On Sun, Oct 11, 2009 at 2:38 AM, Joakim Tjernlund joakim.tjernlund@transmode.se wrote:
Graeme Russ graeme.russ@gmail.com wrote on 10/10/2009 13:21:10:
On Sat, Oct 10, 2009 at 9:47 PM, Joakim Tjernlund joakim.tjernlund@transmode.se wrote:
Graeme Russ graeme.russ@gmail.com wrote on 10/10/2009 12:38:19:
On Sat, Oct 10, 2009 at 8:27 PM, Joakim Tjernlund joakim.tjernlund@transmode.se wrote:
Graeme Russ graeme.russ@gmail.com wrote on 10/10/2009 10:46:52:
On Sat, Oct 10, 2009 at 7:07 PM, Joakim Tjernlund joakim.tjernlund@transmode.se wrote: > Graeme Russ graeme.russ@gmail.com wrote on 10/10/2009 06:43:52: >> >> On Fri, Oct 9, 2009 at 10:12 AM, Joakim Tjernlund >> joakim.tjernlund@transmode.se wrote: >> >> >> >> On Fri, Oct 9, 2009 at 9:27 AM, J. William Campbell >> >> jwilliamcampbell@comcast.net wrote: >> >> > Graeme Russ wrote: >> >> >> >> >> >> On Fri, Oct 9, 2009 at 2:58 AM, J. William Campbell >> >> >> jwilliamcampbell@comcast.net wrote: >> >> >> >> >> >>> >> >> >>> Graeme Russ wrote: >> >> >>> >> >> >>>> >> >> >>>> Out of curiosity, I wanted to see just how much of a size penalty I am >> >> >>>> incurring by using gcc -fpic / ld -pic on my x86 u-boot build. Here are >> >> >>>> the results (fixed width font will help - its space, not tab, >> >> >>>> formatted): >> >> >>>> >> >> >>>> Section non-reloc reloc >> >> >>>> --------------------------------------- >> >> >>>> .text 000118c4 000137fc <- 0x1f38 bytes (~8kB) bigger >> >> >>>> .rodata 00005bad 000059d0 >> >> >>>> .interp n/a 00000013 >> >> >>>> .dynstr n/a 00000648 >> >> >>>> .hash n/a 00000428 >> >> >>>> .eh_frame 00003268 000034fc >> >> >>>> .data 00000a6c 000001dc >> >> >>>> .data.rel n/a 00000098 >> >> >>>> .data.rel.ro.local n/a 00000178 >> >> >>>> .data.rel.local n/a 000007e4 >> >> >>>> .got 00000000 000001f0 >> >> >>>> .got.plt n/a 0000000c >> >> >>>> .rel.got n/a 000003e0 >> >> >>>> .rel.dyn n/a 00001228 >> >> >>>> .dynsym n/a 00000850 >> >> >>>> .dynamic n/a 00000080 >> >> >>>> .u_boot_cmd 000003c0 000003c0 >> >> >>>> .bss 00001a34 00001a34 >> >> >>>> .realmode 00000166 00000166 >> >> >>>> .bios 0000053e 0000053e >> >> >>>> ======================================= >> >> >>>> Total 0001d5dd 00022287 <- 0x4caa bytes (~19kB) bigger >> >> >>>> >> >> >>>> Its more than a 16% increase in size!!! >> >> >>>> >> >> >>>> .text accounts for a little under half of the total bloat, and of that, >> >> >>>> the crude dynamic loader accounts for only 341 bytes >> >> >>>> >> >> >>>> >> >> >>> >> >> >>> Hi Graeme, >> >> >>> I would be interested in a third option (column), the x86 build with >> >> >>> just -mrelocateable but NOT -fpic. It will not be definitive because >> >> >>> there >> >> >>> will be extra code that references the GOT and missing code todo some of >> >> >>> the relocation, but it would still be interesting. >> >> >>> >> >> >> >> >> >> x86 does not have -mrelocatable. This is a PPC only option :( >> >> >> >> >> > >> >> > Hi Graeme, >> >> > You are unfortunately correct. However, I wonder if we can get >> >> > essentially the same result by executing the final ld step with the >> >> > --emit-relocs switch included. This may also include some "extra" sections >> >> > that we would want to strip out, but if it works, it could give all >> >> > ELF-based systems a way to a relocatable u-boot. >> >> > >> >> >> >> I don't think --emit-relocs is necessary with -pic. I haven't gone through >> >> all the permutations to see if there is a smaller option, but gcc -fpic and >> >> ld -pie creates enough information to perform relocation on the x86 >> >> platform >> > >> > Try -fvisibility=hidden >> >> Thanks - Shaved another 2539 bytes off the binary >> >> Also found out how to get rid of .eh_frame (crept in when I upgraded to >> gcc 4.4.1) with -fno-dwarf2-cfi-asm, so that shaves another 13452 bytes >> >> Total saving of 15.6k > > Great, so now you are back at just a few percent added I guess? > >
Not really - The .eh_frame saving applies to both relocated and non relocated builds
OK, so you didn't use PIC before at all?
