[tooling] Teach Tooling to understand compilation with offloading.

[hliao]

Summary

• So far, we only recognize the compilation with offloading and skip the offloading part.

[tra]

Added Manuel as someone familiar with tooling.

clang/lib/Tooling/Tooling.cpp

126

Is this still the right job for us to pick? I think we want this to be the host compilation.

As things are right now my guess is that we're probably picking the first device-side compilation.

[hliao]

clang/lib/Tooling/Tooling.cpp

126

yes, OffloadAction is appended to the original host action. The corresponding first job is the host compilation. The frontend has similar handling in lib/Frontend/CreateInvocationFromCommandLine.cpp

[hliao]

PING for review

[tra]

clang/lib/Tooling/Tooling.cpp

126

Can we add an assert or diagnostics for that? I think the check below would not always be able to tell if it unintentionally picked a wrong kind of job.

[hliao]

clang/lib/Tooling/Tooling.cpp

126

It seems to me that the cast<> and check against "clang" @ L118 already did what I asked.

[tra]

clang/lib/Tooling/Tooling.cpp

126

Device-side compilation would also use clang, so the check for clang is insufficient for ensuring that it's a *host* compilation.

Furthermore, the order of jobs *will* change depending on whether -fsyntax-only is in effect.

@klimek -- Manuel, Is tooling allowed to create an invocation w/o -fsyntax-only? AFAICT, runToolOnCodeWithArgs always adds it, but one could technically create an Invocation without it.

[klimek]

clang/lib/Tooling/Tooling.cpp

126

Generally, tooling is allowed to do it, but it's not super well supported or tested.

[hliao]

add more assertions on offload compilation.

[hliao]

clang/lib/Tooling/Tooling.cpp

126

I added more offload compilation based on actions, which is more reliable that jobs. Is that sufficient?

[tra]

clang/lib/Tooling/Tooling.cpp

102

Why 2? Will it not be different if user targeted multiple GPUs?

103

I'd add a comment describing that we expect the first jub to be a (host-side)compilation.
It would not be obvious to someone w/o familiarity of cuda/hip compilation.

[hliao]

revice comment.

[hliao]

clang/lib/Tooling/Tooling.cpp

102

that's the number of top-level actions. no matter how many GPUs are specified, that offload action is the top-level action masters all of them.

[thakis]

This makes tests assert on Mac: http://45.33.8.238/mac/1415/step_6.txt

[hliao]

In D68660#1705087, @thakis wrote:

This makes tests assert on Mac: http://45.33.8.238/mac/1415/step_6.txt

I have no access to MacOS but try to fix that in r374478. Maybe that's a good excuse for a MacBook, ;)

[hliao]

In D68660#1705087, @thakis wrote:

This makes tests assert on Mac: http://45.33.8.238/mac/1415/step_6.txt

please let me know whether that works for you. thanks.

clang/lib/Tooling/Tooling.cpp

 /// Retrieves the clang CC1 specific flags out of the compilation's jobs.
 ///
 /// Returns nullptr on error.
 static const llvm::opt::ArgStringList *getCC1Arguments(
     DiagnosticsEngine *Diagnostics, driver::Compilation *Compilation) {
   // We expect to get back exactly one Command job, if we didn't something
   // failed. Extract that job from the Compilation.
   const driver::JobList &Jobs = Compilation->getJobs();
-  if (Jobs.size() != 1 || !isa<driver::Command>(*Jobs.begin())) {
+  const driver::ActionList &Actions = Compilation->getActions();
+  bool OffloadCompilation = false;
+  if (Jobs.size() > 1) {
+    for (auto A : Actions){
+      // On MacOSX real actions may end up being wrapped in BindArchAction
+      if (isa<driver::BindArchAction>(A))
+        A = *A->input_begin();
+      if (isa<driver::OffloadAction>(A)) {
+        // Offload compilation has 2 top-level actions, one (at the front) is
+        // the original host compilation and the other is offload action
+        // composed of at least one device compilation. For such case, general
+        // tooling will consider host-compilation only. For tooling on device
+        // compilation, device compilation only option, such as
+        // `--cuda-device-only`, needs specifying.
+        assert(Actions.size() == 2);
+        assert(isa<driver::CompileJobAction>(Actions.front()));
+        OffloadCompilation = true;
+        break;
+      }
+    }
+  }
+  if (Jobs.size() == 0 || !isa<driver::Command>(*Jobs.begin()) ||
+      (Jobs.size() > 1 && !OffloadCompilation)) {
     SmallString<256> error_msg;
     llvm::raw_svector_ostream error_stream(error_msg);
     Jobs.Print(error_stream, "; ", true);
     Diagnostics->Report(diag::err_fe_expected_compiler_job)
         << error_stream.str();
     return nullptr;
   }
 
   // The one job we find should be to invoke clang again.
   const auto &Cmd = cast<driver::Command>(*Jobs.begin());
   if (StringRef(Cmd.getCreator().getName()) != "clang") {
     Diagnostics->Report(diag::err_fe_expected_clang_command);
     return nullptr;
   }
 
   return &Cmd.getArguments();
 }
 

clang/test/Tooling/clang-check-offload.cpp

+// RUN: not clang-check "%s" -- -c -x hip -nogpulib 2>&1 | FileCheck %s
+
+// CHECK: C++ requires
+invalid;