ChuanqiXu9 wrote: > The 10x increase in the time spent reading modules does look surprising and I > would need to check if it's related to PCMs becoming larger or is localized > to that particular compile. (Even if on a single example). It should not related to the size of the PCMs otherwise we will see the regression on X86 too. Another point may be, the proportion is not correct, 25% size increase shouldn't introduce such a big change.
BTW, I am working on a patch to mitigate the size issue. And for the different size increase number, it should be related to the specific code bases. Generally, the more references to decls, the more increases. > > @ChuanqiXu9 do you expect alignment to have such a drastic effect on > deserialization time? That would be very surprising to me. The reproducer > will take time, could we revert in the meantime? I won't say 'expected'. I just met some performance regression in AArch64 related to alignments. So I am just guessing it is related to alignments. And for this particular issue, besides reproducer, it should be pretty helpful to have some hotspots data. It might not take too long. For reverting, how about giving it a few days to locate the problem? https://github.com/llvm/llvm-project/pull/92083 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
