> Hi, > this patch adds znver2 scheduler model. Znver2 is close enough to znver1 > that I have decided to implement it in one automaton. The main > difference is extra AGU unit that seems to be used for store only > (according to CPU diagram), the fact that 256bit vector operations are > no longer split (and thus they behave like 128bit) and reduced latency > of fp multiply and conversion operations. > > The patch seems to have very little effect on overall performance but > since we do not model the out of order core and thus we think that the > CPU is mostly starved by not having enough parallelism to exectue. > Still it is better to be precise. > > Bootstrapped/regtested x86_64-linux, commited. > > * common/config/i386/i386-common.c: Use PROCESSOR_ZNVER2 scheduler for > znver2. > * config/i386/znver1.md: Enable patterns for znver2 and add store > variants which use extra AGU unit.
Hello, I have backported this patch to gcc 9 branch now too. Honza