How to compile and execute from memory directly?

Possible? Not the way you seem to wish. The task has two parts:

1) How to get the binary into memory

When we specify /dev/stdout as output file in Linux we can then pipe into our program x0 that reads an executable from stdin and executes it:

 gcc -pipe YourFiles1.cpp YourFile2.cpp -o/dev/stdout -Wall | ./x0 

In x0 we can just read from stdin until reaching the end of the file:

int main(int argc, const char ** argv) { const int stdin = 0; size_t ntotal = 0; char * buf = 0; while(true) { /* increasing buffer size dynamically since we do not know how many bytes to read */ buf = (char*)realloc(buf, ntotal+4096*sizeof(char)); int nread = read(stdin, buf+ntotal, 4096); if (nread<0) break; ntotal += nread; } memexec(buf, ntotal, argv); } 

It would also be possible for x0 directly execute the compiler and read the output. This question has been answered here: Redirecting exec output to a buffer or file

Caveat: I just figured out that for some strange reason this does not work when I use pipe | but works when I use the x0 < foo.

Note: If you are willing to modify your compiler or you do JIT like LLVM, clang and other frameworks you could directly generate executable code. However for the rest of this discussion I assume you want to use an existing compiler.

Note: Execution via temporary file

Other programs such as UPX achieve a similar behavior by executing a temporary file, this is easier and more portable than the approach outlined below. On systems where /tmp is mapped to a RAM disk for example typical servers, the temporary file will be memory based anyway.

#include<cstring> // size_t #include <fcntl.h> #include <stdio.h> // perror #include <stdlib.h> // mkostemp #include <sys/stat.h> // O_WRONLY #include <unistd.h> // read int memexec(void * exe, size_t exe_size, const char * argv) { /* random temporary file name in /tmp */ char name[15] = "/tmp/fooXXXXXX"; /* creates temporary file, returns writeable file descriptor */ int fd_wr = mkostemp(name, O_WRONLY); /* makes file executable and readonly */ chmod(name, S_IRUSR | S_IXUSR); /* creates read-only file descriptor before deleting the file */ int fd_ro = open(name, O_RDONLY); /* removes file from file system, kernel buffers content in memory until all fd closed */ unlink(name); /* writes executable to file */ write(fd_wr, exe, exe_size); /* fexecve will not work as long as there in a open writeable file descriptor */ close(fd_wr); char *const newenviron[] = { NULL }; /* -fpermissive */ fexecve(fd_ro, argv, newenviron); perror("failed"); } 

Caveat: Error handling is left out for clarities sake. Includes for sake of brevity.

Note: By combining step main() and memexec() into a single function and using splice(2) for copying directly between stdin and fd_wr the program could be significantly optimized.

2) Execution directly from memory

One does not simply load and execute an ELF binary from memory. Some preparation, mostly related to dynamic linking, has to happen. There is a lot of material explaining the various steps of the ELF linking process and studying it makes me believe that theoretically possible. See for example this closely related question on SO however there seems not to exist a working solution.

Update UserModeExec seems to come very close.

Writing a working implementation would be very time consuming, and surely raise some interesting questions in its own right. I like to believe this is by design: for most applications it is strongly undesirable to (accidentially) execute its input data because it allows code injection.

What happens exactly when an ELF is executed? Normally the kernel receives a file name and then creates a process, loads and maps the different sections of the executable into memory, performs a lot of sanity checks and marks it as executable before passing control and a file name back to the run-time linker ld-linux.so (part of libc). The takes care of relocating functions, handling additional libraries, setting up global objects and jumping to the executables entry point. AIU this heavy lifting is done by dl_main() (implemented in libc/elf/rtld.c).

Even fexecve is implemented using a file in /proc and it is this need for a file name that leads us to reimplement parts of this linking process.

Libraries

• UserModeExec
• libelf -- read, modify, create ELF files
• eresi -- play with elfes
• OSKit (seems like a dead project though)

Reading

• http://www.linuxjournal.com/article/1060?page=0,0 -- introduction
• http://wiki.osdev.org/ELF -- good overview
• http://s.eresi-project.org/inc/articles/elf-rtld.txt -- more detailed Linux-specific explanation
• http://www.codeproject.com/Articles/33340/Code-Injection-into-Running-Linux-Application -- how to get to hello world
• http://www.acsu.buffalo.edu/~charngda/elf.html -- nice reference of ELF structure
• Loaders and Linkers by John Levine -- deeoer explanation of linking

Related Questions at SO

• Linux user-space ELF loader
• ELF Dynamic loader symbol lookup ordering
• load-time ELF relocation
• How do global variables get initialized by the elf loader

So it seems possible, you decide whether is also practical.

