When object files containing LTO sections are linked together with ld -r the linker merges all sections with the same name. The LTO reader cannot deal with this. Currently it will just ignore excessive data behind the first entry for most sections, which leads to all kinds of errors. I have been working on a patch kit to solve this, this is just a tracking bug for it.
And partial linking support will break mixed LTO / non-LTO objects. Unless we drop all non-LTO sections from LTO objects and thus the .text sections of partially linked mixed LTO / non-LTO objects will be still used in the final link.
Not sure I understand the comment. The case I've been looking at is ld -r (without a LTO code generation stage) to combine existing object and then using gold for the final linking/LTO code generation based on the combined objects. For me it seems like gold handles this correctly (with my patches) Your scenario sounds like ld -r with code generation? That's probably harder.
I must say I don't like your solution. IMHO much better is instead add a header to LTO sections, which says the length of the LTO chunk (similarly e.g. to how .debug_info section chunks have length in the header), perhaps version, followed by that length of bytes of compressed stream, and change LTO reader to iterate over all chunks in the LTO section. This will also have advantage that it doesn't require non-ELF object formats to have section length not a multiple of 4, etc.
This is actually what I tried first, but it turned out to be quite complicated, had to change a lot of code and my patch was growing and growing and it didn't fit clearly with the different readers etc. That is why I ended up with the post fixes instead which is much simpler and isolated in a single part of the reader. I don't understand the comment about section lengths? There's no dependency on section lengths being a multiple of 4
I was refering to a situation like gcc -c -flto t1.c gcc -c t2.c gcc -o t.o -r -nostdlib t1.o t2.o [-flto] gcc -o t t.o -flto which would break with your solution (it's broken right now as well, of course). We could make it work by not emitting code but only LTO sections for t1.o. Then the partial link would cause us to have both LTO sections and regular sections, so at link-time we'd pick up the LTO parts for re-optimization and link the code parts without re-optimization. We'd still need a way to merge LTO sections (or mangle them like you do), of course.
Subject: Bug 44992 Author: ak Date: Fri Jul 23 05:33:51 2010 New Revision: 162443 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=162443 Log: gcc: 2010-07-10 Andi Kleen <ak@linux.intel.com> PR lto/44992 * lto-opts.c (lto_write_options): Add NULL file_data argument to lto_get_section_name. * lto-section-out.c (lto_destroy_simple_output_block): Likewise. * lto-streamer-out.c (produce_asm): Likewise. (copy_function): Likewise. (produce_symtab): Likewise. (produce_asm_for_decls): Likewise. * lto-streamer.c (lto_get_section_name): Add file_data argument. Rewrite to add random postfix to LTO sections. * lto-streamer.h (lto_file_decl_data): Add next, id, resolutions. (lto_get_section_name): Add file_data argument to prototype. lto: 2010-07-10 Andi Kleen <ak@linux.intel.com> PR lto/44992 * lto.c: Include splay-tree.h (lto_resolution_read): Change to walk file_ids tree and parse extra file_id in resolution file. (lto_section_with_id): Add. (create_subid_section_table): Add. (lwstate): Add. (lto_create_files_from_ids): Add. (lto_file_read): Change to handle sub file ids and create list of file_datas. Add output argument for count. (get_section_data): Pass file_data to lto_get_section_name. (lto_flatten_file): Add. (read_cgraph_and_symbols): Handle linked lists of file_datas. lto-plugin: 2010-07-10 Andi Kleen <ak@linux.intel.com> PR lto/44992 * lto-plugin.c (sym_aux): Add. (plugin_symtab): Remove slots. Add aux and id. (parse_table_entry): Change to use aux instead of slots. (LTO_SECTION_PREFIX): Add. (translate): Improve buffer allocation. Change to append symbols to existing out buffer. (get_section): Remove. (process_symtab): Add. (free_2): Free symtab->aux. (write_resolution): Handle aux instead of slots. Print sub id to resolution file. (claim_file_handler): Clear lto_file. Replace get_symtab/translate calls with call to process_symtab. Modified: trunk/gcc/ChangeLog trunk/gcc/lto-opts.c trunk/gcc/lto-section-out.c trunk/gcc/lto-streamer-out.c trunk/gcc/lto-streamer.c trunk/gcc/lto-streamer.h trunk/gcc/lto/ChangeLog trunk/gcc/lto/lto.c trunk/lto-plugin/ChangeLog trunk/lto-plugin/lto-plugin.c
Fixed.
Sorry this is not fixed yet, only partially. Still working on the last bits, in particular passthrough of non LTOed code like assembler functions.
GCC 4.6.0 is being released, adjusting target milestone.
I consider this fixed now because everything I need works together with HJ's binutils. Mainstream binutils will hopefull catch up eventually.
*** Bug 43576 has been marked as a duplicate of this bug. ***