clang generates ill-formed PTX (parsing error with ptxas)

[ptal]

Compiling my code in Debug mode leads to a parse error in the PTX. Compiling in mode Release works but crashes at the execution (while nvcc doesn't), perhaps due to the limited support of CUDA 12.1, I have not investigated.

Replicate debug issue:

$ git clone git@github.com:ptal/turbo.git
$ git checkout clang-bug64365
$ cmake -DCMAKE_BUILD_TYPE=Debug -DGPU=ON -Bbuild/gpu-debug-clang -DCMAKE_CUDA_ARCHITECTURES=75 -DCMAKE_CUDA_COMPILER=clang++-18 -DCMAKE_CXX_COMPILER=clang++-18 -DCMAKE_VERBOSE_MAKEFILE=ON -DREDUCE_PTX_SIZE=ON
$ cmake --build build/gpu-debug-clang/
...
ptxas /tmp/turbo-sm_75-e1f897.s, line 5502094; fatal   : Parsing error near '.': syntax error
ptxas fatal   : Ptx assembly aborted due to errors
clang++-18: error: ptxas command failed with exit code 255 (use -v to see invocation)
Ubuntu clang version 18.0.0 (++20230805031655+ce2ec065165b-1~exp1~20230805151815.1104)
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin
clang++-18: note: diagnostic msg: 
********************

PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:
Preprocessed source(s) and associated run script(s) are located at:
clang++-18: note: diagnostic msg: /tmp/turbo-dcd464.cu
clang++-18: note: diagnostic msg: /tmp/turbo-sm_75-343cfe.cu
clang++-18: note: diagnostic msg: /tmp/turbo-dcd464.sh
clang++-18: note: diagnostic msg: 

********************

[arsenm]

Needs reproducer attached. If I were to guess this is the new alias support?

[ptal]

Here you go, but I tried to run the .sh and it says a file is missing.

turbo-reproducer.zip

$ chmod +x turbo-703b62.sh 
$ ./turbo-703b62.sh 
ptxas fatal   : Input file '/tmp/turbo-sm_75-e1e5bf.s' could not be opened
ptxas fatal   : Ptx assembly aborted due to errors

[amitbagate]

@ptal have u received solution to reported issue yet?

[ptal]

Not really although not blocking for me, I'm using nvcc instead.

[rbberger]

This is probably a duplicate of #58491 which was fixed in 20.1.0.