[clang][Driver][HIP] Add support for mixing AMDGCNSPIRV & concrete offload-archs. (#113509)

This removes the temporary ban on mixing AMDGCN flavoured SPIR-V and
concrete targets (e.g. `gfx900`) in the same HIPAMD compilation. This is
done primarily by tweaking the effective / observable triple when the
target is `amdgcnspirv`, which seamlessly composes with the existing
infra. The test is stolen from #75357.

Author: AlexVlx

clang/lib/Driver/Driver.cpp

@@ -149,13 +149,9 @@ static std::optional<llvm::Triple>
 getHIPOffloadTargetTriple(const Driver &D, const ArgList &Args) {
   if (!Args.hasArg(options::OPT_offload_EQ)) {
     auto OffloadArchs = Args.getAllArgValues(options::OPT_offload_arch_EQ);
-    if (llvm::is_contained(OffloadArchs, "amdgcnspirv")) {
-      if (OffloadArchs.size() == 1)
-        return llvm::Triple("spirv64-amd-amdhsa");
-      // Mixing specific & SPIR-V compilation is not supported for now.
-      D.Diag(diag::err_drv_only_one_offload_target_supported);
-      return std::nullopt;
-    }
+    if (llvm::is_contained(OffloadArchs, "amdgcnspirv") &&
+        OffloadArchs.size() == 1)
+      return llvm::Triple("spirv64-amd-amdhsa");
     return llvm::Triple("amdgcn-amd-amdhsa"); // Default HIP triple.
   }
   auto TT = getOffloadTargetTriple(D, Args);
@@ -3477,9 +3473,11 @@ class OffloadingActionBuilder final {
       llvm::StringMap<bool> Features;
       // getHIPOffloadTargetTriple() is known to return valid value as it has
       // been called successfully in the CreateOffloadingDeviceToolChains().
-      auto ArchStr = parseTargetID(
-          *getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()), IdStr,
-          &Features);
+      auto T =
+          (IdStr == "amdgcnspirv")
+              ? llvm::Triple("spirv64-amd-amdhsa")
+              : *getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs());
+      auto ArchStr = parseTargetID(T, IdStr, &Features);
       if (!ArchStr) {
         C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << IdStr;
         C.setContainsError();
@@ -5755,7 +5753,7 @@ InputInfoList Driver::BuildJobsForActionNoCache(
     // We only have to generate a prefix for the host if this is not a top-level
     // action.
     std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix(
-        A->getOffloadingDeviceKind(), TC->getTriple().normalize(),
+        A->getOffloadingDeviceKind(), EffectiveTriple.normalize(),
         /*CreatePrefixForHost=*/isa<OffloadPackagerJobAction>(A) ||
             !(A->getOffloadingHostActiveKinds() == Action::OFK_None ||
               AtTopLevel));

clang/lib/Driver/ToolChain.cpp

@@ -1099,6 +1099,12 @@ std::string ToolChain::ComputeLLVMTriple(const ArgList &Args,
   }
   case llvm::Triple::aarch64_32:
     return getTripleString();
+  case llvm::Triple::amdgcn: {
+    llvm::Triple Triple = getTriple();
+    if (Args.getLastArgValue(options::OPT_mcpu_EQ) == "amdgcnspirv")
+      Triple.setArch(llvm::Triple::ArchType::spirv64);
+    return Triple.getTriple();
+  }
   case llvm::Triple::arm:
   case llvm::Triple::armeb:
   case llvm::Triple::thumb:

clang/lib/Driver/ToolChains/HIPAMD.cpp

@@ -205,7 +205,7 @@ void AMDGCN::Linker::ConstructJob(Compilation &C, const JobAction &JA,
   if (JA.getType() == types::TY_LLVM_BC)
     return constructLlvmLinkCommand(C, JA, Inputs, Output, Args);
 
-  if (getToolChain().getTriple().isSPIRV())
+  if (getToolChain().getEffectiveTriple().isSPIRV())
     return constructLinkAndEmitSpirvCommand(C, JA, Inputs, Output, Args);
 
   return constructLldCommand(C, JA, Inputs, Output, Args);
@@ -264,12 +264,14 @@ void HIPAMDToolChain::addClangTargetOptions(
     CC1Args.push_back("-fapply-global-visibility-to-externs");
   }
 
-  // For SPIR-V we embed the command-line into the generated binary, in order to
-  // retrieve it at JIT time and be able to do target specific compilation with
-  // options that match the user-supplied ones.
-  if (getTriple().isSPIRV() &&
-      !DriverArgs.hasArg(options::OPT_fembed_bitcode_marker))
-    CC1Args.push_back("-fembed-bitcode=marker");
+  if (getEffectiveTriple().isSPIRV()) {
+    // For SPIR-V we embed the command-line into the generated binary, in order
+    // to retrieve it at JIT time and be able to do target specific compilation
+    // with options that match the user-supplied ones.
+    if (!DriverArgs.hasArg(options::OPT_fembed_bitcode_marker))
+      CC1Args.push_back("-fembed-bitcode=marker");
+    return; // No DeviceLibs for SPIR-V.
+  }
 
   for (auto BCFile : getDeviceLibs(DriverArgs)) {
     CC1Args.push_back(BCFile.ShouldInternalize ? "-mlink-builtin-bitcode"
@@ -361,8 +363,7 @@ llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12>
 HIPAMDToolChain::getDeviceLibs(const llvm::opt::ArgList &DriverArgs) const {
   llvm::SmallVector<BitCodeLibraryInfo, 12> BCLibs;
   if (DriverArgs.hasArg(options::OPT_nogpulib) ||
-      (getTriple().getArch() == llvm::Triple::spirv64 &&
-       getTriple().getVendor() == llvm::Triple::AMD))
+      getGPUArch(DriverArgs) == "amdgcnspirv")
     return {};
   ArgStringList LibraryPaths;
 
