_L2_normal):
+	prefetcht0	0x380(%edx)
+	prefetcht0	0x3c0(%edx)
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movaps	%xmm0, 0x10(%edx)
+	movaps	%xmm0, 0x20(%edx)
+	movaps	%xmm0, 0x30(%edx)
+	movaps	%xmm0, 0x40(%edx)
+	movaps	%xmm0, 0x50(%edx)
+	movaps	%xmm0, 0x60(%edx)
+	movaps	%xmm0, 0x70(%edx)
+	add	$128, %edx
+	cmp	$128, %ecx 	
+	jae	L(128bytes_L2_normal)
+
+L(128bytesless_L2_normal):
+	add	%ecx, %edx
+	shr	$2, %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	RESTORE_EBX_STATE
+L(128bytesormore_nt_start):
+	sub	%ebx, %ecx
+	mov	%ebx, %eax
+	and	$0x7f, %eax
+	add	%eax, %ecx
+	movd	%xmm0, %eax
+	ALIGN (4)
+L(128bytesormore_shared_cache_loop):
+	prefetcht0	0x3c0(%edx)
+	prefetcht0	0x380(%edx)
+	sub	$0x80, %ebx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	add	$0x80, %edx
+	cmp	$0x80, %ebx
+	jae	L(128bytesormore_shared_cache_loop)
+	cmp	$0x80, %ecx
+	jb	L(shared_cache_loop_end)
+
+	ALIGN (4)
+L(128bytesormore_nt):
+	sub	$0x80, %ecx
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm0, 0x10(%edx)
+	movntdq	%xmm0, 0x20(%edx)
+	movntdq	%xmm0, 0x30(%edx)
+	movntdq	%xmm0, 0x40(%edx)
+	movntdq	%xmm0, 0x50(%edx)
+	movntdq	%xmm0, 0x60(%edx)
+	movntdq	%xmm0, 0x70(%edx)
+	add	$0x80, %edx
+	cmp	$0x80, %ecx
+	jae	L(128bytesormore_nt)
+	sfence
+L(shared_cache_loop_end):
+#if defined DATA_CACHE_SIZE || !defined SHARED
+	POP (%ebx)
+#endif
+	add	%ecx, %edx
+	shr	$2, %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	.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(%edx)
+L(aligned_16_96bytes):
+	movdqa	%xmm0, -96(%edx)
+L(aligned_16_80bytes):
+	movdqa	%xmm0, -80(%edx)
+L(aligned_16_64bytes):
+	movdqa	%xmm0, -64(%edx)
+L(aligned_16_48bytes):
+	movdqa	%xmm0, -48(%edx)
+L(aligned_16_32bytes):
+	movdqa	%xmm0, -32(%edx)
+L(aligned_16_16bytes):
+	movdqa	%xmm0, -16(%edx)
+L(aligned_16_0bytes):
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_116bytes):
+	movdqa	%xmm0, -116(%edx)
+L(aligned_16_100bytes):
+	movdqa	%xmm0, -100(%edx)
+L(aligned_16_84bytes):
+	movdqa	%xmm0, -84(%edx)
+L(aligned_16_68bytes):
+	movdqa	%xmm0, -68(%edx)
+L(aligned_16_52bytes):
+	movdqa	%xmm0, -52(%edx)
+L(aligned_16_36bytes):
+	movdqa	%xmm0, -36(%edx)
+L(aligned_16_20bytes):
+	movdqa	%xmm0, -20(%edx)
+L(aligned_16_4bytes):
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_120bytes):
+	movdqa	%xmm0, -120(%edx)
+L(aligned_16_104bytes):
+	movdqa	%xmm0, -104(%edx)
+L(aligned_16_88bytes):
+	movdqa	%xmm0, -88(%edx)
+L(aligned_16_72bytes):
+	movdqa	%xmm0, -72(%edx)
+L(aligned_16_56bytes):
+	movdqa	%xmm0, -56(%edx)
+L(aligned_16_40bytes):
+	movdqa	%xmm0, -40(%edx)
+L(aligned_16_24bytes):
+	movdqa	%xmm0, -24(%edx)
+L(aligned_16_8bytes):
+	movq	%xmm0, -8(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_124bytes):
+	movdqa	%xmm0, -124(%edx)
+L(aligned_16_108bytes):
+	movdqa	%xmm0, -108(%edx)
+L(aligned_16_92bytes):
+	movdqa	%xmm0, -92(%edx)
+L(aligned_16_76bytes):
+	movdqa	%xmm0, -76(%edx)
+L(aligned_16_60bytes):
+	movdqa	%xmm0, -60(%edx)
+L(aligned_16_44bytes):
+	movdqa	%xmm0, -44(%edx)
+L(aligned_16_28bytes):
+	movdqa	%xmm0, -28(%edx)
+L(aligned_16_12bytes):
+	movq	%xmm0, -12(%edx)
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+END (sse2_memset32_atom)