[Driver] Enable SYCL AOT compilation using --offload-arch

clang/include/clang/Basic/DiagnosticDriverKinds.td

@@ -396,6 +396,8 @@ def err_drv_sycl_missing_amdgpu_arch : Error<
   "missing AMDGPU architecture for SYCL offloading; specify it with '-Xsycl-target-backend%select{|=%1}0 --offload-arch=<arch-name>'">;
 def err_drv_sycl_thinlto_split_off: Error<
   "'%0' is not supported when '%1' is set with '-fsycl'">;
+def err_drv_sycl_offload_arch_new_driver: Error<
+  "'--offload-arch' is supported when '-fsycl' is set with '--offload-new-driver'">;  
 def warn_drv_sycl_offload_target_duplicate : Warning<
   "SYCL offloading target '%0' is similar to target '%1' already specified; "
   "will be ignored">, InGroup<SyclTarget>;

clang/lib/Driver/Driver.cpp

@@ -1022,7 +1022,7 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
 
       for (const auto &TripleAndArchs : DerivedArchs)
         OpenMPTriples.insert(TripleAndArchs.first());
-    }
+    } // end of offload-arch
 
     for (StringRef Val : OpenMPTriples) {
       llvm::Triple TT(ToolChain::getOpenMPTriple(Val));
@@ -1182,12 +1182,13 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
   llvm::StringMap<llvm::DenseSet<StringRef>> DerivedArchs;
   llvm::StringMap<StringRef> FoundNormalizedTriples;
   llvm::SmallVector<llvm::Triple, 4> UniqueSYCLTriplesVec;
+  llvm::StringSet<> SYCLTriples;
   if (HasSYCLTargetsOption) {
     // At this point, we know we have a valid combination
     // of -fsycl*target options passed
     Arg *SYCLTargetsValues = SYCLTargets;
     if (SYCLTargetsValues) {
-      llvm::StringSet<> SYCLTriples;
+
       if (SYCLTargetsValues->getNumValues()) {
 
