Michal Babej

I've rebased this on master, did some changes (i've changed it to report fine-grained SVM, to support coarse-grained only it'd have to implement `svm_map` & `svm_unmap` device ops), and did...

because libpocl needs to be linked to the target environment's libraries (which ones, depends on your build configuration).

Encountered this issue with LLVM 16 on ARM 64 (raspberry pi 4). Seems to be reliably reproducible. The backtrace is almost identical: ``` (gdb) bt #0 0x0000ffffb6695f18 in llvm::LexicalScopes::initialize (this=this@entry=0xffffeb56c620,...

`LLVM_ENABLE_THREADS=OFF` did not make any difference.

Related issue: SPIR-V input fails certain vector types. E.g. this kernel (one of CTS printf tests): ``` __kernel void test10(void) { ushort2 tmp = (ushort2)(0x1234,0x8765); printf("%#v2hx\n",tmp); } ``` when compiled...

AFAICT the incorrect CPU identification is a [bug in LLVM.](https://lists.llvm.org/pipermail/llvm-dev/2015-October/091667.html) The `llc` tool identifies the CPU incorrectly, which is why the build is failing.

Thanks, yeah this is aborting in PoCL. I'm a bit surprised it did not trigger an error much earlier in the compilation pipeline, it aborts in the final step.

Renamed the user-facing strings to 'cpu' for pthread and 'cpu-minimal' for basic; the source file renaming can happen after we have merged everything in the backlog.

IIRC I've seen this warning also on 128bit vector CPUs using 256bit (and larger) vectors. My understanding is that this warning is generated when source code uses Clang's "extended vector"...

Perhaps [the documentation](https://github.com/pocl/pocl/blob/release_5_0/doc/sphinx/source/cuda.rst) is unclear: "image types and samplers are unimplemented" means there is no image support *at all* in the CUDA driver. Actually it's also in the [report](https://opencl.gpuinfo.org/compare.php?reports=2745,2754#deviceinfo) :...