A project that I’m working on at the moment calls for a very small footprint. This post is about how to make code really small for the ARM architecture.

As you can probably guess, I’m interested in operating system code, so as an example, I’ve taken a very simple piece of operating system code, and stripped it right back to demonstrate some of the techniques to use when optimising for space. So here is the snippet of code we are going to optimise:

01 struct thread {
02     unsigned int notify_bits;
03 };
04 
05 unsigned int
06 poll(struct thread *thread, unsigned int mask)
07 {
08     unsigned int result;
09     result = thread->notify_bits & mask;
10     if (result) {
11         thread->notify_bits &= ~result;
12     }
13     return result;
14 }

In this very simple operating system we have threads (thread data is stored in struct thread). Each thread has a set of 32 signals (encoded in a single word notify_bits). This poll is used by a thread to determine if it has been sent any signals. The mask parameter is the set of signals that the thread is interested in checking. So, a thread can check if a single signal has been thread, or if any signal has been set, or if a specific subset of signals has been set. The function returns the signals that are available (which is simply the bit-wise and of notify_bits and mask). It is important that the function clears any signals that have been returned. This makes sure that if poll is called twice the same signals are not returned. This is achieved in lines 10—12.

So, our goal is to try and get this code as small as possible. And every byte counts! First off, we just try and compile this with the standard ARM gcc compiler. I’m using version 4.2.2. So we start with: $ arm-elf-gcc -c poll.c -o poll.o. We can then use object dump to work out what the compiler did: $ arm-elf-objdump -dl poll.o.

00000000 <poll&ht;:
poll():
   0:	e1a0c00d 	mov	ip, sp
   4:	e92dd800 	push	{fp, ip, lr, pc}
   8:	e24cb004 	sub	fp, ip, #4	; 0x4
   c:	e24dd00c 	sub	sp, sp, #12	; 0xc
  10:	e50b0014 	str	r0, [fp, #-20]
  14:	e50b1018 	str	r1, [fp, #-24]
  18:	e51b3014 	ldr	r3, [fp, #-20]
  1c:	e5932000 	ldr	r2, [r3]
  20:	e51b3018 	ldr	r3, [fp, #-24]
  24:	e0023003 	and	r3, r2, r3
  28:	e50b3010 	str	r3, [fp, #-16]
  2c:	e51b3010 	ldr	r3, [fp, #-16]
  30:	e3530000 	cmp	r3, #0	; 0x0
  34:	0a000006 	beq	54 <poll+0x54>
  38:	e51b3014 	ldr	r3, [fp, #-20]
  3c:	e5932000 	ldr	r2, [r3]
  40:	e51b3010 	ldr	r3, [fp, #-16]
  44:	e1e03003 	mvn	r3, r3
  48:	e0022003 	and	r2, r2, r3
  4c:	e51b3014 	ldr	r3, [fp, #-20]
  50:	e5832000 	str	r2, [r3]
  54:	e51b3010 	ldr	r3, [fp, #-16]
  58:	e1a00003 	mov	r0, r3
  5c:	e24bd00c 	sub	sp, fp, #12	; 0xc
  60:	e89da800 	ldm	sp, {fp, sp, pc}

So, our first go gets us 100 bytes of code. Which is 10 bytes (or 2.5 instructions) for each of our lines of code. We should be able to do better. Well, the first thing is we should try is to use optimisation: $ arm-elf-gcc -c poll.c -o poll.o -O2. This gives us a much better code generation output:

00000000 <poll>:
poll():
   0:	e5902000 	ldr	r2, [r0]
   4:	e1a0c000 	mov	ip, r0
   8:	e0110002 	ands	r0, r1, r2
   c:	11e03000 	mvnne	r3, r0
  10:	10033002 	andne	r3, r3, r2
  14:	158c3000 	strne	r3, [ip]
  18:	e12fff1e 	bx	lr

So this got us down to 28 bytes. A factor 4 improvement for one compiler flag, not bad. Now, -O2 does some standard omptimisations, but -Os, will do optimisations specifically for reducing the amount of code. So trying: $ arm-elf-gcc -c poll.c -o poll.o -Os. This gives a little bit better code-gen:

00000000 <poll>:
poll():
   0:	e5903000 	ldr	r3, [r0]
   4:	e1a02000 	mov	r2, r0
   8:	e0110003 	ands	r0, r1, r3
   c:	11c33000 	bicne	r3, r3, r0
  10:	15823000 	strne	r3, [r2]
  14:	e12fff1e 	bx	lr

Down to 24 bytes (6 instructions), is pretty good. Now, as you can see the generated code has 32-bits per instruction. The some of the ARM architectures have two distinct instruction sets, ARM and Thumb. The Thumb instruction set uses 16-bit per instruction, instead of 32-bit. This denser instruction set can enable much smaller code sizes. Of course there is a trade-off here. The functionality of the 16-bit instructions is going to be less than the 32-bit instructions. But lets give it a try. At the same time, we will tell the compiler the exact CPU we want to compile for (which is the ARM7TDMI-S) in our case. The compiler line is: $ arm-elf-gcc -c poll.c -o poll.o -Os -mcpu=arm7tdmi -mthumb. Which produces code like:

00000000 <poll>:
poll():
   0:	6803      	ldr	r3, [r0, #0]
   2:	1c02      	adds	r2, r0, #0
   4:	1c08      	adds	r0, r1, #0
   6:	4018      	ands	r0, r3
   8:	d001      	beq.n	e <poll+0xe>
   a:	4383      	bics	r3, r0
   c:	6013      	str	r3, [r2, #0]
   e:	4770      	bx	lr

So, now we are down to 16 bytes, so in Thumb we need 8 instructions (2 more than ARM), but each is only 2 bytes, not 4, so we end up with a 1/3 improvement. To get any further, we need to start looking at our code again, and see if there are ways of improving the code. Looking at the code again:

00 unsigned int
01 poll(struct thread *thread, unsigned int mask)
02 {
03     unsigned int result;
04     result = thread->notify_bits & mask;
05     if (result) {
06         thread->notify_bits &= ~result;
07     }
08     return result;
09 }

You may notice that the branch instruction on line 5, you may notice that this is actually redundant. If result is zero, then ~result well be 0xffffffff. Given this thread->notify_bits &= 0xffffffff will not change the value of thread->notify_bits. So, we can reduce this to:

00 unsigned int
01 poll(struct thread *thread, unsigned int mask)
02 {
03     unsigned int result;
04     result = thread->notify_bits & mask;
05     thread->notify_bits &= ~result;
06     return result;
07 }

When we compile this we get down to:

00000000 <poll>:
poll():
   0:	6803      	ldr	r3, [r0, #0]
   2:	4019      	ands	r1, r3
   4:	438b      	bics	r3, r1
   6:	6003      	str	r3, [r0, #0]
   8:	1c08      	adds	r0, r1, #0
   a:	4770      	bx	lr

This gets us down to 6 instructions (12 bytes). Pretty good since we started at 100 bytes. Now lets look at the object code in a little bit more detail. If you look at address 0x8, the instruction simply moves register r1 into register r0 so that it in the right place for return. (Note: The ARM ABI has the return value stored in r0). This seems like a bit of a waste, it would be good if there was a way we could have the value stored in r0 and not waste an instruction just moving values between registers. Now, to get a better understanding of what the instructions are doing, I’m going to slightly rewrite the code, and then compile with debugging, so we can see how the generated code matches up with the source code. So first, lets rewrite the code a little bit:

00 unsigned int
01 poll(struct thread *thread, unsigned int mask)
02 {
03     unsigned int tmp_notify_bits = thread->notify_bits;
04     mask &= tmp_notify_bits;
05     tmp_notify_bits &= ~mask;
06     thread->notify_bits = tmp_notify_bits;
07     return mask;
08 }

