[CUDA] Allow using -o with -fsyntax-only

[tra]

Summary

-fsyntax-only breaks down CUDA compilation pipeline and make it look like
multiple independent subcompilations and that trips the multiple arguments check
when -o is specified.

We do want to allow -fsyntax-only to be used with otherwise unmodified clang
options as it's commonly used by various tooling.

[tra]

Added a test.

[tra]

Added new driver tests.

[jhuber6]

LG, thanks.

clang/lib/Driver/Driver.cpp

4555

Might be cleaner since we only use this once and it's explained by the comment above, up to you.

[yaxunl]

clang/test/Driver/cuda-bindings.cu

99

should we check there is no binding for Assembler and Linker?

[tra]

clang/test/Driver/cuda-bindings.cu

99

@jhuber6 : Interestingly enough, this exposed what looks like an issue with the new driver -- it will apparently attempt to run linker on empty inputs.
I believe it's an issue independent of this patch and will need to be fixed separately.

bin/clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 --offload-new-driver /usr/local/google/home/tra/work/llvm/repo/clang/test/Driver/cuda-bindings.cu
# "nvptx64-nvidia-cuda" - "clang", inputs: ["clang/test/Driver/cuda-bindings.cu"], output: (nothing)
# "nvptx64-nvidia-cuda" - "clang", inputs: ["clang/test/Driver/cuda-bindings.cu"], output: (nothing)
# "nvptx64-nvidia-cuda" - "NVPTX::Linker", inputs: [(nothing), (nothing)], output: "/tmp/cuda-bindings-0bbbdf.fatbin"
# "powerpc64le-ibm-linux-gnu" - "clang", inputs: ["clang/test/Driver/cuda-bindings.cu", "/tmp/cuda-bindings-0bbbdf.fatbin"], output: (nothing)

[tra]

Updated tests and addressed comments.

[tra]

clang/test/Driver/cuda-bindings.cu

99

-fsyntax-only indeed crashes the new driver: https://godbolt.org/z/vqs4b8Mv1

[jhuber6]

clang/test/Driver/cuda-bindings.cu

99

Yeah it seems to be because we don't quit when there's a TY_Nothing output. The expected behaviour of fsyntax-only should have no output, but we use the output from some previous phases when we build. Will need to find a way to still run both the host and device phases without input.

clang/lib/Driver/Driver.cpp

Action *Driver::ConstructPhaseAction(
 
   llvm_unreachable("invalid phase in ConstructPhaseAction");
 }
 
 void Driver::BuildJobs(Compilation &C) const {
   llvm::PrettyStackTraceString CrashInfo("Building compilation jobs");
 
   Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o);
+  bool SyntaxOnly = C.getArgs().getLastArg(options::OPT_fsyntax_only);
 
   // It is an error to provide a -o option if we are making multiple output
   // files. There are exceptions:
   //
   // IfsMergeJob: when generating interface stubs enabled we want to be able to
   // generate the stub file at the same time that we generate the real
   // library/a.out. So when a .o, .so, etc are the output, with clang interface
   // stubs there will also be a .ifs and .ifso at the same location.
   //
   // CompileJob of type TY_IFS_CPP: when generating interface stubs is enabled
   // and -c is passed, we still want to be able to generate a .ifs file while
   // we are also generating .o files. So we allow more than one output file in
   // this case as well.
   //
   // OffloadClass of type TY_Nothing: device-only output will place many outputs
   // into a single offloading action. We should count all inputs to the action
-  // as outputs.
+  // as outputs. Also ignore device-only outputs if we're compiling with
+  // -fsyntax-only.
   if (FinalOutput) {
     unsigned NumOutputs = 0;
     unsigned NumIfsOutputs = 0;
     for (const Action *A : C.getActions()) {
       if (A->getType() != types::TY_Nothing &&
           !(A->getKind() == Action::IfsMergeJobClass ||
             (A->getType() == clang::driver::types::TY_IFS_CPP &&
              A->getKind() == clang::driver::Action::CompileJobClass &&
              0 == NumIfsOutputs++) ||
             (A->getKind() == Action::BindArchClass && A->getInputs().size() &&
              A->getInputs().front()->getKind() == Action::IfsMergeJobClass)))
         ++NumOutputs;
       else if (A->getKind() == Action::OffloadClass &&
-               A->getType() == types::TY_Nothing)
+               A->getType() == types::TY_Nothing && !SyntaxOnly)
         NumOutputs += A->size();
     }
 
     if (NumOutputs > 1) {
       Diag(clang::diag::err_drv_output_argument_with_multiple_files);
       FinalOutput = nullptr;
     }
   }

clang/test/Driver/cuda-bindings.cu

 // RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings --cuda-gpu-arch=sm_30 %s -S 2>&1 \
 // RUN: | FileCheck -check-prefix=ASM %s
 // ASM-DAG: # "nvptx64-nvidia-cuda" - "clang",{{.*}} output: "cuda-bindings-cuda-nvptx64-nvidia-cuda-sm_30.s"
 // ASM-DAG: # "powerpc64le-ibm-linux-gnu" - "clang",{{.*}} output: "cuda-bindings.s"
 
