Currently , I have doing some experimental work by using llvm, Is it possible to drop the machine code once it has been generated for particular function while program executing. For example some void test(int) function has been executed on native machine , I want to drop the code before I start execute some other function in my long running program.
Currently , I have doing some experimental work by using llvm,
Is it possible to drop the machine code once it has been generated for
particular function while program executing. For example some *void
test(int)* function has been executed on native machine , I want to drop
the code before I start execute some other function in my long running
You can disassemble the code once it has been compiled. Mesa's llvmpipe
driver does this:
Thanks for the reference , is there any llvm function that I could directly call to drop the generated machine code ? Currently I am having compiled llvm IR function and I could able to get generated native function pointer by using getpointertofunction, /
This isn’t currently supported directly. It depends on what you’re doing, which JIT you’re using, how you use modules, and to what extent you’re relying on LLVM to do linking for you.
You can’t safely drop a function’s code if you have other functions in that module.
You can’t safely drop a module if there are other modules that have calls that you’ve already resolved to functions in the module you’re dropping unless you have your own mechanism for unlinking those calls.
The MCJIT currently does support dropping the memory for a module, but it involves destroying the MCJIT execution engine object. This works best if you use your own JIT memory manager and you steal the executable memory from the MCJIT, and delete the MCJIT after code is generated. Then your own memory manager can manage the memory however you like. This depends on not having LLVM call instructions resolve to any of the functions you would be dropping. WebKit is an example of a system that does this. Each function gets it’s own module and all LLVM data structures are dropped once the code is compiled. Call instructions are only used for intrinsics and for runtime calls; source level calls are implemented as patchpoints and WebKit does all of the linking (and unlinking).
Long story short, there is no shrink-wrapped solution but it’s doable if you’re willing to get dirty.
Thank you for your detailed explanation, I was actually looking to implement an adaptive approach which is basically when some function executed more frequently, I was trying to drop that function and compiled and linked with new optimized function. I just did the following -
whenever some function executed more times , I called-back to program, so I that I could be able to call freeMachineCodeForFunction (F) then I compiled that more frequent function with some kind of optimization. But , still I am getting previous function signature and not newest one. Could you please explain , why we can not use this freeMachineCodeForFunction for this purpose. If not, how we can hook some instruction in emitted machine code that will call back to our code in llvm.
Addition to my previous mail, llvm has recompileAndRelinkFunction function , so, once we modified the llvm function, and pass IR to recompileAndRelinkFunction , I hope it should be compiled and linked with previous one.
That’s a good point. But it’s worth noting that recompileAndRelinkFunction() and freeMachineCodeForFunction() are both vestiges of the old JIT (i.e. the “JIT” as opposed to the “MCJIT”). The old JIT is no longer actively supported.
Thank you for your clarification, Actually I compiled VMKit 3.2 with llvm 3.2 and trying do recompileAndRelinkFunction() . Could you please verity the following things to me. For instance , code need to be executed as follows
as soon as I executed the first test() function , I modified the llvm IR ( modified some instruction) of test , then when I execute differentTest(), I used recompileAndRelinkFunction() to link the modified test() before I execute the last test() function. Actually , I was expecting the different result when I was executing the last test() function but unfortunately it execute the previous test() function not modified one. How can I use this recompileAndRelinkFunction() to modify the native code during the execution.
If main is compiled before the call to recompile, it will call test with a direct call. After the update of the test function, the second call site is not updated. To call the new test, you can either call test with an indirect call, recompile also main, or use a patchpoint to update the second call. The last solution is probably the best one.
Thanks for your guidance . I will try patchpoint method to solve this issue .