You should convince yourself that this code is equivalent to the existing code. (The code generated is identical). So, we can now line up the source code with the generated code. On line 03 (unsigned int tmp_notify_bits = thread->notify_bits), this matches up with address 0x0 (ldr r3, [r0, #0]). So, register r3 is used to store variable tmp_notify_bits. The parameter thread is stored in register r0. Now, line 04 (mask &= tmp_notify_bits) matches directly with address 0x2 (ands r1, r3). Register r1 matches directly with the mask parameter. The important part of restructuring the code as we have done, is that it becomes obvious that we can directly using the mask parameter instead of needing an extra variable like in the previous code. As we continue line 05 (tmp_notify_bits &= ~mask), matches directly to 0x4 (bics r3, r1). The bics instruction is quite neat in that it can essentially do the &= ~ in one 16-bit instruction. Line 06 (thread->notify_bits = tmp_notify_bits) stores the result back to memory, matches directly to 0x6. Now the final line of code (return mask;), needs two instructions 0x8 and 0xa (adds r0, r1, #0 and bx lr). Now the reason we need to instructions is because mask is stored in register r1, and the return value needs to be in register r0. So how can we get mask to be stored in r0 all along? Well, if we switch the parameters around poll(unsigned int mask, struct thread *thread), the mask will instead be stored in r0 instead of r1. (Note: We don’t necessarily have to change the interface of the function. If we want to support keep the interface we can use a simple macro to textually swap the parameters.) If we compile this, we get the following generated code:

00000000 <poll>:
poll():
   0:	680b      	ldr	r3, [r1, #0]
   2:	4018      	ands	r0, r3
   4:	4383      	bics	r3, r0
   6:	600b      	str	r3, [r1, #0]
   8:	4770      	bx	lr

So, we have got this down to 5 instructions, 10 bytes. This is a factor 10 improvement, not bad!

So a bit of a review:

• Compile to Thumb instructions.
• Use -Os to optimise for space.
• Look for unnecessary code.
• Look for examples of registers used in the wrong place.

Introduction

The UNIX® shell is one of the tools (along with Emacs) that I never feel like I’m using to the fullest. This morning was another opportunity to expand my shell repertoire. Most of my source code is stored in mercurial repositories, and I was sick of doing % find . | grep -v \.hg to strip out the meta-data. Surely there is a better way? Note: I’m using zsh and GNU tools on Mac OS X, although I’m pretty sure that most of the examples should work fine in Bourne-like on other UNIX-like operating systems.

find and -prune

So the thing that sucks about the find . | grep -v <foo> type pattern is that find will do all the work of descending into the foo directory and scanning it. It also means you can’t take advantage of other find niceties such as the -exec flag. So, if only there was a way to prune our directory tree search. Now, the find expression language is a mighty strange language to code in, or at least it is to me. The language consists of a set of primaries and a set of operators; an expression is made up of a primaries linked together with operators. Now, there are a few things that make this less than obvious, especially if you have been na̮ïvely using. The first is that in almost all cases, the find utility automatically adds a -print primary to your expression. The other big one (for me at least) is that expression expression is short-hand for expression -and expression. To me, grokking this made a huge difference to my understanding of how find actually works.

The other part to grokking find is that some expression are what I would call filters, that is the expression returns true if the file matches some property, e.g: -name, -newer, etc. Then there is a case of command expressions that have some type of external side-effect. Examples are: -print, -print0, -exec, -ok. Next is the -prune command, which is in a special category all on its own because it modifies the find internal search process. Finally there are what I would call global expression, such as -maxdepth and -mindepth. These are global because each applies to the entire expression even if it would not normally be evaluated. This is pretty strange behaviour if you ask me, surely they would have been better as options rather than expressions!

So with this in mind, lets build up an expression to do a find, but ignoring my .hg. So we start with an expression that will just print out the .hg directory. We need something like: find . -type d -name .hg. -type d returns true if the current path is a directory, and -name .hg returns true if that last element of the path is .hg. Since our -and is implicit, this expression only returns true for directories named .hg. And, since we do not specify and command like expression, find automatically assumes we want to print the path names, so the full expression is really somethings like: find . \( -type d -and -name .hg \) -print.

So far, this does the exact opposite of what we ant, it simply prints the .hg directory. So instead lets use -prune, which tells find to avoid descending into the directory.

Note: find usually does a pre-order traversal (i.e: directory path comes before directory content), if you use the -d option, the behaviour is changed to a post-order traversal (i.e: directory path comes after all the directory content). It should be clear that the -prune expression is only going to work with pre-order traversal. find doesn’t have anyway to do a bread-first traversal.

Since -prune doesn’t count as one of the -find commands, find will still auto-magically mess with out expression to add -print, so our expression becomes: find . \( -type d -and -name .hg -and -prune \) -print. Unsurprisingly it continues to only output the .hg directory. To get the result we wan, we need to do something in the other case, when the path is not a directory named .hg. So, in the other case lets do a -print, so we end up with a command line: find . -type d -name .hg -prune -o -print. Since there is a command in the expression, find doesn’t auto-magically add a -print. This now does pretty much what we want. Of course typing out find . -type d -name .hg -prune -o -print is a lot more work than find . | grep -v .hg, so it would be nice to make an alias for this command.

I’m going to have an alias called sfind (short for source find, not the best name, but short and easy for the muscle memory). So first attempt would be to just alias the command we had last, but I’d really like the alias to work mostly like find, so that I can add new arguments and do things like sfind -name "*.c" and have it just work™. Simply aliasing the above command would end up with the final expression being find . -type d -name .hg -prune -o -print -name "*.c", which will do all the printing before it tests if the final ends in .c. Another alternative would be to alias find . -type d -name .hg -prune -o, but this forces me to always have to type an expression; just sfind by itself won’t work. What I really need a fully formed expression, which still works when adding new constraints. What would be ideal is something like: find . -type d -name .hg -prune -o -true, so the other case always evaluates to true. In the case where no arguments are added this will end up being the equivalent of find . \( -type d -name .hg -prune -o -true \) -print, and in the case where further expression are added it will end up being: find . \( -type d -name .hg -prune -o -true <expression> \) -print. Unfortunately, there is no expression -true. So, how can we have an expression that always returns true? The best I’ve come up with is -name "*". Other possibilities might be depth +0. This works pretty well, but you might notice that now the .hg directory is being printed! Why? Because now instead of just printing files in the other case, find is printing whenever the entire expression is true, and since the last part of the left-hand-side is -prune, and -prune always returns true, the pruned directory evaluates to true. So how can we stop that? Well, something like find . -type d -name .hg -prune -false -o -name * should do the trick. But of course, just as there is no unconditional true, there isn’t an unconditional false either, thankfully, it is trivial to construct one: -not -name *. So, what we end up with is: find . -type d -name .hg -prune -not -name "*" -o -name "*". Easy!

Grepping code and shell functions

Now the next thing I often do is find . | xargs grep "foo" or grep -r "foo" .. (Which really should be find -print0 | xargs -0, but hey I’m lazy and my file names don’t contain white-space). Since I was learning more about find I figured I should work out how to do this without piping to xargs, especially since my new found power of -o means I might want to run different commands on different files, which wouldn’t work with a simple pipe to stdout. So the standard approach is using the cryptic expression: -exec grep "foo" {} \;. It is actually pretty straight forward. The braces are replaced by the current pathname and the command needs to be terminated in with a semi-colon (which needs to be escaped in most interactive shells). Now this isn’t quite as efficient as using xargs since xargs will run the command on multiple pathnames at once, whereas -exec runs the command once per pathname. Thankfully modern find seems to have an alternative representation: -exec grep "foo" {} +. The plus-sign makes the behaviour of -exec similar to that of xargs, processing multiple pathnames per invocation.

