As I mentioned in my previous post, we recently built a prototype that would only uncompress the bits we do require from our libraries, as opposed to the complete library. Or at least try to.
The original idea was to do selective decompression of blocks as they are required. Such that instead of uncompressing the whole file, and then use it, we’d only uncompress blocks we do need, at the moment we need them. So, for example, instead of uncompressing 32KB, and then use the first and last 4KB (completely ignoring 24KB in the middle), we’d only uncompress and use the first and last 4KB.
Unfortunately, compressed streams in ZIP files are not seekable, so with the compressed data we have currently, we just can’t do that. Which is why we’re going to work on figuring out how to best do it with an index and how the zlib library will allow us to do so. Fortunately, one of the zlib author already has some working code doing something similar, though it’s probably not usable as-is.
Anyways, as we needed to put things together for a prototype quite rapidly and had experience with binary reordering in the past, we figured a way to get the same kind of benefit would be to improve the binary layout so that we can uncompress subsets of the file progressively, as it is used. So for example, if the order in which we need blocks looks like [10, 11, 7, 13, 17], we’d first uncompress up to the 10th blocks, then the 11th, then use the 7th, that was uncompressed earlier, then uncompress up to the 13th, and so on.
In fact, the way the prototype worked was by having a background thread doing decompression while other thread(s) would do the actual work, and when it hits a block that hasn’t been uncompressed yet, wait for the required data to be reached by the background thread. The implementation was however probably racy, which led to multiple random crashes. There may also be some issues with the signal handler being used. These are details that we are expecting to figure in time.
The first obstacle is that when a library is loaded, relocations are performed. Relocations are basically what allows pointers in a library to point to the right place in memory. They usually apply to data sections. They can apply to code sections when the code is not PIC (Position Independent Code), but this is fortunately not the case anymore on Firefox for Android. So they apply to data sections, and data sections are at the end of libraries. Which means, under our progressive decompression scheme as explained above, we’d still have to uncompress almost the entire file.
On the long term, we should be able to apply these relocations as we uncompress each blocks, but that requires our linker to handle elfhack relocations itself, and some more tweaks to access the appropriate relocations without having to scan the entire list. But even then, data is accessed during startup. Thus until we can seek in the compressed data stream, data needs to be placed before code.
As I was using the gold linker for various other reasons, and as gold doesn’t have much flexibility for these things, I added a little something to elfhack to rewrite the ELF file such that data sections would be before code sections in the file, but still after when loaded in memory. Nice trick, but unfortunately, a lot of tools are not very happy with it (even less than what elfhack already does).
Once data is dealt with by being placed before code, we need to make sure code required during startup is grouped together. The linker, however, is not able to reorder functions without some special treatment when compiling the source to object files: it requires the
-ffunction-sections option to be given to GCC.
Normally, when compiling a source file to an object file, GCC creates a unique
.text section containing all the code. If in a given source file, there are two functions,
bar, the function call from
bar will be hard-coded to use the distance from the call location to the start of
-ffunction-sections, each function is given a separate
.text.function_name section, and the compiler places relocations for calls between them. Simply put, instead of having recorded the distance from the call location to the start of
bar, we’d have a relocation, applying on the call location, with a symbolic information telling the destination is the start of
The linker will then apply the relocations, and the resulting code will be using the distance between the functions, as without
-ffunction-sections. The difference comes from the fact that the linker now is able to move these function sections independently and adjust the various calls accordingly.
Unfortunately, there are several parts that
-ffunction-sections doesn’t affect, and thus can’t be reordered without some effort:
- Assembler source
- (Static) system libraries (e.g. libgcc, libstlport)
For the former, the various inline assembly and separate assembler sources must be annotated with individual sections for each function defined. For the latter, the static libraries can be rebuilt with
ARM processors we target with Firefox for Android have two instruction sets: ARM and Thumb-2. Switching from one of them to the other is called interwork. While we build Firefox for Android with the Thumb-2 instruction set, system libraries may use the ARM instruction set. As such, they must be able to call system library functions with interworking.
Technically, interworking can be triggered from Thumb-2 or ARM, to call either one. However, the Procedure Linkage Table, containing code that either calls the linker to resolve symbols, or the system library functions, contains ARM code. Theoretically, in a program or library that is entirely Thumb-2, it should be possible to use Thumb-2 code, but neither GNU ld nor gold do that. They however both (currently) have a different behavior.
GNU ld generates PLT entries that start as Thumb-2, switch to ARM and then call whatever they need to call.
Gold generates PLT entries that are fully ARM, but adds trampolines into which Thumb-2 code jumps to interwork and jump back to the ARM PLT.
