-S and -fsycl-device-only

Author: Jakub Chlanda

clang/include/clang/Basic/DiagnosticDriverKinds.td

@@ -420,6 +420,11 @@ def err_drv_fsycl_wrong_optimization_options : Error<
 def warn_drv_fsycl_add_default_spec_consts_image_flag_in_non_AOT : Warning<
   "-fsycl-add-default-spec-consts-image flag has an effect only in Ahead of Time Compilation mode (AOT)">,
   InGroup<SyclTarget>;
+def warn_drv_fsycl_device_obj_asm_device_only : Warning<
+  "-fsycl-device-obj=asm flag has an effect only when compailing device code "
+  "and emitting assembly, make sure both -fsycl-device-only and -S flags are "
+  "present. The option will be ignored">,
+  InGroup<SyclTarget>;
 def warn_drv_ftarget_register_alloc_mode_pvc : Warning<
   "using '%0' to set GRF mode on PVC hardware is deprecated; use '-ftarget-register-alloc-mode=pvc:%1'">,
   InGroup<Deprecated>;

clang/lib/Driver/Driver.cpp

@@ -1303,11 +1303,15 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
   // to using llvmir.  This can be improved to be more obvious in usage.
   if (Arg *DeviceObj = C.getInputArgs().getLastArgNoClaim(
           options::OPT_fsycl_device_obj_EQ)) {
+    const bool SYCLDeviceOnly = C.getDriver().offloadDeviceOnly();
+    const bool EmitAsm = C.getInputArgs().getLastArgNoClaim(options::OPT_S);
     StringRef ArgValue(DeviceObj->getValue());
     SmallVector<StringRef, 3> DeviceObjValues = {"spirv", "llvmir", "asm"};
     if (llvm::find(DeviceObjValues, ArgValue) == DeviceObjValues.end())
       Diag(clang::diag::warn_ignoring_value_using_default)
           << DeviceObj->getSpelling().split('=').first << ArgValue << "llvmir";
+    else if (ArgValue == "asm" && (!SYCLDeviceOnly || !EmitAsm))
+      Diag(clang::diag::warn_drv_fsycl_device_obj_asm_device_only);
   }
 
   Arg *SYCLForceTarget =
@@ -5321,7 +5325,7 @@ class OffloadingActionBuilder final {
                                                        types::TY_SPIRV);
             if (SYCLDeviceOnly)
               continue;
-          } else if ((SYCLDeviceOnly || FinalPhase != phases::Link) &&
+          } else if (SYCLDeviceOnly && Args.hasArg(options::OPT_S) &&
                      Args.getLastArgValue(options::OPT_fsycl_device_obj_EQ)
                          .equals_insensitive("asm")) {
             auto *CompileAction =

clang/test/Driver/sycl-device-obj-ptx.cpp renamed to clang/test/Driver/sycl-device-obj-asm.cpp

@@ -14,16 +14,26 @@
 // CHECK-AMD: 4: offload, "device-sycl (amdgcn-amd-amdhsa:gfx90a)" {3}, assembler
 
 /// Check -fsycl-device-obj=asm for Nvidia.
-// RUN:   %clang -fsycl-device-only -fsycl -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend=nvptx64-nvidia-cuda "--cuda-gpu-arch=sm_50" -fsycl-device-obj=asm -S %s 2>&1 -ccc-print-phases -o - | FileCheck %s --check-prefix=CHECK-PTX
+// RUN:   %clang -fsycl-device-only -fsycl -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend=nvptx64-nvidia-cuda --cuda-gpu-arch=sm_50 -fsycl-device-obj=asm -S %s 2>&1 -ccc-print-phases -o - | FileCheck %s --check-prefix=CHECK-PTX
 // CHECK-PTX: 0: input, "{{.+\.cpp}}", c++, (device-sycl, sm_50)
 // CHECK-PTX: 1: preprocessor, {0}, c++-cpp-output, (device-sycl, sm_50)
 // CHECK-PTX: 2: compiler, {1}, ir, (device-sycl, sm_50)
 // CHECK-PTX: 3: backend, {2}, assembler, (device-sycl, sm_50)
 // CHECK-PTX: 4: offload, "device-sycl (nvptx64-nvidia-cuda:sm_50)" {3}, assembler
 
 /// Check -fsycl-device-obj option when emitting llvm IR.
-// RUN:   %clang -fsycl-device-only -fsycl -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend=nvptx64-nvidia-cuda "--cuda-gpu-arch=sm_50" -fsycl-device-obj=llvmir -S %s 2>&1 -ccc-print-phases -o - | FileCheck %s --check-prefix=CHECK-LLVMIR
+// RUN:   %clang -fsycl-device-only -fsycl -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend=nvptx64-nvidia-cuda --cuda-gpu-arch=sm_50 -fsycl-device-obj=llvmir -S %s 2>&1 -ccc-print-phases -o - | FileCheck %s --check-prefix=CHECK-LLVMIR
 // CHECK-LLVMIR: 0: input, "{{.+\.cpp}}", c++, (device-sycl, sm_50)
 // CHECK-LLVMIR: 1: preprocessor, {0}, c++-cpp-output, (device-sycl, sm_50)
 // CHECK-LLVMIR: 2: compiler, {1}, ir, (device-sycl, sm_50)
 // CHECK-LLVMIR: 3: offload, "device-sycl (nvptx64-nvidia-cuda:sm_50)" {2}, ir
+
+/// -fsycl-device-obj=asm should always be accompanied by -fsycl-device-only
+/// and -S, check that the compiler issues a correct warning message:
+// RUN:   %clang -fsycl-device-only -fsycl -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend=nvptx64-nvidia-cuda --cuda-gpu-arch=sm_50 -fsycl-device-obj=asm %s 2>&1 -o - | FileCheck %s --check-prefix=CHECK-NO-DEV-ONLY-NO-S
+// CHECK-NO-DEV-ONLY-NO-S: warning: -fsycl-device-obj=asm flag has an effect only when compailing device code and emitting assembly, make sure both -fsycl-device-only and -S flags are present. The option will be ignored [-Wsycl-target]
+
+/// -fsycl-device-obj=asm will finish at generating assembly stage, hence
+/// inform users that generating library will not be possible (ignore -c)
+// RUN:   %clang -fsycl-device-only -fsycl -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend=nvptx64-nvidia-cuda --cuda-gpu-arch=sm_50 -fsycl-device-obj=asm %s 2>&1 -fsycl-device-only -S -c -o - | FileCheck %s --check-prefix=CHECK-DASH-C-IGNORE
+// CHECK-DASH-C-IGNORE: warning: argument unused during compilation: '-c' [-Wunused-command-line-argument]