So what I really want to do is grep for a given pattern, in a certain subset of my files. And I’d like to do this without needing to type-out the verbose syntax each time. Something like srcfind <regexp&> [find-expression], would be ideal. Now a simple alias isn’t going to work for this, and I thought I was up for writing yet another tiny shell script. Luckily, I was clued on to shell functions. I feel very ignorant for not knowing about these before hand, but hey, I’m a systems programmer, not a shell-scripting sysadmin guru. Anyway, long story short, adding an appropriate shell function, such as:

srcgrep() { 
    if [ $# -gt 0 ]; then {
	    RE=$1; 
	    shift; 
	    if [ $# != 0 ]; then
		O="("; C=")"; 
	    else
		O=""; C="";
	    fi;
	    sfind -type f $O $@ $C -exec grep $RE {} \+ ;
    } else {
	echo "Usage: srcgrep  [expression]";
    } fi;
}

to my .zshrc does the trick. This shell function ended up being more complex than I initially thought it would. Briefly it checks there is the correct number of arguments, printing a usage if necessary. Next it check if there are any find expressions. When there are it is necessary to enclose them in brackets, or else the precedence rules breaks expression that might have an -or in them. Finally it runs the previously defined sfind alias. The -type f avoid running grep on directories.

Colourised grep

Now grep is nice and all, but it can sometimes be hard to pick out the actual search result from the results. This is where colour grep comes in. Grep has an option --color, which will use ANSI escape codes to markup the output. E.g:

bodyjar:/Users/benno/work/3.2/okl4-mainline% grep --color  SORT coverage.sh 
declare SORT=""
    F)  SORT="-F"
    tools/coverage.py -f fifo $SORT -s -S $SCOPE $IMAGE &
    tools/coverage.py -f fifo $SORT -s -S $SCOPE $IMAGE 2>&1 | tee $OUTFILE &

Now this is good as far as it goes, but when you start having more complex regular expressions, you get to a point where you only want to colourise part of the match, not the whole thing. For example: ^[A-Za-z_][A-Za-z0-9_]*\( matches lines with function declarations (at least when they follow the coding standard), but what I would really like hi-lighted is only the function-name, not the function-name plus opening bracket. This is where I ended up hitting the limits of shell and standard UNIX tools, and ended up back with good old familiar python. I created a simple grep replacement, colgrep, that by default will colourise in the same manner as grep, but if you use groups, it will only color the matched groups (and will give each group a different colour). Groups are part of the Python regular expression syntax (and probably Perl too). Basically if a portion of the regular expression is enclosed in brackets, that part of the regexp forms a group. E.g: In SO(R)T the 'R' will be matched as group #1. So for the previous example, if I write instead: ^([A-Za-z_][A-Za-z0-9_])*\(, only the function name will be hi-lighted. The other neat thing about python regular expressions is that you can name groups with the syntax (?P<name>expr). I take advantage of this in colgrep to color groups with the name "err*" with a background of red to hi-light errors.

Conclusion

So in conclusion, we’ve added some powerful commands to our toolbox which make it much more effective to quickly browse and search large revision controlled code bases.

When traveling with my laptop, I find myself on all manner of different wireless networks, and these diverse range of networks always seem to do seomthing funky with SMTP. To get around this I’ve resorted to sending my e-mail through an SSH tunnel. I had previously tried using both SSL and TLS as methods of securely connecting to my upstream mail server, but neither of these methods appeared to work as reliably as a good old SSH tunnel.

On my laptop I use mutt which, unlike most GUI email clients, relies on a mail server running locally on the machine to deliver the mail upstream. The default local mail server on Mac OS is postfix, which works pretty well. In postfix’s main configuration file (main.cf), you can set the relayhost option to use a specific mail server as a relay (instead of delivery mail directly form the laptop). Usually you would set this to your ISP’s outgoing mail server. E.g:

relayhost = [smtp.example.com]

Now, instead of directly contacting smtp.example.com, I want to go through an SSH tunnel. So I change the relay host to:

relayhost = [127.0.0.1]:8025

I use the dotted decimal rather than localhost to avoid an DNS lookup, and force things to use IPv4. The square brackets force it to directly use the host, rather than performing any MX lookups. Finally the last part is the port number. So, now rather than trying to contact smpt.example.com the local mailserver will try to send any email to port 8025 on the localhost.

So, the next question is how to actually make it so that port 8025 is a tunnel to port 25 on the real mailserver. One option is to use ssh’s -L option. Something like: ssh mail.example.org -L 8025:mail.example.org:25. This works fine, except that it is not very robust. The ssh session will drop out from time-to-time, and it gets up when moving between networks. The solution to this is to create the tunnel on-demand. Each time a program tries access port 8025 on localhost, we set up the tunnel then. If things were really hammering port 8025 this might be a bit of overhead, but realistically, this isn’t a problem, and things become much more robust and reliable.

On a traditional UNIX system you would use (x)inetd to automatically run a program when a socket is opened. Mac OS X instead uses launchd. A file, /System/Library/LaunchDaemon/mailtunnel.plist is created, which is our config file for creating the SSH tunnel on demand. It looks something like:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple Computer//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
	<key>Label</key>
	<string>au.id.benno.mail</string>
        <key>UserName</key>
        <string>benno</string>
	<key>ProgramArguments</key>
	<array>
		<string>/usr/bin/ssh</string>
                <string>-i</string>
                <string>tunnel_key</string>
                <string>-q</string>
                <string>-T</string>
                <string>mail.example.com</string>
	</array>
	<key>inetdCompatibility</key>
	<dict>
		<key>Wait</key>
		<false/>
	</dict>
	<key>Sockets</key>
	<dict>
		<key>Listeners</key>
		<dict>
			<key>SockServiceName</key>
			<string>8025</string>
		</dict>
	</dict>
	<key>SessionCreate</key>
	<true/>
</dict>
</plist>

The import bits here are the Sockets fragment, where we say that we are going to run the service whenever something connect to port 8025. The other important part is the ProgramArguments part, which is a verbose way of specifying what command to run to create the tunnel. In this case we run ssh -i tunnel_key -q -T mail.example-com.

The final part is actually setting up the command to create the tunnel on demand. First we create an ssh key-pair (using ssh-keygen). This is going to be a key without a passphrase so that we can ssh to our mailserver without any interaction requiring a password to be entered. One you create the public keypair, you copy the .pub file across to your mail server, and keep the private one locally. Now, ordinarily, you don’t want a passphraseless key, because it would be very powerful and if someone got access to the key, they would have full access to your account on the mailserver. The next trick we play is to limit what the key can be used for. Specifically, we ensure that the key can only be used to access port 25 on the mailserver. We do this by setting some options when adding the public key to the list of authorized keys on the mailserver.

The .ssh/authorized_keys2 file on the mailserver contains the list of keys which can be used to gain access to the server via ssh. We want to add our newly create public key to this list, but add some options to limit the scope of what can be done when using the key. The format of the authorized_keys2 file is: <options> ssh-dss <key-data> <comment>. So, instead of just adding the key, we are going to put some options in place:

command="nc localhost 25",no-X11-forwarding,no-agent-forwarding,no-port-forwarding 

What these options do is disable any forwarding using the key, and then specifically set the command to run when using the key to be nc localhost 25. nc a.k.a netcat is a very nifty little program that is useful for these scenarios. So with this set up, when using ssh, it will connect to the mailserver, but you won't get a console, or be able to use any other programs. We can test this out:

$  SSH_AUTH_SOCK= ssh  -i tunnel_key  mail.example.com 
220 mail.example.com ESMTP
quit
221 mail.example.com closing connection

The

-i

command tells ssh to use a specific key, instead of the default key in .ssh. The SSH_AUTH_SOCK= is important as it disables the use of an ssh agent that is running, which would otherwise override your key choice.