         // Multiple targets are currently not supported when using
@@ -1276,9 +1277,10 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
           if (!Arch.empty())
             DerivedArchs[DeviceTriple.getTriple()].insert(Arch);
         }
+
         if (!SYCLTriples.empty()) {
           for (const auto &SYCLTriple : SYCLTriples) {
-            llvm::Triple Triple(SYCLTriple.getKey());
+            llvm::Triple Triple(MakeSYCLDeviceTriple(SYCLTriple.getKey()));
             UniqueSYCLTriplesVec.push_back(Triple);
           }
         }
@@ -1287,7 +1289,95 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
         Diag(clang::diag::warn_drv_empty_joined_argument)
             << SYCLTargetsValues->getAsString(C.getInputArgs());
     }
-  } else {
+  }
+  // If the user specified --offload-arch, deduce the offloading
+  // target triple(s) from the set of architecture(s).
+  // Create a toolchain for each valid triple.
+  else if (HasValidSYCLRuntime &&
+           C.getInputArgs().hasArg(options::OPT_offload_arch_EQ) && !IsHIP &&
+           !IsCuda) {
+    const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>();
+    auto AMDTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs());
+    auto NVPTXTriple = getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(),
+                                                    HostTC->getTriple());
+
+    // Attempt to deduce the offloading triple from the set of architectures.
+    // We need to temporarily create these toolchains so that we can access
+    // tools for inferring architectures.
+    llvm::DenseSet<StringRef> Archs;
+    if (NVPTXTriple) {
+      auto TempTC = std::make_unique<toolchains::CudaToolChain>(
+          *this, *NVPTXTriple, *HostTC, C.getInputArgs(), Action::OFK_None);
+      for (StringRef Arch :
+           getOffloadArchs(C, C.getArgs(), Action::OFK_SYCL, &*TempTC, true))
+        Archs.insert(Arch);
+    }
+    if (AMDTriple) {
+      auto TempTC = std::make_unique<toolchains::AMDGPUOpenMPToolChain>(
+          *this, *AMDTriple, *HostTC, C.getInputArgs());
+      for (StringRef Arch :
+           getOffloadArchs(C, C.getArgs(), Action::OFK_SYCL, &*TempTC, true))
+        Archs.insert(Arch);
+    }
+    if (!AMDTriple && !NVPTXTriple) {
+      for (StringRef Arch :
+           getOffloadArchs(C, C.getArgs(), Action::OFK_SYCL, nullptr, true))
+        Archs.insert(Arch);
+    }
+    for (StringRef Arch : Archs) {
+      if (NVPTXTriple && IsNVIDIAOffloadArch(StringToOffloadArch(
+                             getProcessorFromTargetID(*NVPTXTriple, Arch)))) {
+        DerivedArchs[NVPTXTriple->getTriple()].insert(Arch);
+      } else if (AMDTriple &&
+                 IsAMDOffloadArch(StringToOffloadArch(
+                     getProcessorFromTargetID(*AMDTriple, Arch)))) {
+        DerivedArchs[AMDTriple->getTriple()].insert(Arch);
+      } else if (IsIntelCPUOffloadArch(StringToOffloadArchIntel(Arch))) {
+        DerivedArchs["spir64_x86_64"].insert(Arch);
+      } else if (IsIntelGPUOffloadArch(StringToOffloadArchIntel(Arch))) {
+        DerivedArchs["spir64_gen"].insert(Arch);
+      } else {
+        Diag(clang::diag::err_drv_failed_to_deduce_target_from_arch) << Arch;
+        return;
+      }
+    }
+    // If the set is empty then we failed to find a native architecture.
+    if (Archs.empty()) {
+      Diag(clang::diag::err_drv_failed_to_deduce_target_from_arch) << "native";
+      return;
+    }
+
+    for (const auto &TripleAndArchs : DerivedArchs)
+      SYCLTriples.insert(TripleAndArchs.first());
+
+    for (const auto &Val : SYCLTriples) {
+      llvm::Triple TT(MakeSYCLDeviceTriple(Val.getKey()));
+      std::string NormalizedName = TT.normalize();
+
+      // Make sure we don't have a duplicate triple.
+      auto Duplicate = FoundNormalizedTriples.find(NormalizedName);
+      if (Duplicate != FoundNormalizedTriples.end()) {
+        Diag(clang::diag::warn_drv_omp_offload_target_duplicate)
+            << Val.getKey() << Duplicate->second;
+        continue;
+      }
+
+      // Store the current triple so that we can check for duplicates in the
+      // following iterations.
+      FoundNormalizedTriples[NormalizedName] = Val.getKey();
+    }
+
+    if (!SYCLTriples.empty()) {
+      for (const auto &SYCLTriple : SYCLTriples) {
+        llvm::Triple Triple(MakeSYCLDeviceTriple(SYCLTriple.getKey()));
+        UniqueSYCLTriplesVec.push_back(Triple);
+      }
+    }
+
+    addSYCLDefaultTriple(C, UniqueSYCLTriplesVec);
+
+  } // end of --offload-arch
+  else {
     // If -fsycl is supplied without -fsycl-targets we will assume SPIR-V.
     // For -fsycl-device-only, we also setup the implied triple as needed.
     if (HasValidSYCLRuntime) {
@@ -7244,6 +7334,22 @@ void Driver::BuildActions(Compilation &C, DerivedArgList &Args,
 
   handleArguments(C, Args, Inputs, Actions);
 
+  bool HasValidSYCLRuntime =
+      C.getInputArgs().hasFlag(options::OPT_fsycl, options::OPT_fno_sycl,
+                               false) ||
+      hasSYCLDeviceOnly(C.getInputArgs());
+  bool IsSYCLOffloadArchEnabled =
+      HasValidSYCLRuntime &&
+      C.getInputArgs().hasArg(options::OPT_offload_arch_EQ);
+  /*
+    if (IsSYCLOffloadArchEnabled &&
+        !C.getInputArgs().hasFlag(options::OPT_offload_new_driver,
+                                  options::OPT_no_offload_new_driver, false)) {
+      Diag(clang::diag::err_drv_sycl_offload_arch_new_driver);
+      return;
+    }
+  */
+
   // If '-fintelfpga' is passed, add '-fsycl' to the list of arguments
   const llvm::opt::OptTable &Opts = getOpts();
   Arg *SYCLFpgaArg = C.getInputArgs().getLastArg(options::OPT_fintelfpga);

clang/lib/Driver/ToolChains/SYCL.cpp

@@ -27,6 +27,31 @@ using namespace clang::driver::tools;
 using namespace clang;
 using namespace llvm::opt;
 
