[webkit-dev] Using OfflineAsm
Hello, I'm looking for some information about the Java script interpreter ASM backend, aka OfflineAsm. First a bit of background information; the code I have been using is about a year old, on several platforms we don't run with JIT and won't be able to enable it. This means that we've been using the LLINT_C_LOOP path in these cases. It was considered that the ASM backend may be able to improve performance in this case, currently I'm taking a look at how things are done in a more recent snapshot of the trunk to be sure this development effort will be going in the right direction on what is done there. As a side note I see that LLINT_C_LOOP has been replaced with #if !ENABLE(JIT) which I guess is pretty much how it has always behaved. So my questions are: 1) How is the ASM backend intended to be used? Is there a doc that covers this? 2) Can I essentially move from the C LOOP to neither C LOOP nor JIT? or is the ASM backend dependent on JIT itself? It appears this would be handled by the defines in Source\JavaScriptCore\llint\LLIntOfflineAsmConfig.h 3) If JIT is expected to use the ASM backend is there a way this can be worked around? 4) Should I actually expect a performance increase compared to the C LOOP? 5) Are there any other suggestions on how to improve performance of Java script code in a non-JIT port? Thanks for any feedback Chris Vienneau ___ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev
Re: [webkit-dev] Using OfflineAsm
On Sep 5, 2014, at 3:23 PM, Vienneau, Christopher cvienn...@ea.com wrote: Hello, I’m looking for some information about the Java script interpreter ASM backend, aka OfflineAsm. First a bit of background information; the code I have been using is about a year old, on several platforms we don’t run with JIT and won’t be able to enable it. This means that we’ve been using the LLINT_C_LOOP path in these cases. It was considered that the ASM backend may be able to improve performance in this case, currently I’m taking a look at how things are done in a more recent snapshot of the trunk to be sure this development effort will be going in the right direction on what is done there. As a side note I see that LLINT_C_LOOP has been replaced with #if !ENABLE(JIT) which I guess is pretty much how it has always behaved. So my questions are: 1) How is the ASM backend intended to be used? As a bottom tier for a mixed mode execution VM. We compile to assembly to be able to get an ABI that is compatible with our JITs and to be able to do OSR. Is there a doc that covers this? No. 2) Can I essentially move from the “C LOOP” to neither “C LOOP” nor “JIT”? or is the ASM backend dependent on JIT itself? You should be able to use the asm backend without the JIT. We do this if we compiled with the JIT but we failed to allocate writable+executable memory at run time. With some hacks you could probably compile webkit to always do this. It appears this would be handled by the defines in Source\JavaScriptCore\llint\LLIntOfflineAsmConfig.h 3) If JIT is expected to use the ASM backend is there a way this can be worked around? The JIT doesn't use offlineasm. 4) Should I actually expect a performance increase compared to the “C LOOP”? Not really, unless your C compiler sucks. 5) Are there any other suggestions on how to improve performance of Java script code in a non-JIT port? Not really but contributions are welcome. Thanks for any feedback Chris Vienneau ___ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev ___ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev
Re: [webkit-dev] Using OfflineAsm
Christopher, Let me add a little to Phil’s comments. As Phil says, you shouldn’t expect much if any performance gain going from the C LOOP to one of the assembler back ends if you also don’t plan on enabling JITs. The architectures / OS combinations that work with the LLInt is nearly the same as what is supported by the baseline JIT. Basically X86 (32 64), ARM (traditional, thumb2 and 64) and to a lessor degree MIPS and SH4. These work as appropriate on Mac OS X, Windows, Linux, iOS and a couple other OSes. If you have a processor and/or OS, or combination of the two that is not supported, it will likely require work in the offline assembler and / or LLInt code for it to work. The effort is probably won’t payoff unless you plan on eventually supporting one or more of the JITs. What processor(s) and OS(es) are you targeting? - Michael On Sep 5, 2014, at 3:47 PM, Filip Pizlo fpi...@apple.com wrote: On Sep 5, 2014, at 3:23 PM, Vienneau, Christopher cvienn...@ea.com wrote: Hello, I’m looking for some information about the Java script interpreter ASM backend, aka OfflineAsm. First a bit of background information; the code I have been using is about a year old, on several platforms we don’t run with JIT and won’t be able to enable it. This means that we’ve been using the LLINT_C_LOOP path in these cases. It was considered that the ASM backend may be able to improve performance in this case, currently I’m taking a look at how things are done in a more recent snapshot of the trunk to be sure this development effort will be going in the right direction on what is done there. As a side note I see that LLINT_C_LOOP has been replaced with #if !ENABLE(JIT) which I guess is pretty much how it has always behaved. So my questions are: 1) How is the ASM backend intended to be used? As a bottom tier for a mixed mode execution VM. We compile to assembly to be able to get an ABI that is compatible with our JITs and to be able to do OSR. Is there a doc that covers this? No. 2) Can I essentially move from the “C LOOP” to neither “C LOOP” nor “JIT”? or is the ASM backend dependent on JIT itself? You should be able to use the asm backend without the JIT. We do this if we compiled with the JIT but we failed to allocate writable+executable memory at run time. With some hacks you could probably compile webkit to always do this. It appears this would be handled by the defines in Source\JavaScriptCore\llint\LLIntOfflineAsmConfig.h 3) If JIT is expected to use the ASM backend is there a way this can be worked around? The JIT doesn't use offlineasm. 