With that working, you should now be able to directly attach to the tunnel using telnet localhost 8025.

I have one of the cute OLPC machines sitting at home, but unfortunately I have not had anytime to make use of this machine and contribute in any meaningful way to the project. So I’d really like to give it to someone who can make better use of it. If you’re interested please e-mail me.

Well, for a change I will be flying for pleasure, not business (although 8 hours in economy is hardly pleasurable). Suzy and I’ll be taking time-off in Vietnam, Cambodia and Singapore. I don’t expect to be contactable or regularly checking e-mail for the next two weeks, so please don’t expect a reply to any correspondence. For those still interesting in talking at the Open Mobile Miniconf please keep the submissions coming!

Well, it started almost a year ago, but I finally now have Android booting on my Neo 1973 phone:

It ain’t exactly running fast yet, and not everything it working 100%, but I think most of the tricky bits are done. I’m starting to push most of these changes back to the android project. It seems that while I’ve been working on this Sean McNeill has been having similar successes getting Android up on the latest Freerunner phones.

So, my earlier post on this was a little premature; anyone who has tried out the code has found out that it pretty much doesn’t work (hey I did warn you!). Now there are a range of fun reasons why this didn’t work, most of which I’ve now solved.

Firstly, it turns out that EABI and ARMv4T are pretty much incompatible. (I’ll post separately about that!). In short, thumb interworking doesn’t (can’t) work, so I’ve reverted back to plain old ARMv4 architecture as my target (the only difference between ARMv4 and ARMv4T is the thumb stuff, which we can’t use until the compiler / spec is fixed.). So I’ve updated the linux-arm.mk to support ARMv4 for now as well.

Of course the next problem that this introduces is that the bx instruction doesn’t exist on ARMv4, and GCC (helpfully) complains and stops the compilation. Now a BX without thumb support is simply a mov pc, instruction, so I went through and provided a BX macro that expands to either bx or mov pc,. This is a little bit nasty/invasive because it touches all the system call bindings, thankfully these are generated anyway, but it makes the diff quite large. (When I have time I’ll make it so that generation is part of the buid system, not a manual process.)

The next problem is that the provided compiler’s libgcc library is build for ARMv5, and has instructions that just don’t exist on ARMv4 (shc as clz), so I went and built a new compiler targeted to ARMv4. There is no reason why this couldn’t be set up as a multi-lib compiler that supports both, but I don’t have enough GCC wizardry in me to work that out right now. So a new compiler.

This got things to a booting stage, but not able to mount /system or /data. Basically, Android by default uses yet another flash file-system (YAFFS), but for some reasons, which I couldn’t fully work out initially, the filesystem just didn’t seem to cleanly initialise and then mount. So, without diving too deep, I figured I could just use jffs2 instead, which I know works on the target. So I upgraded the Android build system to support allowing you to choose which filesystem type to use, and providing jffs2 as an option. This was going much better, and I got a lot further, far enough that I needed to recompile my kernel with support for some of the Android specific drivers like ashmem, binder and logger. Unfortunately I was getting a hang on an mmap call, for reasons that I couldn’t quite work out. After a lot of tedious debugging (my serial console is broken, so I have to rely on graphics console, which is really just an insane way to try and debug anything), anyway, it turns out that part of what the Dalvik virtual machine does when optimising class files is to mmap the file as writable memory. This was what was failing, with the totally useless error invalid argument. Do you know how many unique paths along the mmap system call can set EINVAL? Well it’s a lot. Anyway, long story short, it turns out that the jffs2 filesystem doesn’t support writable mmaps! %&!#.

After I finished cursing, I decided to go back to using yaffs and working out what the real problem is. After upgrading u-boot (in a pointless attempt to fix my serial console), I noticed a new write yaffs[1] command. This wasn’t there in the old version. Ok, cool, maybe this has something do to with the problem. But what is this the deal with yaffs versus yaffs1? Well it turns out that NAND has different pagesize, 512 bytes, and 2k (or multiples thereof, maybe??). And it turns out that YAFFS takes advantage of this and has different file systems for different sized NAND pages, and of course, everything that can go wrong will so, the filesystem image that the build system creates is YAFFS2 which is for 2k pages not 512b pages. So, I again updated the build system to firstly build both the mkyaffs2image and the mkyaffsimage tool, and then set off building a YAFFS file system.

Now, while u-boot supports yaffs filesystem, device firmware update doesn’t (appear to). So this means I need to copy the image to memory first, then on the device copy it from memory to flash. Now, the other fun thing is that dfu can only copy 2MB or so to RAM at a time, and the system.img file is around 52MB or so, which means that it takes around 26 individual copies of 2MB sections.... very, very painful. But in the end this more or less worked. So now I have a 56MB partition for the system, and a 4MB partition for the user and things are looking good.

Good that is, right up until the point where dalvik starts up and writes out cached version of class files to /data. You see, it needs more than 4MB, a lot more, so I’m kind of back to square one. I mean, if I’d looked at the requirements I would have read 128MB of flash, but meh, who reads requirements? The obvious option would be some type of MMC card, but as it turns out the number of handy Fry’s stores on Boeing 747 from Sydney to LA number in the zeroes.

So the /system partition is read-only, and since the only problem with jffs2 was when we were writing to it, it seems that we could use jffs2 for the read-only system partition, which has the advantage of jffs2 doing compression, and fitting in about 30MB, not about 50MB, leaving plenty of room for the user data partition, which is where the Dalvik cached files belong. This also has the advantage of being able to use normal DFU commands to install the image (yay!). So after more updates to the build system to now support individually setting the system filesystem type and the user filesystem type things seem a lot happier.

Currently, I have a system that boots init, starts up most of the system services, including the Dalvik VM, runs a bunch of code, but bombs out with an out-of-memory error in the pixelflinger code which I’m yet to have any luck tracing. Currently my serial console is fubar, so I can’t get any useful logging, which makes things doubly painful. The next step is to get adb working over USB so I have at least an output of the errors and warning, which should give me half a chance of tracking down the problem.

So if you want to try and get up to this point, what are the steps? Well, firstly go and download the android toolchain source code. and compile it for a v4 target. You use the --target=armv4-android-eabi argument to configure if I remember correctly.

Once you have that done, grab my latest patch and apply it to the Android source code base. (That is tar file with diffs for each individual project, apply these correctly is left as an exercise for the reader). Then you want to compile it with the new toolchain. I use a script like this:

#!/bin/sh

make TARGET_ARCH_VERSION=armv4 \
     MKJFFS2_CMD="ssh nirvana -x \"cd `pwd`; mkfs.jffs2\""  \
     SYSTEM_FSTYPE=jffs2 \
     USERDATA_FSTYPE=yaffs \
     TARGET_TOOLS_PREFIX=/opt/benno/bin/armv4-android-eabi- $@

Things you will need to change it the tools prefix, and the mkjffs2 command. The evil-hackery above is to run it on my linux virtual machine (I’m compiling the rest under OS X, and I can’t get mkfs.jffs2 to compile under it yet.)

After some time passes you should end up with a ramdisk.img, userdata.img and system.img files. The next step is to get a usable kernel.

