Am 09.01.2015 um 12:54 schrieb Jakub Jelinek:
On i?86 it isn't around 10%, but more like 10%-30%.
i doubt that it's always 30% - real workloads matter not worst cases and what are the 100% - if 100% is below a second 30% don't matter - there are millions of tasks where even 50% would not matter
what you also ignore is that most tasks are already in DSO libraries and the executeable binaries itself are only a small part these days
however, i686 is a dying architecture, otherwise RHEL would not have stopped tu support it at all and i personally did not see any i686 machine in the past 6 years
That said, even on x86_64 it isn't anything close to no overhead. Tried last night to rebuild GCC's cc1plus as -fpie -pie, and then rebuild stage3 of GCC with make -j1 separately with the original stage3 cc1plus (ET_EXEC binary) and PIE cc1plus (ET_DYN). The build (which included still time for various other tools being not PIE, make, ld, as) got 2.1% slower user time. Also, the number of relocations and memory consumption got up. Non-PIE cc1plus: Relocation section '.rela.dyn' at offset 0x187d30 contains 190 entries: Relocation section '.rela.plt' at offset 0x188f00 contains 284 entries: GNU_RELRO 0x1d14730 0x0000000002314730 0x0000000002314730 0x0058d0 0x0058d0 R 0x1 PIE cc1plus: Relocation section '.rela.dyn' at offset 0x187d90 contains 75803 entries: Relocation section '.rela.plt' at offset 0x344018 contains 230 entries: GNU_RELRO 0x1e18cf0 0x0000000002018cf0 0x0000000002018cf0 0x10e310 0x10e310 R 0x1 That means e.g. on the startup of each cc1plus process, that means 1MB extra COW wastage (executable has 24KB of pages written and then made non-writable, while PIE over 1MB)
that may all be true while 2% don't impress me that much*but* since *mobile phones* and other operating systems in the meantime are full PIE and it improves security how can someone justify the reason performance on a desktop/server distribution with much more powerful hardware?
signature.asc
Description: OpenPGP digital signature
-- devel mailing list devel@lists.fedoraproject.org https://admin.fedoraproject.org/mailman/listinfo/devel Fedora Code of Conduct: http://fedoraproject.org/code-of-conduct