[Clang] Determine offloading architectures at Toolchain creation

jhuber6 · jhuber6 · commit 904f4b8f5635 · 2025-06-26T11:09:54.000-05:00
Summary:
Previously we had this weird disconnect where we would get some
offloading architectures beforehand and some later. This patch changes
it to where we just generate this information at Toolchain creation.
There's a few edge cases that will need to be cleaned up. Namely, we
don't handle the strange SPIR-V handling that mixes two separate
toolchains and we needed a pre-check to reject errors when inferring the
toolchain from `--offload-arch` in OpenMP.

Possible we could also use this information for some host defines if
needed.
diff --git a/clang/include/clang/Driver/Driver.h b/clang/include/clang/Driver/Driver.h
@@ -367,10 +367,9 @@ class Driver {
   /// stored in it, and will clean them up when torn down.
   mutable llvm::StringMap<std::unique_ptr<ToolChain>> ToolChains;
 
-  /// Cache of known offloading architectures for the ToolChain already derived.
-  /// This should only be modified when we first initialize the offloading
-  /// toolchains.
-  llvm::DenseMap<const ToolChain *, llvm::DenseSet<llvm::StringRef>> KnownArchs;
+  /// The associated offloading architectures with each toolchain.
+  llvm::DenseMap<const ToolChain *, llvm::SmallVector<llvm::StringRef>>
+      OffloadArchs;
 
 private:
   /// TranslateInputArgs - Create a new derived argument list from the input
@@ -535,11 +534,11 @@ class Driver {
 
   /// Returns the set of bound architectures active for this offload kind.
   /// If there are no bound architctures we return a set containing only the
-  /// empty string. The \p SuppressError option is used to suppress errors.
-  llvm::DenseSet<StringRef>
+  /// empty string.
+  llvm::SmallVector<StringRef>
   getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
                   Action::OffloadKind Kind, const ToolChain *TC,
-                  bool SuppressError = false) const;
+                  bool SpecificToolchain = true) const;
 
   /// Check that the file referenced by Value exists. If it doesn't,
   /// issue a diagnostic and return false.
diff --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
@@ -988,6 +988,8 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
     if (CudaInstallation.isValid())
       CudaInstallation.WarnIfUnsupportedVersion();
     C.addOffloadDeviceToolChain(&TC, Action::OFK_Cuda);
+    OffloadArchs[&TC] = getOffloadArchs(C, C.getArgs(), Action::OFK_Cuda, &TC,
+                                        /*SpecificToolchain=*/true);
   } else if (IsHIP && !UseLLVMOffload) {
     if (auto *OMPTargetArg =
             C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) {
@@ -1004,6 +1006,12 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
         getOffloadToolChain(C.getInputArgs(), Action::OFK_HIP, *HIPTriple,
                             C.getDefaultToolChain().getTriple());
     C.addOffloadDeviceToolChain(&TC, Action::OFK_HIP);
+
+    // TODO: Fix 'amdgcnspirv' handling with the new driver.
+    if (C.getInputArgs().hasFlag(options::OPT_offload_new_driver,
+                                 options::OPT_no_offload_new_driver, false))
+      OffloadArchs[&TC] = getOffloadArchs(C, C.getArgs(), Action::OFK_HIP, &TC,
+                                          /*SpecificToolchain=*/true);
   }
 
   if (IsCuda || IsHIP)
@@ -1069,40 +1077,43 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
         auto &TC = getOffloadToolChain(C.getInputArgs(), Action::OFK_OpenMP, TT,
                                        C.getDefaultToolChain().getTriple());
         C.addOffloadDeviceToolChain(&TC, Action::OFK_OpenMP);
+        OffloadArchs[&TC] =
+            getOffloadArchs(C, C.getArgs(), Action::OFK_OpenMP, &TC,
+                            /*SpecificToolchain=*/true);
       }
     } else if (C.getInputArgs().hasArg(options::OPT_offload_arch_EQ) &&
                ((!IsHIP && !IsCuda) || UseLLVMOffload)) {
       llvm::Triple AMDTriple("amdgcn-amd-amdhsa");
       llvm::Triple NVPTXTriple("nvptx64-nvidia-cuda");
 
