libmir
2 x slower with LLVM 4.X - 5.X
- gemm micro kernel, simple
- gemm micro kernel, generic
- LLVM 3.9, broadwell, assembler output
- LLVM 4.0, broadwell, assembler output
- LLVM 4.0, avx512f, assembler output
LLVM 4.0, avx512f has the same issue as LLVM 4.0, broadwell.
LDC - the LLVM D compiler (1.1.1):
based on DMD v2.071.2 and LLVM 4.0.0
built with LDC - the LLVM D compiler (0.17.3)
Default target: x86_64-apple-darwin16.4.0
Host CPU: haswell
http://dlang.org - http://wiki.dlang.org/LDC
vs
LDC - the LLVM D compiler (1.1.1):
based on DMD v2.071.2 and LLVM 3.9.1
built with LDC - the LLVM D compiler (1.1.1)
Default target: x86_64-apple-darwin16.4.0
Host CPU: haswell
http://dlang.org - http://wiki.dlang.org/LDC
Could you spend some time on minimizing this? Thanks!
Done, please see gists list in the issue.
@9il: What would probably be most useful is to generate the LLVM IR for the kernel function (-output-ll) and post it at bugs.llvm.org together with the assembler listings, possibly CCing some of the people who have worked on x86 codegen before (if you happen to know someone). Am I correct in the assumption that simple is just a manually expanded version of generic?
Am I correct in the assumption that simple is just a manually expanded version of generic?
Yes.
Okay, thanks. If there are any issues with the account, I can also make the report for you.
Any update on this? And does LLVM 4.0 only decrease performance on your code or are there improvements as well? We'll have to decide on the compiler to use to ship the pre-built 1.2 binaries soon.
Only LLVM 4.0 has no bad AVX512 support. Feel free to update compiler to the LLVM 4.0. GLAS is not completely implemented for BLAS API. I will fill the issue and send link to the LDC gitter
@9il: See Gitter – you might need to turn off detection of LLD headers explicitly with LLVM 5.0 until this is fixed.