在使用 Clang 编译 CUDA 文件时无法检测 flags

[OpportunityLiu]

Xmake 版本

3.0.5+20251129

操作系统版本和架构

ubuntu 20.04

描述问题

使用 clang-18 / CUDA-12.9

所有 flags 检测失败，单独运行检测指令输出如下：

root@38dcb593c78d:/app/app/app# clang "-fdiagnostics-color=always" "-Qunused-arguments" "--cuda-path=/usr/local/cuda" src/main.cu -o main.o -c -S
clang: warning: CUDA version is newer than the latest supported version 12.3 [-Wunknown-cuda-version]
clang: error: cannot specify -o when generating multiple output files
clang: error: GPU arch sm_35 is supported by CUDA versions between 7.0 and 11.8 (inclusive), but installation at /usr/local/cuda is ; use '--cuda-path' to specify a different CUDA install, pass a different GPU arch with '--cuda-gpu-arch', or pass '--no-cuda-version-check'

期待的结果

可以正常检测

考虑检测 *.cu 时移除 -S 并为 CUDA>=12 添加 --cuda-gpu-arch=sm_75

工程配置

xmake create -l cuda 创建的默认工程，修改 xmake.lua 使用 clang 编译

add_rules("mode.debug", "mode.release")
set_toolchains("clang")
set_toolset("cu", "clang")

target("app")
    set_kind("binary")
    add_files("src/*.cu")

    -- generate SASS code for SM architecture of current host
    add_cugencodes("native")

    -- generate PTX code for the virtual architecture to guarantee compatibility
    add_cugencodes("compute_35")

    -- -- generate SASS code for each SM architecture
    -- add_cugencodes("sm_35", "sm_37", "sm_50", "sm_52", "sm_60", "sm_61", "sm_70", "sm_75")

    -- -- generate PTX code from the highest SM architecture to guarantee forward-compatibility
    -- add_cugencodes("compute_75")

附加信息和错误日志

root@38dcb593c78d:/app/app/app# xmake -rvD -P .
checking for platform ... linux
checking for architecture ... x86_64
checking for gcc ... /usr/bin/gcc
checkinfo: cannot runv(nim --version), No such file or directory
checking for nim ... no
checking for Cuda SDK directory ... /usr/local/cuda
checking for clang ... /usr/bin/clang
checking for nvcc ... /usr/local/cuda/bin/nvcc
checking for cuda devices
checking for flags (-fPIC) ... no
> clang "-fPIC" "-Qunused-arguments" "--cuda-path=/usr/local/cuda"
checkinfo: @programdir/core/sandbox/modules/os.lua:273: clang: warning: CUDA version is newer than the latest supported version 12.3 [-Wunknown-cuda-version]
clang: error: cannot specify -o when generating multiple output files
clang: error: GPU arch sm_35 is supported by CUDA versions between 7.0 and 11.8 (inclusive), but installation at /usr/local/cuda is ; use '--cuda-path' to specify a different CUDA install, pass a different GPU arch with '--cuda-gpu-arch', or pass '--no-cuda-version-check'

stack traceback:
    [C]: in function 'error'
    [@programdir/core/base/os.lua:1125]:
    [@programdir/core/sandbox/modules/os.lua:273]: in function 'runv'
    [@programdir/modules/core/tools/gcc/has_flags.lua:43]:
checking for flags (-O3) ... ok
> clang "-O3" "-Qunused-arguments" "--cuda-path=/usr/local/cuda"
checking for flags (--cuda-gpu-arch=sm_35) ... no
> clang "--cuda-gpu-arch=sm_35" "-Qunused-arguments" "--cuda-path=/usr/local/cuda"
checkinfo: @programdir/core/sandbox/modules/os.lua:273: clang: warning: CUDA version is newer than the latest supported version 12.3 [-Wunknown-cuda-version]
clang: error: cannot specify -o when generating multiple output files
clang: error: GPU arch sm_35 is supported by CUDA versions between 7.0 and 11.8 (inclusive), but installation at /usr/local/cuda is ; use '--cuda-path' to specify a different CUDA install, pass a different GPU arch with '--cuda-gpu-arch', or pass '--no-cuda-version-check'

stack traceback:
    [C]: in function 'error'
    [@programdir/core/base/os.lua:1125]:
    [@programdir/core/sandbox/modules/os.lua:273]: in function 'runv'
    [@programdir/modules/core/tools/gcc/has_flags.lua:43]:
checking for flags (-DNDEBUG) ... ok
> clang "-DNDEBUG" "-Qunused-arguments" "--cuda-path=/usr/local/cuda"
[ 19%]: compiling.release src/main.cu
clang -c -Qunused-arguments --cuda-path=/usr/local/cuda -fvisibility=hidden -O3 -I/usr/local/cuda/include -m64 -fcuda-rdc -DNDEBUG -o build/.objs/app/linux/x86_64/release/src/main.cu.o src/main.cu
checking for flags (-MMD -MF) ... ok
> clang "-MMD" "-MF" "/dev/null" "-Qunused-arguments" "--cuda-path=/usr/local/cuda"
checking for flags (-fdiagnostics-color=always) ... no
> clang "-fdiagnostics-color=always" "-Qunused-arguments" "--cuda-path=/usr/local/cuda"
checkinfo: @programdir/core/sandbox/modules/os.lua:273: clang: warning: CUDA version is newer than the latest supported version 12.3 [-Wunknown-cuda-version]
clang: error: cannot specify -o when generating multiple output files
clang: error: GPU arch sm_35 is supported by CUDA versions between 7.0 and 11.8 (inclusive), but installation at /usr/local/cuda is ; use '--cuda-path' to specify a different CUDA install, pass a different GPU arch with '--cuda-gpu-arch', or pass '--no-cuda-version-check'

