All opencl kernel fail to build returning a build error about conflicting printf types. Easiest to reproduce is use `clinfo` and only look at stderr output : ``` $ clinfo > /dev/null === CL_PROGRAM_BUILD_LOG === In file included from <built-in>:2: ./CTHeader.h:929:7: error: conflicting types for 'printf' int printf(__constant const char* st, ...); ^ ./opencl-c-base.h:656:5: note: previous declaration is here int printf(__constant const char* st, ...) __attribute__((format(printf, 1, 2))); ``` (clinfo tries to build a small kernel to query some runtime info. But the same issue crops up in any opencl app). I'm using these versions : =dev-util/intel-graphics-compiler-2.10.2 =dev-libs/intel-compute-runtime-25.09.32961.8 (yes, they are masked by default but for another reason because intel dropped support for gen8-11 iGPU and I have a gen12 one which is supported and I need some features that came in recent releases. )
I went back to : dev-util/intel-graphics-compiler-1.0.17791.18:0/1.0.1::gentoo dev-libs/intel-compute-runtime-24.35.30872.32:0/1.5.30872::gentoo and that did _NOT_ solve the issue ... I tried various other version too and same problem. Tried working around the problem by removing the defainition from intel-graphics-compiler and then some CL software would just _hang_ ( gdb shows they're in some LLVM code indefinitely ). If someone has a working version, knowing exaclty which version of these package and their dependency is used would be nice.
Damnit ... problem was a conflict with rusticl (which was loading LLVM19 in the same address space and ... issues). Moving /etc/OpenCL/vendors/rusticl.icd away fixed the issue. Alternatively making it so that this icd is loaded second and not first by the ICD loader also seems to work around the issue. But if rusticl loads first, intel-compute-runtime doesn't work. Sorry for the noise ...
