Crash when running pass 'Adapt OCL types for SPIR-V' on module 'builtins.link.spirv64-mesa3d-.bc'

[illwieckz]

I was compiling libclc from commit 5bd8735, llvm-spirv was also built using recent commit KhronosGroup/SPIRV-LLVM-Translator@55d1de8 .

llvm-spirv: include/llvm/IR/Type.h:384: llvm::Type *llvm::Type::getNonOpaquePointerElementType() const: Assertion `NumContainedTys && "Attempting to get element type of opaque pointer"' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.	Program arguments: llvm-spirv --spirv-max-version=1.1 -o spirv64-mesa3d-.spv builtins.link.spirv64-mesa3d-.bc
1.	Running pass 'Adapt OCL types for SPIR-V' on module 'builtins.link.spirv64-mesa3d-.bc'.
 #0 0x000000000102b6ba llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) llvm-project/llvm/lib/Support/Unix/Signals.inc:565:11
 #1 0x000000000102b86b PrintStackTraceSignalHandler(void*) llvm-project/llvm/lib/Support/Unix/Signals.inc:632:1
 #2 0x0000000001029f06 llvm::sys::RunSignalHandlers() llvm-project/llvm/lib/Support/Signals.cpp:103:5
 #3 0x000000000102bfa5 SignalHandler(int) llvm-project/llvm/lib/Support/Unix/Signals.inc:407:1
 #4 0x00007fe00f595520 (/lib/x86_64-linux-gnu/libc.so.6+0x42520)
 #5 0x00007fe00f5e9828 __pthread_kill_implementation ./nptl/pthread_kill.c:44:76
 #6 0x00007fe00f5e9828 __pthread_kill_internal ./nptl/pthread_kill.c:80:10
 #7 0x00007fe00f5e9828 pthread_kill ./nptl/pthread_kill.c:91:10
 #8 0x00007fe00f595476 gsignal ./signal/../sysdeps/posix/raise.c:27:6
 #9 0x00007fe00f57b7b7 abort ./stdlib/abort.c:81:7
#10 0x00007fe00f57b6db get_sysdep_segment_value ./intl/loadmsgcat.c:509:8
#11 0x00007fe00f57b6db _nl_load_domain ./intl/loadmsgcat.c:970:34
#12 0x00007fe00f58ce26 (/lib/x86_64-linux-gnu/libc.so.6+0x39e26)
#13 0x000000000045f0b7 llvm::Type::getNonOpaquePointerElementType() const llvm-project/llvm/include/llvm/IR/Type.h:0:5
#14 0x0000000000507395 llvm::Type::getPointerElementType() const (llvm-spirv+0x507395)
#15 0x000000000052ea9d SPIRV::isPointerToOpaqueStructType(llvm::Type*) (llvm-spirv+0x52ea9d)
#16 0x00000000006e380d SPIRV::OCLTypeToSPIRVBase::adaptFunctionArguments(llvm::Function*) (llvm-spirv+0x6e380d)
#17 0x00000000006e3293 SPIRV::OCLTypeToSPIRVBase::runOCLTypeToSPIRV(llvm::Module&) (llvm-spirv+0x6e3293)
#18 0x00000000006e30d1 SPIRV::OCLTypeToSPIRVLegacy::runOnModule(llvm::Module&) (llvm-spirv+0x6e30d1)
#19 0x0000000000e1c6e9 (anonymous namespace)::MPPassManager::runOnModule(llvm::Module&) llvm-project/llvm/lib/IR/LegacyPassManager.cpp:1545:23
#20 0x0000000000e1c25d llvm::legacy::PassManagerImpl::run(llvm::Module&) llvm-project/llvm/lib/IR/LegacyPassManager.cpp:535:16
#21 0x0000000000e20f61 llvm::legacy::PassManager::run(llvm::Module&) llvm-project/llvm/lib/IR/LegacyPassManager.cpp:1672:3
#22 0x0000000000581832 llvm::writeSpirv(llvm::Module*, SPIRV::TranslatorOpts const&, std::ostream&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >&) (llvm-spirv+0x581832)
#23 0x000000000042a69d convertLLVMToSPIRV(SPIRV::TranslatorOpts const&) llvm-spirv.cpp:0:0
#24 0x0000000000428212 main (llvm-spirv+0x428212)
#25 0x00007fe00f57cfd0 __libc_start_call_main ./csu/../sysdeps/nptl/libc_start_call_main.h:58:16
#26 0x00007fe00f57d07d call_init ./csu/../csu/libc-start.c:128:20
#27 0x00007fe00f57d07d __libc_start_main ./csu/../csu/libc-start.c:379:5
#28 0x0000000000424ea5 _start (llvm-spirv+0x424ea5)

[llvmbot]

@llvm/issue-subscribers-mlir-spirv

[joker-eph]

Not MLIR related as far as I can tell

[illwieckz]

Purging build directories of llvm, clang, lib-clc and llvm-spriv and installation directory fixed it, it may be a cmake issue where it doesn't rebuild everything that is needed to be rebuilt on new commits. I won't investigate this more.

