对 printf 的调用完全支配了即使是效率极低的循环的运行时间。 (您是否注意到即使您从未在任何地方使用过 rcx 也会推送/弹出 rcx?也许这是使用 the slow LOOP instruction 时遗留下来的)。
要了解有关编写高效 x86 asm 的更多信息,请参阅 Agner Fog's Optimizing Assembly guide。 (还有他的微架构指南,如果你想真正了解特定 CPU 的详细信息以及它们的不同之处:一个 uarch CPU 上的最佳方案可能不在另一个 CPU 上。例如,IMUL r64 在 Intel 上具有更好的吞吐量和延迟CPU 比 AMD 上的高,但 CMOV 和 ADC 在 Intel pre-Broadwell 上是 2 uop,具有 2 个周期延迟。而在 AMD 上是 1,因为 3 输入 ALU m-ops(FLAGS + 两个寄存器)对 AMD 来说不是问题.) 另请参阅x86 标签 wiki 中的其他链接。
在不更改对 printf 的 5M 调用的情况下纯粹优化循环仅作为如何正确编写循环的示例有用,而不是实际加速此代码。但让我们从这个开始:
; trivial fixes to loop efficiently while calling the same slow function
global main
extern printf
main:
push rbx
mov ebx, 5000000 ; don't waste a REX prefix for constants that fit in 32 bits
.print:
;; removed the push/pops from inside the loop.
; Use call-preserved regs instead of saving/restoring stuff inside a loop yourself.
mov edi, format ; static data / code always has a 32-bit address
mov esi, ebx
xor eax, eax ; The x86-64 SysV ABI requires al = number of FP args passed in FP registers for variadic functions
call printf
dec ebx
jnz .print
pop rbx ; restore rbx, the one call-preserved reg we actually used.
xor eax,eax ; successful exit status.
ret
section .rodata ; it's usually best to put constant data in a separate section of the text segment, not right next to code.
format:
db "%ld", 10, 0
为了加快速度,我们应该利用冗余将连续整数转换为字符串。由于"5000000\n" 只有 8 个字节长(包括换行符),所以字符串表示适合 64 位寄存器。
我们可以将该字符串存储到缓冲区中,并将指针增加字符串长度。 (因为对于较小的数字它会变得更短,所以只需将当前字符串长度保存在寄存器中,您可以在它更改的特殊情况分支中更新它。)
我们可以就地递减字符串表示,以避免(重新)执行除以 10 以将整数转换为十进制字符串的过程。
由于进位/借位不会自然地在寄存器内传播,并且AAS 指令在 64 位模式下不可用(并且仅适用于 AX,甚至不适用于 EAX,而且速度很慢),我们必须我们自己做。我们每次都减 1,所以我们知道会发生什么。我们可以通过展开 10 次来处理最低有效数字,因此没有分支来处理它。
还要注意,由于我们想要按打印顺序排列数字,进位无论如何都会走错方向,因为 x86 是 little-endian。如果有一个很好的方法来利用我们的字符串以其他字节顺序,我们可以使用 BSWAP 或 MOVBE。 (但请注意,MOVBE r64 是 Skylake 上的 3 个融合域微指令,其中 2 个是 ALU 微指令。BSWAP r64 也是 2 个微指令。)
也许我们应该在 XMM 向量寄存器的两半中并行处理奇数/偶数计数器。但是一旦字符串短于 8B,它就会停止工作。一次存储一个数字字符串,我们可以很容易地重叠。不过,我们可以在向量 reg 中进行进位传播,并使用 MOVQ 和 MOVHPS 分别存储两半。或者由于从 0 到 5M 的数字中有 4/5 是 7 位数字,因此对于可以存储两个数字的整个 16B 向量的特殊情况,可能值得编写代码。
处理较短字符串的更好方法:SSSE3 PSHUFB 将两个字符串打乱到左包装在向量寄存器中,然后一个 MOVUPS 一次存储两个。 shuffle 掩码只需要在字符串长度(位数)发生变化时更新,因此不经常执行的进位处理特殊情况代码也可以做到这一点。
循环的热部分的矢量化应该非常简单且成本低廉,并且性能应该几乎翻倍。
;;; Optimized version: keep the string data in a register and modify it
;;; instead of doing the whole int->string conversion every time.
section .bss
printbuf: resb 1024*128 + 4096 ; Buffer size ~= half L2 cache size on Intel SnB-family. Or use a giant buffer that we write() once. Or maybe vmsplice to give it away to the kernel, since we only run once.