Yes, though doing it properly requires designing significant parts of the compiler with this in mind. The LLVM guys have done this, first with a kinda-separate JIT, and later with the MC subproject. I don't think there's a ready-made tool doing it. But in principle, it's just a matter of linking to clang and llvm, passing the source to clang, and passing the IR it creates to MCJIT. Maybe a demo does this (I vaguely recall a basic C interpreter that worked like this, though I think it was based on the legacy JIT).

Edit: Found the demo I recalled. Also, there's cling, which seems to do basically what I described, but better.

Linux can create virtual file systems in RAM using tempfs. For example, I have my tmp directory set up in my file system table like so:

tmpfs /tmp tmpfs nodev,nosuid 0 0 

Using this, any files I put in /tmp are stored in my RAM.

Windows doesn't seem to have any "official" way of doing this, but has many third-party options.

Without this "RAM disk" concept, you would likely have to heavily modify a compiler and linker to operate completely in memory.

If you are not specifically tied to C++, you may also consider other JIT based solutions:

• in Common Lisp SBCL is able to generate machine code on the fly
• you could use TinyCC and its libtcc.a which emits quickly poor (i.e. unoptimized) machine code from C code in memory.
• consider also any JITing library, e.g. libjit, GNU Lightning, LLVM, GCCJIT, asmjit
• of course emitting C++ code on some tmpfs and compiling it...

But if you want good machine code, you'll need it to be optimized, and that is not fast (so the time to write to a filesystem is negligible).

If you are tied to C++ generated code, you need a good C++ optimizing compiler (e.g. g++ or clang++); they take significant time to compile C++ code to optimized binary, so you should generate to some file foo.cc (perhaps in a RAM file system like some tmpfs, but that would give a minor gain, since most of the time is spent inside g++ or clang++ optimization passes, not reading from disk), then compile that foo.cc to foo.so (using perhaps make, or at least forking g++ -Wall -shared -O2 foo.cc -o foo.so, perhaps with additional libraries). At last have your main program dlopen that generated foo.so. FWIW, MELT was doing exactly that, and on Linux workstation the manydl.c program shows that a process can generate then dlopen(3) many hundred thousands of temporary plugins, each one being obtained by generating a temporary C file and compiling it. For C++ read the C++ dlopen mini HOWTO.

^{Alternatively, generate a self-contained source program foobar.cc, compile it to an executable foobarbin e.g. with g++ -O2 foobar.cc -o foobarbin and execute with execve that foobarbin executable binary}

When generating C++ code, you may want to avoid generating tiny C++ source files (e.g. a dozen lines only; if possible, generate C++ files of a few hundred lines at least; unless lots of template expansion happens thru extensive use of existing C++ containers, where generating a small C++ function combining them makes sense). For instance, try if possible to put several generated C++ functions in the same generated C++ file (but avoid having very big generated C++ functions, e.g. 10KLOC in a single function; they take a lot of time to be compiled by GCC). You could consider, if relevant, to have only one single #include in that generated C++ file, and pre-compile that commonly included header.

Jacques Pitrat's book Artificial Beings, the conscience of a conscious machine (ISBN 9781848211018) explains in details why generating code at runtime is useful (in symbolic artificial intelligence systems like his CAIA system). The RefPerSys project is trying to follow that idea and generate some C++ code (and hopefully, more and more of it) at runtime. Partial evaluation is a relevant concept.

Your software is likely to spend more CPU time in generating C++ code than GCC in compiling it.

tcc compiler "-run" option allows for exactly this, compile into memory, run there and finally discard the compiled stuff. No filesystem space needed. "tcc -run" can be used in shebang to allow for C script, from tcc man page:

#!/usr/local/bin/tcc -run #include <stdio.h> int main() { printf("Hello World\n"); return 0; } 

C scripts allow for mixed bash/C scripts, with "tcc -run" not needing any temporary space:

#!/bin/bash echo "foo" sed -n "/^\/\*\*$/,\$p" $0 | tcc -run - exit /** */ #include <stdio.h> int main() { printf("bar\n"); return 0; } 

Execution output:

$ ./shtcc2 foo bar $ 

C scripts with gcc are possible as well, but need temporary space like others mentioned to store executable. This script produces same output as the previous one:

#!/bin/bash exc=/tmp/`basename $0` if [ $0 -nt $exc ]; then sed -n "/^\/\*\*$/,\$p" $0 | gcc -x c - -o $exc; fi echo "foo" $exc exit /** */ #include <stdio.h> int main() { printf("bar\n"); return 0; } 

C scripts with suffix ".c" are nice, headtail.c was my first ".c" file that needed to be executable:

$ echo -e "1\n2\n3\n4\n5\n6\n7" | ./headtail.c 1 2 3 6 7 $ 

I like C scripts, because you just have one file, you can easily move around, and changes in bash or C part require no further action, they just work on next execution.