These trampolines are unfortunately “randomly” placed in the resulting binary. They’re not actually randomly positioned, but the result is that these trampolines end up in places that are definitely not near the beginning of the file. In my Firefox for Android builds, they were usually grouped in various places, some of which were at the end of the code section, which went completely against our goal. For the purpose of making the prototype quickly available, we had to compile Firefox for Android as ARM, which effectively makes the binary larger (ARM instructions are larger than Thumb-2 instructions), thus longer to uncompress.
Identical Code Folding
Identical Code Folding is one of the features we were using gold for. It allows to replace various identical implementations of functions with a single occurrence. This is particularly interesting on C++ code, where templates are heavily used. For example, a templated function used for bools and for ints will likely result in the same machine code.
In some other cases, simple enough functions can end up completely identical at the machine code level despite not looking as such in C++.
The problem is that gold handles function (sections) reordering after ICF. Once the ICF pass is done, only one of the original functions is known of gold for function reordering (even if in the end it does expose all the symbols appropriately). And it may well be the variant that does not appear in the function order list.
The current workaround is to link libxul.so once, with an option printing the functions being folded together, and adding all of them in the order list, at the position where it makes most sense. For instance, if
bar are folded and
bar appears in the order list,
bar would be put in place of
bar in the list.
As it may have transpired from above explanations, gold doesn’t actually do function reordering. It does section reordering. In the simple cases, sections for functions are just named
.text.function_name. However, there are also other variants, such as
.text.unlikely.function_name. Even more subtle, there can also be
.text.function_name.part.number sections, but for these, the
.part.number is actually part of the function name as seen in the symbols table.
Tools that do the kind of profiling required to do function ordering like we wanted to do will give function names, not section names, because the section names are only present in the object files, not in the final libraries and programs. Thus after getting a list of function names as they are called during startup, cross-referencing with the object files to get the corresponding section names is required.
We used Valgrind in order to get the order in which functions were called during startup. While our icegrind plugin would have allowed it, various changes in Valgrind in the past year apparently broke icegrind. But ARM support for Valgrind is work in progress and is based on current trunk, so we needed a recent build, which excludes using icegrind, except if we update it.
But it also turns out Valgrind has a built-in feature that more or less allows to get the order in which functions are called. Valgrind can show a trace of each basic blocks the first time it sees them executed (
--trace-flags=10000000, and to be on the safe side,
--vex-guest-chase-thresh=0). Basic blocks are subsets of functions, so Valgrind likely will report the same function several times, but it’s pretty easy to filter out those duplicates.
Unfortunately, we encountered a couple problems with this built-in tracing, not reporting names for some functions when the binary is compiled as Thumb-2, and not reporting at all some functions that we know are being called when the binary is compiled as ARM (it may also be the case for Thumb-2, haven’t verified). In the end, the order list was updated manually for the prototype.
Hardware-specific code paths
There are various places in the Mozilla code base where different implementations of the same functions are available for different kind of hardware, and the most suitable one is chosen at runtime, depending on the hardware detected. On x86, we have some implementations using SSE2 instructions, and fall-back implementations in plain x86. On ARM, there are implementations using NEON instructions, and fall-back implementations in plain ARM or Thumb-2. Sometimes, there aren’t even fall-backs.
This means that which function is going to be used at runtime will very much depend on the hardware it will be executed on. So if you profile Firefox for Android on a tablet with a Tegra-2 CPU (which doesn’t have NEON), you won’t be exercising the same code paths as when running the same Firefox for Android build on a Nexus S (which CPU does have NEON).
If we get the order of functions called at startup on the Tegra-2 tablet, we won’t be getting any of the NEON-specific functions in the functions list, and the resulting reordered binary may have the NEON-specific functions at an inconvenient position for progressive decompression. So the runtime benefits we’d see on the Tegra-2 tablet would go away on the Nexus S. And vice-versa.
The GCC google branch has patches adding a new interesting feature: callgraph-based reordering. It makes the compiler add call graph information extracted from PGO/FDO profile data to object files, and the linker use that information to reorder the binary. Theoretically, this would allow us to reorder our binaries without relying on valgrind or something similar, and even take advantage of the fact that we already do PGO/FDO (except we currently don’t, for Android builds).
The main drawback, however, is exactly the same as for the
-ffunction-sections option: assembler sources and static (system) libraries are not going to be covered. And unlike with
-ffunction-sections, it’s hard to work around.
It also most likely suffers the same problem as section ordering w.r.t identical code folding, and also has the same problem with ARM/Thumb-2 trampolines and hardware-specific code paths.
It’s still a long way before having efficiently reordered binaries for Android builds. It involves toolchain fixes and tweaks, and build engineering tricks to allow running the build on some device and during the cross-compiled build. We should however be able to benefit from incremental decompression with seekable compressed streams well before that. Stay tuned.