Saturday, July 27, 2013

The problem with userdefined boolean types in C.
It's become apparent that a lot of people prefer to define their own
boolean types in C, expecially if the target compiler lacks a _Bool
type. The problem is all versions of it are fundamentally flawed.
Lets look at some typical example of this:

    typedef int bool; // wrong
    typedef unsigned int bool; // wrong
    typedef char bool; // wrong
    typedef unsigned char bool; // wrong

Infact bool cannot be represented at all by a typedef, and I'll explain
why shortly, but lets first assume the most typical implementation, and
the problems that can arise from it. The most typical typedef I've seen
for boolean is int. Lets consider the following code (that is intended to
operate between C/C++).

    // I see this ALL the time
    #ifndef __cplusplus
    typedef int bool; // harmless looking

    struct foo { 
        bool a;
        bool b;

When this is compiled in C, the sizeof this structure (assuming 4-byte
ints) is 8. However in C++ sizeof(bool) == 1, so you'd assume this
structure is 2 bytes. That would be correct, these structures do not match
at all in memory. The thing is people get away with this sort of stuff
because of typical computing architectures. On my x86, structure elements
are all aligned on 4-byte boundaries, the low-order byte always comes first
as well, these are the only two architectural-specific factors keeping this
structure from blowing up in your face.

Some would argue that is a far reach, but there is an even bigger issue
with this typedef, an issue that wouldn't technically manifest unless
you're trying to converse memory, mainly bit-fields. Lets look at a typical
example in C.

    typedef int bool;
    struct {
        bool a : 1;
    } foo;

    foo.a = 1; // whats wrong with this code?

This innocent looking code has more problems than you'd expect, mainly
there is an implicit overflow of the constant expression `foo.a=1`.
If you enable pedantic compiler errors on GCC you'll get:

foo.c:7:5: warning: overflow in implicit constant conversion [-Woverflow]
     foo.a = 1;

The reason is the semantics of signed integer truncated to 1-bit, you get
just the sign-bit, so you can have -1 or 0, not 0 and 1. This comes as much
to surprise to someone because the use of bool in bit-fields is explicitly
allowed by C. To quote:

`A bit-field is interpreted as a signed or unsigned integer type consisting
 of the specified number of bits. If the value 0 or 1 is stored into a 
 nonzero-width bit-field of type _Bool, the value of the bit-field shall
 compare equal to the value stored.`

You could technically fix that issue by using:
    typedef unsigned int bool;

But then you still have the C++ issue with sizeof(bool) for forward
compatibility, and another issue that just causes more pain, now that
bool is unsigned int, your code becomes littered with comparisons between
signed (bool in C++ is signed) and unsigned (typedef bool is unsigned)
boolean values. All these comparisons cause a ton of warnings with any
decent C++ compiler, and if your project is pedantic about code correctness
will even error.

As for the other variations of typedef, they break bit-fields. The only
valid types that can be used in a bit-field are signed/unsigned int and
_Bool, everything else is undefined (GCC however does accept it as an

But to reiterate on the main premise here, _Bool simply cannot be represented
by a typedef, if you want the semantics of it to be correct. Which is why
it was added as a type in the first place, so if you want to do booleans in
C, just stick to being explicit about integers, for bitfields be explicit
with unsigned, unless you're targeting C99, then you're guaranteed _Bool.

Sunday, October 7, 2012

C++'s bad things for performance critical applications

One of the most often disputed things about C++ are its exceptions, virtuals, and runtime type information (RTTI).  Some of the arguments centered around these are often subjected to bias, and I should disclaimer that this article may be subjected to my bias as well.  But I'd like to take the time and reflect a few of the common statements about these thing.

  1. Virtual functions are slow!
  2. Exceptions are slow!
  3. RTTI is slow!

Virtual functions are slow:

Back when C++ introduced virtual functions the state of computing in general was (for a lack of a better word) shitty.  The cost of a virtual table look up on a 100MHz CPU was negligible, but a general rule of thumb now is if it cost more than allocating from main memory, it's slow.  Virtual functions aren't that slow, they're a little more expensive than a non-virtual member function, if they're called in a predictable pattern, if that pattern changes at any-time you will get a miss-prediction penalty of 10 to 30 clock cycles [reference] (page 44 of Agners Frog's "Optimizing Software in C++" ).  This performance penalty isn't really terrible on modern computers, but if you find yourself working with embedded systems or need to barge a little more speed out of your application you might want to reconsider your use of virtual functions.

