Performance of Unaligned Memory Access in Raspberry Pi

SoC of Raspberry pi is a Broadcom BCM2835. This contains an ARM1176JZFS, ARMv6 architecture.

ARMv6 adds unaligned word(4 bytes) and halfword load and store data access support.
For detail, please check.

• ARM1176JZF-S Technical Reference Manual: 4.2. Unaligned access support
• ARM unaligned data access and floating point in Linux 

I would like to test performance when a user space process invokes a unaligned memory access (only load).
Here is my test environment and cases.

[Environment]

---

Kernel: Linux xbian 3.6.1 #4 PREEMPT Thu Nov 8 18:54:20 CET 2012 armv6l GNU/Linux
GCC: gcc version 4.6.3 (Debian 4.6.3-12+rpi1)

# cat /proc/cpuinfo
Processor       : ARMv6-compatible processor rev 7
                  (v6l)
BogoMIPS        : 697.95
Features        : swp half thumb fastmult vfp edsp
                  java tls
CPU implementer : 0x41
CPU architecture: 7
CPU variant     : 0x0
CPU part        : 0xb76
CPU revision    : 7

Hardware        : BCM2708
Revision        : 0002
Serial          : 0000000009b752ff

[Program]

---

unaligned.c

#include <stdio.h>
#include <stdint.h>

#ifdef USE_GCC_FIXUP
struct __una_u32 {
    uint32_t x;  
} __attribute__((packed));

static inline uint32_t get_unaligned_32(const void *p)
{
    const struct __una_u32 *ptr = 
        (const struct __una_u32 *)p;
    return ptr->x;
}
#endif

int main()
{
    uint8_t buf[16];
    uint32_t i, j;

    for (i = 0; i < sizeof(buf); i++)
        buf[i] = i;

    for (j = 0; j < 100000000; j++) {
        /* unaligned access */
#ifndef USE_GCC_FIXUP
        i = *(unsigned int*)(&buf[1]);
#else
        i = get_unaligned_32(&buf[1]);
#endif
    }

    printf("0x%X\n", i);

    return 0;
}

[Case]

---

1. unaligned word access
   
   • fix up by hardware
     # gcc -c unaligned.c
     # time ./unaligned
     0x4030201
     
     real    0m2.053s
     user    0m2.040s
     sys     0m0.010s
   
   
   • fix up by software (gcc)
     # gcc -DUSE_GCC_FIXUP -o unaligned_gcc unaligned.c
     # time ./unaligned_gcc
     0x4030201
     
     real    0m6.384s
     user    0m6.370s
     sys     0m0.010s
   
   The result is very clear.
   
   Here I add a case is aligned access.
   Only modify:
   

   #ifndef USE_GCC_FIXUP
           i = *(unsigned int*)(&buf[0]);
   #else
   

   # time ./aligned
   0x3020100
   
   real    0m1.934s
   user    0m1.920s
   sys     0m0.000s


2. unaligned double words access
   Modify unaligned.c to support double words access
   • fix up by kernel
     # time ./unaligned64
     0x807060504030201
     
     real    1m8.754s
     user    0m7.800s
     sys     1m0.830s
     
     
   • fix up by software (gcc)
     # time ./unaligned64_gcc
     0x807060504030201
     
     real    0m9.753s
     user    0m9.700s
     sys     0m0.030s
   
   Also add a case for aligned access.
   # time ./aligned64
   0x706050403020100
   
   real    0m2.413s
   user    0m2.400s
   sys     0m0.000s

So, to avoid unaligned memory access if possible!