[Driver][SYCL] Improvements to -fsycl-device-only behaviors

Make some improvements to -fsycl-device-only to better match the device
compilations for regular -fsycl device compiles.

 - Do not force the device triple at the initial target setting, allowing
   for improved capabilities with user provided -target <triple>
 - Allow for proper C++ assumptions (source and standard)

Author: mdtoguchi

clang/lib/Driver/Driver.cpp

@@ -1263,17 +1263,6 @@ Compilation *Driver::BuildCompilation(ArrayRef<const char *> ArgList) {
     T.setObjectFormat(llvm::Triple::COFF);
     TargetTriple = T.str();
   }
-  if (Args.hasArg(options::OPT_fsycl_device_only)) {
-    // -fsycl-device-only implies spir arch and SYCL Device
-    llvm::Triple T(TargetTriple);
-    // FIXME: defaults to spir64, should probably have a way to set spir
-    // possibly new -sycl-target option
-    T.setArch(llvm::Triple::spir64);
-    T.setVendor(llvm::Triple::UnknownVendor);
-    T.setOS(llvm::Triple(llvm::sys::getProcessTriple()).getOS());
-    T.setEnvironment(llvm::Triple::SYCLDevice);
-    TargetTriple = T.str();
-  }
   if (const Arg *A = Args.getLastArg(options::OPT_target))
     TargetTriple = A->getValue();
   if (const Arg *A = Args.getLastArg(options::OPT_ccc_install_dir))
@@ -2384,7 +2373,9 @@ void Driver::BuildInputs(const ToolChain &TC, DerivedArgList &Args,
   // actually use it, so we warn about unused -x arguments.
   types::ID InputType = types::TY_Nothing;
   Arg *InputTypeArg = nullptr;
-  bool IsSYCL = Args.hasFlag(options::OPT_fsycl, options::OPT_fno_sycl, false);
+  bool IsSYCL =
+      Args.hasFlag(options::OPT_fsycl, options::OPT_fno_sycl, false) ||
+      Args.hasArg(options::OPT_fsycl_device_only);
 
   // The last /TC or /TP option sets the input type to C or C++ globally.
   if (Arg *TCTP = Args.getLastArgNoClaim(options::OPT__SLASH_TC,

clang/lib/Driver/ToolChains/Clang.cpp

@@ -3978,14 +3978,24 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
                          const ArgList &Args, const char *LinkingOutput) const {
   const auto &TC = getToolChain();
   const llvm::Triple &RawTriple = TC.getTriple();
-  const llvm::Triple &Triple = TC.getEffectiveTriple();
-  const std::string &TripleStr = Triple.getTriple();
+  llvm::Triple Triple = TC.getEffectiveTriple();
 
   bool KernelOrKext =
       Args.hasArg(options::OPT_mkernel, options::OPT_fapple_kext);
   const Driver &D = TC.getDriver();
   ArgStringList CmdArgs;
 