Exceptions are slow:

Exceptions are a really handy feature to have in C++, they've got their wonderful uses, but again suffer from some serious overhead and might not work well in performance critical situations.  To understand why they're so slow let me explain to you how a typical implementation of exception handling (which is used in visual studio) works. I should note that no compiler for C++ currently implements exceptions in a zero-overhead fashion (that actually works), for both visual studio and GCC the compiler adds prologue and epilogue instructions to track information about the currently executing scope.  This information allows for faster stack unwinding, although this has one problem and that is, it greatly imposes a larger run time overhead when a exception isn't thrown.  In visual studios explicit case, the compiler creates hidden local variables in every function that references back to an exception handler function, and a table of data specific for that function [reference] ("How a C++ compiler implements exception handling.").  The overhead of exception handling in this example for when there is an absence of a throw is literally the cost of registering a handler as each function is entered, as well as the unregistering when the function is left.  If you observe some visual studio produced assembler you can instantly see that's about three push instructions, two movs for the prologue, and about three more movs for the epilogue.  This is (as you can already tell) rather expensive.  Of course there is yet another problem we derive from all this stuff the compiler splices into the instruction stream, that is cache locality.  With the presence of all these exception handlers (scattered through the instruction space) our cache locality is severely affected.  So when you consider all the variables in this particular example, the overhead of exception handling when an exception is thrown is the overall cost of unwinding the stack, iterating over all the function information tables for every function.  Each of these info tables must also be searched to find stack objects that require destruction, and to check the appropriate type catch blocks for what ever was thrown, handle the destruction accordingly and jump to that scope.  This is all very slow.  So you might want to reconsider your use of exceptions.

Runtime Type Information is slow:

RTTI is often abused for things like dynamic_cast, a clever cast that does a type check to see if the cast between two types is legal, however some might not recognize that this operation is very expensive both in time and space complexity.  The space cost itself for RTTI is very high, literally every class that has at least one virtual method will be given a vtable with some information about it's name, and information about it's base classes, this information grows exponentially for deep hierarchies, as you can already tell.  There is no point in further discussing this, RTTI is just plain wrong.

Tuesday, July 31, 2012

Accelerate your SIMD knowledge to accelerate your code: Tricks to blaze through SSE code

Often when dealing with SSE code you find yourself referring to the documentation to understand what an instruction does, and it can impede on your time.  This is a part of a mini series I'm writing "Accelerate your SIMD knowledge to accelerate your code".  Lets begin shall we? 

1) Eat those instruction mnemonics
Learn the instruction mnemonic rules, doing this will greatly improve your speed in hand writing SSE code.
SSE Has five categories of instructions:
  • Data movement instructions
  • Arithmetic instructions
  • Comparison instructions
  • Logical instructions
  • Shuffle instruction 
All data movement instructions start with MOV, and have either [U/A/H/L] after MOV. These four letters describe something.
U = unaligned
A = aligned
H = high (64 bits high)
L = low (64 bits low)

There is combination HL (high to low) and combination, LH (low to high) which moves SRC[H/L] -> DST[H/L]; finally PS is postfixed after the choice of instruction. Knowing this you can easily think of something to yourself: "I need to move the upper 64bit of this register to the lower 64bits of this other register"; which is easy as MOV (move) H(high) L(low) PS, or [MOVHLPS].  If you need an unalligned move for 128 bits, we know unalligned is U so MOVUPS is the instruction we need.

2) Order your memory, and fries with that.
SSE greatly benefits from predictable memory patterns, it prefers linear over arbitrary data.  Iterating over a linked list of vectors to normalize is faster than incrementing a float pointer for every four elements of the vector and performing the normalize on the next four elements.

Aliasing memory strangely stalls the pipeline: i.e converting back from float to _m128 etc are wasteful operations.  However aliasing memory in respect to SSE is fine.  __mm_set_ps for example composes a _m128 from four floats, aliasing an aligned structure of four elements , or an array to _m128 would work the same.

/* allowed */
float x = 1,y = 2,z = 3, w = 0;
typedef struct { float x,y,z,w; } vec3 __attribute__((aligned(16)));
vec3 v3 = {x,y,z};
_m128 p = *(m128*)&v3;
/* same as this */
float x = 1,y = 2,z = 3, w = 0;
_m128 p = _mm_set_ps(x,y,z,w);
Aliasing memory like this allows for flexibility.

