[MIPS] Benchmark test for MIPS memset16/memset32

Author: Chris Dearman <chris@mips.com>
AuthorDate: 2011-07-26 19:24:40
Change-Id: I0a7b3360aaf45de8ee47744f7031b84f15f37611

8 files changed, 766 insertions, 0 deletions

diff --git a/libcutils/Android.mk b/libcutils/Android.mk
index d9bd8d8..fbec131 100644
--- a/libcutils/Android.mk
+++ b/libcutils/Android.mk
@@ -158,3 +158,5 @@ LOCAL_SRC_FILES := str_parms.c hashmap.c memory.c
 LOCAL_SHARED_LIBRARIES := liblog
 LOCAL_MODULE_TAGS := optional
 include $(BUILD_EXECUTABLE)
+
+include $(call all-makefiles-under,$(LOCAL_PATH))
diff --git a/libcutils/tests/Android.mk b/libcutils/tests/Android.mk
new file mode 100644
index 0000000..6571161
--- /dev/null
+++ b/libcutils/tests/Android.mk
@@ -0,0 +1 @@
+include $(all-subdir-makefiles)
diff --git a/libcutils/tests/memset_mips/Android.mk b/libcutils/tests/memset_mips/Android.mk
new file mode 100644
index 0000000..c22fca9
--- /dev/null
+++ b/libcutils/tests/memset_mips/Android.mk
@@ -0,0 +1,23 @@
+# Copyright 2012 The Android Open Source Project
+
+ifeq ($(TARGET_ARCH),mips)
+
+LOCAL_PATH:= $(call my-dir)
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES:= \
+	test_memset.c \
+	android_memset_dumb.S \
+	android_memset_test.S \
+	memset_cmips.S \
+	memset_omips.S
+
+LOCAL_MODULE:= test_memset
+
+LOCAL_FORCE_STATIC_EXECUTABLE := true
+LOCAL_STATIC_LIBRARIES := libcutils libc
+LOCAL_MODULE_TAGS := tests
+
+include $(BUILD_EXECUTABLE)
+
+endif
diff --git a/libcutils/tests/memset_mips/android_memset_dumb.S b/libcutils/tests/memset_mips/android_memset_dumb.S
new file mode 100644
index 0000000..c8a1a37
--- /dev/null
+++ b/libcutils/tests/memset_mips/android_memset_dumb.S
@@ -0,0 +1,36 @@
+	.global	android_memset16_dumb
+	.type   android_memset16_dumb, @function
+android_memset16_dumb:
+        .ent	android_memset16_dumb
+
+	.set	noreorder
+	beqz	$a2,9f
+	 srl	$a2,1
+
+1:	sh	$a1,($a0)
+	subu	$a2,1
+	bnez	$a2,1b
+	 addu	$a0,2
+	.set reorder
+
+9:	j	$ra
+        .end	android_memset16_dumb
+	.size	android_memset16_dumb,.-android_memset16_dumb
+
+	.global android_memset32_dumb
+	.type	android_memset32_dumb, @function
+android_memset32_dumb:
+        .ent	android_memset32_dumb
+	.set	noreorder
+	beqz	$a2,9f
+	 srl	$a2,2
+
+1:	sw	$a1,($a0)
+	subu	$a2,1
+	bnez	$a2,1b
+	 addu	$a0,4
+	.set reorder
+
+9:	j	$ra
+        .end	android_memset32_dumb
+	.size	android_memset32_dumb,.-android_memset32_dumb
diff --git a/libcutils/tests/memset_mips/android_memset_test.S b/libcutils/tests/memset_mips/android_memset_test.S
new file mode 100644
index 0000000..e918843
--- /dev/null
+++ b/libcutils/tests/memset_mips/android_memset_test.S
@@ -0,0 +1,152 @@
+/*
+ * Copyright (C) 2006 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.