-      // Attempt to deduce the offloading triple from the set of architectures.
-      // We can only correctly deduce NVPTX / AMDGPU triples currently.
-      for (const llvm::Triple &TT : {AMDTriple, NVPTXTriple}) {
-        auto &TC = getOffloadToolChain(C.getInputArgs(), Action::OFK_OpenMP, TT,
-                                       C.getDefaultToolChain().getTriple());
-
-        llvm::DenseSet<StringRef> Archs =
-            getOffloadArchs(C, C.getArgs(), Action::OFK_OpenMP, &TC, true);
-        llvm::DenseSet<StringRef> ArchsForTarget;
-        for (StringRef Arch : Archs) {
+      for (StringRef A :
+           C.getInputArgs().getAllArgValues(options::OPT_offload_arch_EQ)) {
+        for (StringRef Arch : llvm::split(A, ",")) {
           bool IsNVPTX = IsNVIDIAOffloadArch(
               StringToOffloadArch(getProcessorFromTargetID(NVPTXTriple, Arch)));
           bool IsAMDGPU = IsAMDOffloadArch(
               StringToOffloadArch(getProcessorFromTargetID(AMDTriple, Arch)));
-          if (!IsNVPTX && !IsAMDGPU && !Arch.equals_insensitive("native")) {
+          if (!IsNVPTX && !IsAMDGPU && !Arch.empty() &&
+              !Arch.equals_insensitive("native")) {
             Diag(clang::diag::err_drv_failed_to_deduce_target_from_arch)
                 << Arch;
             return;
           }
-
-          if (TT.isNVPTX() && IsNVPTX)
-            ArchsForTarget.insert(Arch);
-          else if (TT.isAMDGPU() && IsAMDGPU)
-            ArchsForTarget.insert(Arch);
         }
-        if (!ArchsForTarget.empty()) {
+      }
+
+      // Attempt to deduce the offloading triple from the set of architectures.
+      // We can only correctly deduce NVPTX / AMDGPU triples currently.
+      for (const llvm::Triple &TT : {AMDTriple, NVPTXTriple}) {
+        auto &TC = getOffloadToolChain(C.getInputArgs(), Action::OFK_OpenMP, TT,
+                                       C.getDefaultToolChain().getTriple());
+
+        llvm::SmallVector<StringRef> Archs =
+            getOffloadArchs(C, C.getArgs(), Action::OFK_OpenMP, &TC,
+                            /*SpecificToolchain=*/false);
+        if (!Archs.empty()) {
           C.addOffloadDeviceToolChain(&TC, Action::OFK_OpenMP);
-          KnownArchs[&TC] = ArchsForTarget;
+          OffloadArchs[&TC] = Archs;
         }
       }
 
@@ -1143,9 +1154,11 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
     // going to create will depend on both.
     const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>();
     for (const auto &TT : UniqueSYCLTriplesVec) {
-      auto SYCLTC = &getOffloadToolChain(C.getInputArgs(), Action::OFK_SYCL, TT,
-                                         HostTC->getTriple());
-      C.addOffloadDeviceToolChain(SYCLTC, Action::OFK_SYCL);
+      auto &TC = getOffloadToolChain(C.getInputArgs(), Action::OFK_SYCL, TT,
+                                     HostTC->getTriple());
+      C.addOffloadDeviceToolChain(&TC, Action::OFK_SYCL);
+      OffloadArchs[&TC] = getOffloadArchs(C, C.getArgs(), Action::OFK_SYCL, &TC,
+                                          /*SpecificToolchain=*/true);
     }
   }
 