3) Nuke the runtime selection
Testing CPUID at runtime to determine if SSE is present is a common trick programs do for CPU compatibility.  These are often fine for selecting the correct optimized routines, but many people get it wrong.
Consider the following example:

void (*add)(const vec3*, const vec3*) = NULL;
int main() {
  add = (HasSSE()) ? &sse_add : &scalar_add;

This code is terribly wrong, because the function pointer calls are indirect calls, SSE doesn't like indirect calls, it instantly prevents any sort of pipelinability, and could very well be slower than the scalar version.
The right way to doing this would involve two separate translation units for each version of the function, and simply calling the correct version at runtime.

4) Punch a hole in your type system (C++ only)
When working with C++ and SSE intrinsics you often see yourself using lots of explicit casts, like reinterpret_cast, const_cast.  It's no wonder this is an issue especially with things like _mm_malloc, which allocates an safely aligned chunk of memory for SSE operations.  The problem is it returns void*, and C++ doesn't allow void* to be implicitly convertible to any pointer type.  To work around this I've often found myself punching a hole in C++'s type system via a simple structure.

class voidptr_t {
  void *ptr;
public: voidptr_t (void *pptr) : ptr (pptr) {}
   template<typename T>
  operator T *() { return (T*) ptr; }

As such when using _mm_malloc the void * will implicitly be converted to a voidptr_t using voidptr_t::voidptr_t(void *), and then the voidptr_t will implicitly convert itself to whatever pointer type you are trying to assign to.

5) Forget about prefetching
Unless you're writing code for CPU's before AMD anthlon XP or Intel Pentium 4, you're
wasting your time. Now virtually every CPU includes automatic cache prefetching. Which is costless and more accurate.

Sunday, June 24, 2012

EDID Extremely Disappointing Identification

Recently one of my online friends was having an issue with getting his display to work as expected.  After system updates his display simply wouldn't go above standard VGA resolutions.  Annoyed by this he came to me for help and after awhile of searching through system logs we finally found the cause of it.

There it sat inside one of those hidden logs that very few know exist, and even fewer know how to find.

 [    11.695] (WW) NVIDIA(GPU-0): The EDID read for display device CRT-0 is invalid: the
 [    11.695] (WW) NVIDIA(GPU-0):     checksum for EDID version 1 is invalid.
 [    11.695] (--) NVIDIA(GPU-0): 
 [    11.695] (--) NVIDIA(GPU-0): Raw EDID bytes:
 [    11.695] (--) NVIDIA(GPU-0): 
 [    11.695] (--) NVIDIA(GPU-0):   00 ff ff ff ff ff ff 00  5a 63 13 35 01 01 01 01
 [    11.695] (--) NVIDIA(GPU-0):   06 0e 01 03 1d 24 1b be  2a bb b8 a3 52 46 98 24
 [    11.695] (--) NVIDIA(GPU-0):   0f 48 4c ff ff 80 81 99  81 59 71 4f 61 59 a9 4f
 [    11.695] (--) NVIDIA(GPU-0):   c1 40 cd 40 d1 40 86 3d  00 c0 51 00 30 40 40 a0
 [    11.695] (--) NVIDIA(GPU-0):   13 00 60 08 11 00 00 1e  00 00 00 ff 00 33 35 44
 [    11.695] (--) NVIDIA(GPU-0):   30 34 30 36 30 30 32 36  31 0a 00 00 00 fd 00 32
 [    11.695] (--) NVIDIA(GPU-0):   b4 1e 61 18 00 0a 20 20  20 20 20 20 00 00 00 fc
 [    11.695] (--) NVIDIA(GPU-0):   00 47 39 30 66 2b 0a 20  20 20 20 20 20 20 00 e4
 [    11.695] (--) NVIDIA(GPU-0): 
 [    11.695] (--) NVIDIA(GPU-0):   00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
 [    11.695] (--) NVIDIA(GPU-0):   00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
 [    11.695] (--) NVIDIA(GPU-0):   00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
 [    11.695] (--) NVIDIA(GPU-0):   00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
 [    11.695] (--) NVIDIA(GPU-0):   00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
 [    11.695] (--) NVIDIA(GPU-0):   00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
 [    11.695] (--) NVIDIA(GPU-0):   00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
 [    11.695] (--) NVIDIA(GPU-0):   00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00

Of course what does this all mean?  After searching through the required documents that are free thanks to VESA ( We can conclude there is already an issue with this file. (If it isn't already apparent with the checksum fail message). Clearly the trailing 0 bytes stand out to anyone, so are they causing a problem?

Thankfully the spec states that the EDID data-chunk should be 128-bytes.  So these trailing zeros are useless (is this what was causing the checksum fail?) Turns out yes.  So lets rewrite it shall we?

