> > > > #define REG_FREQ_FROM_BB(bb) ((optimize_function_for_size_p (cfun) > > > > \ > > > > || !cfun->cfg->count_max.initialized_p > > > > ()) \ > > > > ? REG_FREQ_MAX > > > > \ > > > > : ((bb)->count.to_frequency (cfun) > > > > \ > > > > * REG_FREQ_MAX / BB_FREQ_MAX) > > > > \ > > > > ? ((bb)->count.to_frequency (cfun) > > > > \ > > > > * REG_FREQ_MAX / BB_FREQ_MAX) > > > > \ > > > > : 1) > > > > > > > > 1000 is the default. If it isn't 1000, it isn't the default. I only > > > > want > > > > to get a more reasonable default scale, instead of 1000. Lower > > > > scale will fail the PR rtl-optimization/111673 test on powerpc64. > > > > > > I see. Why not adjust the above macro then? That would be a bit more > > > obvious. Like use MAX/2 or so? > > > > commit 3b9b8d6cfdf59337f4b7ce10ce92a98044b2657b > > Author: Surya Kumari Jangala <jskum...@linux.ibm.com> > > Date: Tue Jun 25 08:37:49 2024 -0500 > > > > ira: Scale save/restore costs of callee save registers with block > > frequency > > > > uses REG_FREQ_FROM_BB as the cost scale. I don't know if it is a misuse. > > I don't want to change REG_FREQ_FROM_BB since it is used in other places, > > not as a cost scale. Maybe the above commit should be reverted and we add > > a target hook for callee-saved register cost scale. Each target can choose > > a proper cost scale, install of increasing the cost by 1000x for everyone. > > I believe testing cfun->cfg->count_max.initialized_p () is a bit odd at least, > as it doesn't seem to be used. The comment talks about profile feedback,
It is used by count.to_frequency, which basically computes count/max_count * REG_FREQ_MAX. It aborts if max_count is uninitialized rather than returning arbitrary value... Honza
