[Clang] Parse toolchain-specific offloading arguments directly

[jhuber6]

Summary

OpenMP supports multiple offloading toolchains and architectures. In
order to support this we originally used getArgsForToolchain to get
the arguments only intended for each toolchain. This allowed users to
manually specify if an --offload-arch= argument was intended for which
toolchain using -Xopenmp-target= or other methods. For example,

clang input.c -fopenmp -fopenmp-targets=nvptx64,amdgcn -Xopenmp-target=nvptx64 --offload-arch=sm_70 -Xopenmp-target=amdgcn --offload-arch=gfx908

However, this was causing problems with the AMDGPU toolchain. This is
because the AMDGPU toolchain for OpenMP uses an amdgpu arch to determine the
architecture. If this tool is not availible the compiler will exit with an error
even when manually specifying the architecture. This patch pulls out the logic in
getArgsForToolchain and specializes it for extracting --offload-arch
arguments to avoid this.

[JonChesterfield]

LG, thanks!

[yaxunl]

need a test

[jhuber6]

In D129435#3642450, @yaxunl wrote:

need a test

There's a test clang/test/Driver/openmp-offload-gpu-new.c already making sure that the following works, so this patch should maintain the same functionality we had before.

clang input.c -fopenmp -fopenmp-targets=nvptx64,amdgcn -Xopenmp-target=nvptx64 --offload-arch=sm_70 -Xopenmp-target=amdgcn --offload-arch=gfx908

[yaxunl]

LGTM. Thanks.

clang/lib/Driver/Driver.cpp

C.getDriver().Diag(diag::err_opt_not_valid_with_opt)
                 ? "--offload-arch"
                 : "--no-offload-arch");
   }
 
   if (KnownArchs.find(TC) != KnownArchs.end())
     return KnownArchs.lookup(TC);
 
   llvm::DenseSet<StringRef> Archs;
-  for (auto &Arg : Args) {
+  for (auto *Arg : Args) {
+    // Extract any '--[no-]offload-arch' arguments intended for this toolchain.
+    std::unique_ptr<llvm::opt::Arg> ExtractedArg = nullptr;
+    if (Arg->getOption().matches(options::OPT_Xopenmp_target_EQ) &&
+        ToolChain::getOpenMPTriple(Arg->getValue(0)) == TC->getTriple()) {
+      Arg->claim();
+      unsigned Index = Args.getBaseArgs().MakeIndex(Arg->getValue(1));
+      ExtractedArg = getOpts().ParseOneArg(Args, Index);
+      Arg = ExtractedArg.get();
+    }
+
     if (Arg->getOption().matches(options::OPT_offload_arch_EQ)) {
       for (StringRef Arch : llvm::split(Arg->getValue(), ","))
         Archs.insert(getCanonicalArchString(C, Args, Arch, TC->getTriple()));
     } else if (Arg->getOption().matches(options::OPT_no_offload_arch_EQ)) {
       for (StringRef Arch : llvm::split(Arg->getValue(), ",")) {
         if (Arch == StringRef("all"))
           Archs.clear();
         else
if (ToolChains.empty())
       continue;
 
     types::ID InputType = Input.first;
     const Arg *InputArg = Input.second;
 
     // Get the product of all bound architectures and toolchains.
     SmallVector<std::pair<const ToolChain *, StringRef>> TCAndArchs;
     for (const ToolChain *TC : ToolChains)
-      for (StringRef Arch : getOffloadArchs(
-               C, C.getArgsForToolChain(TC, "generic", Kind), Kind, TC))
+      for (StringRef Arch : getOffloadArchs(C, Args, Kind, TC))
         TCAndArchs.push_back(std::make_pair(TC, Arch));
 
     for (unsigned I = 0, E = TCAndArchs.size(); I != E; ++I)
       DeviceActions.push_back(C.MakeAction<InputAction>(*InputArg, InputType));
 
     if (DeviceActions.empty())
       return HostAction;