The first problem, where do we get a copy of this data to modify?  Turns out there are tons of extraction programs, and ways to dump this data, but all of them are very specific to the driver, and even more specific to the operating system.  So I simply wrote each byte inside a hexeditor. Was I worried about making a physical mistake while retyping it? Of course I was which is why I had to be sure that the data I retyped was valid.

The EDID data spec states that the checksum of the file equals all bytes MOD 256, and the byte given from that operation is stored at the 128 byte mark.  So I did just that, I added up every byte from 0 to 127, and MOD 256'ed to get 0xE4.  Which is what was at the end of the file so I knew from this operation that the retype was successful.

So after all was said and done the binary blob I had looked something like this:

00000000  00 ff ff ff ff ff ff 00  5a 63 13 35 01 01 01 01  |........Zc.5....|
00000010  06 0e 01 03 1d 24 1b be  2a bb b8 a3 52 46 98 24  |.....$..*...RF.$|
00000020  0f 48 4c ff ff 80 81 99  81 59 71 4f 61 59 a9 4f  |.HL......YqOaY.O|
00000030  c1 40 cd 40 d1 40 86 3d  00 c0 51 00 30 40 40 a0  |.@.@.@.=..Q.0@@.|
00000040  13 00 60 08 11 00 00 1e  00 00 00 ff 00 33 35 44  |..`..........35D|
00000050  30 34 30 36 30 30 32 36  31 0a 00 00 00 fd 00 32  |040600261......2|
00000060  b4 1e 61 18 00 0a 20 20  20 20 20 20 00 00 00 fc  |..a...      ....|
00000070  00 47 39 30 66 2b 0a 20  20 20 20 20 20 20 00 e4  |.G90f+.       ..|

After finishing this I sent the file to my friend and we suffered from another issue that I quickly resolved before he even realized I had no clue what I was getting into.  How exactly do we set this EDID file without overwriting the monitors EEPROM (since there is nothing wrong, this is merely a driver issue, which we concluded was true as everything worked fine in Windows).  The solution is to override it in the driver through that really annoying xorg.conf file.  Oh yes so where's the nice trusty document for that?  There isn't any.  At least nothing as official as a nicely wrote and formatted PDF.  Instead a few man page searching, and some google search to offset the pain resulted in a nice Option override to specify your own EDID file.

So after asking him to move the file into /etc/X11 and giving it the name edid.bin I told him to add
Option "CustomEDID" "CRT-0:/etc/X11/edid.bin"
to his xorg.conf gile under the "Device" section (as per the documentation states).  After starting X everything was promising at first.  It was in a higher resolution by default.  But not as high as he expected it should be.  After he fiddled inside nvidia-settings finally the resolution options were available and it was trivial to change (something that he couldn't do before)

After all was said and done, a job well done was expected, but instead I got this:
<hirato> I am so gay for you right now

Thursday, June 21, 2012

Writing a primitive dynamic loader for Linux

The process of dynamic loading libraries is a complex one.  Infact too complex to elaborate and explain on a blog site.  Thousands of man hours and research has been wasted on dynamic loading.  Things like PIC (Position independant code) take time to perfect and implement and are often subjected to bugs and performance issues. Of course you're not here to listen to me ramble about it.

What I'm going to show you is the bread and butter of writing your own user-space dynamic loader by getting down to the bits and bytes of it.  Before we continue I should suggest you have a working toolchain which includes a compiler (gcc, clang, pathscale will work) and binutils (objdump and objcopy).

Lets start with a simple library function:

int print_hello_world() {
    return printf("Hello World!\n");

This is the function we're going to want to _dynamically_ load from our dynamic loader and execute.  Lets begin by compiling this into an object with gcc.

# gcc hello.c -c hello.o

I assume you know assembly (because you're going to need it) if not here's a small crash course.  Assembly language is hardware-level language.  Every thing done in code is directly translated to a byte-code encoding of various instructions that move arguments in and off the stack, move values and memory around and perform basic arithmetic.  It's very low-level and working with it is nearly impossible.  Which is why there exists assemblers that take an input assembly file (human readable version of the bytes) and assembles it to a (non human readable but computer readable) format.

