> On Tue, Sep 2, 2025 at 11:22 PM Michal Jireš <mji...@suse.cz> wrote: > > > > On 8/28/25 7:40 PM, Andi Kleen wrote: > > > > > > Just 1:1 is slow, isn't it? > > > > > > > > > Surprisingly not by much: > > > > rm vmlinux.o; time make -j16 vmlinux > > > > # -flto-partition=1to1 > > ________________________________________________________ > > Executed in 235.38 secs fish external > > usr time 37.97 mins 0.00 millis 37.97 mins > > sys time 2.12 mins 1.69 millis 2.12 mins > > > > # -flto-partition=cache > > ________________________________________________________ > > Executed in 231.91 secs fish external > > usr time 36.71 mins 0.40 millis 36.71 mins > > sys time 1.12 mins 1.12 millis 1.12 mins > > What about -flto-partition=balanced though? I'd be also interested in > the temporary space requirements for the LTRANS IR files and object files.
1:1 used to be terrible but really got better over time since a lot of work went into making WPA->ltrans streaming more scalable by streaming less contxt. linking cc1plus with balanced (and release checking) on CPU with 256 hyperthreads gets me: real 0m54.644s user 17m46.968s sys 0m13.082s While 1to1: real 1m4.690s user 23m13.399s sys 0m32.854s and max: real 1m33.235s user 43m26.934s sys 156m1.788s tmp file usage is comparable too 2750420 balanced 2856968 1to1 6253232 max I am not quite sure why user time grows considerably, I think it is simply because more lto1 invocations are needed (786 lto1, 127 balanced and 67846 max). So it is a collateral damage of starting compiler. Which also reminds me of increasing default number of partitions for modern CPUs. Overall getting initial version of kernel LTO with -flto-partition=1to1 as first step is IMO acceptable. Honza
