> On Tue, 4 Oct 2011, Andi Kleen wrote: > > > > Well, I'm not sure we should jump through too much hoops to make > > > -flto work with -fno-toplevel-reorder. Simply because I think nothing > > > defines any "toplevel order" for multiple object files. So, I think > > > > In practice it seems to work because real programs rely on it.
I don't see much jump in undefinedness in between single file and LTO definition of -fno-toplevel-reorder. -fno-toplevel reorder is "please outout functions as they are written in the source file because it is what GCC used to do. Ignore inlines and comdats" at IPA level we just concatenate the order in order how files are passed to linker. Again there is bit of undefinedness in how static libraries and garbage collected sections are handled, but in the common cases people care about it just does the job. > > > > I can just say with this patch and Honza's my program works > > (with -flto-partition=none), without it it doesn't. > > > > Also as Honza pointed out it has other benefits, like making > > compiles more reproducible. For example if you have a memory corruption > > somewhere the random order currently will randomly move it from > > run to run and make it harder to debug. > > > > > both ld and gcc/collect2 need to first document they will never > > > break toplevel order (what about ld with -ffunction-sections and > > > -gc-sections? Or -relax?) > > > > -ffunction-sections and -gc-sections do not reorder as far as I know. > > Not sure about -relax. > > > > But the programs that rely on order should "don't do it when it hurts" > > so not set too magic options. I'm sure there are other creative ways to > > break > > order too, but it doesn't seem to be too hard in practice to avoid > > them. > > Sure, the question is if "-flto" counts as magic and thus > "don't do it when it hurts" ;)) I suppose with -flto-partition=none > (or 1to1) it would be reasonable to make -fno-toplevel-reorder work > (and thus maybe -fno-toplevel-reorder should simply force 1to1 > partitioning). Note that 1to1 partitioning won't give you -fno-toplevel-reorder as currently implemented. It assigns the partitions in random order and then the partitions get sorted by size. The default partitioner is written to honor predefined order. In fact I do have patch to make it honnor -fno-toplevel-reorder from very beggining, just I have never submitted it. it is how my ipa-reorder pass works (it simply decides on "optimal" function order and then assigns them ->order indexes). It does not handle variables and asm statements. Neither is terribly difficult to do so I will post patches for those. Honza > > Richard.