The first thing we want to do is look at the dissasembly of  our hello.o object file.  To do this we use objdump

# objdump -d hello.o

0000000000000000 <print_hello_world>:
   0:    55                  push  %rbp
   1:    48 89 e5            mov    %rsp,%rbp
   4:    bf 00 00 00 00      mov    $0x0,%edi
   9:    e8 00 00 00 00      callq   e <print_hello_world+0xe>
   e:    5d                  pop    %rbp
   f:    c3                  retq

Note: the output may vary depending on your architecture.
Lets get a few things figured out first.  As you can see we have some basic stack manipulation in this function in accordance to the SysV ABI.  But the thing worth nothing is the `callq`.  The printf function doesn't exist at this point, during link the linker will patch the call with a call to library printf.  We don't want that, why?  because printf itself is a library function which exists in a dynamically loaded library.  We can't use it.  So how do we perform a printf without a printf?

Welcome to the world of system calls.  System calls are the process of invoking the kernel.  Every major OS has system calls of some sort.  But for the sake of linux there is a very specific way at performing a systemcall.  Linux expects arguments for a systemcall to exist in registers in the order of EAX,EBX,ECX,EDX etc etc.  Once the registers contain the values required simply calling int 128 (int 0x80) will perform the systemcall.

So what systemcall does printf?  None.  Instead there is write, which writes to a file descriptor.  The write systemcall expects the fd to write to, the data, and the size of the data.  So lets rewrite our test.c file to accomodate that.

int print_hello_world() {
    const char hello[] = "Hello World!\n";
    const int  size       = sizeof(hello);
    int        ret;
    __asm__ __volatile__ (
        "movl $4,  %%eax\n\t"
        "movl $1,  %%ebx\n\t"
        "movl %1, %%ecx\n\t"
        "movl %2, %%edx\n\t"
        "int $0x80" :
            "=a"(ret) : "g"(hello), "g"(size) :
            "%ebx", "%ecx", "%edx", "%esi", "%edi"
    return ret;

Looks fancy doesn't it?  Lets work out this assembly first
the movl for 4 is the syscall number for write
the movl for 1 is the file descriptor for stdout

This code will output "Hello World!\n" to stdout.  Pretty awesome eh?
Of course now we need to take a look at the assembly of this.

0000000000000000 <print_hello_world>:
   0:    55                        push   %rbp
   1:    48 89 e5                  mov    %rsp,%rbp
   4:    41 54                     push   %r12
   6:    53                        push   %rbx
   7:    c7 45 d0 48 65 6c 6c      movl   $0x6c6c6548,-0x30(%rbp)
   e:    c7 45 d4 6f 20 57 6f      movl   $0x6f57206f,-0x2c(%rbp)
  15:    c7 45 d8 72 6c 64 21      movl   $0x21646c72,-0x28(%rbp)
  1c:    66 c7 45 dc 0a 00         movw   $0xa,-0x24(%rbp)
  22:    c7 45 ec 0e 00 00 00      movl   $0xe,-0x14(%rbp)
  29:    48 8d 45 d0               lea    -0x30(%rbp),%rax
  2d:    48 89 45 c8               mov    %rax,-0x38(%rbp)
  31:    b8 04 00 00 00            mov    $0x4,%eax
  36:    bb 01 00 00 00            mov    $0x1,%ebx
  3b:    8b 4d c8                  mov    -0x38(%rbp),%ecx
  3e:    8b 55 ec                  mov    -0x14(%rbp),%edx
  41:    cd 80                     int    $0x80
  43:    41 89 c4                  mov    %eax,%r12d
  46:    44 89 65 e8               mov    %r12d,-0x18(%rbp)
  4a:    8b 45 e8                  mov    -0x18(%rbp),%eax
  4d:    5b                        pop    %rbx
  4e:    41 5c                     pop    %r12
  50:    5d                        pop    %rbp
  51:    c3                        retq 

Pay special attention to the encoding not the instructions, this is the bytecode the CPU understands and executes, it all encodes to a special meaning, this meaning is on the right.

Normally there would be a problem with this code.  Currently it's PIC, but it might not be PIC (for example if you rely on data outside the function).  Note the three successive movl's at the top, these are encoding for hello world and the size (but they could be addresses if you're not paying special attention).  Simply loading code that uses addresses to read from memory and executing it would cause a segmentation fault (since we don't have access to that memory this).  Instead we need a surfire way to make sure the code is position independant (even if it already is).   Doing it requires a lot of work, you need relocation tables, and patching of all those addresses.  Instead we can pass this information by registers from the dynamic loader (this isn't very usefull but this tutorial is only to explain and implement a primitive dynamic loader).

