JIT compiler, Windows, external functions like cos

Hello LLVM team,

our software FluidSIM (www.fluidsim.de) simulates pneumatic, hydraulic and electric circuits. For the mathematical models we use the language Modelica (www.modelica.org). We developed our own Modelica simulator which solves the dynamical created algebraic differential equation systems. One tool is our small JIT compiler, which compiles mathematical expressions like “2*x0 + sin(x1)” at runtime.

In the future we want to compile more complex expressions, especially entire blocks, like
X1 = 2*x0 +4
X2 = 3*sin(x1)
X3 = x1 + x2

At the moment, we are evaluating the LLVM suite to replace our JIT compiler. Our software runs under the Windows platform (32 and 64 Bit). We are using the Microsoft Visual Studio 2017 Version 15.9.4.
For our evaluation we use LLVM 7.0.1. We created the LLVM suite for 32 Bit with the cmake tools. We wrote a small parser and took the LLVM tutorials as starting point to implement our LLVM JIT compiler. All arithmetic expressions work fine. But we failed to integrate external functions like “cos” and “sin”. Only “sqrt” worked. After many trials and errors we found out that no LLVM example works (“4. Kaleidoscope: Adding JIT and Optimizer Support”, “The Fibonacci project“).

It seems that the LLVM JIT compiler cannot find external functions like “sin” and “cos” nor intrinsic functions under Windows. Every (tutorial) example crashes using them. Is there any solution for this problem (compiler switches, libraries to include, a simple Visual Studio project)?

The following code works for “sqrt” but not for “cos”. It just encapsulates an external call. We use
Microsoft Visual Studio 2017 Version 15.9.4.
LLVM 7.0.1
Solution Configuration: Debug, 32 Bit, Switch: /MTd

using namespace llvm;

typedef double(__cdecl *JitCompiledFn)(double);

int main()
{
// "sqrt" works.
//const char externalFnName = "sqrt";

// "cos", "sin", etc. fails.
const char externalFnName = "cos";

InitializeNativeTarget();
InitializeNativeTargetAsmPrinter();
InitializeNativeTargetAsmParser();

LLVMContext context;

IRBuilder<> builder(context);

std::unique_ptr<llvm::Module> module(new Module("TestModule", context));
Module* pModule = module.get();

auto externalFn_IR = cast<Function>(pModule->getOrInsertFunction("externalFn", Type::getDoubleTy(context), Type::getDoubleTy(context)));
Value* x = externalFn_IR->arg_begin();
x->setName("x");

BasicBlock *entryBlock = BasicBlock::Create(context, "EntryBlock", externalFn_IR);
builder.SetInsertPoint(entryBlock);

std::vector<Type *> args(1, Type::getDoubleTy(context));
FunctionType *FT = FunctionType::get(Type::getDoubleTy(context), args, false);
auto externalFn_llvm = Function::Create(FT, Function::ExternalLinkage, externalFnName, pModule);
auto ret = builder.CreateCall(externalFn_llvm, x);

builder.CreateRet(ret);

errs() << "Created Module:\n\n" << *pModule;

auto jitCompiler = EngineBuilder(std::move(module)).setOptLevel(CodeGenOpt::Level::Default).create();

JitCompiledFn externalFn = (JitCompiledFn)jitCompiler->getFunctionAddress(externalFn_IR->getName());

errs() << "\n\nexternalFn(9.0) = ";

double y = externalFn(9.0);

errs() << y << "\n\n";

return 0;
}

Kind regards
Marcus Hoffmann

Hi Marcus

If I understand correctly, you are creating a module with a declaration for an external function. Then you compile and load the module with the ExecutionEngine JIT interface. The loader is supposed to resolve your declaration to an existing definition in the host program, right? Do you get the error message “Program used external function ‘…’ which could not be resolved!”?

There is a variety of potential reasons here. The mystery that it works for sqrt, but not for cos or sin, may be that something in your host program exposes a definition with its unmangled name only for sqrt, while the others are mangled or not exposed at all. After all, your host program links to the C++ stdlib, which provides multiple versions of all these functions with different signatures!

In general, using host-process symbols directly can be dangerous and hard to debug. Basically, you are relying on details of your host compiler and linker. The correct way may be adding another module with the stdlib implementation that your JITed code should use. The simplest way may be keeping the host-process calls, but expose your own wrappers explicitly by address. Using the Orc JIT interface directly, a QND implementation with LLVM 5.0 looked like this:
Note that on Windows you want to declare the wrappers as:

Hi Marcus (+list)

Have you ever successfully compiled and run your project with the
current Visual Studio 2017 under Windows 10?

Well, it's possible I have missed a detail like this during the LLVM 5.0
port, but I am sure it worked on Windows with LLVM 4.0 at the time of
writing the examples. I do not have an archived version or anything, but
the original branches are still there:
https://github.com/weliveindetail/JitFromScratch/tree/master/llvm40

LLVM ERROR: Program used external function '_abs' which could not be
resolved!

On Windows there should be no underscore as mangling prefix. Maybe
that's the issue in the example.

Maybe I just miss the right linker or compiler options for using the
standard library?

What you want is a pure C stdlib and what your host program brings is a
C++ stdlib. In order to use the latter you must mangle your names, which
gets complicated quickly. If you go with custom wrappers in your host
program, you can declare them extern "C" and avoid the mangling headache!

Let's keep the conversation on the list.