[llvm-bugs] [Bug 41597] New: Certain CUDA codes produce "invalid device function" - appears to be fixed in trunk

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=41597

            Bug ID: 41597
           Summary: Certain CUDA codes produce "invalid device function" -
                    appears to be fixed in trunk
           Product: clang
           Version: 8.0
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: -New Bugs
          Assignee: unassignedclangbugs at nondot.org
          Reporter: philip.salzmann at uibk.ac.at
                CC: htmldeveloper at gmail.com, llvm-bugs at lists.llvm.org,
                    neeilans at live.com, richard-llvm at metafoo.co.uk

For certain CUDA codes Clang 8 will produce an executable that causes the error

> cudaErrorInvalidDeviceFunction (error 8) due to "invalid device function" on CUDA API call to cudaLaunch.

It's difficult to pinpoint exactly what conditions cause this behavior, as it
strikes seemingly "at random". I've encountered this in the context of hipSYCL,
a SYCL implementation based on HIP (however the code is being compiled as CUDA
in this case). Here you can find the original issue with a smallish demo:
https://github.com/illuhad/hipSYCL/issues/49. Of course this unfortunately
includes all of the added complexity surrounding hipSYCL.

This has so far been reproduced using Clang 8 and CUDA 9.2 on Arch, as well as
using Clang 8 and CUDA 10.0 on Ubuntu.

I'd like to provide a self-contained demo, however compiling the test case with
`-save-temps` produces two files, one `*-cuda-nvptx64-nvidia-cuda-sm_52.cui`
and another `-host-x86_64-pc-linux-gnu.cui`. These files are rather large, and
I'm not sure if there is a way of feeding them both back into Clang to be able
to reduce their size with delta. Any advice is welcome.

The good news: The issue appears to have been fixed in trunk, and I've narrowed
the fix down to https://reviews.llvm.org/D58163. That being said, the assertion
at the top of `CGNVCUDARuntime::emitDeviceStub` still fails (i.e., the demo
works when compiled with a release build), which is not surprising as the
commit was meant to address something seemingly unrelated.

The assertion in question is

> llvm-project/clang/lib/CodeGen/CGCUDANV.cpp:228: virtual void {anonymous}::CGNVCUDARuntime::emitDeviceStub(clang::CodeGen::CodeGenFunction&, clang::CodeGen::FunctionArgList&): Assertion `getDeviceSideName(CGF.CurFuncDecl) == CGF.CurFn->getName() || getDeviceSideName(CGF.CurFuncDecl) + ".stub" == CGF.CurFn->getName() || CGF.CGM.getContext().getTargetInfo().getCXXABI() != CGF.CGM.getContext().getAuxTargetInfo()->getCXXABI()' failed.

and it fails in the degenerate case (see GitHub issue) because the mangled
names differ right after the $:

> _ZN2cl4sycl6detail8dispatch19parallel_for_kernelILi1EZZ4mainENK3$_0clERNS0_7handlerEEUlNS0_4itemILi1ELb1EEEE_EEvT0_NS0_5rangeIXT_EEE
> _ZN2cl4sycl6detail8dispatch19parallel_for_kernelILi1EZZ4mainENK3$_1clERNS0_7handlerEEUlNS0_4itemILi1ELb1EEEE_EEvT0_NS0_5rangeIXT_EEE

whereas for the working example they are the same.

Please let me know if there are any additional steps I can take to provide more
context.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20190425/8947dc0e/attachment.html>