 //
 // Test two gpu architectures with complete compilation.
 //
-// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 %s 2>&1 \
-// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings --offload-arch=sm_30,sm_35 %s 2>&1 \
-// RUN: | FileCheck -check-prefix=BIN2 %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings \
+// RUN:        --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=BIN2,AOUT %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings \
+// RUN:       --offload-arch=sm_30,sm_35 %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=BIN2,AOUT %s
+// .. same, but with explicitly specified output.
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings \
+// RUN:       --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 %s -o %t/out 2>&1 \
+// RUN: | FileCheck -check-prefixes=BIN2,TOUT %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings \
+// RUN:        --offload-arch=sm_30,sm_35 %s -o %t/out 2>&1 \
+// RUN: | FileCheck -check-prefixes=BIN2,TOUT %s
 // BIN2: # "nvptx64-nvidia-cuda" - "clang",{{.*}} output:
 // BIN2-NOT: cuda-bindings-device-cuda-nvptx64
 // BIN2: # "nvptx64-nvidia-cuda" - "NVPTX::Assembler",{{.*}} output:
 // BIN2-NOT: cuda-bindings-device-cuda-nvptx64
 // BIN2: # "nvptx64-nvidia-cuda" - "clang",{{.*}} output:
 // BIN2-NOT: cuda-bindings-device-cuda-nvptx64
 // BIN2: # "nvptx64-nvidia-cuda" - "NVPTX::Assembler",{{.*}} output:
 // BIN2-NOT: cuda-bindings-device-cuda-nvptx64
 // BIN2: # "nvptx64-nvidia-cuda" - "NVPTX::Linker",{{.*}} output:
 // BIN2-NOT: cuda-bindings-device-cuda-nvptx64
 // BIN2: # "powerpc64le-ibm-linux-gnu" - "clang",{{.*}}  output:
 // BIN2-NOT: cuda-bindings-device-cuda-nvptx64
-// BIN2: # "powerpc64le-ibm-linux-gnu" - "GNU::Linker", inputs:{{.*}}, output: "a.out"
+// AOUT: # "powerpc64le-ibm-linux-gnu" - "GNU::Linker", inputs:{{.*}}, output: "a.out"
+// TOUT: # "powerpc64le-ibm-linux-gnu" - "GNU::Linker", inputs:{{.*}}, output: "{{.*}}/out"
 
+// .. same, but with -fsyntax-only
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only \
+// RUN:       --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 %s 2>&1 \
+// RUN: | FileCheck -check-prefix=SYN %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only \
+// RUN:        --offload-arch=sm_30,sm_35 %s -o %t/out 2>&1 \
+// RUN: | FileCheck -check-prefix=SYN %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only \
+// RUN:       --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 %s 2>&1 \
+// RUN: | FileCheck -check-prefix=SYN %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only \
+// RUN:        --offload-arch=sm_30,sm_35 %s -o %t/out 2>&1 \
+// RUN: | FileCheck -check-prefix=SYN %s
+// .. and with --offload-new-driver
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only \
+// RUN:       --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 --offload-new-driver %s 2>&1 \
+// RUN: | FileCheck -check-prefix=SYN %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only \
+// RUN:        --offload-arch=sm_30,sm_35 %s --offload-new-driver -o %t/out 2>&1 \
+// RUN: | FileCheck -check-prefix=SYN %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only \
+// RUN:       --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 %s --offload-new-driver 2>&1 \
+// RUN: | FileCheck -check-prefix=SYN %s
+// RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings -fsyntax-only \
+// RUN:        --offload-arch=sm_30,sm_35 %s --offload-new-driver -o %t/out 2>&1 \
+// RUN: | FileCheck -check-prefix=SYN %s
+// SYN-DAG: # "powerpc64le-ibm-linux-gnu" - "clang", inputs: [{{.*}}], output: (nothing)
+// SYN-DAG: # "nvptx64-nvidia-cuda" - "clang", inputs: [{{.*}}], output: (nothing)
+// SYN-DAG: # "nvptx64-nvidia-cuda" - "clang", inputs: [{{.*}}], output: (nothing)
 //
 // Test two gpu architectures up to the assemble phase.
 //
 // RUN: %clang -target powerpc64le-ibm-linux-gnu -ccc-print-bindings \
 // RUN:        --cuda-gpu-arch=sm_30 --cuda-gpu-arch=sm_35 %s -S 2>&1 \
 // RUN: | FileCheck -check-prefix=ASM2 %s
 // ASM2-DAG: # "nvptx64-nvidia-cuda" - "clang",{{.*}} output: "cuda-bindings-cuda-nvptx64-nvidia-cuda-sm_30.s"
 // ASM2-DAG: # "nvptx64-nvidia-cuda" - "clang",{{.*}} output: "cuda-bindings-cuda-nvptx64-nvidia-cuda-sm_35.s"
 // ASM2-DAG: # "powerpc64le-ibm-linux-gnu" - "clang",{{.*}} output: "cuda-bindings.s"