@@ -4703,20 +4716,22 @@ static StringRef getCanonicalArchString(Compilation &C,
                                         const llvm::opt::DerivedArgList &Args,
                                         StringRef ArchStr,
                                         const llvm::Triple &Triple,
-                                        bool SuppressError = false) {
+                                        bool SpecificToolchain) {
   // Lookup the CUDA / HIP architecture string. Only report an error if we were
   // expecting the triple to be only NVPTX / AMDGPU.
   OffloadArch Arch =
       StringToOffloadArch(getProcessorFromTargetID(Triple, ArchStr));
-  if (!SuppressError && Triple.isNVPTX() &&
+  if (Triple.isNVPTX() &&
       (Arch == OffloadArch::UNKNOWN || !IsNVIDIAOffloadArch(Arch))) {
-    C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch)
-        << "CUDA" << ArchStr;
+    if (SpecificToolchain)
+      C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch)
+          << "CUDA" << ArchStr;
     return StringRef();
-  } else if (!SuppressError && Triple.isAMDGPU() &&
+  } else if (Triple.isAMDGPU() &&
              (Arch == OffloadArch::UNKNOWN || !IsAMDOffloadArch(Arch))) {
-    C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch)
-        << "HIP" << ArchStr;
+    if (SpecificToolchain)
+      C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch)
+          << "HIP" << ArchStr;
     return StringRef();
   }
 
@@ -4725,13 +4740,9 @@ static StringRef getCanonicalArchString(Compilation &C,
 
   if (IsAMDOffloadArch(Arch)) {
     llvm::StringMap<bool> Features;
-    auto HIPTriple = getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs());
-    if (!HIPTriple)
-      return StringRef();
-    auto Arch = parseTargetID(*HIPTriple, ArchStr, &Features);
+    std::optional<StringRef> Arch = parseTargetID(Triple, ArchStr, &Features);
     if (!Arch) {
       C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << ArchStr;
-      C.setContainsError();
       return StringRef();
     }
     return Args.MakeArgStringRef(getCanonicalTargetID(*Arch, Features));
@@ -4754,10 +4765,10 @@ getConflictOffloadArchCombination(const llvm::DenseSet<StringRef> &Archs,
   return getConflictTargetIDCombination(ArchSet);
 }
 
-llvm::DenseSet<StringRef>
+llvm::SmallVector<StringRef>
 Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
                         Action::OffloadKind Kind, const ToolChain *TC,
-                        bool SuppressError) const {
+                        bool SpecificToolchain) const {
   if (!TC)
     TC = &C.getDefaultToolChain();
 
@@ -4772,9 +4783,6 @@ Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
                 : "--no-offload-arch");
   }
 
-  if (KnownArchs.contains(TC))
-    return KnownArchs.lookup(TC);
-
   llvm::DenseSet<StringRef> Archs;
   for (auto *Arg : C.getArgsForToolChain(TC, /*BoundArch=*/"", Kind)) {
     // Add or remove the seen architectures in order of appearance. If an
@@ -4784,7 +4792,7 @@ Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
         if (Arch == "native" || Arch.empty()) {
           auto GPUsOrErr = TC->getSystemGPUArchs(Args);
           if (!GPUsOrErr) {
-            if (SuppressError)
+            if (!SpecificToolchain)
               llvm::consumeError(GPUsOrErr.takeError());
             else
               TC->getDriver().Diag(diag::err_drv_undetermined_gpu_arch)
@@ -4794,16 +4802,21 @@ Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
           }
 
