On Wed, Jan 31, 2018 at 12:46 AM, Andreas Beckmann <a...@debian.org> wrote: > On 2018-01-31 01:16, Punit Agrawal wrote: >>> 7/121 Test #22: kernel/test_shuffle_char >>> ............................................................***Failed >>> Required regular expression found.Regex=[OK > >> The attached patch (on top of the package repository) fixes it for me. > > Thanks. > >> While investigating this failure, I noticed an unrelated issue with >> the debian build for arm64 - instead of choosing a generic armv8 as >> the target cpu, the rules file chooses to build with cortex-a53 (value >> picked from debian/supported-arch). Just thought I'd mention it. > > Thanks. Do the shuffle tests still succeed for armv8?
All the tests (except for test_fabs) passed when LLVM_ARCH was set to "generic" on arm64. I did this by modifying debian/supported_archs. > It's damned hard to find the generic default -march= setting from llvm > for each architecture - we have to set it explicitly for pocl, otherwise > it will internally use the equivalent of -march=native. I am not familiar with llvm configuration so could be misunderstanding but I was just worried that choosing "cortex-a53" might assume micro-architectural features that may not be present on other ARMv8 implementations. It's not so bad if the value is used as a hint to choose optimised code sequences. Punit