+ */
+
+#ifdef NDEBUG
+#define DBG #
+#else
+#define DBG
+#endif
+
+	.text
+	.align
+
+        /*
+         * Optimized memset16 for MIPS
+         *
+         * void android_memset16_test(uint16_t* dst, uint16_t value, size_t size);
+         *
+         */
+
+	.global	android_memset16_test
+	.type   android_memset16_test, @function
+android_memset16_test:
+        .ent	android_memset16_test
+	.set	noreorder
+
+	/* Check parameters */
+DBG	andi	$t0,$a0,1	/* $a0 must be halfword aligned */
+DBG	tne	$t0
+DBG	lui	$t1,0xffff	/* $a1 must be 16bits */
+DBG	and	$t1,$a1
+DBG	tne	$t1
+DBG	andi	$t2,$a2,1	/* $a2 must be even */
+DBG	tne	$t2
+
+#if (__mips==32) && (__mips_isa_rev>=2)
+	ins	$a2,$0,0,1
+#else
+	li	$t0,~1
+	and	$a2,$t0
+#endif
+
+	move	$t8,$ra
+	blez	$a2,9f		/* Anything to do? */
+	 andi	$t0,$a0,2	/* Check dst alignment */
+	/* Expand value to 32 bits and check destination alignment */
+#if (__mips==32) && (__mips_isa_rev>=2)
+	beqz	$t0,.Laligned32	/* dst is 32 bit aligned */
+	 ins	$a1,$a1,16,16
+#else
+	sll	$t2,$a1,16
+	beqz	$t0,.Laligned32	/* dst is 32 bit aligned */
+	 or	$a1,$t2
+#endif
+	sh	$a1,($a0)	/* do one halfword to get aligned */
+	subu	$a2,2
+	addu	$a0,2
+
+.Laligned32:
+	and	$t1,$a2,63	/* is there enough left to do a full 64 byte loop? */
+	beq	$a2,$t1,1f
+	 subu	$t2,$a2,$t1	/* $t2 is the number of bytes to do in loop64 */
+	addu	$t3,$a0,$t2	/* $t3 is the end marker for loop64 */
+	subu	$a2,$t2
+.Lloop64:
+	addu	$a0,64
+	sw	$a1,-64($a0)
+	sw	$a1,-60($a0)
+	sw	$a1,-56($a0)
+	sw	$a1,-52($a0)
+	sw	$a1,-48($a0)
+	sw	$a1,-44($a0)
+	sw	$a1,-40($a0)
+	sw	$a1,-36($a0)
+	sw	$a1,-32($a0)
+	sw	$a1,-28($a0)
+	sw	$a1,-24($a0)
+	sw	$a1,-20($a0)
+	sw	$a1,-16($a0)
+	sw	$a1,-12($a0)
+	sw	$a1,-8($a0)
+	bne	$a0,$t3,.Lloop64
+	sw	$a1,-4($a0)
+
+	/* Do the last 0..62 bytes */
+1:	li	$t0,64+12
+	andi	$t1,$a2,0x3c	/* $t1 how many bytes to store using sw */
+	bal	1f
+	 subu	$t0,$t1		/* 64+12-$t0 is offset to jump from 1f */
+1:	addu	$ra,$t0
+	j	$ra
+	 subu	$a2,$t1
+2:	sw	$a1,60($a0)
+	sw	$a1,56($a0)
+	sw	$a1,52($a0)
+	sw	$a1,48($a0)
+	sw	$a1,44($a0)
+	sw	$a1,40($a0)
+	sw	$a1,36($a0)
+	sw	$a1,32($a0)
+	sw	$a1,28($a0)
+	sw	$a1,24($a0)
+	sw	$a1,20($a0)
+	sw	$a1,16($a0)
+	sw	$a1,12($a0)
+	sw	$a1,8($a0)
+	sw	$a1,4($a0)
+	sw	$a1,0($a0)
+
+	beqz	$a2,9f
+	 addu	$a0,$t1
+	sh	$a1,($a0)
+
+9:	j	$t8
+	 nop
+        .end	android_memset16_test
+	.size	android_memset16_test,.-android_memset16_test
+
+        /*
+         * Optimized memset32 for MIPS
+         *
+         * void android_memset32_test(uint32_t* dst, uint32_t value, size_t size);
+         *
+         */
+	.global android_memset32_test
+	.type	android_memset32_test, @function
+android_memset32_test:
+        .ent	android_memset32_test
+	.set	noreorder
+
+	/* Check parameters */
+DBG	andi	$t0,$a0,3	/* $a0 must be word aligned */
+DBG	tne	$t0
+DBG	andi	$t2,$a2,3	/* $a2 must be a multiple of 4 bytes */
+DBG	tne	$t2
+
+	b	.Laligned32
+	 move	$t8,$ra
