provide c and c++ compilers for nvptx targets

Author: pietroalbini

src/ci/docker/host-x86_64/test-various/Dockerfile

@@ -60,7 +60,11 @@ ENV WASM_SCRIPT python3 /checkout/x.py --stage 2 test --host='' --target $WASM_T
   tests/assembly \
   library/core
 
-ENV NVPTX_TARGETS=nvptx64-nvidia-cuda
+ENV NVPTX_TARGETS=nvptx64-nvidia-cuda \
+    CC_nvptx64_nvidia_cuda=clang-11 \
+    CXX_nvptx64_nvidia_cuda=clang++-11 \
+    CFLAGS_nvptx64_nvidia_cuda=-flto \
+    CXXFLAGS_nvptx64_nvidia_cuda=-flto \
 ENV NVPTX_SCRIPT python3 /checkout/x.py --stage 2 test --host='' --target $NVPTX_TARGETS \
   tests/run-make \
   tests/assembly

tests/auxiliary/rust_test_helpers.c

@@ -268,6 +268,11 @@ float get_c_exhaust_sysv64_ints(
     return h.c;
 }
 
+// Variadic arguments are broken when using clang to compile nvptx-nvidia-cuda,
+// causing clang to ICE. As no test run on CI currently relies on these
+// functions, we don't compile them when building for nvptx-nvidia-cuda.
+#ifdef __CUDA_ARCH__
+
 // Calculates the average of `(x + y) / n` where x: i64, y: f64. There must be exactly n pairs
 // passed as variadic arguments. There are two versions of this function: the
 // variadic one, and the one that takes a `va_list`.
@@ -290,6 +295,8 @@ double rust_interesting_average(uint64_t n, ...) {
     return sum;
 }
 
+#endif
+
 int32_t rust_int8_to_int32(int8_t x) {
     return (int32_t)x;
 }