[illwieckz]

It appeared the bug disappearing on my end was not caused by building from scratch but by building older LLVM commit to workaround the compat-breaking changes of LLVM turning on opaque pointer by default that broke Mesa, see #54970. So it can be tagged opaque-pointers.

I assume it's now an llvm-spirv issue, see:

• How to deal with opaque pointers in the translator KhronosGroup/SPIRV-LLVM-Translator#1444
• Support opaque pointers KhronosGroup/LLVM-SPIRV-Backend#168

The error is now printed this way:

llvm-spirv: include/llvm/IR/Type.h:384: llvm::Type *llvm::Type::getNonOpaquePointerElementType() const: Assertion `NumContainedTys && "Attempting to get element type of opaque pointer"' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.	Program arguments: bin/llvm-spirv --spirv-max-version=1.1 -o spirv64-mesa3d-.spv builtins.link.spirv64-mesa3d-.bc
1.	Running pass 'Adapt OCL types for SPIR-V' on module 'builtins.link.spirv64-mesa3d-.bc'.
Build libclc: [ 99%] Built target opt.spirv-mesa3d-.bc
Build libclc: [ 99%] Built target prepare-clspv--.bc
Build libclc: [ 99%] Built target prepare-clspv64--.bc
 #0 0x00007f74b67c97e2 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) llvm-project/llvm/lib/Support/Unix/Signals.inc:565:22
 #1 0x00007f74b67c98a9 PrintStackTraceSignalHandler(void*) llvm-project/llvm/lib/Support/Unix/Signals.inc:632:1
 #2 0x00007f74b67c7454 llvm::sys::RunSignalHandlers() llvm-project/llvm/lib/Support/Signals.cpp:103:20
 #3 0x00007f74b67c9113 SignalHandler(int) llvm-project/llvm/lib/Support/Unix/Signals.inc:407:1
 #4 0x00007f74b5011520 (/lib/x86_64-linux-gnu/libc.so.6+0x42520)
 #5 0x00007f74b5065828 __pthread_kill_implementation ./nptl/pthread_kill.c:44:76
 #6 0x00007f74b5065828 __pthread_kill_internal ./nptl/pthread_kill.c:80:10
 #7 0x00007f74b5065828 pthread_kill ./nptl/pthread_kill.c:91:10
 #8 0x00007f74b5011476 gsignal ./signal/../sysdeps/posix/raise.c:27:6
 #9 0x00007f74b4ff77b7 abort ./stdlib/abort.c:81:7
#10 0x00007f74b4ff76db get_sysdep_segment_value ./intl/loadmsgcat.c:509:8
#11 0x00007f74b4ff76db _nl_load_domain ./intl/loadmsgcat.c:970:34
#12 0x00007f74b5008e26 (/lib/x86_64-linux-gnu/libc.so.6+0x39e26)
#13 0x000055f6364f4e97 llvm::Type::getNonOpaquePointerElementType() const (bin/llvm-spirv+0x10be97)
#14 0x000055f636594c65 llvm::Type::getPointerElementType() const (bin/llvm-spirv+0x1abc65)
#15 0x000055f6365b91bd SPIRV::isPointerToOpaqueStructType(llvm::Type*) (bin/llvm-spirv+0x1d01bd)
#16 0x000055f63675e93d SPIRV::OCLTypeToSPIRVBase::adaptFunctionArguments(llvm::Function*) (bin/llvm-spirv+0x37593d)
#17 0x000055f63675e3c3 SPIRV::OCLTypeToSPIRVBase::runOCLTypeToSPIRV(llvm::Module&) (bin/llvm-spirv+0x3753c3)
#18 0x000055f63675e201 SPIRV::OCLTypeToSPIRVLegacy::runOnModule(llvm::Module&) (bin/llvm-spirv+0x375201)
#19 0x00007f74b6abc94d (anonymous namespace)::MPPassManager::runOnModule(llvm::Module&) llvm-project/llvm/lib/IR/LegacyPassManager.cpp:1545:20
#20 0x00007f74b6ab767a llvm::legacy::PassManagerImpl::run(llvm::Module&) llvm-project/llvm/lib/IR/LegacyPassManager.cpp:535:13
#21 0x00007f74b6abd25f llvm::legacy::PassManager::run(llvm::Module&) llvm-project/llvm/lib/IR/LegacyPassManager.cpp:1673:1
#22 0x000055f63660aba2 llvm::writeSpirv(llvm::Module*, SPIRV::TranslatorOpts const&, std::ostream&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >&) (bin/llvm-spirv+0x221ba2)
#23 0x000055f6364c1db5 convertLLVMToSPIRV(SPIRV::TranslatorOpts const&) llvm-spirv.cpp:0:0
#24 0x000055f6364bf942 main (bin/llvm-spirv+0xd6942)
#25 0x00007f74b4ff8fd0 __libc_start_call_main ./csu/../sysdeps/nptl/libc_start_call_main.h:58:16
#26 0x00007f74b4ff907d call_init ./csu/../csu/libc-start.c:128:20
#27 0x00007f74b4ff907d __libc_start_main ./csu/../csu/libc-start.c:379:5
#28 0x000055f6364bc5e5 _start (bin/llvm-spirv+0xd35e5)