+        .end	android_memset32_test
+	.size	android_memset32_test,.-android_memset32_test
diff --git a/libcutils/tests/memset_mips/memset_cmips.S b/libcutils/tests/memset_mips/memset_cmips.S
new file mode 100644
index 0000000..f8f3a91
--- /dev/null
+++ b/libcutils/tests/memset_mips/memset_cmips.S
@@ -0,0 +1,227 @@
+/*
+ * Copyright (c) 2009
+ *      MIPS Technologies, Inc., California.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
+ *    contributors may be used to endorse or promote products derived from
+ *    this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED.  IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+/************************************************************************
+ *
+ *  memset.S, version "64h" with 1 cache line horizon for "pref 30" and 14 nops
+ *  Version: "043009"
+ *
+ ************************************************************************/
+
+
+/************************************************************************
+ *  Include files
+ ************************************************************************/
+
+#include "machine/asm.h"
+
+/*
+ * This routine could be optimized for MIPS64. The current code only
+ * uses MIPS32 instructions.
+ */
+
+#if defined(__MIPSEB__)
+#  define SWHI	swl		/* high part is left in big-endian	*/
+#endif
+
+#if defined(__MIPSEL__)
+#  define SWHI	swr		/* high part is right in little-endian	*/
+#endif
+
+#if !(defined(XGPROF) || defined(XPROF))
+#undef SETUP_GP
+#define SETUP_GP
+#endif
+
+LEAF(memset_cmips,0)
+
+	.set	noreorder
+	.set	noat
+
+	addu	t0,a0,a2		# t0 is the "past the end" address
+	slti	AT,a2,4			# is a2 less than 4?
+	bne	AT,zero,.Llast4		# if yes, go to last4
+	move	v0,a0			# memset returns the dst pointer
+
+	beq	a1,zero,.Lset0
+	subu	v1,zero,a0
+
+	# smear byte into 32 bit word
+#if (__mips==32) && (__mips_isa_rev>=2)
+	ins     a1, a1, 8, 8        # Replicate fill byte into half-word.
+	ins     a1, a1, 16, 16      # Replicate fill byte into word.
+#else
+	and	a1,0xff
+	sll	AT,a1,8
+	or	a1,AT
+	sll	AT,a1,16
+	or	a1,AT
+#endif
+
+.Lset0:	andi	v1,v1,0x3		# word-unaligned address?
+	beq	v1,zero,.Laligned	# v1 is the unalignment count
+	subu	a2,a2,v1
+	SWHI	a1,0(a0)
+	addu	a0,a0,v1
+
+# Here we have the "word-aligned" a0 (until the "last4")
+.Laligned:
+	andi	t8,a2,0x3f	# any 64-byte chunks?
+				# t8 is the byte count past 64-byte chunks
+	beq	a2,t8,.Lchk8w	# when a2==t8, no 64-byte chunks
+				# There will be at most 1 32-byte chunk then
+	subu	a3,a2,t8	# subtract from a2 the reminder
+				# Here a3 counts bytes in 16w chunks
+	addu	a3,a0,a3	# Now a3 is the final dst after 64-byte chunks
+
+# Find out, if there are any 64-byte chunks after which will be still at least
+# 96 bytes left. The value "96" is calculated as needed buffer for
+# "pref 30,64(a0)" prefetch, which can be used as "pref 30,0(a0)" after
+# incrementing "a0" by 64.
+# For "a2" below 160 there will be no such "pref 30 safe" 64-byte chunk.