I’m using the OpenMoko stable kernel, which is 2.6.24 based. I’ve patched this with bits of the Android kernel (enough, I think, to make it run). Make sure you configure support for yaffs, binder, logger and ashmem. Here is the kernel config I’m currently using.

At this stage it is important you have a version of u-boot supporting the yaffs write commands, if you don’t your next step is to install that. After this the next step is to re-partition your flash device. In case it isn’t obvious this will trash your current OS. The useful parts from my uboot environment are:

mtdids=nand0=neo1973-nand
bootdelay=-1
mtdparts=mtdparts=neo1973-nand:256k(uboot)ro,16k(uboot-env),752k(ramdisk),2m(kernel),36m(system),24m(userdata)
rdaddr=0x35000000
kaddr=0x32000000
bootcmd=setenv bootargs ${bootargs_base} ${mtdparts} initrd=${rdaddr},${rdsize}; nand read.e ${kaddr} kernel; nand read.e ${rdaddr} ramdisk; bootm ${kaddr}
bootargs_base=root=/dev/ram rw console=tty0 loglevel=8

Note the mtdparts which defines the partitions, and the bootcmd. (I’m not entirely happy with the boot command, mostly because when I install new RAM image I need to manually update $rdsize, which is a pain).

With this in place you are ready to start. The first image to move across is your userdata image. Now to make this happen we first copy it into memory using dfu-util:

sudo dfu-util -a 0 -R -D source/out/target/product/generic/userdata.img  -R

Then you need to use the nand write.yaffs1 command to copy it to the data partition. Note, at this stage I get weird behaviour, I’m not convinced that the yaffs support truly works yet! Afterwards I get some messed up data in other parts of the flash (which is why we are doing it first). After you have copied it in, I suggest reseting the device, and you may find you need to reinitialise u-boot (using dyngen, and resetting up the environment as above.

After this you are good to use dfu-util to copy accross the kernel, system.img and ramdisk.img. After copying the ramdisk.img across update the rdsize variable with the size of the ramdisk.

Once all this is done, you are good to boot, I wish you luck! If you have a working serial console you can probably try the logcat command to see why graphics aren’t working. If you get this far please email me the results!

After a lot of stuffing around installing new hard drives so I had enough space to actually play with the source code, getting screwed by Time Machine when trying to convert my filesystem from case-insenstive to case-insensitive (I gave up and am now usuing a case-sensitive disk image on top of my case-insenstive file system.. sigh), I finally have the Android source code compiling, yay!.

Compiling is fairly trivial, just make and away it goes. The fun thing is trying to work out exactly what the hell the build system is actually doing. I’ve got to admit though, it is a pretty clean build system, although it isn’t going to win any speed records. I’m going to go into more details on the build sstem when i have more time, and I’ve actually worked out what the hell is happening.

Anyway, after a few false starts I now have the build system compiling for ARMv4T processors (such as the one inside the Neo1973), and hopefully at the same time I haven’t broken compilation from ARMv5TE.

For those interested I have a patch available. Simply apply this to the checked out code, and the build using make TARGET_ARCH_VERSION=armv4t. Now, of course I haven’t actually tried to run this code yet, so it might not work, but it seems to compile fine, so that is a good start! Now once I work out how to make git play nice I'll actually put this into a branch and make it available, but the diff will have to suffice for now. Of course I’m not the only one looking at this, check out Christopher’s page for more information. (Where he actually starts solving some problems instead of just working around them ;)

The rest of this post documents the patch. For those interested it should give you some idea of the build system and layout, and hopefully it is something that can be applied to mainline.

The first changes made are to the linux-arm.mk file. A new make variable TARGET_ARCH_VERSION is added. For now this is defaulted to armv5te, but it can be overridden on the command line as shown above.

project build/
diff --git a/core/combo/linux-arm.mk b/core/combo/linux-arm.mk
index adb82d3..a43368f 100644
--- a/core/combo/linux-arm.mk
+++ b/core/combo/linux-arm.mk
@@ -7,6 +7,8 @@ $(combo_target)TOOLS_PREFIX := \
 	prebuilt/$(HOST_PREBUILT_TAG)/toolchain/arm-eabi-4.2.1/bin/arm-eabi-
 endif
 
+TARGET_ARCH_VERSION ?= armv5te
+
 $(combo_target)CC := $($(combo_target)TOOLS_PREFIX)gcc$(HOST_EXECUTABLE_SUFFIX)
 $(combo_target)CXX := $($(combo_target)TOOLS_PREFIX)g++$(HOST_EXECUTABLE_SUFFIX)
 $(combo_target)AR := $($(combo_target)TOOLS_PREFIX)ar$(HOST_EXECUTABLE_SUFFIX)

The next thing is to make the GLOBAL_CFLAGS variable dependent on the architecture version. The armv5te defines stay in place, but an armv4t architecture version is added. Most of the cflags are pretty similar, except we change the -march flag, and change the pre-processor defines. These will become important later in the patch as they provide the mechanism for distinguishing between versions in the code.

@@ -46,6 +48,7 @@ ifneq ($(wildcard $($(combo_target)CC)),)
 $(combo_target)LIBGCC := $(shell $($(combo_target)CC) -mthumb-interwork -print-libgcc-file-name)
 endif
 
+ifeq ($(TARGET_ARCH_VERSION), armv5te)
 $(combo_target)GLOBAL_CFLAGS += \
 			-march=armv5te -mtune=xscale \
 			-msoft-float -fpic \
@@ -56,6 +59,21 @@ $(combo_target)GLOBAL_CFLAGS += \
 			-D__ARM_ARCH_5__ -D__ARM_ARCH_5T__ \
 			-D__ARM_ARCH_5E__ -D__ARM_ARCH_5TE__ \
 			-include $(call select-android-config-h,linux-arm)
+else
+ifeq ($(TARGET_ARCH_VERSION), armv4t)
+$(combo_target)GLOBAL_CFLAGS += \
+			-march=armv4t \
+			-msoft-float -fpic \
+			-mthumb-interwork \
+			-ffunction-sections \
+			-funwind-tables \
+			-fstack-protector \
+			-D__ARM_ARCH_4__ -D__ARM_ARCH_4T__ \
+			-include $(call select-android-config-h,linux-arm)
+else
+$(error Unknown TARGET_ARCH_VERSION=$(TARGET_ARCH_VERSION))
+endif
+endif
 
 $(combo_target)GLOBAL_CPPFLAGS += -fvisibility-inlines-hidden

The next bit we update is the prelink-linux-arm.map file. The dynamic libraries in android are laid out explicitly in virtual memory according to this map file. If I’m not mistaken those address look suspiciously 1MB aligned, which means they should fit nicely in the pagetable, and provides some opportunity to use fast-address-space-switching techniques. In the port to ARMv4 I have so far been lazy and instead of fixing up any assembler code I’ve just gone with existing C code. One outcome of this is that I need the libffi.so for my foreign function interface, so I’ve added this to the map for now. I’m not 100% sure that when compiling for ARMv5 this won’t cause a problem. Will need to see. Fixing up the code to avoid needing libffi is probably high on the list of things to do.

diff --git a/core/prelink-linux-arm.map b/core/prelink-linux-arm.map
index d4ebf43..6e0bc43 100644
--- a/core/prelink-linux-arm.map
+++ b/core/prelink-linux-arm.map
@@ -113,3 +113,4 @@ libctest.so             0x9A700000
 libUAPI_jni.so          0x9A500000
 librpc.so               0x9A400000 
 libtrace_test.so        0x9A300000 
+libffi.so               0x9A200000