Anyway I think you can do more. Using -Bsymbolic you should get away with RELATIVE relocs only and be able to skip a lot of segments above. Have a look at uClibc ldso/ldso/dl-startup.c
My build options thus far are:
PLATFORM_RELFLAGS += -fpie -fvisibility=hidden PLATFORM_CPPFLAGS += -fno-dwarf2-cfi-asm PLATFORM_LDFLAGS += -pie
-fpic / -pic make no difference
not on x86, on ppc it is a big difference.
Interestingly, -Bsymbolic adds exactly 8 bytes to .dynamic, but doesn't change the size of any other section
Pulling apart the relocation sections, it seems that all relocations are already RELATIVE even without -Bsymbolic
Ah, that is because you built an exe with -pie Then you should be able to drop everything but the RELATIVE from the linking, or almost in any case.
Jocke
Hmm, so its seems I may have hit the limit. I tried:
PLATFORM_LDFLAGS += -r --emit-relocs
but there is not enough information left to complete the relocation. It seems as though I need .rel.got, .got.plt, .dynsym and .rel.dyn in order to find the actual bytes that need modifying (it also seems to mess with the size of the stripped binary for some reason)
Looks like I'll have to proceed with my original plan - a bit bloated, but it works
Relocation costs :(
I am not sure why you need .got.plt, it should be empty, what is in it? Same with dynsym, what is in it?
Memory fails me, but since u-boot is a freestanding app it I think these two might not be needed. Perhaps there are weak unresolved syms in there?
Jocke
Well, I'm in the middle of a pretty intense analysis of what is going on.
Compile flags are:
PLATFORM_RELFLAGS += -fpic -fvisibility=hidden PLATFORM_CPPFLAGS += -fno-dwarf2-cfi-asm PLATFORM_LDFLAGS += -pic -Bsymbolic
So far I have found that the only sections that have changes as a result of a change in TEXT_BASE are: .text .rodata .data.rel .got .got.plt .rel.text .rel.got .rel.dyn .dynsym .dynamic .u_boot_cmd
Changes in .text are covered by .rel.text (see below) or as a result of CONFIG_SYS_MONITOR_BASE being equal to TEXT_BASE (used in cfi_flash.c)
Changes in .rodata are a result of version_string changing for each compile
.rel.text - Contains a list of pointers into .got - All entries are R_386_RELATIVE - All entries (8 of) are in cpu/i386/start.o - cpu/i386/start.o only used during initial bootstrap - not needed after execution starts in RAM - Can be safely discarded
.rel.got - Contains a list of pointers into .got - All entries are R_386_RELATIVE - Not all entries change with TEXT_BASE. Some entries are symbols exported from the linker script (in particular section size exports) while the others are in the somewhat 'special' BIOS and Real Mode sections which are located in a fixed RAM location (these sections are used for real-mode trampolining into Linux by providing a limited PC 'BIOS' - All entries that are not linked to TEXT_BASE are easily identified because they are 'located' below TEXT_BASE (specically between 0x00000000 and 0x00001A34) - This section is not needed in the final binary - Direct processing of .got will achieve the required end result
.rel.dyn - Contains a list of pointers into .data.rel and .u_boot_cmd - Like .rel.got, not all entries in .data.rel need relocating. Again, like .rel.got, these are easily identified - This section not needed
Another 5.5k saved
So, all that is left are .dynsym and .dynamic ... .dynsym - Contains 70 entries (16 bytes each, 1120 bytes) - 44 entries mimic those entries in .got which are not relocated - 21 entries are the remaining symbols exported from the linker script - 4 entries are labels defined in inline asm and used in C - 1 entry is a NULL entry
.dynamic - 88 bytes - Array of Elf32_Dyn - typedef struct { Elf32_Sword d_tag; union { Elf32_Word d_val; Elf32_Addr d_ptr; } d_un; } Elf32_Dyn; - 0x11 entries [00] 0x00000010, 0x00000000 DT_SYMBOLIC, (ignored) [01] 0x00000004, 0x38059994 DT_HASH, points to .hash [02] 0x00000005, 0x380595AB DT_STRTAB, points to .dynstr [03] 0x00000006, 0x3805BDCC DT_SYMTAB, points to .dynsym [04] 0x0000000A, 0x000003E6 DT_STRSZ, size of .dynstr [05] 0x0000000B, 0x00000010 DT_SYMENT, ??? [06] 0x00000015, 0x00000000 DT_DEBUG, ??? [07] 0x00000011, 0x3805A8F4 DT_REL, points to .rel.text [08] 0x00000012, 0x000014D8 DT_RELSZ, ??? [09] 0x00000013, 0x00000008 DT_RELENT, ??? [0a] 0x00000016, 0x00000000 DT_TEXTREL, ??? [0b] 0x6FFFFFFA, 0x00000236 ???, Entries in .rel.dyn [0c] 0x00000000, 0x00000000 DT_NULL, End of Array [0d] 0x00000000, 0x00000000 DT_NULL, End of Array [0e] 0x00000000, 0x00000000 DT_NULL, End of Array [0f] 0x00000000, 0x00000000 DT_NULL, End of Array [10] 0x00000000, 0x00000000 DT_NULL, End of Array
I think some more investigation into the need for .dynsym and .dynamic is still required...
Regards,
Graeme