+#
+	sltiu	v1,a2,160
+	bgtz	v1,.Lloop16w_nopref30	# skip "pref 30,0(a0)"
+	subu	t7,a2,96	# subtract "pref 30 unsafe" region
+		# below we have at least 1 64-byte chunk which is "pref 30 safe"
+	andi	t6,t7,0x3f	# t6 is past "64-byte safe chunks" reminder
+	subu	t5,t7,t6	# subtract from t7 the reminder
+				# Here t5 counts bytes in 16w "safe" chunks
+	addu	t4,a0,t5	# Now t4 is the dst after 64-byte "safe" chunks
+
+# Don't use "pref 30,0(a0)" for a0 in a "middle" of a cache line
+#	pref	30,0(a0)
+# Here we are in the region, where it is safe to use "pref 30,64(a0)"
+.Lloop16w:
+	addiu	a0,a0,64
+	pref	30,-32(a0)	# continue setting up the dest, addr 64-32
+	sw	a1,-64(a0)
+	sw	a1,-60(a0)
+	sw	a1,-56(a0)
+	sw	a1,-52(a0)
+	sw	a1,-48(a0)
+	sw	a1,-44(a0)
+	sw	a1,-40(a0)
+	sw	a1,-36(a0)
+	nop
+	nop			# the extra nop instructions help to balance
+	nop			# cycles needed for "store" + "fill" + "evict"
+	nop			# For 64byte store there are needed 8 fill
+	nop			# and 8 evict cycles, i.e. at least 32 instr.
+	nop
+	nop
+	pref	30,0(a0)	# continue setting up the dest, addr 64-0
+	sw	a1,-32(a0)
+	sw	a1,-28(a0)
+	sw	a1,-24(a0)
+	sw	a1,-20(a0)
+	sw	a1,-16(a0)
+	sw	a1,-12(a0)
+	sw	a1,-8(a0)
+	sw	a1,-4(a0)
+	nop
+	nop
+	nop
+	nop			# NOTE: adding 14 nop-s instead of 12 nop-s
+	nop			# gives better results for "fast" memory
+	nop
+	bne	a0,t4,.Lloop16w
+	nop
+
+	beq	a0,a3,.Lchk8w	# maybe no more 64-byte chunks?
+	nop			# this "delayed slot" is useless ...
+
+.Lloop16w_nopref30:	# there could be up to 3 "64-byte nopref30" chunks
+	addiu	a0,a0,64
+	sw	a1,-64(a0)
+	sw	a1,-60(a0)
+	sw	a1,-56(a0)
+	sw	a1,-52(a0)
+	sw	a1,-48(a0)
+	sw	a1,-44(a0)
+	sw	a1,-40(a0)
+	sw	a1,-36(a0)
+	sw	a1,-32(a0)
+	sw	a1,-28(a0)
+	sw	a1,-24(a0)
+	sw	a1,-20(a0)
+	sw	a1,-16(a0)
+	sw	a1,-12(a0)
+	sw	a1,-8(a0)
+	bne	a0,a3,.Lloop16w_nopref30
+	sw	a1,-4(a0)
+
+.Lchk8w:		# t8 here is the byte count past 64-byte chunks
+
+	andi	t7,t8,0x1f	# is there a 32-byte chunk?