           for (auto ArchStr : *GPUsOrErr) {
-            Archs.insert(
+            StringRef CanonicalStr =
                 getCanonicalArchString(C, Args, Args.MakeArgString(ArchStr),
-                                       TC->getTriple(), SuppressError));
+                                       TC->getTriple(), SpecificToolchain);
+            if (!CanonicalStr.empty())
+              Archs.insert(CanonicalStr);
+            else if (SpecificToolchain)
+              return llvm::SmallVector<StringRef>();
           }
         } else {
-          StringRef ArchStr = getCanonicalArchString(
-              C, Args, Arch, TC->getTriple(), SuppressError);
-          if (ArchStr.empty())
-            return Archs;
-          Archs.insert(ArchStr);
+          StringRef CanonicalStr = getCanonicalArchString(
+              C, Args, Arch, TC->getTriple(), SpecificToolchain);
+          if (!CanonicalStr.empty())
+            Archs.insert(CanonicalStr);
+          else if (SpecificToolchain)
+            return llvm::SmallVector<StringRef>();
         }
       }
     } else if (Arg->getOption().matches(options::OPT_no_offload_arch_EQ)) {
@@ -4812,27 +4825,20 @@ Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
           Archs.clear();
         } else {
           StringRef ArchStr = getCanonicalArchString(
-              C, Args, Arch, TC->getTriple(), SuppressError);
-          if (ArchStr.empty())
-            return Archs;
+              C, Args, Arch, TC->getTriple(), SpecificToolchain);
           Archs.erase(ArchStr);
         }
       }
     }
   }
 
   if (auto ConflictingArchs =
-          getConflictOffloadArchCombination(Archs, TC->getTriple())) {
+          getConflictOffloadArchCombination(Archs, TC->getTriple()))
     C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo)
         << ConflictingArchs->first << ConflictingArchs->second;
-    C.setContainsError();
-  }
 
   // Skip filling defaults if we're just querying what is availible.
-  if (SuppressError)
-    return Archs;
-
-  if (Archs.empty()) {
+  if (SpecificToolchain && Archs.empty()) {
     if (Kind == Action::OFK_Cuda) {
       Archs.insert(OffloadArchToString(OffloadArch::CudaDefault));
     } else if (Kind == Action::OFK_HIP) {
@@ -4858,12 +4864,13 @@ Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
         }
       }
     }
-  } else {
-    Args.ClaimAllArgs(options::OPT_offload_arch_EQ);
-    Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ);
   }
+  Args.ClaimAllArgs(options::OPT_offload_arch_EQ);
+  Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ);
 
-  return Archs;
+  SmallVector<StringRef> Sorted(Archs.begin(), Archs.end());
+  llvm::sort(Sorted);
+  return Sorted;
 }
 
 Action *Driver::BuildOffloadingActions(Compilation &C,
@@ -4927,10 +4934,7 @@ Action *Driver::BuildOffloadingActions(Compilation &C,
     // Get the product of all bound architectures and toolchains.
     SmallVector<std::pair<const ToolChain *, StringRef>> TCAndArchs;
     for (const ToolChain *TC : ToolChains) {
-      llvm::DenseSet<StringRef> Arches = getOffloadArchs(C, Args, Kind, TC);
-      SmallVector<StringRef, 0> Sorted(Arches.begin(), Arches.end());
-      llvm::sort(Sorted);
-      for (StringRef Arch : Sorted) {
+      for (StringRef Arch : OffloadArchs.lookup(TC)) {
         TCAndArchs.push_back(std::make_pair(TC, Arch));
         DeviceActions.push_back(
             C.MakeAction<InputAction>(*InputArg, InputType, CUID));
diff --git a/clang/test/Driver/openmp-offload-gpu.c b/clang/test/Driver/openmp-offload-gpu.c
@@ -307,7 +307,7 @@
 // DRIVER_EMBEDDING: -fembed-offload-object={{.*}}.out
 
 // RUN:   %clang -### --target=x86_64-unknown-linux-gnu -ccc-print-bindings -fopenmp=libomp -fopenmp-targets=nvptx64-nvidia-cuda \
-// RUN:     --offload-host-only -nogpulib %s 2>&1 | FileCheck %s --check-prefix=CHECK-HOST-ONLY
+// RUN:     -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_52 --offload-host-only -nogpulib %s 2>&1 | FileCheck %s --check-prefix=CHECK-HOST-ONLY
 // CHECK-HOST-ONLY: "x86_64-unknown-linux-gnu" - "clang", inputs: ["[[INPUT:.*]]"], output: "[[OUTPUT:.*]]"
 // CHECK-HOST-ONLY: "x86_64-unknown-linux-gnu" - "Offload::Linker", inputs: ["[[OUTPUT]]"], output: "a.out"
 
