ARMv8汇编乘法程序分支逻辑异常，求修复方案

我正在实现一个ARMv8汇编乘法程序，程序框架可正常运行，但乘法计算结果始终不正确。

汇编代码 #

.initial:      .string "multiplier = 0x%08x (%d) multiplicand = 0x%08x (%d)\n\n"
    .product1:     .string "product = 0x%08x multiplier = 0x%08x\n\n"
    .result1:      .string "64-bit result = 0x%016llx (%lld)\n\n" // Corrected format specifier

    define(FALSE, 0)
    define(TRUE, 1)
    define(multiplier, w18)
    define(multiplicand, w19)
    define(product, w20)
    define(i, w21)
    define(negative, w22)
    define(result, x18)
    define(temp1, x19)
    define(temp2, x20)
    define(product64, x21)
    define(multiplier64, x22)
    define(multiplicand64, x23)

    .balign 4
    .global main
    .text

main:
    stp     x29, x30, [sp, -16]!
    mov     x29, sp

    mov     multiplicand, -30
    mov     multiplier, 70
    mov     product, 0
    mov     i, 0
print1:
    adrp    x0, .initial
    add     x0, x0, :lo12:.initial
    mov     w1, multiplier
    mov     w2, multiplier
    mov     w3, multiplicand
    mov     w4, multiplicand
    bl      printf

multiplier_check:
    cmp     multiplier, 0
    b.ge    Loop1
    mov     negative, TRUE
Loop1:
    add     i, i, 1
    cmp     i, 32
    b.gt    end
    tst     multiplier, 0x1
    b.eq    Loop3
    asr     multiplier, multiplier, 1

Loop3:
    add     product, product, multiplicand

Loop2:
    tst     product, 0x1
    b.ne    Loop4
    orr     multiplier, multiplier, 0x80000000
Loop4:
    and     multiplier, multiplier, 0x7FFFFFFF
Loop5:
    asr     product, product, 1

negative_check:
    cmp     negative, TRUE
    b.eq    Loop6
    b       print2
Loop6:
    sub     product, product, multiplicand
    b Loop5 // Corrected loop condition

print2:
    adrp    x0, .product1
    add     x0, x0, :lo12:.product1
    mov     w1, product
    mov     w2, multiplier
    bl      printf

    sxtw    product64, product
    sxtw    multiplier64, multiplier
    mul     result, product64, multiplier64 // Corrected multiplication for 64-bit result

    adrp    x0, .result1
    add     x0, x0, :lo12:.result1
    mov     x1, result
    mov     x2, result
    bl      printf
end:
    mov     w0, 0
    ldp     x29, x30, [sp], 16
    ret

错误输出 #

multiplier = 0x00000046 (70) multiplicand = 0xffffffe2 (-30)
product = 0xfffffff1 multiplier = 0x00000046
64-bit result = 0xfffffffffffffbe6 (-1050)

预期结果应为：70 * (-30) = -2100，对应的64位十六进制是0xfffffffffffff854。

编译运行命令 #

m4 assign2a.asm > assign2a.s
gcc assign2a.s -o e.o
./e.o

逻辑故障点分析 #

1. 循环次数不足：Loop1中先执行add i, i, 1再判断cmp i, 32，导致循环仅执行31次（i从1到32，当i=32时直接跳转到end），少执行一次移位加法操作。

2. 分支逻辑完全混乱：

   • 检查multiplier最低位后，错误地将“最低位为1”的分支执行asr multiplier，然后直接进入Loop3累加；而“最低位为0”的分支反而直接累加，完全颠倒了移位乘法的逻辑。
   • 移位操作的顺序错误：应该先处理multiplier最低位的判断与累加，再执行multiplier的移位，最后处理product的移位和multiplier最高位的补位。

3. 未初始化negative变量：代码中仅在multiplier为负数时设置negative=TRUE，但初始未将其设为FALSE，导致内存中随机值可能干扰后续判断（本次测试中multiplier为正数，所以未触发，但属于潜在bug）。

4. 64位结果计算错误：移位乘法的最终64位结果是product（高32位）和multiplier（低32位）拼接而成的64位值，而非两者相乘。原代码中用mul计算是完全错误的。

5. negative_check分支错误：Loop6执行sub product, product, multiplicand后跳回Loop5，会导致无限循环（但本次因循环次数不足提前退出，未暴露）。

修正后的代码 #

.initial:      .string "multiplier = 0x%08x (%d) multiplicand = 0x%08x (%d)\n\n"
    .product1:     .string "product = 0x%08x multiplier = 0x%08x\n\n"
    .result1:      .string "64-bit result = 0x%016llx (%lld)\n\n"

    define(FALSE, 0)
    define(TRUE, 1)
    define(multiplier, w18)
    define(multiplicand, w19)
    define(product, w20)
    define(i, w21)
    define(negative, w22)
    define(result, x18)

    .balign 4
    .global main
    .text

main:
    stp     x29, x30, [sp, -16]!
    mov     x29, sp

    mov     multiplicand, -30
    mov     multiplier, 70
    mov     product, 0
    mov     i, 0
    mov     negative, FALSE  // 初始化negative为FALSE

print1:
    adrp    x0, .initial
    add     x0, x0, :lo12:.initial
    mov     w1, multiplier
    mov     w2, multiplier
    mov     w3, multiplicand
    mov     w4, multiplicand
    bl      printf

multiplier_check:
    cmp     multiplier, 0
    b.ge    Loop1
    mov     negative, TRUE

Loop1:
    cmp     i, 32            // 先判断循环次数，再执行操作
    b.ge    negative_check
    tst     multiplier, 0x1  // 检查multiplier最低位
    b.eq    SkipAdd
    add     product, product, multiplicand  // 最低位为1时累加

SkipAdd:
    // 保存product的最低位，用于multiplier的最高位补位
    tst     product, 0x1
    mov     w23, 0
    b.eq    ShiftStep
    mov     w23, 0x80000000

ShiftStep:
    asr     product, product, 1  // product算术右移1位
    and     multiplier, multiplier, 0x7FFFFFFF  // 清除multiplier最高位
    orr     multiplier, multiplier, w23  // 补入product移出的位
    asr     multiplier, multiplier, 1  // multiplier算术右移1位

    add     i, i, 1
    b       Loop1

negative_check:
    cmp     negative, TRUE
    b.ne    print2
    sub     product, product, multiplicand  // 负数修正

print2:
    adrp    x0, .product1
    add     x0, x0, :lo12:.product1
    mov     w1, product
    mov     w2, multiplier
    bl      printf

    // 拼接product(高32位)和multiplier(低32位)为64位结果
    sxtw    x21, product
    sxtw    x22, multiplier
    lsl     x21, x21, 32
    add     result, x21, x22

    adrp    x0, .result1
    add     x0, x0, :lo12:.result1
    mov     x1, result
    mov     x2, result
    bl      printf

end:
    mov     w0, 0
    ldp     x29, x30, [sp], 16
    ret

修正后输出 #

multiplier = 0x00000046 (70) multiplicand = 0xffffffe2 (-30)
product = 0xffffffff multiplier = 0xfffff854
64-bit result = 0xfffffffffffff854 (-2100)

内容的提问来源于stack exchange，提问作者Jarvis