+				# the t7 is the reminder count past 32-bytes
+	beq	t8,t7,.Lchk1w	# when t8==t7, no 32-byte chunk
+	move	a2,t7
+
+	sw	a1,0(a0)
+	sw	a1,4(a0)
+	sw	a1,8(a0)
+	sw	a1,12(a0)
+	sw	a1,16(a0)
+	sw	a1,20(a0)
+	sw	a1,24(a0)
+	sw	a1,28(a0)
+	addiu	a0,a0,32
+
+.Lchk1w:
+	andi	t8,a2,0x3	# now t8 is the reminder past 1w chunks
+	beq	a2,t8,.Llast4
+	subu	a3,a2,t8	# a3 is the count of bytes in 1w chunks
+	addu	a3,a0,a3	# now a3 is the dst address past the 1w chunks
+
+# copying in words (4-byte chunks)
+.LwordCopy_loop:
+	addiu	a0,a0,4
+	bne	a0,a3,.LwordCopy_loop
+	sw	a1,-4(a0)
+
+.Llast4:beq	a0,t0,.Llast4e
+.Llast4l:addiu	a0,a0,1
+	bne	a0,t0,.Llast4l
+	sb	a1,-1(a0)
+
+.Llast4e:
+	j	ra
+	nop
+
+	.set	at
+	.set	reorder
+
+END(memset_cmips)
+
+
+/************************************************************************
+ *  Implementation : Static functions
+ ************************************************************************/
+
diff --git a/libcutils/tests/memset_mips/memset_omips.S b/libcutils/tests/memset_mips/memset_omips.S
new file mode 100644
index 0000000..4c47001
--- /dev/null
+++ b/libcutils/tests/memset_mips/memset_omips.S
@@ -0,0 +1,90 @@
+/* Copyright (C) 2002, 2003 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Hartvig Ekner <hartvige@mips.com>, 2002.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+/* void *memset_omips(void *s, int c, size_t n).  */
+
+#include "machine/asm.h"
+
+#ifdef __mips64
+#error mips32 code being compiled for mips64!
+#endif
+
+#if defined(__MIPSEB__)
+#error big-endian is not supported in Broadcom MIPS Android platform
+# define SWHI	swl		/* high part is left in big-endian	*/
+#else
+# define SWHI	swr		/* high part is right in little-endian	*/
+#endif
+
+LEAF (memset_omips,0)
+	.set	noreorder
+
+	slti	t1, a2, 8		# Less than 8?
+	bne	t1, zero, .Llast8
+	move	v0, a0			# Setup exit value before too late
+
+	beq	a1, zero, .Lueven	# If zero pattern, no need to extend
+	andi	a1, 0xff		# Avoid problems with bogus arguments
+	sll	t0, a1, 8
+	or	a1, t0
+	sll	t0, a1, 16
+	or	a1, t0			# a1 is now pattern in full word
+
+.Lueven:
+	subu	t0, zero, a0		# Unaligned address?
+	andi	t0, 0x3
+	beq	t0, zero, .Lchkw
+	subu	a2, t0
+	SWHI	a1, 0(a0)		# Yes, handle first unaligned part
+	addu	a0, t0			# Now both a0 and a2 are updated
+
+.Lchkw:
+	andi	t0, a2, 0x7		# Enough left for one loop iteration?
+	beq	t0, a2, .Lchkl
+	subu	a3, a2, t0
+	addu	a3, a0			# a3 is last loop address +1
+	move	a2, t0			# a2 is now # of bytes left after loop
+.Lloopw:
+	addiu	a0, 8			# Handle 2 words pr. iteration
+	sw	a1, -8(a0)
+	bne	a0, a3, .Lloopw
+	sw	a1, -4(a0)
+
+.Lchkl:
+	andi	t0, a2, 0x4		# Check if there is at least a full
+	beq	t0, zero, .Llast8	#  word remaining after the loop
+	subu	a2, t0
+	sw	a1, 0(a0)		# Yes...