The next module is the bionic module which is the light-weight C library that is part of Android. This has some nice optimised routines for memory copy and compare, but unfortunately they rely on ARMv5 instructions. I’ve changed the build system to only use the optimised assembler when compiling with ARMv5TE, and falling back to C routines in the other cases. (The strlen implementation isn’t pure assembly, but the optimised C implementation has inline asm, so again it needs to drop back to plain old dumb strlen.)

project bionic/
diff --git a/libc/Android.mk b/libc/Android.mk
index faca333..3fb3455 100644
--- a/libc/Android.mk
+++ b/libc/Android.mk
@@ -206,13 +206,9 @@ libc_common_src_files := \
 	arch-arm/bionic/_setjmp.S \
 	arch-arm/bionic/atomics_arm.S \
 	arch-arm/bionic/clone.S \
-	arch-arm/bionic/memcmp.S \
-	arch-arm/bionic/memcmp16.S \
-	arch-arm/bionic/memcpy.S \
 	arch-arm/bionic/memset.S \
 	arch-arm/bionic/setjmp.S \
 	arch-arm/bionic/sigsetjmp.S \
-	arch-arm/bionic/strlen.c.arm \
 	arch-arm/bionic/syscall.S \
 	arch-arm/bionic/kill.S \
 	arch-arm/bionic/tkill.S \
@@ -274,6 +270,18 @@ libc_common_src_files := \
 	netbsd/nameser/ns_print.c \
 	netbsd/nameser/ns_samedomain.c
 
+
+ifeq ($(TARGET_ARCH),arm)
+ifeq ($(TARGET_ARCH_VERSION),armv5te)
+libc_common_src_files += arch-arm/bionic/memcmp.S \
+		arch-arm/bionic/memcmp16.S \
+		arch-arm/bionic/memcpy.S \
+		arch-arm/bionic/strlen.c.arm
+else
+libc_common_src_files += string/memcmp.c string/memcpy.c string/strlen.c string/ffs.c
+endif
+endif
+
 # These files need to be arm so that gdbserver
 # can set breakpoints in them without messing
 # up any thumb code.

Unfortunately, it is clear that this C only code hasn’t been used in a while as there was a trivial bug as fixed by the patch below. This makes me worry about what other bugs that aren’t caught by the compiler may be lurking.

diff --git a/libc/string/memcpy.c b/libc/string/memcpy.c
index 4cd4a80..dea78b2 100644
--- a/libc/string/memcpy.c
+++ b/libc/string/memcpy.c
@@ -25,5 +25,5 @@
  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  * SUCH DAMAGE.
  */
-#define MEM_COPY
+#define MEMCOPY
 #include "bcopy.c"

Finally, frustratingly, the compiler’s ffs() implementation appears to fallback to calling the C library’s ffs() implementation if it can’t doing something optimised. This happens when compiling for ARMv4, so I’ve added an ffs() implementation (stolen from FreeBSD).

#include 
#include 

/*
 * Find First Set bit
 */
int
ffs(int mask)
{
        int bit;

        if (mask == 0)
                return (0);
        for (bit = 1; !(mask & 1); bit++)
                mask = (unsigned int)mask >> 1;
        return (bit);
}

The next module for attention is the dalvik virtual machine. Again this has some code that relies on ARMv5, but there is a C version that we fall back on. In this case it also means pulling in libffi. This is probably the module that needs to most attention in actually updating the code to be ARMv4 assembler in the near future.

project dalvik/
diff --git a/vm/Android.mk b/vm/Android.mk
index dfed78d..c66a861 100644
--- a/vm/Android.mk
+++ b/vm/Android.mk
@@ -189,6 +189,7 @@ ifeq ($(TARGET_SIMULATOR),true)
 endif
 
 ifeq ($(TARGET_ARCH),arm)
+ifeq ($(TARGET_ARCH_VERSION),armv5te)
 	# use custom version rather than FFI
 	#LOCAL_SRC_FILES += arch/arm/CallC.c
 	LOCAL_SRC_FILES += arch/arm/CallOldABI.S arch/arm/CallEABI.S
@@ -204,6 +205,16 @@ else
 		mterp/out/InterpC-desktop.c \
 		mterp/out/InterpAsm-desktop.S
 	LOCAL_SHARED_LIBRARIES += libffi
+	LOCAL_SHARED_LIBRARIES += libdl
+endif
+else
+	# use FFI
+	LOCAL_C_INCLUDES += external/libffi/$(TARGET_OS)-$(TARGET_ARCH)
+	LOCAL_SRC_FILES += arch/generic/Call.c
+	LOCAL_SRC_FILES += \
+		mterp/out/InterpC-desktop.c \
+		mterp/out/InterpAsm-desktop.S
+	LOCAL_SHARED_LIBRARIES += libffi
 endif
 
 LOCAL_MODULE := libdvm

Next is libjpeg, which again, has assembler optimisation that we can’t easily use without real porting work, so we fall back to the C

project external/jpeg/
diff --git a/Android.mk b/Android.mk
index 9cfe4f6..3c052cd 100644
--- a/Android.mk
+++ b/Android.mk
@@ -19,6 +19,12 @@ ifneq ($(TARGET_ARCH),arm)
 ANDROID_JPEG_NO_ASSEMBLER := true
 endif
 
+# the assembler doesn't work for armv4t
+ifeq ($(TARGET_ARCH_VERSION),armv4t)
+ANDROID_JPEG_NO_ASSEMBLER := true
+endif
+
+
 # temp fix until we understand why this broke cnn.com
 #ANDROID_JPEG_NO_ASSEMBLER := true
 

For some reason compiling with ARMv4 doesn’t allow the prefetch loop array compiler optimisation, so we turn it off for ARMv4.

@@ -29,7 +35,10 @@ LOCAL_SRC_FILES += jidctint.c jidctfst.S
 endif
 
 LOCAL_CFLAGS += -DAVOID_TABLES 
-LOCAL_CFLAGS += -O3 -fstrict-aliasing -fprefetch-loop-arrays
+LOCAL_CFLAGS += -O3 -fstrict-aliasing
+ifeq ($(TARGET_ARCH_VERSION),armv5te)
+LOCAL_FLAGS += -fprefetch-loop-arrays
+endif
 #LOCAL_CFLAGS += -march=armv6j
 
 LOCAL_MODULE:= libjpeg

Next up is libffi, which is just a case of turning it on since we now need it for ARMv4.

project external/libffi/
diff --git a/Android.mk b/Android.mk
index f4452c9..07b5c2f 100644
--- a/Android.mk
+++ b/Android.mk
@@ -6,7 +6,7 @@
 # We need to generate the appropriate defines and select the right set of
 # source files for the OS and architecture.
 
-ifneq ($(TARGET_ARCH),arm)
+ifneq ($(TARGET_ARCH_VERSION),armv5te)
 
 LOCAL_PATH:= $(call my-dir)
 include $(CLEAR_VARS)

The external module opencore contains a lot of software implemented codecs. (I wonder about the licensing restrictions on these things...). Not surprisingly these too are tuned for ARMv4, but again we fall back to plain old C.

project external/opencore/
diff --git a/codecs_v2/audio/aac/dec/Android.mk b/codecs_v2/audio/aac/dec/Android.mk
index ffe0089..6abdc2d 100644
--- a/codecs_v2/audio/aac/dec/Android.mk
+++ b/codecs_v2/audio/aac/dec/Android.mk
@@ -150,7 +150,7 @@ LOCAL_SRC_FILES := \
 LOCAL_MODULE := libpv_aac_dec
 
