On Tue, 11 Feb 2014, Jakub Jelinek wrote: > Hi! > > Right now we spent several minutes in verify_tree. The problems I see > is that merge_tlist does the exact opposite of what should it be doing with > copy flag (most merge_tlist calls are with copy=0, thus that means mostly > unnecessary copying, but for the single one for SAVE_EXPR it actually > probably breaks things or can break), the middle-hunk is about one spot > which IMHO uses copy=1 unnecessarily (nothing uses tmp_list2 afterwards). > > The most important is the last hunk, the SAVE_EXPR handling was doing > merge_tlist first on the whole tmp_nosp chain (with copy=0 which mistakenly > copied everything, see above), and then doing that again with the whole > tmp_nosp list except the first entry, then except the first two entries etc. > O(n^2) complexity, but IMHO none of the calls but the first one actually > would do anything, because after the first merge_tlist call all entries > should be actually in tmp_list3 (except for duplicates), and so for all > further entries it would be finding a matching entry already (found=1). > > With this patch pr60101.c compiles within less than a second. > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK. -- Joseph S. Myers jos...@codesourcery.com