So lets make this code PIC by rewriting it, to something like this:

int print_hello_world(const char *hello, int size, int *ret) {
    __asm__ __volatile__ (
        "movl $4, %%eax\n\t"
        "movl $1, %%ebx\n\t"
        "movl %1, %%ecx\n\t"
        "movl %2, %%edx\n\t"
        "int $0x80" :
            "=a"(*ret) : "g"(hello), "g"(size) :
            "%ebx", "%ecx", "%edx", "%esi", "%edi"
    return *ret;

as you can see no data on the heap.  Everything fits in registers and is easily PIC.  Of course you cannot make this assumption, lets check the dissasembly.

#objdump -d hello.o

   0:    55                       push   %rbp
   1:    48 89 e5                 mov    %rsp,%rbp
   4:    53                       push   %rbx
   5:    48 89 7d f0              mov    %rdi,-0x10(%rbp)
   9:    89 75 ec                 mov    %esi,-0x14(%rbp)
   c:    48 89 55 e0              mov    %rdx,-0x20(%rbp)
  10:    b8 04 00 00 00           mov    $0x4,%eax
  15:    bb 01 00 00 00           mov    $0x1,%ebx
  1a:    8b 4d f0                 mov    -0x10(%rbp),%ecx
  1d:    8b 55 ec                 mov    -0x14(%rbp),%edx
  20:    cd 80                    int    $0x80
  22:    41 89 c0                 mov    %eax,%r8d
  25:    48 8b 45 e0              mov    -0x20(%rbp),%rax
  29:    44 89 00                 mov    %r8d,(%rax)
  2c:    48 8b 45 e0              mov    -0x20(%rbp),%rax
  30:    8b 00                    mov    (%rax),%eax
  32:    5b                       pop    %rbx
  33:    5d                       pop    %rbp
  34:    c3                       retq 

Hey Look! Not bad, everything is happening between the registers! This is PIC!

Now we need to break this down further! Currently these object files we're working with are in ELF format and not really easy to load.  What if we just rewrote those bytes on the left side of that dissasembly in a file manually we'd have a flat binary right?  Wishfull thinking, that would take too long, this is what objcopy is for.

#objcopy --only-section=.text --output-target binary hello.o hello.bin

We specify we only want the .text section (other sections are useless) .text = code.  We can thus validate the instructions to the objdump using hexdump

# hexdump -C hello.bin

00000000  55 48 89 e5 53 48 89 7d  f0 89 75 ec 48 89 55 e0  |UH..SH.}..u.H.U.|
00000010  b8 04 00 00 00 bb 01 00  00 00 8b 4d f0 8b 55 ec  |...........M..U.|
00000020  cd 80 41 89 c0 48 8b 45  e0 44 89 00 48 8b 45 e0  |..A..H.E.D..H.E.|
00000030  8b 00 5b 5d c3                                    |..[].|

As we can validate the encode is the same as the objdump.  We now have loadable code that we can EXECUTE!  Lets write our dynamic executor! 

#include <stdio.h>
#include <stdlib.h>

unsigned char *mem;
int            i;
long int       size;
int            ret;
int main() {
    FILE *fp = fopen("hello.bin", "rb");
    if  (!fp) {
        printf("[ld] Failed to load hello.bin\n");
        return -1;
    fseek(fp, 0, SEEK_END);
    size = ftell(fp);
    fseek(fp, 0, SEEK_SET);
    printf("[ld] Loading %d byte program\n", size);
    if (!(mem = malloc(size))) {
        printf("[ld] Failed to allocate memory for code\n");
        return -1;
    } else {
        printf("[ld] Reserved %d bytes for program\n", size);
    fread(mem, 1, size, fp);
    printf("[ld] Loaded program\n");
    printf("    ");
    for(i = 1; i < size+1; i++)
        printf((i%8==0)?"0x%02X\n    ":"0x%02X, ", mem[i-1]);
    printf("\n[ld] Executing ...\n");
    int (*call)(const char *, int, int*) = mem;
    (int)(*call)("Hello World\n", sizeof("Hello World\n"), &ret);
    printf("[ld] Called function; returned %d strlen(\"Hello World\")\n", ret);
    return ret;

Now here's the problem if we compile and execute this code:

# gcc run.c -o run
# ./run

[ld] Loading 53 byte program
[ld] Reserved 53 bytes for program
[ld] Loaded program
    0x55, 0x48, 0x89, 0xE5, 0x53, 0x48, 0x89, 0x7D
    0xF0, 0x89, 0x75, 0xEC, 0x48, 0x89, 0x55, 0xE0
    0xB8, 0x04, 0x00, 0x00, 0x00, 0xBB, 0x01, 0x00
    0x00, 0x00, 0x8B, 0x4D, 0xF0, 0x8B, 0x55, 0xEC
    0xCD, 0x80, 0x41, 0x89, 0xC0, 0x48, 0x8B, 0x45
    0xE0, 0x44, 0x89, 0x00, 0x48, 0x8B, 0x45, 0xE0
    0x8B, 0x00, 0x5B, 0x5D, 0xC3,
[ld] Executing ...
Segmentation fault

Wait what! Segmentation fault? There is some issues preventing us from executing code from the stack.  This is a security feature, and simple to fix.  First we need to tell GCC to not build with stack protection and pass the execstack flag to ld so we can execute code from the stack.  Lets try again

# gcc -fno-stack-protector -z execstack run.c -o run
# ./run
[ld] Loading 53 byte program
[ld] Reserved 53 bytes for program
[ld] Loaded program
    0x55, 0x48, 0x89, 0xE5, 0x53, 0x48, 0x89, 0x7D
    0xF0, 0x89, 0x75, 0xEC, 0x48, 0x89, 0x55, 0xE0
    0xB8, 0x04, 0x00, 0x00, 0x00, 0xBB, 0x01, 0x00
    0x00, 0x00, 0x8B, 0x4D, 0xF0, 0x8B, 0x55, 0xEC
    0xCD, 0x80, 0x41, 0x89, 0xC0, 0x48, 0x8B, 0x45
    0xE0, 0x44, 0x89, 0x00, 0x48, 0x8B, 0x45, 0xE0
    0x8B, 0x00, 0x5B, 0x5D, 0xC3,
[ld] Executing ...
Hello World
[ld] Called function; returned 13 strlen("Hello World")

Look at that we printed HELLO WORLD! I should note this isn't the same as dynamic loading but very close in princible.  I hope you learn how all this works and can adapt it for future logic if you ever implement a dynamic loader or evn linker!.  I should disclaimer here that dynamic linkers are a lot more complex, they handle relocation tables to relocate data, and do PIC via binary patching and other complex methods.  But this is a good start!

Some other subtle notes that aren't covered here.

1 The data cache on the CPU should be cleared as well as the instruction cache since there is no gurantee the data and instruction cache will be in sync.

2 The memory for the code should be allocated with mmap and set unprotected (instead of using GCC trickery like I did)

Friday, April 6, 2012

Addition the bit hack way

Disclaimer:  I would ask that none of this code actually be used in software; it doesn't offer anything, if your CPU has an ADD instruction, use it.  Otherwise this is just a teaching aid.

Addition is a basic mathematical operation which can be emulated with some bitwise operations.  Lets look at a canonical implementation:

uint8_t add(uint8_t x, uint8_t y) {
  uint8_t a;
  uint8_t b;

  do {
    a = x & y;
    b = x ^ y;
    x = a << 1;
    y = b;
  } while (a);

  return b;

To understand what is going on lets follow out with a basic example:
Let in:x = 00000011 (3)
Let in:y = 00000100 (4)

When we invoke this add function the steps taking place are:
  • AND   -- a = (x&y) 
  • XOR   -- b = (x^y)
  • LSHFT -- x = (a<<1)
Resulting in the following:
  • x = 00000011
  • y = 00000100
  • a = 00000000
  • b = 00000100
  • x = 00000000
  • y = 00000111  -- XOR a:b = 7
This canonical implementation works, but it's slow.  We can make it slightly faster, to see why, lets take a look at the assembly that code translates to (note: all code is x86_64 non-optimized)
  pushq %rbp
  movq  %rsp,   %rbp
  movl  %edi,   %edx
  movl  %esi,   %eax
  movb  %dl -20(%rbp) ; y
  movb  %al -24(%rbp) ; x
    movzbl -24(%rbp),   %eax 
    movzbl -20(%rbp),   %edx
    andl       %edx,    %eax
    movb       %al,  -1(%rbp)
    movzbl -24(%rbp),   %eax
    movzbl -20(%rbp),   %edx
    xorl       %edx,    %eax
    movb       %al,  -2(%rbp)
    movzbl  -1(%rbp),   %eax
    addl       %eax,    %eax
    movb       %al, -20(%rbp)
    movzbl  -2(%rbp),   %eax
    movb       %al, -24(%rbp)
    cmpb       $0x0, -1(%rbp) 
    jne .loop
  movzbl -2(%rbp), %eax
  popq             %rbp

There is 9+(15*iterations) worth instructions here.  Assuming the best case senerio, that is 24 instructions.  We can refine our implementation to be slightly faster.
void add(uint8_t *x, uint8_t *y) {
  do {
    *y = (*x^*y);
    *x = (*x&(*x^*y)) << 1;

  } while (*x);

In this version we've removed the return, and replaced our arguments with pointers to our variables, storing the result of the addition in *y, and using no temporaries!.  This is (best case) 14! instruction less:
  movzbl (%rsi), %eax
    movzbl (%rdi),      %ecx
    movl    %eax,       %edx
    andl    %ecx,       %edx
    xorl    %ecx,       %eax
    leal   (%rdx,%rdx), %ecx
    testb   %dl,        %dl
    movb    %cl      (%rdi)
    movb    %al      (%rsi)
    jne .loop

Of course since this function is small we can most likely make it inline, (which could allow for better optimization).  I'd personally perfer a macro before the inline keyword.

#define add(x,y) do{*y=(*x^*y),*x=(*x&(*x^*y))<<1;} while(*x)

That concludes addition by bit hacking. I've played around with many other methods, but this is 3+(9*iterations) worth instructions, or best case (11 instructions). I challenge anyone to beat this for best case.  Must be C code (no assembly) .. post solution in the comments.  Hint: abuse stack.

Thursday, April 5, 2012

Checking if divisible by 10 or 5 the bitwise way.

My Problem: check if something was divisble by 10, the rules where I couldn't use the modulus, divison or multiplication operator. The check had to be in the form of a single expression, no branches or loop constructs could be used, pointer arithmetic wasn't allowed, neither was the use of floating point data types. The code had to work on a x86 CPU; and had to be ANSI C.

After spending a few hours I've come up with the following solution:

Now, this isn't exactly a fast solution, expecially on a P4 which lacks a barrel shifter, however this solution led me finding a faster soluton. The rules where I couldn't use the multiplication operator. But after removing those shifted adds and replacing it with a constant 0xCCCCCCCD, I had the following:

(x*0xCCCCCCCD) < 0x33333334 &&!(x&1)

much to my surprise this is faster than using modulus (x%10) as well as the division operator. MUL on the x86 is actually the fastest way to do division for NPOT (non-power-of-two) constants. This can be proofed with:
(x * 5^-1) < 0x33333334

Explaining how this works is quite simple: the constant 0xCCCCCCCD is the inverse of 5 modulo 2^32. Of course I couldn't stop here. I took it a bit further and came up with the following code:

!((0x19999999 * x + (x >> 1) + (x >> 3)) >> 28) [

This is a lot more complicated to proof, mainly because it's simply odd code. To get the most obvious odd bit sorted out C allows array[idx] and idx[array] because both cases decay to *(idx+array) or *(array+idx) which are equivlent. Now for the explination:

In the table we have here: the following: 0,1,2,2,3,3,4,5,5,6,7,7,8,8,9,0
this maps the multiples of 5 to either 0 or 8. Better expressed as: (8/5 * x) % 16 (the error term is never too big) -- The real equivalency being:
((8/5 + 0x0.00000002/5) * x - delta) % 16 for delta in [0, 0.625/2^28]
This one uses a lot more operations to do the same thing, 3 shifts, 2 adds, and one MUL, still a lot faster on a pentium D than the former one. Of course I took it two more steps further.

The rules where it had to be ANSI C, but I sort of broke that bit, by exposing some x86 assembly (which entailed the best solutions).

ROR x, 0x19999999
CMP x, 0x19999999
JLE hit_table ; same table {0,1,2,2,3,3,4,5,5,6,7,7,8,8,9,0}

This it the same to the CPU as the one above, but expressed in a simple fashion. Essentially the observation is:

x * 0xCCCCCCCD is even exactly if x is even, because 0xCCCCCCCD is odd, so the right rotation to move the last bit to the first and compare against 0x19999999 will give you the same answer, that is if you have ROR. So I rewrote it another way for ROR-less systems.

This method is the same as the first: which I find is a lot more nicer, too many MOVs though, but the real power comes from the x86 LEA instruction.

MOV %edx, 0x33333334
MOV %ebx, %eax
MUL %edx
AND %edx, x
LEA %edx, [%edx+%edx*4]

The reason this works is the x86 MUL instruction puts it's double width result in two registers, the sad part is you cannot write this in C.

But in the end the winner for performance ended up being:
(x * 0xCCCCCCCD) < 0x33333334 && !(x&1)