https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97445
--- Comment #37 from Jan Hubicka <hubicka at ucw dot cz> --- Hi, this implements the heuristics increasing bounds for functions having __builtin_constant_p on parameters. Note that for get_order this is still not enough, since we increase the bound twice if hint applies, so it goes from 70 to 140 and not to 190 needed, however it will handle ohter similar cases. If hint weight is increased to 300%, so 210 I get: hubicka@lomikamen-jh:/aux/hubicka/trunk-git/build10/gcc$ ./xgcc -B ./ -O2 -Winline pipe.i --param inline-heuristics-hint-percent=300 In file included from fs/pipe.c:11: ./include/linux/slab.h: In function ‘alloc_pipe_info’: ./include/linux/slab.h:586:121: warning: inlining failed in call to ‘kmalloc_array.constprop’: --param max-inline-insns-single limit reached [-Winline] ./include/linux/slab.h:605:9: note: called from here ./include/linux/slab.h: In function ‘pipe_resize_ring’: ./include/linux/slab.h:586:121: warning: inlining failed in call to ‘kmalloc_array.constprop’: --param max-inline-insns-single limit reached [-Winline] ./include/linux/slab.h:605:9: note: called from here So the problem only shifts to not inlininig kmalloc_array. (that is why it would be nice to update kernel with the easier get_order) However it shows different problem: ipa-cp produces cone of kmalloc_array since it is always used by constant size, but the clone does not update the predicates, so we lose track about the parameter being constant and that is why we optimize out only late. Martin, I think this is caused by long lasting TODO in ipa_fn_summary_t::duplicate and probably we should implement it: based on the known partial assignment of params to constant we should fold the conditions in predicates. Indeed with ./xgcc -B ./ -O2 -Winline pipe.i -fno-ipa-cp --param inline-heuristics-hint-percent=300 the warning goes away. We still need the stronger hint though.