 LOCAL_CFLAGS := -DAAC_PLUS -DHQ_SBR -DPARAMETRICSTEREO  $(PV_CFLAGS)
-ifeq ($(TARGET_ARCH),arm)
+ifeq ($(TARGET_ARCH_VERSION),armv5te)
  LOCAL_CFLAGS += -D_ARM_GCC
  else
  LOCAL_CFLAGS += -DC_EQUIVALENT
diff --git a/codecs_v2/audio/gsm_amr/amr_wb/dec/Android.mk b/codecs_v2/audio/gsm_amr/amr_wb/dec/Android.mk
index e184178..3223841 100644
--- a/codecs_v2/audio/gsm_amr/amr_wb/dec/Android.mk
+++ b/codecs_v2/audio/gsm_amr/amr_wb/dec/Android.mk
@@ -48,7 +48,7 @@ LOCAL_SRC_FILES := \
 LOCAL_MODULE := libpvamrwbdecoder
 
 LOCAL_CFLAGS :=   $(PV_CFLAGS)
-ifeq ($(TARGET_ARCH),arm)
+ifeq ($(TARGET_ARCH_VERSION),armv5te)
  LOCAL_CFLAGS += -D_ARM_GCC
  else
  LOCAL_CFLAGS += -DC_EQUIVALENT
diff --git a/codecs_v2/audio/mp3/dec/Android.mk b/codecs_v2/audio/mp3/dec/Android.mk
index 254cb6b..c2430fe 100644
--- a/codecs_v2/audio/mp3/dec/Android.mk
+++ b/codecs_v2/audio/mp3/dec/Android.mk
@@ -28,8 +28,8 @@ LOCAL_SRC_FILES := \
 	src/pvmp3_seek_synch.cpp \
 	src/pvmp3_stereo_proc.cpp \
 	src/pvmp3_reorder.cpp
-	
-ifeq ($(TARGET_ARCH),arm)
+
+ifeq ($(TARGET_ARCH_VERSION),armv5te)
 LOCAL_SRC_FILES += \
 	src/asm/pvmp3_polyphase_filter_window_gcc.s \
 	src/asm/pvmp3_mdct_18_gcc.s \
@@ -46,7 +46,7 @@ endif
 LOCAL_MODULE := libpvmp3
 
 LOCAL_CFLAGS :=   $(PV_CFLAGS)
-ifeq ($(TARGET_ARCH),arm)
+ifeq ($(TARGET_ARCH_VERSION),armv5te)
  LOCAL_CFLAGS += -DPV_ARM_GCC
  else
  LOCAL_CFLAGS += -DC_EQUIVALENT

Unfortunately it is not just the build file that needs updating in this module. I need to manually go and update the headers so that some optimised inline assembler is only used in the ARMv5 case. To be honest this messes these files up a little bit, so a nicer solution would be preferred.

diff --git a/codecs_v2/video/m4v_h263/enc/src/dct_inline.h b/codecs_v2/video/m4v_h263/enc/src/dct_inline.h
index 86474b2..41a3297 100644
--- a/codecs_v2/video/m4v_h263/enc/src/dct_inline.h
+++ b/codecs_v2/video/m4v_h263/enc/src/dct_inline.h
@@ -22,7 +22,7 @@
 #ifndef _DCT_INLINE_H_
 #define _DCT_INLINE_H_
 
-#if !defined(PV_ARM_GCC)&& defined(__arm__)
+#if !(defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_5TE__))
 
 #include "oscl_base_macros.h"
 
@@ -109,7 +109,7 @@ __inline int32 sum_abs(int32 k0, int32 k1, int32 k2, int32 k3,
 #elif defined(__CC_ARM)  /* only work with arm v5 */
 
 #if defined(__TARGET_ARCH_5TE)
-
+#error
 __inline int32 mla724(int32 op1, int32 op2, int32 op3)
 {
     int32 out;
@@ -266,7 +266,7 @@ __inline int32 sum_abs(int32 k0, int32 k1, int32 k2, int32 k3,
     return abs_sum;
 }
 
-#elif defined(PV_ARM_GCC) && defined(__arm__) /* ARM GNU COMPILER  */
+#elif defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_5TE__) /* ARM GNU COMPILER  */
 
 __inline int32 mla724(int32 op1, int32 op2, int32 op3)
 {
diff --git a/codecs_v2/video/m4v_h263/enc/src/fastquant_inline.h b/codecs_v2/video/m4v_h263/enc/src/fastquant_inline.h
index 6a35d43..fbfeddf 100644
--- a/codecs_v2/video/m4v_h263/enc/src/fastquant_inline.h
+++ b/codecs_v2/video/m4v_h263/enc/src/fastquant_inline.h
@@ -25,7 +25,7 @@
 #include "mp4def.h"
 #include "oscl_base_macros.h"
 
-#if !defined(PV_ARM_GCC) && defined(__arm__) /* ARM GNU COMPILER  */
+#if !(defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_V5TE__)) /* ARM GNU COMPILER  */
 
 __inline int32 aan_scale(int32 q_value, int32 coeff, int32 round, int32 QPdiv2)
 {
@@ -423,7 +423,7 @@ __inline int32 coeff_dequant_mpeg_intra(int32 q_value, int32 tmp)
     return q_value;
 }
 
-#elif defined(PV_ARM_GCC) && defined(__arm__) /* ARM GNU COMPILER  */
+#elif defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_V5TE__) /* ARM GNU COMPILER  */
 
 __inline int32 aan_scale(int32 q_value, int32 coeff,
                          int32 round, int32 QPdiv2)
diff --git a/codecs_v2/video/m4v_h263/enc/src/vlc_encode_inline.h b/codecs_v2/video/m4v_h263/enc/src/vlc_encode_inline.h
index 69857f3..b0bf46d 100644
--- a/codecs_v2/video/m4v_h263/enc/src/vlc_encode_inline.h
+++ b/codecs_v2/video/m4v_h263/enc/src/vlc_encode_inline.h
@@ -18,7 +18,7 @@
 #ifndef _VLC_ENCODE_INLINE_H_
 #define _VLC_ENCODE_INLINE_H_
 
-#if !defined(PV_ARM_GCC)&& defined(__arm__)
+#if !(defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_V5TE__))
 
 __inline  Int zero_run_search(UInt *bitmapzz, Short *dataBlock, RunLevelBlock *RLB, Int nc)
 {
@@ -208,7 +208,7 @@ __inline  Int zero_run_search(UInt *bitmapzz, Short *dataBlock, RunLevelBlock *R
     return idx;
 }
 
-#elif defined(PV_ARM_GCC) && defined(__arm__) /* ARM GNU COMPILER  */
+#elif defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_V5TE__) /* ARM GNU COMPILER  */
 
 __inline Int m4v_enc_clz(UInt temp)
 {

A similar approach is needed in the skia graphics library.

project external/skia/
diff --git a/include/corecg/SkMath.h b/include/corecg/SkMath.h
index 76cf279..5f0264f 100644
--- a/include/corecg/SkMath.h
+++ b/include/corecg/SkMath.h
@@ -162,7 +162,7 @@ static inline int SkNextLog2(uint32_t value) {
     With this requirement, we can generate faster instructions on some
     architectures.
 */
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARM_ARCH_5TE__) && !defined(__thumb__)
     static inline int32_t SkMulS16(S16CPU x, S16CPU y) {
         SkASSERT((int16_t)x == x);
         SkASSERT((int16_t)y == y);

The sonivox module (no idea what that is!), has the same requirement of updating the build to avoid building ARMv5 specific code.

project external/sonivox/
diff --git a/arm-wt-22k/Android.mk b/arm-wt-22k/Android.mk
index 565c233..a59f917 100644
--- a/arm-wt-22k/Android.mk
+++ b/arm-wt-22k/Android.mk
@@ -73,6 +73,7 @@ LOCAL_COPY_HEADERS := \
 	host_src/eas_reverb.h
 
 ifeq ($(TARGET_ARCH),arm)
+ifeq (($TARGET_ARCH),armv5)
 LOCAL_SRC_FILES+= \
 	lib_src/ARM-E_filter_gnu.s \
 	lib_src/ARM-E_interpolate_loop_gnu.s \

