pouët.net

Haxxoring the elf format for 1k/4k stuff

category: code [glöplog]
greetings to iblis!
Quote:
339bytes + (strlen(libraryname)+9) bytes per library + 4 bytes per symbol

wow! we want a paper please *_*
added on the 2008-07-03 18:33:55 by pera pera
Yup I'll write something about it and release sauce once I've actually made a prod with it. I'd rather not disclose the source just yet because if I did someone would surely beat me to it :)
So what does import-by-hash refer to if you're storing the full names of the libraries you're importing?
added on the 2008-07-03 20:11:34 by doomdoom doomdoom
Full names for the libraries, hashes of the function names you want to use. As far as I know there's no way to import the libraries themselves by hash, and it wouldn't represent much of a saving (if any) because generally you only import 2 or 3 libraries and the standard elf way of importing libraries is small anyway.
parapete: Linux kernel makes a few automatic assumptions if the ELF header is modified to an incompatible form so basically you can save(by means of a better compression rate) a few bytes by changing certain information bytes (like for example architecture information byte iirc) to 0x00. Also changing header's section access rights to RWX or equal probably helps a bit with the compression.

Kudos for your&leblane's research. Shall buy you guys a beer at some demoparty.
added on the 2008-07-04 07:07:47 by waffle waffle
waffle, on the contrary we owe you a beer for doing 1k on Linux in the first place :)

We fiddled about with redundant bytes in the header a bit. I found that you get the best pack rate by zeroing most of them, although there might be an opportunity to store hashes in the header, I've not tested this yet though.

I'm ashamed to admit I did have a quick look at "Paeaeministeri Vaeyrynen greets Accession only" in a hex editor and it looks like you're using the traditional dlopen+dlsym technique. Is that right? Are there any more tricks at work there?
For anyone who's interested, here's my port of flow2 to Linux. 143 bytes smaller than the original windows version, and it even includes proper timer code :)

It's written in 100% asm, and the source is included. It's really meant as a proof of concept of the technique. It requires hand crafted elf headers, so an exe packer or crinkler-style replacement linker needs to be written to make the technique useful for intros written in C/C++. I was hoping to use an ld linker script to do it, but ld seems to be far less flexible than I'd hoped.
added on the 2008-07-05 01:58:02 by leblane leblane
parapete, leblane: I love you, guys :)
added on the 2008-07-05 04:00:19 by slack slack
And just to piss on leblane's bonfire a bit, I have 794 bytes now :D
1024 - 143 -749 = 132 bytes gain?
added on the 2008-07-05 17:56:20 by Tigrou Tigrou
wow flow2 on linux! Very cool work.
added on the 2008-07-05 20:39:09 by auld auld
*bump*

@leblane: thx for the flow2 source code =)
added on the 2008-07-15 18:29:15 by pera pera
Quote:
143 bytes smaller than the original windows version
To be fair I guess you dont have to bother with all that nasty getprocaddress crap I had to deal with under windows to get the shaders up and running. Right?
added on the 2008-07-17 22:35:51 by auld auld
I got a self compilable version of crings and it's only 585 bytes.

problem is you need gcc and glut dev... for those who are interested:
crings in 585 bytes

I'm sure this can be be done better, but this was thrown together in 2 hours.
added on the 2008-10-03 12:48:39 by LiraNuna LiraNuna
Self compiling? You might want to put that in a different thread then ;)

The self compiling technique does have potential for making tiny "executables" but I find it really impure. I guess it's only as impure as using a gzip dropper but then I plan on eradicating the need for that soon enough.
Why not write the intro in Python/Perl/etc instead of compiling C?
added on the 2008-10-03 17:04:49 by Deltafire Deltafire
Deltafire: Because Python/Perl/etc is usually a few orders of maginude slower than C. For the typical "one GL_QUAD and a huge fragment shader" 1k intros that might not matter, though.
added on the 2008-10-04 19:27:21 by KeyJ KeyJ
I just had that silly idea.. why not write a perl script that expands to C code that gets compiled ;)
added on the 2008-10-04 20:58:09 by _-_-__ _-_-__
why not write a quine that outputs it's source code and compiles itself?
added on the 2008-10-04 21:52:09 by LiraNuna LiraNuna
Write a perl script to perform inverse BWT on your C source code!
This is the fruit of my labours from this thread (sauce included). </spam>
did same pouet topic/discussion/article exists for win32 ??? the online thing i found on the web is this http://www.phreedom.org/solar/code/tinype/

i believe that this stuff with a algo that do some LZ , decompress the stuff in memory then execute it would be nice (is this possible in win32, or is it easier to write the stuff to hdd then execute it (like bat/cab do) ?)


seems some top 1k (like himalaya or tracie) use much more advanced techniques (like context mixing) but i have no idea how to implement this properly
added on the 2008-10-13 15:00:40 by Tigrou Tigrou
I'm currently trying a bit arround too, the first step was to improve the make_it_4k by string loader from fit (can be done smaller).

Code: bits 32 extern dlopen, dlsym global import global oglBegin global oglEnd global sdlInit global sdlQuit ; constant definitions RTLD_NOW equ 2 section .text import: ;-1... we'll inc that. mov edi, dword (data-1) _lib: inc edi mov al, byte [edi] ;done? test al, al jz _zero ;--- we have a library... push dword RTLD_NOW push edi call dlopen mov ebx, eax pop eax pop eax ;--- done! _func: ;goto next string xor eax, eax xor ecx, ecx not cl repne scasb mov al, byte [edi] test al, al ;two zeros found? -> new lib follows! jz _lib ;--- we have a function we want to import. push edi push ebx call dlsym ;store it... all function names have to be longer than 4 (or 5 with trailing '\0') stosd pop eax pop eax ;--- done - goto next func jmp _func _zero: ret section .data data: db "libGL.so",0 oglBegin db "glBegin",0 oglEnd db "glEnd",0,0 ; two zeros, new lib follows db "libSDL.so",0 sdlInit db "SDL_Init",0 sdlQuit db "SDL_Quit",0,0,0 ; three zeros -> DONE!


with a nice import.h & import.o you can now use the exported symbols within normal C code.
parapete... now, can we do that with import by hash too? the by string method is so 2001.

Let's make linux more interesting for 4k intros again!
added on the 2008-12-14 14:41:20 by las las
Okay, I'm currently writing "HOW DO I 1K IN LINUX?" (or a commented source at least). The whole technique I used is too complicated to sum up sufficiently in a paragraph, what I can tell you now though is that you want to avoid dlopen and dlsym altogether.

login