[illwieckz]

I reopen the issue because libclc cannot be built with llvm from the same repository using the same commit (through llvm-spirv).

Said otherwise, libclc cannot be built with llvm-spirv built with llvm from the same repository and commit as libclc.

Also llvm-spirv now requires clang 15 to be built, so it needs this repository and this branch.

A workaround or a fix is needed so libclc becomes buildable again, either by implementing a compatibility layer in llvm before llvm-spirv becomes opaque pointer compatible, either by making llvm-spirv opaque pointer compatible (see KhronosGroup/SPIRV-LLVM-Translator#1444 ).

It looks like one can do this:

• build old llvm and clang after 15 branch started but before opaque pointers were turned on by default
• build llvm-spirv with this old llvm and clang
• rebuild llvm on main branch
• build mesa clover with this llvm

but I don't think it's an acceptable workaround to actually need to build two different llvm revisions to build latest libclc from the same repository.

[llvmbot]

@llvm/issue-subscribers-opencl

[AnastasiaStulova]

Do you know on which source file in libclc this failure occurs? My guess is that it will likely be something specific to images or other OpenCL specific types?

[jcranmer-intel]

#15 0x000055f6365b91bd SPIRV::isPointerToOpaqueStructType(llvm::Type*) (bin/llvm-spirv+0x1d01bd)
#16 0x000055f63675e93d SPIRV::OCLTypeToSPIRVBase::adaptFunctionArguments(llvm::Function*) (bin/llvm-spirv+0x37593d)
#17 0x000055f63675e3c3 SPIRV::OCLTypeToSPIRVBase::runOCLTypeToSPIRV(llvm::Module&) (bin/llvm-spirv+0x3753c3)

This specific stack trace is I believe fixed by KhronosGroup/SPIRV-LLVM-Translator#1499, not that it actually finishes the conversion process for SPIR-V emission with opaque pointers (there is another PR I'm working on which should make much more code work, although "much more" is itself a far cry from "all").

What you can probably do to mitigate the damage until the translator is fixed is to compile libclc with opaque pointers disabled.

[nikic]

This issue can probably be closed now?

[illwieckz]

I'll try to reproduce the issue in the coming days to see if it's fixed.

[jcranmer-intel]

At this point, the SPIR-V translator should be capable of handling everything as opaque pointers, save for some of the fancier extensions (e.g., Intel's function pointer extensions) that I'm pretty sure libclc doesn't use.

The SPIR-V backend is I think not yet supporting opaque pointers, but I don't think libclc is even configurable to use the native SPIR-V backend instead of the translator.

[illwieckz]

I'm not able anymore to compile amdcn libclc targets with either llvm 17-git, 16.1.0 and 16.0.0 (I haven't tried older git references) because the build of a dependency of those target fails to build: #62018

So I cannot verify if that error is fixed or not.

[illwieckz]

I fixed the build error preventing to build libclc (see #62018 for the build error and D150871 for the fix), so I'm able to build this again.

I don't experience this bug anymore. Thank you very much!