+struct StringToOffloadArchIntelMap {
+  const char *ArchName;
+  OffloadArchIntel IntelArch;
+};
+
+static const StringToOffloadArchIntelMap StringToArchNamesMap[] = {
+    {"broadwell", OffloadArchIntel::BROADWELL},
+    {"coffeelake", OffloadArchIntel::COFFEELAKE},
+    {"icelake-client", OffloadArchIntel::ICELAKECLIENT},
+    {"bdw", OffloadArchIntel::BDW},
+    {"cfl", OffloadArchIntel::CFL},
+    {"icl", OffloadArchIntel::ICL}};
+
+OffloadArchIntel
+clang::driver::StringToOffloadArchIntel(llvm::StringRef ArchNameAsString) {
+  auto result = std::find_if(
+      std::begin(StringToArchNamesMap), std::end(StringToArchNamesMap),
+      [ArchNameAsString](const StringToOffloadArchIntelMap &map) {
+        return ArchNameAsString == map.ArchName;
+      });
+  if (result == std::end(StringToArchNamesMap))
+    return OffloadArchIntel::UNKNOWN;
+  return result->IntelArch;
+}
+
 SYCLInstallationDetector::SYCLInstallationDetector(const Driver &D)
     : D(D), InstallationCandidates() {
   InstallationCandidates.emplace_back(D.Dir + "/..");

clang/lib/Driver/ToolChains/SYCL.h

@@ -16,6 +16,82 @@
 namespace clang {
 namespace driver {
 
+// List of supported Intel values for CPUs
+// and GPUs.
+enum class OffloadArchIntel {
+  // CPUs
+  UNKNOWN,
+  SKYLAKEAVX512,
+  COREAVX2,
+  COREI7AVX,
+  COREI7,
+  WESTMERE,
+  SANDYBRIDGE,
+  IVYBRIDGE,
+  BROADWELL,
+  COFFEELAKE,
+  ALDERLAKE,
+  SKYLAKE,
+  SKX,
+  CASCADELAKE,
+  ICELAKECLIENT,
+  ICELAKESERVER,
+  SAPPHIRERAPIDS,
+  GRANITERAPIDS,
+  // GPUs
+  BDW,
+  SKL,
+  KBL,
+  CFL,
+  APL,
+  BXT,
+  GLK,
+  WHL,
+  AML,
+  CML,
+  ICLLP,
+  ICL,
+  EHL,
+  JSL,
+  TGLLP,
+  TGL,
+  RKL,
+  ADL_S,
+  RPL_S,
+  ADL_P,
+  ADL_N,
+  DG1,
+  ACM_G10,
+  DG2_G10,
+  ACM_G11,
+  DG2_GLL,
+  ACM_G12,
+  DG2_G12,
+  PVC,
+  PVC_VG,
+  MTL_U,
+  MTL_S,
+  ARL_U,
+  ARL_S,
+  MTL_H,
+  ARL_H,
+  BMG_G21,
+  LNL_M
+};
+
+// Check if the given Arch value is a valid Intel CPU.
+static inline bool IsIntelCPUOffloadArch(OffloadArchIntel Arch) {
+  return Arch >= OffloadArchIntel::SKYLAKEAVX512 &&
+         Arch < OffloadArchIntel::BDW;
+}
+
+// Check if the given Arch value is a valid Intel GPU.
+static inline bool IsIntelGPUOffloadArch(OffloadArchIntel Arch) {
+  return Arch >= OffloadArchIntel::BDW && Arch <= OffloadArchIntel::LNL_M;
+}
+
+OffloadArchIntel StringToOffloadArchIntel(llvm::StringRef ArchNameAsString);
+
 class SYCLInstallationDetector {
 public:
   SYCLInstallationDetector(const Driver &D);

clang/test/Driver/sycl-offload-arch-intelgpu.cpp

@@ -0,0 +1,17 @@
+/// Tests the behaviors of using -fsycl --offload-new-driver 
+//  --offload-arch=<intel-gpu-values>.
+
+// RUN: %clangxx -### --offload-new-driver -fsycl --offload-arch=bdw %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefixes=TARGET-TRIPLE,CLANG-OFFLOAD-PACKAGER -DDEV_STR=bdw -DMAC_STR=BDW
+
+// RUN: %clangxx -### --offload-new-driver -fsycl --offload-arch=cfl %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefixes=TARGET-TRIPLE,CLANG-OFFLOAD-PACKAGER -DDEV_STR=cfl -DMAC_STR=CFL
+
+
+///If Arch is icl, map it to icllp internally to create D__SYCL_TARGET_INTEL_GPU_
+
+// TARGET-TRIPLE: clang{{.*}} "-triple" "spir64_gen-unknown-unknown"
+// TARGET-TRIPLE: "-D__SYCL_TARGET_INTEL_GPU_[[MAC_STR]]__"
+// CLANG-OFFLOAD-PACKAGER: clang-offload-packager{{.*}} "--image={{.*}}triple=spir64_gen-unknown-unknown,arch=[[DEV_STR]],kind=sycl"
+
+