Inline assembly

Remarks#

Inline assembly is the practice of adding assembly instructions in the middle of C source code. No ISO C standard requires support of inline assembly. Since it is not required, the syntax for inline assembly varies from compiler to compiler. Even though it is typically supported there are very few reasons to use inline assembly and many reasons not to.

Pros

1. Performance By writing the specific assembly instructions for an operation, you can achieve better performance than the assembly code generated by the compiler. Note that these performance gains are rare. In most cases you can achieve better performance gains just by rearranging your C code so the optimizer can do its job.
2. Hardware interface Device driver or processor startup code may need some assembly code to access the correct registers and to guarantee certain operations occur in a specific order with a specific delay between operations.

Cons

1. Compiler Portability Syntax for inline assembly is not guaranteed to be the same from one compiler to another. If you are writing code with inline assembly that should be supported by different compilers, use preprocessor macros (#ifdef) to check which compiler is being used. Then, write a separate inline assembly section for each supported compiler.
2. Processor Portability You can’t write inline assembly for an x86 processor and expect it to work on an ARM processor. Inline assembly is intended to be written for a specific processor or processor family. If you have inline assembly that you want supported on different processors, use preprocessor macros to check which processor the code is being compiled for and to select the appropriate assembly code section.
3. Future Performance Changes Inline assembly may be written expecting delays based upon a certain processor clock speed. If the program is compiled for a processor with a faster clock, the assembly code may not perform as expected.

gcc Basic asm support

Basic assembly support with gcc has the following syntax:

asm [ volatile ] ( AssemblerInstructions )

where AssemblerInstructions is the direct assembly code for the given processor. The volatile keyword is optional and has no effect as gcc does not optimize code within a basic asm statement. AssemblerInstructions can contain multiple assembly instructions. A basic asm statement is used if you have an asm routine that must exist outside of a C function. The following example is from the GCC manual:

 /* Note that this code will not compile with -masm=intel */
 #define DebugBreak() asm("int $3")

In this example, you could then use DebugBreak() in other places in your code and it will execute the assembly instruction int $3. Note that even though gcc will not modify any code in a basic asm statement, the optimizer may still move consecutive asm statements around. If you have multiple assembly instructions that must occur in a specific order, include them in one asm statement.

gcc Extended asm support

Extended asm support in gcc has the following syntax:

asm [volatile] ( AssemblerTemplate
                  : OutputOperands
                  [ : InputOperands
                  [ : Clobbers ] ])
 
 asm [volatile] goto ( AssemblerTemplate
                       :
                       : InputOperands
                       : Clobbers
                       : GotoLabels)

where AssemblerTemplate is the template for the assembler instruction, OutputOperands are any C variables that can be modified by the assembly code, InputOperands are any C variables used as input parameters, Clobbers are a list or registers that are modified by the assembly code, and GotoLabels are any goto statement labels that may be used in the assembly code.

The extended format is used within C functions and is the more typical usage of inline assembly. Below is an example from the Linux kernel for byte swapping 16-bit and 32-bit numbers for an ARM processor:

/* From arch/arm/include/asm/swab.h in Linux kernel version 4.6.4 */
#if __LINUX_ARM_ARCH__ >= 6

static inline __attribute_const__ __u32 __arch_swahb32(__u32 x)
{
    __asm__ ("rev16 %0, %1" : "=r" (x) : "r" (x));
    return x;
}
#define __arch_swahb32 __arch_swahb32
#define __arch_swab16(x) ((__u16)__arch_swahb32(x))

static inline __attribute_const__ __u32 __arch_swab32(__u32 x)
{
    __asm__ ("rev %0, %1" : "=r" (x) : "r" (x));
    return x;
}
#define __arch_swab32 __arch_swab32

#endif

Each asm section uses the variable x as its input and output parameter. The C function then returns the manipulated result.

With the extended asm format, gcc may optimize the assembly instructions in an asm block following the same rules it uses for optimizing C code. If you want your asm section to remain untouched, use the volatile keyword for the asm section.

gcc Inline assembly in macros

We can put assembly instructions inside a macro and use the macro like you would call a function.

#define mov(x,y) \
{ \
    __asm__ ("l.cmov %0,%1,%2" : "=r" (x) : "r" (y), "r" (0x0000000F)); \
}

/// some definition and assignment
unsigned char sbox[size][size];
unsigned char sbox[size][size];

///Using
mov(state[0][1], sbox[si][sj]);

Using inline assembly instructions embedded in C code can improve the run time of a program. This is very helpful in time critical situations like cryptographic algorithms such as AES. For example, for a simple shift operation that is needed in the AES algorithm, we can substitute a direct Rotate Right assembly instruction with C shift operator >>.

In an implementation of ‘AES256’, in ‘AddRoundKey()’ function we have some statements like this:

unsigned int w;          // 32-bit
unsigned char subkey[4]; // 8-bit, 4*8 = 32 

subkey[0] = w >> 24;     // hold 8 bit, MSB, leftmost group of 8-bits 
subkey[1] = w >> 16;     // hold 8 bit, second group of 8-bit from left    
subkey[2] = w >> 8;      // hold 8 bit, second group of 8-bit from right
subkey[3] = w;           // hold 8 bit, LSB, rightmost group of 8-bits

/// subkey <- w

They simply assign the bit value of w to subkey array.

We can change three shift + assign and one assign C expression with only one assembly Rotate Right operation.

__asm__ ("l.ror  %0,%1,%2" : "=r" (* (unsigned int *) subkey)  : "r" (w), "r" (0x10));

The final result is exactly same.