@@ -437,8 +438,8 @@ HIPAMDToolChain::getDeviceLibs(const llvm::opt::ArgList &DriverArgs) const {
 void HIPAMDToolChain::checkTargetID(
     const llvm::opt::ArgList &DriverArgs) const {
   auto PTID = getParsedTargetID(DriverArgs);
-  if (PTID.OptionalTargetID && !PTID.OptionalGPUArch) {
+  if (PTID.OptionalTargetID && !PTID.OptionalGPUArch &&
+      PTID.OptionalTargetID != "amdgcnspirv")
     getDriver().Diag(clang::diag::err_drv_bad_target_id)
         << *PTID.OptionalTargetID;
-  }
 }

clang/lib/Driver/ToolChains/HIPUtility.cpp

@@ -304,10 +304,14 @@ void HIP::constructHIPFatbinCommand(Compilation &C, const JobAction &JA,
   for (const auto &II : Inputs) {
     const auto *A = II.getAction();
     auto ArchStr = llvm::StringRef(A->getOffloadingArch());
-    BundlerTargetArg +=
-        "," + OffloadKind + "-" + normalizeForBundler(TT, !ArchStr.empty());
+    BundlerTargetArg += ',' + OffloadKind + '-';
+    if (ArchStr == "amdgcnspirv")
+      BundlerTargetArg +=
+          normalizeForBundler(llvm::Triple("spirv64-amd-amdhsa"), true);
+    else
+      BundlerTargetArg += normalizeForBundler(TT, !ArchStr.empty());
     if (!ArchStr.empty())
-      BundlerTargetArg += "-" + ArchStr.str();
+      BundlerTargetArg += '-' + ArchStr.str();
   }
   BundlerArgs.push_back(Args.MakeArgString(BundlerTargetArg));
 

clang/test/Driver/hip-toolchain-no-rdc.hip

@@ -36,6 +36,11 @@
 // RUN:   %t/a.o %t/b.o \
 // RUN: 2>&1 | FileCheck -check-prefixes=LKONLY %s
 
+// RUN: %clang -### --target=x86_64-linux-gnu \
+// RUN:   --offload-arch=amdgcnspirv --offload-arch=gfx900 \
+// RUN:   %s -nogpuinc -nogpulib \
+// RUN: 2>&1 | FileCheck -check-prefixes=AMDGCNSPIRV %s
+
 //
 // Compile device code in a.cu to code object for gfx803.
 //
@@ -177,3 +182,16 @@
 // LKONLY-NOT: {{".*/llc"}}
 // LKONLY: [[LD:".*ld.*"]] {{.*}} "{{.*/a.o}}" "{{.*/b.o}}"
 // LKONLY-NOT: "-T" "{{.*}}.lk"
+
+//
+// Check mixed AMDGCNSPIRV and concrete GPU arch.
+//
+
+// AMDGCNSPIRV: "-cc1" "-triple" "spirv64-amd-amdhsa" {{.*}}"-emit-obj" {{.*}} "-o" "[[AMDGCNSPV_OBJ:.*o]]"
+// AMDGCNSPIRV: {{".*llvm-link.*"}} "-o" "[[AMDGCNSPV_TMP:.*out]]" "[[AMDGCNSPV_OBJ]]"
+// AMDGCNSPIRV: {{".*llvm-spirv.*"}} "--spirv-max-version=1.6" "--spirv-ext=+all" {{.*}} "[[AMDGCNSPV_TMP]]" {{.*}}"-o" "[[AMDGCNSPV_CO:.*out]]"
+// AMDGCNSPIRV: "-cc1" "-triple" "amdgcn-amd-amdhsa" {{.*}}"-emit-obj" {{.*}}"-target-cpu" "gfx900"{{.*}} "-o" "[[GFX900_OBJ:.*o]]"
+// AMDGCNSPIRV: {{".*lld.*"}} {{.*}}"-plugin-opt=mcpu=gfx900" {{.*}} "-o" "[[GFX900_CO:.*out]]" {{.*}}"[[GFX900_OBJ]]"
+// AMDGCNSPIRV: {{".*clang-offload-bundler.*"}} "-type=o"
+// AMDGCNSPIRV-SAME: "-targets={{.*}}hipv4-spirv64-amd-amdhsa--amdgcnspirv,hipv4-amdgcn-amd-amdhsa--gfx900"
+// AMDGCNSPIRV-SAME: "-input=[[AMDGCNSPV_CO]]" "-input=[[GFX900_CO]]"