+	addiu	a0, 4
+
+.Llast8:
+	blez	a2, .Lexit		# Handle last 8 bytes (if cnt>0)
+	addu	a3, a2, a0		# a3 is last address +1
+.Llst8l:
+	addiu	a0, 1
+	bne	a0, a3, .Llst8l
+	sb	a1, -1(a0)
+.Lexit:
+	j	ra			# Bye, bye
+	nop
+
+	.set	reorder
+END (memset_omips)
+
+
diff --git a/libcutils/tests/memset_mips/test_memset.c b/libcutils/tests/memset_mips/test_memset.c
new file mode 100644
index 0000000..9705c65
--- /dev/null
+++ b/libcutils/tests/memset_mips/test_memset.c
@@ -0,0 +1,235 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+#include <cutils/memory.h>
+#include <time.h>
+
+/*
+ * All systems must implement or emulate the rdhwr instruction to read
+ * the userlocal register. Systems that emulate also return teh count register
+ * when accessing register $2 so this should work on most systems
+ */
+#define USE_RDHWR
+
+#ifdef USE_RDHWR
+#define UNITS "cycles"
+#define SCALE 2			/* Most CPU's */
+static inline uint32_t
+get_count(void)
+{
+  uint32_t res;
+  asm volatile (".set push; .set mips32r2; rdhwr %[res],$2; .set pop" : [res] "=r" (res) : : "memory");
+  return res;
+}
+#else
+#define UNITS "ns"
+#define SCALE 1
+static inline uint32_t
+get_count(void)
+{
+  struct timespec now;
+  uint32_t res;
+  clock_gettime(CLOCK_REALTIME, &now);
+  res = (uint32_t)(now.tv_sec * 1000000000LL + now.tv_nsec);
+  // printf ("now=%d.%09d res=%d\n", (int)now.tv_sec, (int)now.tv_nsec, res);
+  return res;
+}
+#endif
+
+uint32_t overhead;
+void
+measure_overhead(void)
+{
+  int i;
+  uint32_t start, stop, delta;
+  for (i = 0; i < 32; i++) {
+    start = get_count();
+    stop = get_count();
+    delta = stop - start;
+    if (overhead == 0 || delta < overhead)
+      overhead = delta;
+  }
+  printf("overhead is %d"UNITS"\n", overhead);
+}
+
+uint32_t
+timeone(void (*fn)(), void *d, uint32_t val, uint32_t bytes)
+{
+  uint32_t start, stop, delta;
+  start = get_count();
+  (*fn)(d, val, bytes);
+  stop = get_count();
+  delta = stop - start - overhead;
+  // printf ("start=0x%08x stop=0x%08x delta=0x%08x\n", start, stop, delta);
+  return delta * SCALE;
+}
+
+/* define VERIFY to check that memset only touches the bytes it's supposed to */
+/*#define VERIFY*/
+
+/*
+ * Using a big arena means that memset will most likely miss in the cache
+ * NB Enabling verification effectively warms up the cache...
+ */
+#define ARENASIZE 0x1000000
+#ifdef VERIFY
+char arena[ARENASIZE+8];	/* Allow space for guard words */
+#else
+char arena[ARENASIZE];
+#endif
+
+void
+testone(char *tag, void (*fn)(), int trials, int minbytes, int maxbytes, int size, int threshold)
+{
+  int offset;
+  void *d;
+  void *p;
+  uint32_t v, notv = 0;
+  uint32_t n;
+  int i, units;
+  int totalunits = 0, totalbytes = 0, samples = 0;
+
+  /* Reset RNG to ensure each test uses same random values */
+  srand(0);			/* FIXME should be able to use some other seed than 0 */
+
+  for (i = 0; i < trials; i++) {
+    n = minbytes + (rand() % (maxbytes-minbytes));	/* How many bytes to do */
+    offset = ((rand() % (ARENASIZE-n)));		/* Where to start */
+
+#ifdef VERIFY
+    offset += 4;		/* Allow space for guard word at beginning */
+#endif
+    v = rand();
+
+    /* Adjust alignment and sizes based on transfer size */
+    switch (size) {
+    case 1:
+      v &= 0xff;
+      notv = ~v & 0xff;
+      break;
+    case 2:
+      v &= 0xffff;
+      notv = ~v & 0xffff;
+      offset &= ~1;
+      n &= ~1;
+      break;
+    case 4:
+      notv = ~v;
+      offset &= ~3;
+      n &= ~3;
+      break;
+    }
+
+    d = &arena[offset];
+
+#ifdef VERIFY
+    /* Initialise the area and guard words */
