[OpenMP] Customize CUDA-based tool chain selection

[gtbercea]

Summary

This patch provides a generic way of selecting CUDA based tool chains as host-device pairs.

[Hahnfeld]

lib/Frontend/CompilerInvocation.cpp

2181–2190 ↗

(On Diff #87454)

This is unrelated to CUDA, please split that into its own patch.

[gtbercea]

Split patch.

[Hahnfeld]

Is there a way to actually test the changed code? The current test changes seem unrelated...

include/clang/Basic/DiagnosticDriverKinds.td

165–166 ↗

(On Diff #93679)

This diagnostic seems unrelated to the code.

test/OpenMP/target_messages.cpp

7–8 ↗

(On Diff #93679)

This test seems unrelated to the changes to the code.

[gtbercea]

Remove tests which belong into a different patch.

[Hahnfeld]

LGTM with one small note

lib/Driver/Driver.cpp

564

The code above uses HostTriple.str(), maybe better align to this?

[ABataev]

Tests?

lib/Driver/Driver.cpp

556

Seems to me you're not using this initial value, so it is better to leave this variable uninitialized

[gtbercea]

Rebase on latest master.

[gtbercea]

lib/Driver/Driver.cpp

564

HostTriple is equivalent with HostTC->getTriple(). HostTC->getTriple() is only used once so no need for a local variable unless you want me to refactor the CUDA code above and take out HostTriple from within the if statement and then re0use HostTriple that way. I'd rather not make changes to the CUDA toolchain above.

[gtbercea]

In D29658#732520, @ABataev wrote:

Tests?

Can't write any meaningful tests. This will be tested implicitely by all future patches that perform offloading using OpenMP.

[Hahnfeld]

lib/Driver/Driver.cpp

564

Sorry, I meant the mismatch between .str() above and .normalize() here

[gtbercea]

Use str()

lib/Driver/Driver.cpp

if (OpenMPTargets->getNumValues()) {
           // Store the current triple so that we can check for duplicates in the
           // following iterations.
           FoundNormalizedTriples[NormalizedName] = Val;
 
           // If the specified target is invalid, emit a diagnostic.
           if (TT.getArch() == llvm::Triple::UnknownArch)
             Diag(clang::diag::err_drv_invalid_omp_target) << Val;
           else {
-            const ToolChain &TC = getToolChain(C.getInputArgs(), TT);
-            C.addOffloadDeviceToolChain(&TC, Action::OFK_OpenMP);
+            const ToolChain *TC = nullptr;
+            // CUDA toolchains have to be selected differently. They pair host
+            // and device in their implementation.
+            if (TT.isNVPTX()) {
+              const ToolChain *HostTC =
+                  C.getSingleOffloadToolChain<Action::OFK_Host>();
+              assert(HostTC && "Host toolchain should be always defined.");
+              auto &CudaTC =
+                  ToolChains[TT.str() + "/" + HostTC->getTriple().normalize()];
+              if (!CudaTC)
+                CudaTC = llvm::make_unique<toolchains::CudaToolChain>(
+                    *this, TT, *HostTC, C.getInputArgs());
+              TC = CudaTC.get();
+            } else
+              TC = &getToolChain(C.getInputArgs(), TT);
+            C.addOffloadDeviceToolChain(TC, Action::OFK_OpenMP);
           }
         }
       } else
         Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets);
     } else
       Diag(clang::diag::warn_drv_empty_joined_argument)
           << OpenMPTargets->getAsString(C.getInputArgs());
   }