[Polly] [PPCGCodeGeneration] Allow PPCGCodeGeneration to run without libdevice.

[bollu]

Summary

We currently _always_ assume that we can have a valid libdevice if we
are using the NVPTX backend. This is the wrong assumption to make,
because it is possible that the end user does not wish to use libdevice.

If the CUDALibDevice string is empty, we do not try to use libdevice.

[singam-sanjay]

lib/CodeGen/PPCGCodeGeneration.cpp

113

Is it possible to initialize CUDALibDevice according to the OS ?

[philip.pfaffe]

lib/CodeGen/PPCGCodeGeneration.cpp

113

Can we at least configure this at compile-time?

[bollu]

lib/CodeGen/PPCGCodeGeneration.cpp

113

Not sure how you plan to do that?

[philip.pfaffe]

lib/CodeGen/PPCGCodeGeneration.cpp

113

Discover the file with cmake, and have it configurable through a cmake variable

lib/CodeGen/PPCGCodeGeneration.cpp

FailOnVerifyModuleFailure("polly-acc-fail-on-verify-module-failure",
                               cl::desc("Fail and generate a backtrace if"
                                        " verifyModule fails on the GPU "
                                        " kernel module."),
                               cl::Hidden, cl::init(false), cl::ZeroOrMore,
                               cl::cat(PollyCategory));
 
 static cl::opt<std::string> CUDALibDevice(
     "polly-acc-libdevice", cl::desc("Path to CUDA libdevice"), cl::Hidden,
     cl::init("/usr/local/cuda/nvvm/libdevice/libdevice.compute_20.10.ll"),
     cl::ZeroOrMore, cl::cat(PollyCategory));
 
 static cl::opt<std::string>
     CudaVersion("polly-acc-cuda-version",
                 cl::desc("The CUDA version to compile for"), cl::Hidden,
                 cl::init("sm_30"), cl::ZeroOrMore, cl::cat(PollyCategory));
 
 static cl::opt<int>
bool runOnScop(Scop &CurrentScop) override {
     DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
     SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
     DL = &S->getRegion().getEntry()->getModule()->getDataLayout();
     RI = &getAnalysis<RegionInfoPass>().getRegionInfo();
 
     DEBUG(dbgs() << "PPCGCodeGen running on : " << getUniqueScopName(S)
                  << " | loop depth: " << S->getMaxLoopDepth() << "\n");
 
+    const bool AllowCUDALibDevice = Architecture == GPUArch::NVPTX64 &&
+        CUDALibDevice != "";
+
     // We currently do not support functions other than intrinsics inside
     // kernels, as code generation will need to offload function calls to the
     // kernel. This may lead to a kernel trying to call a function on the host.
     // This also allows us to prevent codegen from trying to take the
     // address of an intrinsic function to send to the kernel.
     if (containsInvalidKernelFunction(CurrentScop,
-                                      Architecture == GPUArch::NVPTX64)) {
+                                      AllowCUDALibDevice)) {
       DEBUG(
           dbgs() << getUniqueScopName(S)
                  << " contains function which cannot be materialised in a GPU "
                     "kernel. Bailing out.\n";);
       return false;
     }
 
     auto PPCGScop = createPPCGScop();