The low-level audio code in audioflinger suffers from the same optimisations, and we need to dive into the code on this occasion to fix things up.

project frameworks/base/
diff --git a/libs/audioflinger/AudioMixer.cpp b/libs/audioflinger/AudioMixer.cpp
index 9f1b17f..4c0890c 100644
--- a/libs/audioflinger/AudioMixer.cpp
+++ b/libs/audioflinger/AudioMixer.cpp
@@ -400,7 +400,7 @@ void AudioMixer::process__validate(state_t* state, void* output)
 static inline 
 int32_t mulAdd(int16_t in, int16_t v, int32_t a)
 {
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
     int32_t out;
     asm( "smlabb %[out], %[in], %[v], %[a] \n"
          : [out]"=r"(out)
@@ -415,7 +415,7 @@ int32_t mulAdd(int16_t in, int16_t v, int32_t a)
 static inline 
 int32_t mul(int16_t in, int16_t v)
 {
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
     int32_t out;
     asm( "smulbb %[out], %[in], %[v] \n"
          : [out]"=r"(out)
@@ -430,7 +430,7 @@ int32_t mul(int16_t in, int16_t v)
 static inline 
 int32_t mulAddRL(int left, uint32_t inRL, uint32_t vRL, int32_t a)
 {
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
     int32_t out;
     if (left) {
         asm( "smlabb %[out], %[inRL], %[vRL], %[a] \n"
@@ -456,7 +456,7 @@ int32_t mulAddRL(int left, uint32_t inRL, uint32_t vRL, int32_t a)
 static inline 
 int32_t mulRL(int left, uint32_t inRL, uint32_t vRL)
 {
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
     int32_t out;
     if (left) {
         asm( "smulbb %[out], %[inRL], %[vRL] \n"
diff --git a/libs/audioflinger/AudioResamplerSinc.cpp b/libs/audioflinger/AudioResamplerSinc.cpp
index e710d16..88b8c22 100644
--- a/libs/audioflinger/AudioResamplerSinc.cpp
+++ b/libs/audioflinger/AudioResamplerSinc.cpp
@@ -62,7 +62,7 @@ const int32_t AudioResamplerSinc::mFirCoefsDown[] = {
 static inline 
 int32_t mulRL(int left, int32_t in, uint32_t vRL)
 {
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
     int32_t out;
     if (left) {
         asm( "smultb %[out], %[in], %[vRL] \n"
@@ -88,7 +88,7 @@ int32_t mulRL(int left, int32_t in, uint32_t vRL)
 static inline 
 int32_t mulAdd(int16_t in, int32_t v, int32_t a)
 {
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
     int32_t out;
     asm( "smlawb %[out], %[v], %[in], %[a] \n"
          : [out]"=r"(out)
@@ -103,7 +103,7 @@ int32_t mulAdd(int16_t in, int32_t v, int32_t a)
 static inline 
 int32_t mulAddRL(int left, uint32_t inRL, int32_t v, int32_t a)
 {
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
     int32_t out;
     if (left) {
         asm( "smlawb %[out], %[v], %[inRL], %[a] \n"

The AndroidConfig.h header file is included on every compile. We mess with it to convince it that we don’t have an optimised memcmp16 function.

project system/core/
diff --git a/include/arch/linux-arm/AndroidConfig.h b/include/arch/linux-arm/AndroidConfig.h
index d7e182a..76f424e 100644
--- a/include/arch/linux-arm/AndroidConfig.h
+++ b/include/arch/linux-arm/AndroidConfig.h
@@ -249,8 +249,9 @@
 /*
  * Do we have __memcmp16()?
  */
+#if defined(__ARCH_ARM_5TE__)
 #define HAVE__MEMCMP16  1
-
+#endif
 /*
  * type for the third argument to mincore().
  */

Next up is the pixelflinger, where things get interesting, because all of a sudden we have armv6 code. I’ve taken the rash decision of wrapping this in conditionals that are only enabled if you actually have an ARMv6 version, not a pesky ARMv5E, but I really need to better understand the intent here. It seems a little strange.

diff --git a/libpixelflinger/Android.mk b/libpixelflinger/Android.mk
index a8e5ee4..077cf47 100644
--- a/libpixelflinger/Android.mk
+++ b/libpixelflinger/Android.mk
@@ -5,7 +5,7 @@ include $(CLEAR_VARS)
 # ARMv6 specific objects
 #
 
-ifeq ($(TARGET_ARCH),arm)
+ifeq ($(TARGET_ARCH_VERSION),armv6)
 LOCAL_ASFLAGS := -march=armv6
 LOCAL_SRC_FILES := rotate90CW_4x4_16v6.S
 LOCAL_MODULE := libpixelflinger_armv6
@@ -39,7 +39,7 @@ PIXELFLINGER_SRC_FILES:= \
 	raster.cpp \
 	buffer.cpp
 
-ifeq ($(TARGET_ARCH),arm)
+ifeq ($(TARGET_ARCH_VERSION),armv5te)
 PIXELFLINGER_SRC_FILES += t32cb16blend.S
 endif
 
@@ -67,7 +67,7 @@ ifneq ($(BUILD_TINY_ANDROID),true)
 LOCAL_MODULE:= libpixelflinger
 LOCAL_SRC_FILES := $(PIXELFLINGER_SRC_FILES)
 LOCAL_CFLAGS := $(PIXELFLINGER_CFLAGS) -DWITH_LIB_HARDWARE
-ifeq ($(TARGET_ARCH),arm)
+ifeq ($(TARGET_ARCH_VERSION),armv6)
 LOCAL_WHOLE_STATIC_LIBRARIES := libpixelflinger_armv6
 endif
 include $(BUILD_SHARED_LIBRARY)

Finally scanline has an optimised asm version it calls in preference to doing the same thing inline with C code. Again, I take the easy way out, and use the C code.

diff --git a/libpixelflinger/scanline.cpp b/libpixelflinger/scanline.cpp
index d24c988..685a3b7 100644
--- a/libpixelflinger/scanline.cpp
+++ b/libpixelflinger/scanline.cpp
@@ -1312,7 +1312,7 @@ void scanline_t32cb16blend(context_t* c)
     const int32_t v = (c->state.texture[0].shade.it0>>16) + y;
     uint32_t *src = reinterpret_cast(tex->data)+(u+(tex->stride*v));
 
-#if ((ANDROID_CODEGEN >= ANDROID_CODEGEN_ASM) && defined(__arm__))
+#if ((ANDROID_CODEGEN >= ANDROID_CODEGEN_ASM) && defined(__arm__) && defined(__ARCH_ARM_5TE__))
     scanline_t32cb16blend_arm(dst, src, ct);
 #else
     while (ct--) {

And that my friends, is that! Now to see if I can actually run this code!

Usually when companies say release 4th quarter 2008 you usually see something around January 2009, and to be honest that was when I was expecting the Android source code to finally drop. So I was a little surprised to see that the code was released early this morning.

Stay tuned, more to come as I start playing.

It’s been almost two years now since I help organise and run linux.conf.au 2007, and I thought it was time to jump back into the fray again. This time I’ll not be doing something as silly as trying to organise the whole conference, but I will be running a small miniconf on the first two days of the conference. So I’d like to invite you all to the Open Mobile Miniconf in January next year. And if you think you’ve got something cool you’d like to share with other developers, please take a look at the call for presentations, and drop me a line.