4) Should I actually expect a performance increase compared to the “C LOOP”? Not really, unless your C compiler sucks. 5) Are there any other suggestions on how to improve performance of Java script code in a non-JIT port? Not really but contributions are welcome. Thanks for any feedback Chris Vienneau ___ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev ___ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev ___ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev
Re: [webkit-dev] Using OfflineAsm
Hi Chris, I’ll try to answer your questions below … On Sep 5, 2014, at 3:23 PM, Vienneau, Christopher cvienn...@ea.com wrote: Hello, I’m looking for some information about the Java script interpreter ASM backend, aka OfflineAsm. First a bit of background information; the code I have been using is about a year old, on several platforms we don’t run with JIT and won’t be able to enable it. This means that we’ve been using the LLINT_C_LOOP path in these cases. It was considered that the ASM backend may be able to improve performance in this case, currently I’m taking a look at how things are done in a more recent snapshot of the trunk to be sure this development effort will be going in the right direction on what is done there. As a side note I see that LLINT_C_LOOP has been replaced with #if !ENABLE(JIT) which I guess is pretty much how it has always behaved. Yes, LLINT_C_LOOP is synonymous with !ENABLE_JIT. The C loop is only meant as stand-in for ports that don’t want to spend the effort to get the JIT running. So my questions are: 1) How is the ASM backend intended to be used? Is there a doc that covers this? The code is the only documentation on this. Some context that may help guide you in understanding how it works: 1. The “ASM code for the interpreter is in Source/JavaScriptCore/llint/LowLevelInterpreter*.asm. 2. This ASM code is assembled using the offlineasm assembler which is a ruby program. 3. The code for the offlineasm assembler is in Source/JavaScriptCore/offlineasm. 4. At build time: 1. LLIntOffsetsExtractor.cpp (which contains a lot of offsets that the LLINT needs) is built for the target platform. 2. The offlineasm will parse the binary generated from step 1 for the offset values. 3. The offlineasm will parse the LLINT ASM files plus the previously computed offsets, and emit target specific assembly code using an appropriate backend. See Source/JavaScriptCore/offlineasm/backends.rb for a list of current backends. 4. The generated target specific “assembly file (typically named LLIntAssembly.h) is assembled with the target build tools. For most platforms, LLIntAssembly.h contains inline assembly that is #include’d into llint/LowLevelInterpreter.cpp and built. On Windows, an actual ASM file is generated (not inline assembly) and is built using MASM. 5. The built assembler is linked with the JSC VM. 2) Can I essentially move from the “C LOOP” to neither “C LOOP” nor “JIT”? or is the ASM backend dependent on JIT itself? It appears this would be handled by the defines in Source\JavaScriptCore\llint\LLIntOfflineAsmConfig.h LLINT_C_LOOP and ENABLE_JIT are mutually exclusive. If you want the JIT, you must support the ASM LLINT. Once you have the ASM LLINT, you will essentially have JIT support (unless you’re working with some exotic CPU that is not currently supported). That said, you can build the ASM LLINT and not run with the JIT. You can set Options::useJIT() = false to only run with the LLINT. 3) If JIT is expected to use the ASM backend is there a way this can be worked around? Already answered above. 4) Should I actually expect a performance increase compared to the “C LOOP”? Yes. The ASM LLINT should be faster than the C LOOP LLINT. 5) Are there any other suggestions on how to improve performance of Java script code in a non-JIT port? I presume options that are current supported. The ASM LLINT is your best bet. Alternatively, you can try optimizing the C Loop LLINT if you don’t want to tackle the ASM LLINT. The C Loop LLINT is intended to be simple and not intended to be super optimized, but any improvements that does not come at the cost of too much added complexity is welcomed. Regards, Mark Thanks for any feedback Chris Vienneau ___ webkit-dev mailing list webkit-dev@lists.webkit.org mailto:webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev https://lists.webkit.org/mailman/listinfo/webkit-dev ___ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev
Re: [webkit-dev] Using OfflineAsm
On Sep 5, 2014, at 4:05 PM, Mark Lam mark@apple.com wrote: 4) Should I actually expect a performance increase compared to the “C LOOP”? Yes. The ASM LLINT should be faster than the C LOOP LLINT. To clarify, the C LOOP LLINT emulated a CPU’s behavior using emitted C instructions. As a result, it incurs a performance penalty that is under 10% (vs the ASM LLINT) if I remember correctly. So, the perf gain you’ll get comes from recuperating the losses due to the C loop being a CPU emulator. You should not expect to see any of the more substantive gains that come from the JITs. Regards, Mark ___ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev