[clang][Driver] Enable the SPIR-V backend for HIP via -use-spirv-backend.

Author: mgcarrasco

clang/lib/Driver/Driver.cpp

@@ -5024,15 +5024,24 @@ Action *Driver::BuildOffloadingActions(Compilation &C,
     // Compiling HIP in device-only non-RDC mode requires linking each action
     // individually.
     for (Action *&A : DeviceActions) {
-      // Special handling for the HIP SPIR-V toolchain because it doesn't use
-      // the SPIR-V backend yet doesn't report the output as an object.
       bool IsAMDGCNSPIRV = A->getOffloadingToolChain() &&
                            A->getOffloadingToolChain()->getTriple().getOS() ==
                                llvm::Triple::OSType::AMDHSA &&
                            A->getOffloadingToolChain()->getTriple().isSPIRV();
+      bool UseSPIRVBackend = Args.hasFlag(options::OPT_use_spirv_backend,
+                                          options::OPT_no_use_spirv_backend,
+                                          /*Default=*/false);
+
+      // Special handling for the HIP SPIR-V toolchain in device-only.
+      // The translator path has a linking step, whereas the SPIR-V backend path
+      // does not to avoid any external dependency such as spirv-link. The
+      // linking step is skipped for the SPIR-V backend path.
+      bool IsAMDGCNSPIRVWithBackend = IsAMDGCNSPIRV && UseSPIRVBackend;
+
       if ((A->getType() != types::TY_Object && !IsAMDGCNSPIRV &&
            A->getType() != types::TY_LTO_BC) ||
-          HIPRelocatableObj || !HIPNoRDC || !offloadDeviceOnly())
+          HIPRelocatableObj || !HIPNoRDC || !offloadDeviceOnly() ||
+          (IsAMDGCNSPIRVWithBackend && offloadDeviceOnly()))
         continue;
       ActionList LinkerInput = {A};
       A = C.MakeAction<LinkJobAction>(LinkerInput, types::TY_Image);
@@ -5258,12 +5267,28 @@ Action *Driver::ConstructPhaseAction(
           Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC;
       return C.MakeAction<BackendJobAction>(Input, Output);
     }
+    bool UseSPIRVBackend = Args.hasFlag(options::OPT_use_spirv_backend,
+                                        options::OPT_no_use_spirv_backend,
+                                        /*Default=*/false);
+
+    auto OffloadingToolChain = Input->getOffloadingToolChain();
+    // For AMD SPIRV, if offloadDeviceOnly(), we call the SPIRV backend unless
+    // LLVM bitcode was requested explicitly or RDC is set. If
+    // !offloadDeviceOnly, we emit LLVM bitcode, and clang-linker-wrapper will
+    // compile it to SPIRV.
+    bool UseSPIRVBackendForHipDeviceOnlyNoRDC =
+        TargetDeviceOffloadKind == Action::OFK_HIP && OffloadingToolChain &&
+        OffloadingToolChain->getTriple().isSPIRV() && UseSPIRVBackend &&
+        offloadDeviceOnly() &&
+        !Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, false);
+
     if (Args.hasArg(options::OPT_emit_llvm) ||
         TargetDeviceOffloadKind == Action::OFK_SYCL ||
         (((Input->getOffloadingToolChain() &&
            Input->getOffloadingToolChain()->getTriple().isAMDGPU() &&
            TargetDeviceOffloadKind != Action::OFK_None) ||
           TargetDeviceOffloadKind == Action::OFK_HIP) &&
+         !UseSPIRVBackendForHipDeviceOnlyNoRDC &&
          ((Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc,
                         false) ||
            (Args.hasFlag(options::OPT_offload_new_driver,
@@ -5285,6 +5310,19 @@ Action *Driver::ConstructPhaseAction(
               : types::TY_LLVM_BC;
       return C.MakeAction<BackendJobAction>(Input, Output);
     }
+
+    // The SPIRV backend compilation path for HIP must avoid external
+    // dependencies. The default compilation path assembles and links its
+    // output, but the SPIRV assembler and linker are external tools. This code
+    // ensures the backend emits binary SPIRV directly to bypass those steps and
+    // avoid failures. Without -save-temps, the compiler may already skip
+    // assembling and linking. With -save-temps, these steps must be explicitly
+    // disabled, as done here. We also force skipping these steps regardless of
+    // -save-temps to avoid relying on optimizations (unless -S is set).
+    // The current HIP bundling expects the type to be types::TY_Image
+    if (UseSPIRVBackendForHipDeviceOnlyNoRDC && !Args.hasArg(options::OPT_S))
+      return C.MakeAction<BackendJobAction>(Input, types::TY_Image);
+
     return C.MakeAction<BackendJobAction>(Input, types::TY_PP_Asm);
   }
   case phases::Assemble:

clang/lib/Driver/ToolChains/Clang.cpp

@@ -5057,6 +5057,10 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
     Args.ClaimAllArgs(options::OPT_femit_dwarf_unwind_EQ);
   }
 
+  bool IsAMDSPIRVForHIPDevice =
+      IsHIPDevice && getToolChain().getTriple().isSPIRV() &&
+      getToolChain().getTriple().getVendor() == llvm::Triple::AMD;
+
   if (isa<AnalyzeJobAction>(JA)) {
     assert(JA.getType() == types::TY_Plist && "Invalid output type.");
     CmdArgs.push_back("-analyze");
@@ -5154,6 +5158,8 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
       rewriteKind = RK_Fragile;
     } else if (JA.getType() == types::TY_CIR) {
       CmdArgs.push_back("-emit-cir");
+    } else if (JA.getType() == types::TY_Image && IsAMDSPIRVForHIPDevice) {
+      CmdArgs.push_back("-emit-obj");
     } else {
       assert(JA.getType() == types::TY_PP_Asm && "Unexpected output type!");
     }
@@ -9084,7 +9090,9 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA,
       OPT_fno_lto,
       OPT_flto,
       OPT_flto_partitions_EQ,
-      OPT_flto_EQ};
+      OPT_flto_EQ,
+      OPT_use_spirv_backend};
+
   const llvm::DenseSet<unsigned> LinkerOptions{OPT_mllvm, OPT_Zlinker_input};
   auto ShouldForwardForToolChain = [&](Arg *A, const ToolChain &TC) {
     // Don't forward -mllvm to toolchains that don't support LLVM.

clang/lib/Driver/ToolChains/HIPAMD.cpp

@@ -159,10 +159,9 @@ void AMDGCN::Linker::constructLldCommand(Compilation &C, const JobAction &JA,
 
 // For SPIR-V the inputs for the job are device AMDGCN SPIR-V flavoured bitcode
 // and the output is either a compiled SPIR-V binary or bitcode (-emit-llvm). It
-// calls llvm-link and then the llvm-spirv translator. Once the SPIR-V BE will
-// be promoted from experimental, we will switch to using that. TODO: consider
-// if we want to run any targeted optimisations over IR here, over generic
-// SPIR-V.
+// calls llvm-link and then the llvm-spirv translator or the SPIR-V BE.
+// TODO: consider if we want to run any targeted optimisations over IR here,
+// over generic SPIR-V.
 void AMDGCN::Linker::constructLinkAndEmitSpirvCommand(
     Compilation &C, const JobAction &JA, const InputInfoList &Inputs,
     const InputInfo &Output, const llvm::opt::ArgList &Args) const {
@@ -173,17 +172,40 @@ void AMDGCN::Linker::constructLinkAndEmitSpirvCommand(
   const char *LinkedBCFilePath = HIP::getTempFile(C, LinkedBCFilePrefix, "bc");
   InputInfo LinkedBCFile(&JA, LinkedBCFilePath, Output.getBaseInput());
 
+  bool UseSPIRVBackend =
+      Args.hasFlag(options::OPT_use_spirv_backend,
+                   options::OPT_no_use_spirv_backend, /*Default=*/false);
+
   constructLlvmLinkCommand(C, JA, Inputs, LinkedBCFile, Args);
 
-  // Emit SPIR-V binary.
-  llvm::opt::ArgStringList TrArgs{
-      "--spirv-max-version=1.6",
-      "--spirv-ext=+all",
-      "--spirv-allow-unknown-intrinsics",
-      "--spirv-lower-const-expr",
-      "--spirv-preserve-auxdata",
-      "--spirv-debug-info-version=nonsemantic-shader-200"};
-  SPIRV::constructTranslateCommand(C, *this, JA, Output, LinkedBCFile, TrArgs);
+  if (UseSPIRVBackend) {
+    // This code handles the case in the new driver when --offload-device-only
+    // is unset and clang-linker-wrapper forwards the bitcode that must be
+    // compiled to SPIR-V.
+
+    llvm::opt::ArgStringList CmdArgs;
+    const char *Triple =
+        C.getArgs().MakeArgString("-triple=spirv64-amd-amdhsa");
+
+    CmdArgs.append({"-cc1", Triple, "-emit-obj", "-disable-llvm-optzns",
+                    LinkedBCFile.getFilename(), "-o", Output.getFilename()});
+
+    const char *Exec = getToolChain().getDriver().getClangProgramPath();
+    C.addCommand(std::make_unique<Command>(JA, *this,
+                                           ResponseFileSupport::None(), Exec,
+                                           CmdArgs, LinkedBCFile, Output));
+  } else {
+    // Emit SPIR-V binary using the translator
+    llvm::opt::ArgStringList TrArgs{
+        "--spirv-max-version=1.6",
+        "--spirv-ext=+all",
+        "--spirv-allow-unknown-intrinsics",
+        "--spirv-lower-const-expr",
+        "--spirv-preserve-auxdata",
+        "--spirv-debug-info-version=nonsemantic-shader-200"};
+    SPIRV::constructTranslateCommand(C, *this, JA, Output, LinkedBCFile,
+                                     TrArgs);
+  }
 }
 
 // For amdgcn the inputs of the linker job are device bitcode and output is