Read path to CUDA from env. variable CUDA_PATH on Windows

[mojca]

Summary

This is heavily related to https://reviews.llvm.org/D114326 and uses a different approach for locating path to CUDA on Windows.

[mojca]

clang/lib/Driver/ToolChains/Cuda.cpp

137–141

Sorry, wrong patch. I'll upload it again.

[tra]

clang/lib/Driver/ToolChains/Cuda.cpp

137

Do we want to keep this as the fall-back for cases when CUDA_PATH is not set?

Otherwise, we're risking to break existing users.

[mojca]

clang/lib/Driver/ToolChains/Cuda.cpp

137

That was going to be my question as well (for which I wanted your input).
From the initial comments in the other review (D114326) I understood that eventually you wanted to get rid of the hardcoded list of versions (but I didn't fully understand the intent).

(Technically speaking this won't break existing users since they were only able to access CUDA versions up to 8.0 until now, and we would just remove support for anything older that CUDA 10.0.)

[tra]

clang/lib/Driver/ToolChains/Cuda.cpp

137

Good point. Then it's a good opportunity to decommission the auto-search by version magic and replace it with something more predictable.

I think this patch would obviate the D114326, too, and we should be able to remove the multi-version search everywhere.

[mojca]

What can/should I do next in order to proceed with this?

[tra]

In D114601#3172199, @mojca wrote:

What can/should I do next in order to proceed with this?

Next step would be to commit the patch. If you do not have commit access, I can help you with that.
If that's the case, please let me know the name and email you want the commit to be attributed to.

[tra]

Ping. @mojca, do you need help landing the patch?

[mojca]

In D114601#3223155, @tra wrote:

Ping. @mojca, do you need help landing the patch?

Yes, please. I don't have commit access yet.
You can attribute it to mojca at macports.org, for example.

We also need a fix for unit tests on the CI in order to find the files, but I cannot help with that as I don't know how that works.

[mojca]

Also, I would like to get to do some further "baby steps" towards better support of CUDA on Windows in particular, but I would need some guidelines.
I requested a special channel that would allow a bit of discussion https://discord.com/channels/636084430946959380/.
What would be the best way to shortly discuss the options?

[tra]

In D114601#3224469, @mojca wrote:

In D114601#3223155, @tra wrote:

Ping. @mojca, do you need help landing the patch?

Yes, please. I don't have commit access yet.
You can attribute it to mojca at macports.org, for example.

Will do once we sort out the tests.

We also need a fix for unit tests on the CI in order to find the files, but I cannot help with that as I don't know how that works.

You may need to adjust clang/test/Driver/cuda-windows.cu and possibly the mock CUDA installation under clang/test/Driver/Inputs/CUDA-windows.

In D114601#3224485, @mojca wrote:

Also, I would like to get to do some further "baby steps" towards better support of CUDA on Windows in particular, but I would need some guidelines.
I requested a special channel that would allow a bit of discussion https://discord.com/channels/636084430946959380/.
What would be the best way to shortly discuss the options?

Email would work. cfe-dev@ list would work, too. Chat does not work well for me -- my reply latency tends to be higher than that of email and it's not as good for keeping a record of the conversation.

clang/lib/Driver/ToolChains/Cuda.cpp

CudaInstallationDetector::CudaInstallationDetector(
   // In decreasing order so we prefer newer versions to older versions.
   std::initializer_list<const char *> Versions = {"8.0", "7.5", "7.0"};
   auto &FS = D.getVFS();
 
   if (Args.hasArg(clang::driver::options::OPT_cuda_path_EQ)) {
     Candidates.emplace_back(
         Args.getLastArgValue(clang::driver::options::OPT_cuda_path_EQ).str());
   } else if (HostTriple.isOSWindows()) {
-    for (const char *Ver : Versions)
-      Candidates.emplace_back(
-          D.SysRoot + "/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v" +
-          Ver);
+    if (Optional<std::string> CudaPathValue =
+            llvm::sys::Process::GetEnv("CUDA_PATH")) {
+      StringRef CudaPath = *CudaPathValue;
+      Candidates.emplace_back(CudaPath.str());
+    }
   } else {
     if (!Args.hasArg(clang::driver::options::OPT_cuda_path_ignore_env)) {
       // Try to find ptxas binary. If the executable is located in a directory
       // called 'bin/', its parent directory might be a good guess for a valid
       // CUDA installation.
       // However, some distributions might installs 'ptxas' to /usr/bin. In that
       // case the candidate would be '/usr' which passes the following checks
       // because '/usr/include' exists as well. To avoid this case, we always