P.S:
The above shown "tcc -run" C script has a problem, C script stdin is not available for executed C code. Reason was that I passed extracted C code via pipe to "tcc -run". New gist run_from_memory_stdin.c does it correctly:

... echo "foo" tcc -run <(sed -n "/^\/\*\*$/,\$p" $0) 42 ... 

"foo" is printed by bash part, "bar 42" from C part (42 is passed argv[⁠1]), and piped script input gets printed from C code then:

$ route -n | ./run_from_memory_stdin.c foo bar 42 Kernel IP routing table Destination Gateway Genmask Flags Metric Ref Use Iface 0.0.0.0 172.29.58.98 0.0.0.0 UG 306 0 0 wlan1 10.0.0.0 0.0.0.0 255.255.255.0 U 0 0 0 wlan0 169.254.0.0 0.0.0.0 255.255.0.0 U 303 0 0 wlan0 172.29.58.96 0.0.0.0 255.255.255.252 U 306 0 0 wlan1 $ 

One can easily modify the compiler itself. It sounds hard first but thinking about it, it seams obvious. So modifying the compiler sources directly expose a library and make it a shared library should not take that much of afford (depending on the actual implementation).

Just replace every file access with a solution of a memory mapped file.

It is something I am about to do with compiling something transparently in the background to op codes and execute those from within Java.

-

But thinking about your original question it seams you want to speed up compilation and your edit and run cycle. First of all get a SSD-Disk you get almost memory speed (use a PCI version) and lets say its C we are talking about. C does this linking step resulting in very complex operations that are likely to take more time than reading and writing from / to disk. So just put everything on SSD and live with the lag.

Finally the answer to OP question is yes!

I found memrun repo from guitmz, that demoed running (x86_64) ELF from memory, with golang and assembler. I forked that, and provided C version of memrun, that runs ELF binaries (verified on x86_64 and armv7l), either from standard input, or via first argument process substitution. The repo contains demos and documentation (memrun.c is 47 lines of code only):
https://github.com/Hermann-SW/memrun/tree/master/C#memrun

Here is simplest example, with "-o /dev/fd/1" gcc compiled ELF gets sent to stdout, and piped to memrun, which executes it:

pi@raspberrypi400:~/memrun/C $ gcc info.c -o /dev/fd/1 | ./memrun My process ID : 20043 argv[0] : ./memrun no argv[1] evecve --> /usr/bin/ls -l /proc/20043/fd total 0 lr-x------ 1 pi pi 64 Sep 18 22:27 0 -> 'pipe:[1601148]' lrwx------ 1 pi pi 64 Sep 18 22:27 1 -> /dev/pts/4 lrwx------ 1 pi pi 64 Sep 18 22:27 2 -> /dev/pts/4 lr-x------ 1 pi pi 64 Sep 18 22:27 3 -> /proc/20043/fd pi@raspberrypi400:~/memrun/C $ 

The reason I was interested in this topic was usage in "C script"s. run_from_memory_stdin.c demonstrates all together:

pi@raspberrypi400:~/memrun/C $ wc memrun.c | ./run_from_memory_stdin.c foo bar 42 47 141 1005 memrun.c pi@raspberrypi400:~/memrun/C $ 

The C script producing shown output is so small ...

#!/bin/bash echo "foo" ./memrun <(gcc -o /dev/fd/1 -x c <(sed -n "/^\/\*\*$/,\$p" $0)) 42 exit /** */ #include <stdio.h> int main(int argc, char *argv[]) { printf("bar %s\n", argc>1 ? argv[1] : "(undef)"); for(int c=getchar(); EOF!=c; c=getchar()) { putchar(c); } return 0; } 

P.S:
I added tcc's "-run" option to gcc and g++, for details see:
https://github.com/Hermann-SW/memrun/tree/master/C#adding-tcc--run-option-to-gcc-and-g

Just nice, and nothing gets stored in filesystem:

pi@raspberrypi400:~/memrun/C $ uname -a | g++ -O3 -Wall -run demo.cpp 42 bar 42 Linux raspberrypi400 5.10.60-v7l+ #1449 SMP Wed Aug 25 15:00:44 BST 2021 armv7l GNU/Linux pi@raspberrypi400:~/memrun/C $