同时循环遍历数组,MIPS汇编

时间:2014-09-14 13:02:14

标签: c assembly mips

我想将 while(w [i] == x)i + = j; 转换为MIPS汇编代码。 假设INTEGERS i,j和x分别为3美元,4美元和5美元。还假设在while循环之前最初i = 0。 w =>整数数组及其基地址存储在$ 6中。到目前为止,我有这个。

Loop:
    sll $10, $3, 2    # $10 = i* 4
    add $10, $6, $10  # $10 has address of w[i]
    lw  $11, 0($10)   # $11 = w[i]
    bne $11, $5, Exit # exit from loop if w[i]!= x
    add $3,  $3, $4   # i= i+ j
    j Loop
Exit:

是否可以通过j * 4移动基址本身来优化此代码,并且还可以删除多个分支指令?因为我不清楚如何做到这一点。

提前致谢!

2 个答案:

答案 0 :(得分:0)

为了摆脱多分支指令,可以使用这个技巧:
警告:不完全等同于您的代码

Loop:
    sll $10, $3, 2    # $10 = i* 4
    add $10, $6, $10  # $10 has address of w[i]
    lw  $11, 0($10)   # $11 = w[i]
    add $3,  $3, $4   # i = i + j
    beq $11, $5, Loop # keep looping if w[i] == x
Exit:
    sub $3,  $3, $4   # i = i - j

诀窍是在测试保持循环之前执行i += j 这确实会引入一个问题:当你的代码没有时,可能会触发一个额外的整数溢出。

EDITED:

这就像重写一样:

while (some_condition())
    do_something();

进入这个:

do
    do_something();
while (some_condition());
undo_something();

EDITED:

好吧,让我试着“这次用j * 4将指针从基地址移动”:)

Start:
    sll $11, $3, 2    # $11 = i * 4
    add $10, $11, $6  # Let $10 be a "cursor" pointing to w[i]        
Loop:
    lw  $11, 0($10)   # $11 = w[i]
    sll $12, $4, 2    # $12 = j * 4
    add $10, $10, $12 # update $10 by 4 * j
    add $3,  $3, $4   # update i by j
    beq $11, $5, Loop # keep looping if w[i] == x
Exit:
    sub $3,  $3, $4   # i = i - j

然而,它并没有比上面给出的版本更优化:它们都在循环体内使用了5条指令。

答案 1 :(得分:0)

为了使比较更容易,我写了一个小的虚函数:

#include <stdint.h>
#include <stdlib.h>

uint32_t fun1(uint32_t const *in, uint32_t cmp, size_t j)
{
  size_t i = 0;
  while (in[i] == cmp)
    {
      i += j;
    }
  return in[i];
}

这可以编译,输出可以与等效函数进行比较:

uint32_t fun2(uint32_t const *in, uint32_t cmp, size_t j)
{
  while (*in == cmp)
    {
      in += j;
    }
  return *in;
}

对于两个函数,gcc(x86-64上的4.8)生成一个只有4条指令的循环。 对于第二个功能,它基本上是:

temp1 = (in)
compare temp1, cmp
if not equal, return temp1
temp2 = j*sizeof(uint32_t)

loop:
in += temp2            #\
temp1 = (in)           # \
compare temp1, cmp     #  - 4 instructions loop
if equal, goto loop    # /

return temp1

这样的伪装配可以像MIPS那样实现:

lw   $v0, 0($a0)
beq  $a1, $v0, end
sll  $t1, $a2, 2

loop:
add  $a0, $a0, $t1      #\
lw   $v0, 0($a0)        # - only 3 instructions in loop, due to test-and-branch
bne  $a1, $v0, loop     #/

end:
jr $ra