Windows nvcc: Cannot cache (multiple input files) triggered by -Xcompiler "/openmp:llvm" #2726

Summary

When sccache wraps nvcc on Windows, passing -Xcompiler "/openmp:llvm" through to
the host MSVC compiler causes sccache to log
Cannot cache(multiple input files, [...]) and return exit code 1, instead of
caching or falling through to a direct compile. The cl.exe argument list shown
in the error contains exactly one source file (the x_0.cudafe1.cpp produced
by nvcc), so the "multiple input files" classification appears wrong.

Likely cause: sccache's MSVC argument parser splitting /openmp:llvm on : and
counting llvm as a second positional (input file) argument.

Environment

• sccache: 0.15.0
• OS: Windows Server Core LTSC2019 (inside a Docker container)
• MSVC: 14.43.34808 (Visual Studio 2022 BuildTools)
• nvcc: CUDA 13.0 (sm_75 used for repro; not arch-specific)
• Host arch: x86_64

Minimal reproducer

echo __global__ void k() {} > hello.cu
echo int main() { k^<^<^<1,1^>^>^>(); return 0; } >> hello.cu

set SCCACHE_DIR=C:\Temp\sc
set SCCACHE_LOG=debug
sccache --stop-server >NUL 2>&1

sccache nvcc -forward-unknown-to-host-compiler ^
  -Xcompiler "/openmp:llvm" ^
  --generate-code arch=compute_75,code=sm_75 ^
  -std=c++17 -c hello.cu -o hello.obj

Exit code: 1. Removing -Xcompiler "/openmp:llvm" makes the same invocation
succeed (exit 0) and cache.

Verbatim error

  "C:\\BuildTools\\VC\\Tools\\MSVC\\14.43.34808\\bin\\HostX64\\x64\\cl.exe"
  ["/std:c++17", "-Fohello.obj", "-D__CUDA_ARCH__=750", "-D__CUDA_ARCH_LIST__=750",
   "-nologo", "-c", "-TP", "-DCUDA_DOUBLE_MATH_FUNCTIONS", "-EHsc", "-MT",
   "/openmp:llvm", "-Ic:/tools/cuda/...", ...,
   "C:/Users/.../sccache_nvccXXXXX\\x_0.cudafe1.cpp"]

The bracketed file list contains exactly one source file.

Bisection

Four -Xcompiler flag combos against the same hello-world .cu:

  ┌──────┬──────────────────────────────┬──────┬──────────────────────┐
  │ Case │       -Xcompiler flags       │ Exit │ Cannot cache emitted │
  ├──────┼──────────────────────────────┼──────┼──────────────────────┤
  │ 1    │ -MT, "/openmp:llvm", /bigobj │ 1    │ yes                  │
  ├──────┼──────────────────────────────┼──────┼──────────────────────┤
  │ 2    │ -MT, "/openmp:llvm"          │ 1    │ yes                  │
  ├──────┼──────────────────────────────┼──────┼──────────────────────┤
  │ 3    │ -MT, /bigobj                 │ 0    │ no                   │
  ├──────┼──────────────────────────────┼──────┼──────────────────────┤
  │ 4    │ -MT only (control)           │ 0    │ no                   │
  └──────┴──────────────────────────────┴──────┴──────────────────────┘

/openmp:llvm alone is sufficient to trigger; /bigobj is innocent.

Hypothesis

The MSVC argument parser appears to split /openmp:llvm on ":" and treat llvm
as a positional argument.

Impact

CUDA projects on Windows that pass -Xcompiler /openmp:llvm cannot be built
through sccache today. Not sure whether Linux is impacted.