global main
extern printf
main:
push rbx
; use some REX-only regs for values that we're always going to use a REX prefix with anyway for 64-bit operand size.
mov rdx, `5000000\n` ; (NASM string constants as integers work like little-endian, so AL = '5' = 0x35 and the high byte holds '\n' = 10). Note that YASM doesn't support back-ticks for C-style backslash processing.
mov r9, 1<<56 ; decrement by 1 in the 2nd-last byte: LSB of the decimal string
;xor r9d, r9d
;bts r9, 56 ; IDK if this code-size optimization outside the loop would help or not.
mov eax, 8 ; string length.
mov edi, printbuf
.storeloop:
;; rdx = "????x9\n". We compute the start value for the next iteration, i.e. counter -= 10 in rdx.
mov r8, rdx
;; r8 = rdx. We modify it to have each last digit from 9 down to 0 in sequence, and store those strings in the buffer.
;; The string could be any length, always with the first ASCII digit in the low byte; our other constants are adjusted correctly for it
;; narrower than 8B means that our stores overlap, but that's fine.
;; Starting from here to compute the next unrolled iteration's starting value takes the `sub r8, r9` instructions off the critical path, vs. if we started from r8 at the bottom of the loop. This gives out-of-order execution more to play with.
;; It means each loop iteration's sequence of subs and stores are a separate dependency chain (except for the store addresses, but OOO can get ahead on those because we only pointer-increment every 2 stores).
mov [rdi], r8
sub r8, r9 ; r8 = "xxx8\n"
mov [rdi + rax], r8 ; defer p += len by using a 2-reg addressing mode
sub r8, r9 ; r8 = "xxx7\n"
lea edi, [rdi + rax*2] ; if we had len*3 in another reg, we could defer this longer
;; our static buffer is guaranteed to be in the low 31 bits of address space so we can safely save a REX prefix on the LEA here. Normally you shouldn't truncate pointers to 32-bits, but you asked for the fastest possible. This won't hurt, and might help on some CPUs, especially with possible decode bottlenecks.
;; repeat that block 3 more times.
;; using a short inner loop for the 9..0 last digit might be a win on some CPUs (like maybe Core2), depending on their front-end loop-buffer capabilities if the frontend is a bottleneck at all here.
;; anyway, then for the last one:
mov [rdi], r8 ; r8 = "xxx1\n"
sub r8, r9
mov [rdi + rax], r8 ; r8 = "xxx0\n"
lea edi, [rdi + rax*2]
;; compute next iteration's RDX. It's probably a win to interleave some of this into the loop body, but out-of-order execution should do a reasonably good job here.
mov rcx, r9
shr rcx, 8 ; maybe hoist this constant out, too
; rcx = 1 in the second-lowest digit
sub rdx, rcx
; detect carry when '0' (0x30) - 1 = 0x2F by checking the low bit of the high nibble in that byte.
shl rcx, 5
test rdx, rcx
jz .carry_second_digit
; .carry_second_digit is some complicated code to propagate carry as far as it needs to go, up to the most-significant digit.
; when it's done, it re-enters the loop at the top, with eax and r9 set appropriately.
; it only runs once per 100 digits, so it doesn't have to be super-fast
; maybe only do buffer-length checks in the carry-handling branch,
; in which case the jz .carry can be jnz .storeloop
cmp edi, esi ; } while(p < endp)
jbe .storeloop
; write() system call on the buffer.
; Maybe need a loop around this instead of doing all 5M integer-strings in one giant buffer.
pop rbx
xor eax,eax ; successful exit status.
ret
这并没有完全充实,但应该让您了解哪些方法可能会奏效。
如果使用 SSE2 进行矢量化,可能会使用一个标量整数寄存器来跟踪何时需要突破和处理进位。即从 10 开始的递减计数器。
即使是这个标量版本也可能接近于每个时钟维持一个存储,这会使存储端口饱和。它们只有 8B 的存储空间(当字符串变短时,有用的部分也会变短),所以如果我们没有缓存未命中的瓶颈,我们肯定会将性能留在桌面上。但是对于 3GHz CPU 和双通道 DDR3-1600(~25.6GB/s 理论最大带宽),每时钟 8B 足以使单核主内存饱和。
我们可以并行化,并将 5M .. 1 的范围分成块。通过一些巧妙的数学运算,我们可以找出写入"2500000\n" 的第一个字符的字节,或者我们可以让每个线程以正确的顺序调用write() 本身。 (或者使用相同的巧妙数学方法让它们以不同的文件偏移量独立调用pwrite(2),这样内核就会负责同一文件的多个写入者的所有同步。)