3 ms·
When you say "reuses the JIT", does it reuse the JIT API that gets JVM bytecode as input, and native code as output? Or is it a lower layer, that JVM bytecode i
by int_19h 8y ago
When you say "reuses the JIT", does it reuse the JIT API that gets JVM bytecode as input, and native code as output? Or is it a lower layer, that JVM bytecode is implemented on top of, and that is not constrained by the Java object and memory model? Like, if it can do unbounded pointer arithmetic without emulating it with arrays (which is slow and inefficient), surely the JIT needs to have the corresponding low-level opcodes for that, like e.g. CLR does?
- pjmlp 8y agoGraal only knows about JVM bytecodes, and there isn't any emulation of pointer arithmetic. You need first to understand the role of Truffle in this. It is a framework to generate AST interpreters, that plugs into Graal. So an interpreter for LLVM gets written in Truffle, which while running a specific LLVM application, generates an interpreter for that specific code, so after awhile the JIT comes into the picture optimizing the generated interpreter that is processing that LLVM code. This goes a few rounds until it becomes almost indistinguishable from generating the code straight from LLVM. It is the same principle behind using PyPY and RPython to add new languages into PyPy. JRuby is one of the projects making use of this infrastructure to run C extensions. Naturally due to the slow startup until everything is finally converted into native code, this approach is only usable for long lived server based applications. You can get more information about it in some of these links, mostly outdated though. https://llvm.org/devmtg/2016-01/slides/Sulong.pdf https://llvm.org/devmtg/2016-01/slides/Sulong.pdf "Using LLVM and Sulong for Language C Extensions - LLVM Cauldron 2016" https://www.youtube.com/watch?v=bJzMfYX6n9A https://www.youtube.com/watch?v=bJzMfYX6n9A "Project Sulong: an LLVM bitcode interpreter on the Graal VM" https://www.youtube.com/watch?v=yyDD_KRdQQU https://www.youtube.com/watch?v=yyDD_KRdQQU
- int_19h 8y agoRegardless of how it gets there in the end - from an interpreter or otherwise - if it produces pure JVM bytecode, that means that the memory model is also JVM. How does that accommodate the C memory model? Emulating heap with giant arrays?
- pjmlp 8y agoMost likely yes, which doesn't matter, because Graal will eventually optimize bounds check away anyway, as it does for regular Java code. The links that I provided go into detail about all of this.