+  // -fsycl-device-only implies a SPIRV arch triple.  Do not set if current
+  // effective triple is SYCLDevice
+  if (Args.hasArg(options::OPT_fsycl_device_only) &&
+      Triple.getEnvironment() != llvm::Triple::SYCLDevice) {
+    const char *SYCLTargetArch = "spir64";
+    if (C.getDefaultToolChain().getTriple().getArch() == llvm::Triple::x86)
+      SYCLTargetArch = "spir";
+    Triple = C.getDriver().MakeSYCLDeviceTriple(SYCLTargetArch);
+  }
+  const std::string &TripleStr = Triple.getTriple();
+
   // Check number of inputs for sanity. We need at least one input.
   assert(Inputs.size() >= 1 && "Must have at least one input.");
   // CUDA/HIP compilation may have multiple inputs (source file + results of
@@ -4000,7 +4010,8 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
   bool IsHIP = JA.isOffloading(Action::OFK_HIP);
   bool IsOpenMPDevice = JA.isDeviceOffloading(Action::OFK_OpenMP);
   bool IsSYCLOffloadDevice = JA.isDeviceOffloading(Action::OFK_SYCL);
-  bool IsSYCL = JA.isOffloading(Action::OFK_SYCL);
+  bool IsSYCL = JA.isOffloading(Action::OFK_SYCL) ||
+                Args.hasArg(options::OPT_fsycl_device_only);
   bool IsHeaderModulePrecompile = isa<HeaderModulePrecompileJobAction>(JA);
   assert((IsCuda || IsHIP || (IsOpenMPDevice && Inputs.size() == 2) || IsSYCL ||
           IsHeaderModulePrecompile || Inputs.size() == 1) &&
@@ -4048,14 +4059,17 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
       (IsSYCL || IsCuda || IsHIP) ? TC.getAuxTriple() : nullptr;
   bool IsWindowsMSVC = RawTriple.isWindowsMSVCEnvironment();
   bool IsIAMCU = RawTriple.isOSIAMCU();
-  bool IsSYCLDevice = (RawTriple.getEnvironment() == llvm::Triple::SYCLDevice);
+  bool IsSYCLDevice = (RawTriple.getEnvironment() == llvm::Triple::SYCLDevice ||
+                       Triple.getEnvironment() == llvm::Triple::SYCLDevice);
   // Using just the sycldevice environment is not enough to determine usage
   // of the device triple when considering fat static archives.  The
   // compilation path requires the host object to be fed into the partial link
   // step, and being part of the SYCL tool chain causes the incorrect target.
   // FIXME - Is it possible to retain host environment when on a target
   // device toolchain.
-  bool UseSYCLTriple = IsSYCLDevice && (!IsSYCL || IsSYCLOffloadDevice);
+  bool UseSYCLTriple =
+      IsSYCLDevice && (!IsSYCL || IsSYCLOffloadDevice ||
+                       Args.hasArg(options::OPT_fsycl_device_only));
 
   // Adjust IsWindowsXYZ for CUDA/HIP/SYCL compilations.  Even when compiling in
   // device mode (i.e., getToolchain().getTriple() is NVPTX/AMDGCN, not
@@ -4144,7 +4158,7 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
                       options::OPT_fno_sycl_early_optimizations,
                       Triple.getSubArch() != llvm::Triple::SPIRSubArch_fpga))
       CmdArgs.push_back("-fno-sycl-early-optimizations");
-    else if (RawTriple.isSPIR()) {
+    else if (IsSYCLDevice) {
       // Set `sycl-opt` option to configure LLVM passes for SPIR target
       CmdArgs.push_back("-mllvm");
       CmdArgs.push_back("-sycl-opt");
@@ -4157,7 +4171,8 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
 
     // Pass the triple of host when doing SYCL
     llvm::Triple AuxT = C.getDefaultToolChain().getTriple();
-    if (Args.hasFlag(options::OPT_fsycl_device_only, OptSpecifier(), false))
+    if (Args.hasArg(options::OPT_fsycl_device_only) &&
+        RawTriple.getEnvironment() == llvm::Triple::SYCLDevice)
       AuxT = llvm::Triple(llvm::sys::getProcessTriple());
     std::string NormalizedTriple = AuxT.normalize();
     CmdArgs.push_back("-aux-triple");
@@ -6812,6 +6827,7 @@ void Clang::AddClangCLArgs(const ArgList &Args, types::ID InputType,
   unsigned RTOptionID = options::OPT__SLASH_MT;
   bool isNVPTX = getToolChain().getTriple().isNVPTX();
   bool isSYCLDevice =
+      Args.hasArg(options::OPT_fsycl_device_only) ||
       getToolChain().getTriple().getEnvironment() == llvm::Triple::SYCLDevice;
   bool isSYCL = Args.hasArg(options::OPT_fsycl) || isSYCLDevice;
   // For SYCL Windows, /MD is the default.
@@ -7823,7 +7839,8 @@ void SPIRVTranslator::ConstructJob(Compilation &C, const JobAction &JA,
 