stack traceback:
    [C]: in function 'error'
    [@programdir/core/base/os.lua:1125]:
    [@programdir/core/sandbox/modules/os.lua:273]: in function 'runv'
    [@programdir/modules/core/tools/gcc/has_flags.lua:43]:
checking for flags (-fcolor-diagnostics) ... ok
> clang "-fcolor-diagnostics" "-Qunused-arguments" "--cuda-path=/usr/local/cuda"
error: @programdir/core/main.lua:274: @programdir/actions/build/main.lua:161: @programdir/modules/async/runjobs.lua:261: @programdir/modules/private/action/build/object.lua:100: @programdir/modules/core/tools/gcc.lua:1047: clang: warning: CUDA version is newer than the latest supported version 12.3 [-Wunknown-cuda-version]
clang: error: GPU arch sm_35 is supported by CUDA versions between 7.0 and 11.8 (inclusive), but installation at /usr/local/cuda is ; use '--cuda-path' to specify a different CUDA install, pass a different GPU arch with '--cuda-gpu-arch', or pass '--no-cuda-version-check'
stack traceback:
    [C]: in function 'error'
    [@programdir/core/base/os.lua:1125]:
    [@programdir/modules/core/tools/gcc.lua:1047]: in function 'catch'
    [@programdir/core/sandbox/modules/try.lua:123]: in function 'try'
    [@programdir/modules/core/tools/gcc.lua:988]:
    [C]: in function 'xpcall'
    [@programdir/core/base/utils.lua:260]:
    [@programdir/core/tool/compiler.lua:285]: in function 'compile'
    [@programdir/modules/private/action/build/object.lua:100]: in function 'script'
    [@programdir/modules/private/action/build/object.lua:131]: in function 'build_object'
    [@programdir/modules/private/action/build/object.lua:171]: in function 'job_func'
    [@programdir/modules/async/runjobs.lua:432]:
    [C]: in function 'xpcall'
    [@programdir/core/base/utils.lua:260]: in function 'trycall'
    [@programdir/core/sandbox/modules/try.lua:117]: in function 'try'
    [@programdir/modules/async/runjobs.lua:401]: in function 'cotask'
    [@programdir/core/base/scheduler.lua:514]:

stack traceback:
        [C]: in function 'error'
        @programdir/core/base/os.lua:1125: in function 'base/os.raiselevel'
        (...tail calls...)
        @programdir/core/main.lua:274: in upvalue 'cotask'
        @programdir/core/base/scheduler.lua:514: in function <@programdir/core/base/scheduler.lua:507>
warning: failed to find cuda devices: cudaErrorInsufficientDriver (CUDA driver version is insufficient for CUDA runtime version)
warning: add_cuflags("--cuda-gpu-arch=sm_35") is ignored, please pass `{force = true}` or call `set_policy("check.auto_ignore_flags", false)` if you want to set it.

[Issues-translate-bot]

Bot detected the issue body's language is not English, translate it automatically.

---

Title: Cannot detect flags when compiling CUDA files with Clang

[waruqi]

目前 clang 没有对 cuda flags 做特别的支持，走的 gcc 常规的 c/c++ 检测逻辑。 https://github.com/xmake-io/xmake/blob/dev/xmake/modules/core/tools/gcc/has_flags.lua

我现在暂时没环境调试这个。你可以配置 check.auto_ignore_flags 去禁用自动 flags 检测，或者提个 pr 过来，直接改进相关的 flags 检测。

[Issues-translate-bot]

Bot detected the issue body's language is not English, translate it automatically.

---

Currently, clang does not provide special support for cuda flags, and uses gcc's regular c/c++ detection logic. https://github.com/xmake-io/xmake/blob/dev/xmake/modules/core/tools/gcc/has_flags.lua

I don't have the environment to debug this right now. You can configure check.auto_ignore_flags to disable automatic flags detection, or submit a PR to directly improve related flags detection.

[OpportunityLiu]

我是把 clang 分出来，还是直接在 gcc 上面改？

[Issues-translate-bot]

Bot detected the issue body's language is not English, translate it automatically.

---

Should I separate clang, or change it directly on gcc?

[waruqi]

我是把 clang 分出来，还是直接在 gcc 上面改？

如果改动不大，就直接 gcc 里面判断下 cuda 做下处理，如果改动大，就分出来吧。

[Issues-translate-bot]

Bot detected the issue body's language is not English, translate it automatically.

---

Should I separate clang, or change it directly on gcc?

If the change is not big, just use gcc to check cuda for processing. If the change is big, separate it out.