User Details
- User Since
- Jul 27 2015, 9:08 PM (590 w, 9 h)
Yesterday
I found this issue about two months ago. Thing is, correcting it also lowers the performance in my benchmarks by a couple of %. Can you do some performance testing? The current behavior doesn't cause any errors and if it's faster, IMO we should keep it.
Sun, Nov 15
Sat, Nov 14
Fri, Nov 13
I thought we generally want to upstream patches to LLVM first and only once it's done backport them? If that is not necessary, I have some patches for LLVM that are waiting for upstream review and could be put to src.
Ping.
Thu, Nov 12
Rebased to main from yesterday.
Wed, Nov 11
Tue, Nov 10
Build-tested: lib/clang/libllvm, libclang, usr.bin/clang/clang and opt built from the patched tree (main, so with LLVM_ASSERTIONS on) on 15.2-BETA1 powerpc64le. The resulting clang compiles the csound TU that crashed, while the stock 15.2 clang 21.1.8 still dies on the same preprocessed file with exit code 133 in the slp-vectorizer pass. No assertion fired. opt -passes=slp-vectorizer from that build vectorizes both the reduced reproducer and upstream's regression test into a load <4 x double> (with the llvm.vector.reduce.fminimum.v4f64 the upstream CHECK lines expect) instead of trapping.
