AMD Still Planning For Zen Scheduler In LLVM 5.0

[phoronix]

Phoronix: AMD Still Planning For Zen Scheduler In LLVM 5.0

Back in March we reported about LLVM 5.0 anticipating better AMD Ryzen performance thanks to a proper scheduling model slated to land. But months later, AMD still hasn't produced the code...

http://www.phoronix.com/scan.php?page=news_item&px=LLVM-5-Scheduler-Plans-Zen

[andrei_me]

What is the GCC situation related to Zen scheduler?

[oleid]

Just to be clear : we are only talking about compile time here, correct? This does not affect generic tuning?

[hubicka]

AMD contributed one to GCC in 2015 https://gcc.gnu.org/viewcvs/gcc/trun...i386/znver1.md

Honza

[seek]

oleid
The post is talking about Instruction Scheduling, which affects sequence of instructions generated by the compiler. It can have quite a big influence on the performance of generated code.

As far as I know the zenver1 architechture for GCC is an early version that still requires similar improvements to the instruction scheduling. Was there any progress since the original contribution to gcc?

[discordian]

Just to be clear : we are only talking about compile time here, correct? This does not affect generic tuning?

No, it affects the generated code. Possibly only if you specify Zen as target architecture.

[carewolf]

AMD contributed one to GCC in 2015 https://gcc.gnu.org/viewcvs/gcc/trun...i386/znver1.md

Honza

Yeah, but apparently it isn't very good (if I remember correctly this one is based on the bulldozer scheduling), and Zen chips run the same or better with binaries optimized for Haswell.

[Soul_keeper]

llvm and gcc tend to get the support pretty much at the same time.
The file with the instruction cost models and other files seem to be near identical between both compilers.