   TranslatorArgs.push_back("-o");
   TranslatorArgs.push_back(Output.getFilename());
-  if (getToolChain().getTriple().isSYCLDeviceEnvironment()) {
+  if (getToolChain().getTriple().isSYCLDeviceEnvironment() ||
+      TCArgs.hasArg(options::OPT_fsycl_device_only)) {
     TranslatorArgs.push_back("-spirv-max-version=1.1");
     TranslatorArgs.push_back("-spirv-debug-info-version=legacy");
     // Prevent crash in the translator if input IR contains DIExpression

clang/test/Driver/sycl.c

@@ -40,6 +40,7 @@
 // DEFAULT: "-triple" "spir64-unknown-{{.*}}-sycldevice{{.*}}" "-fsycl-is-device"{{.*}} "-emit-llvm-bc"
 // DEFAULT: "-internal-isystem" "{{.*}}bin{{[/\\]+}}..{{[/\\]+}}include{{[/\\]+}}sycl"
 // DEFAULT: "-internal-isystem" "{{.*lib.*clang.*include}}"
+// DEFAULT: "-std=c++17"
 // DEFAULT-NOT: "{{.*}}llvm-spirv"{{.*}} "-spirv-max-version=1.1"{{.*}} "-spirv-ext=+all,-SPV_INTEL_usm_storage_classes"
 // DEFAULT-NOT: "-std=c++11"
 // DEFAULT-NOT: "-std=c++14"
@@ -49,15 +50,28 @@
 // COMBINED: "-triple" "spir64-unknown-{{.*}}-sycldevice"{{.*}} "-fsycl-is-device"{{.*}} "-emit-llvm-bc"
 // TEXTUAL: "-triple" "spir64-unknown-{{.*}}-sycldevice{{.*}}" "-fsycl-is-device"{{.*}} "-emit-llvm"
 
+/// -fsycl-device-only triple checks
+// RUN: %clang -fsycl-device-only -target x86_64-unknown-linux-gnu -### %s 2>&1 \
+// RUN:  | FileCheck --check-prefix=DEVICE-64 %s
+// RUN: %clang_cl -fsycl-device-only --target=x86_64-unknown-linux-gnu -### %s 2>&1 \
+// RUN:  | FileCheck --check-prefix=DEVICE-64 %s
+// DEVICE-64: clang{{.*}} "-triple" "spir64-unknown-unknown-sycldevice" {{.*}} "-aux-triple" "x86_64-unknown-linux-gnu"
+
+// RUN: %clang -fsycl-device-only -target i386-unknown-linux-gnu -### %s 2>&1 \
+// RUN:  | FileCheck --check-prefix=DEVICE-32 %s
+// RUN: %clang_cl -fsycl-device-only --target=i386-unknown-linux-gnu -### %s 2>&1 \
+// RUN:  | FileCheck --check-prefix=DEVICE-32 %s
+// DEVICE-32: clang{{.*}} "-triple" "spir-unknown-unknown-sycldevice" {{.*}} "-aux-triple" "i386-unknown-linux-gnu"
+
 /// Verify that the sycl header directory is before /usr/include
 // RUN: %clangxx -### -fsycl-device-only %s 2>&1 | FileCheck %s --check-prefix=HEADER_ORDER
 // RUN: %clangxx -### -fsycl %s 2>&1 | FileCheck %s --check-prefix=HEADER_ORDER
 // HEADER_ORDER-NOT: clang{{.*}} "/usr/include"{{.*}} "-internal-isystem" "{{.*}}bin{{[/\\]+}}..{{[/\\]+}}include{{[/\\]+}}
 
 /// Verify -fsycl-device-only phases
 // RUN: %clang -### -ccc-print-phases -fsycl-device-only %s 2>&1 | FileCheck %s --check-prefix=DEFAULT-PHASES
-// DEFAULT-PHASES: 0: input, "{{.*}}", c
-// DEFAULT-PHASES: 1: preprocessor, {0}, cpp-output
+// DEFAULT-PHASES: 0: input, "{{.*}}", c++
+// DEFAULT-PHASES: 1: preprocessor, {0}, c++-cpp-output
 // DEFAULT-PHASES: 2: compiler, {1}, ir
 // DEFAULT-PHASES: 3: backend, {2}, ir
 // DEFAULT-PHASES-NOT: linker