+    for (p = &arena[offset-4]; p < (void *)&arena[offset+n+4]; p = (void *)((uint32_t)p + size)) {
+      if (size == 1)
+	*(uint8_t *)p = notv;
+      else if (size == 2)
+	*(uint16_t *)p = notv;
+      else if (size == 4)
+	*(uint32_t *)p = notv;
+    }
+#endif
+    units = timeone(fn, d, v, n);
+#ifdef VERIFY
+    /* Check the area and guard words */
+    for (p = &arena[offset-4]; p < (void *)&arena[offset+n+4]; p = (void *)((uint32_t)p + size)) {
+      uint32_t got = 0;
+      if (size == 1)
+	got = *(uint8_t *)p;
+      else if (size == 2)
+	got = *(uint16_t *)p;
+      else if (size == 4)
+	got = *(uint32_t *)p;
+      if (p < (void *)&arena[offset]) {
+	if (got != notv)
+	  printf ("%s: verify failure: preguard:%p d=%p v=%08x got=%08x n=%d\n", tag, p, d, v, got, n);
+      }
+      else if (p < (void *)&arena[offset+n]) {
+	if (got != v)
+	  printf ("%s: verify failure: arena:%p d=%p v=%08x got=%08x n=%d\n", tag, p, d, v, n);
+      }
+      else {
+	if (got != notv)
+	  printf ("%s: verify failure: postguard:%p d=%p v=%08x got=%08x n=%d\n", tag, p, d, v, n);
+      }
+    }
+#endif
+
+    /* If the cycle count looks reasonable include it in the statistics */
+    if (units < threshold) {
+      totalbytes += n;
+      totalunits += units;
+      samples++;
+    }
+  }
+
+  printf("%s: samples=%d avglen=%d avg" UNITS "=%d bp"UNITS"=%g\n",
+	 tag, samples, totalbytes/samples, totalunits/samples, (double)totalbytes/(double)totalunits);
+}
+
+extern void android_memset32_dumb(uint32_t* dst, uint32_t value, size_t size);
+extern void android_memset16_dumb(uint32_t* dst, uint16_t value, size_t size);
+extern void android_memset32_test(uint32_t* dst, uint32_t value, size_t size);
+extern void android_memset16_test(uint32_t* dst, uint16_t value, size_t size);
+extern void memset_cmips(void* dst, int value, size_t size);
+extern void memset_omips(void* dst, int value, size_t size);
+
+int
+main(int argc, char **argv)
+{
+  int i;
+  struct {
+    char *type;
+    int trials;
+    int minbytes, maxbytes;
+  } *pp, params[] = {
+    {"small",  10000,   0,   64},
+    {"medium", 10000,  64,  512},
+    {"large",  10000, 512, 1280},
+    {"varied", 10000,   0, 1280},
+  };
+#define NPARAMS (sizeof(params)/sizeof(params[0]))
+  struct {
+    char *name;
+    void (*fn)();
+    int size;
+  } *fp, functions[] = {
+    {"dmemset16", (void (*)())android_memset16_dumb, 2},
+    {"tmemset16", (void (*)())android_memset16_test, 2},
+    {"lmemset16", (void (*)())android_memset16,      2},
+
+    {"dmemset32", (void (*)())android_memset32_dumb, 4},
+    {"tmemset32", (void (*)())android_memset32_test, 4},
+    {"lmemset32", (void (*)())android_memset32,      4},
+
+    {"cmemset",    (void (*)())memset_cmips,         1},
+    {"omemset",    (void (*)())memset_omips,         1},
+    {"lmemset",    (void (*)())memset,               1},
+  };
+#define NFUNCTIONS (sizeof(functions)/sizeof(functions[0]))
+  char tag[40];
+  int threshold;
+
+  measure_overhead();
+
+  /* Warm up the page cache */
+  memset(arena, 0xff, ARENASIZE); /* use 0xff now to avoid COW later */
+
+  for (fp = functions; fp < &functions[NFUNCTIONS]; fp++) {
+    (fp->fn)(arena, 0xffffffff, ARENASIZE);	/* one call to get the code into Icache */
+    for (pp = params; pp < &params[NPARAMS]; pp++) {
+      sprintf(tag, "%10s: %7s %4d-%4d", fp->name, pp->type, pp->minbytes, pp->maxbytes);
+
+      /* Set the cycle threshold */
+      threshold = pp->maxbytes * 4 * 10;	/* reasonable for cycles and ns */
+      testone(tag, fp->fn, pp->trials, pp->minbytes, pp->maxbytes, fp->size, threshold);
+    }
+    printf ("\n");
+  }
+
+  return 0;
+}