错误:在操作数 1 处无效使用向量寄存器 [英] Error: invalid use of vector register at operand 1

问题描述

我正在 64 位 Aarch64 设备上的 ARM 下学习 GCC 内联汇编器.我看到一条我不太明白的错误消息.GCC 内联汇编器中的错误消息:

$ gcc -DNDEBUG -g3 -O1 -march=armv8-a+crc+crypto test.cc -o test.exe/tmp/ccCHOWrn.s:汇编程序消息:/tmp/ccCHOWrn.s:19: 错误:在操作数 1 处无效使用向量寄存器 -- `pmull v0,v0,v0'

示例程序只是尝试练习多项式乘法:

$ cat test.cc#include int main(int argc, char* argv[]){uint64x2_t r = {0,0}, a = {2,4};__asm__ __volatile__("pmull %0, %1, %1;": "=w" (r): "w" (a), "w" (a):抄送");返回 (int)r[0];}

"w" 是一个 Aarch64 机器约束.在这种情况下，它被描述为浮点或 SIMD 向量寄存器"，这似乎是我想要的.

uint64x2_t 类型通常与 ARM 内在函数一起使用.但它是一个 128 位类型并为 SIMD 协处理器对齐，因此它似乎是示例的不错选择.

该设备是一个 LeMaker HiKey一个 Linaro 图像和 GCC 4.9.2 编译器.看起来这个错误是几年前修复的，但我不确定它是否相关:arm64 上的 fbb ftbfs.

我有两个问题:

• 该错误是什么意思，我该如何解决?
• pmull 和 pmull2 是否有内在函数?

<小时>

我尝试添加排列说明符，但我并不惊讶它不起作用，因为我不知道语法:

$ gcc -DNDEBUG -g3 -O1 -march=armv8-a+crc+crypto test.cc -o test.exetest.cc: 在函数‘int main(int, char**)’中:test.cc:8:15: 错误:数字常量前的预期‘)’:=w"(r.1q)^test.cc:8:15: 错误:数字常量前的预期‘)’test.cc:9:6: 错误: 预期的‘;’在‘:’标记之前: "w" (a.1d), "w" (a.1d)^test.cc:9:6: 错误:':' 标记前的预期主表达式

<小时>

我还尝试添加双百分号(即 %%0 和 %%1)，因为汇编器在使用 .att_stntax 和 .intel_syntax:

$ gcc -DNDEBUG -g3 -O1 -march=armv8-a+crc+crypto test.cc -o test.exe/tmp/ccPpnvUP.s:汇编程序消息:/tmp/ccPpnvUP.s:19: 错误:操作数 1 应该是 SIMD 向量寄存器——`pmull %0,%1,%1'

解决方案

该错误是什么意思，我该如何解决?

我猜这个错误意味着发生了违反约束的情况.看起来以下内容可以解决问题:

$ cat test.cc#include #include int main(int argc, char* argv[]){uint64x2_t r = {0,0}, a = {2,4};__asm__ __volatile__("pmull %0.1q, %1.1d, %1.1d;": "=w" (r): "w" (a[0]), "w" (a[1]):抄送");fprintf(stdout, "%d, %d\n", r[0], r[1]);返回0；}

还有:

$ gcc -march=armv8-a+crc+crypto test.cc -o test.exe$ ./test.exe4, 0

和:

$ gdb -batch -ex 'disassemble main' ./test.exe转储函数 main 的汇编代码:0x00000000004005f0 <+0>: stp x29, x30, [sp,#-64]!0x00000000004005f4 <+4>: mov x29, sp0x00000000004005f8 <+8>: str w0, [x29,#28]0x00000000004005fc <+12>: str x1, [x29,#16]0x0000000000400600 <+16>:movi v0.4s，#0x00x0000000000400604 <+20>: mov x0, v0.d[0]0x0000000000400608 <+24>: mov x1, v0.d[1]0x000000000040060c <+28>: fmov d0, x00x0000000000400610 <+32>: mov v0.d[1], x10x0000000000400614 <+36>: str q0, [x29,#48]0x0000000000400618 <+40>: ldr q0, 0x4006a00x000000000040061c <+44>: mov x0, v0.d[0]0x0000000000400620 <+48>: mov x1, v0.d[1]0x0000000000400624 <+52>: fmov d0, x00x0000000000400628 <+56>: mov v0.d[1], x10x000000000040062c <+60>: str q0, [x29,#32]0x0000000000400630 <+64>: ldr x0, [x29,#32]0x0000000000400634 <+68>: ldr x1, [x29,#40]0x0000000000400638 <+72>: fmov d0, x00x000000000040063c <+76>: fmov d1, x10x0000000000400640 <+80>: pmull v0.1q, v0.1d, v0.1d0x0000000000400644 <+84>: mov x0, v0.d[0]0x0000000000400648 <+88>: mov x1, v0.d[1]0x000000000040064c <+92>: fmov d0, x00x0000000000400650 <+96>: mov v0.d[1], x10x0000000000400654 <+100>: str q0, [x29,#48]0x0000000000400658 <+104>: adrp x0, 0x4100000x000000000040065c <+108>:添加 x0、x0、#0x9f00x0000000000400660 <+112>: ldr x4, [x0]0x0000000000400664 <+116>: ldr x1, [x29,#48]0x0000000000400668 <+120>: ldr x2, [x29,#56]0x000000000040066c <+124>: adrp x0, 0x4000000x0000000000400670 <+128>: 添加 x0, x0, #0x7480x0000000000400674 <+132>: mov x3, x20x0000000000400678 <+136>: mov x2, x10x000000000040067c <+140>: mov x1, x00x0000000000400680 <+144>: mov x0, x40x0000000000400684 <+148>:bl 0x4004a0 <fprintf@plt>0x0000000000400688 <+152>: mov w0, #0x0//#00x000000000040068c <+156>: ldp x29, x30, [sp],#640x0000000000400690 <+160>: ret汇编程序转储结束.

<小时><块引用>

pmull 和 pmull2 是否有内在函数?

看起来有一些内在函数:

$ gcc -march=armv8-a+crc+crypto -E test.cc |grep -B 4 pmull__extension__ 静态 __inline poly16x8_t __attribute__ ((__always_inline__))vmull_high_p8 (poly8x16_t a, poly8x16_t b){poly16x8_t 结果；__asm__ ("pmull2 %0.8h,%1.16b,%2.16b"——__extension__ 静态 __inline poly16x8_t __attribute__ ((__always_inline__))vmull_p8 (poly8x8_t a, poly8x8_t b){poly16x8_t 结果；__asm__ ("pmull %0.8h, %1.8b, %2.8b"——静态 __inline poly128_tvmull_p64 (poly64_t a, poly64_t b){返回__builtin_aarch64_crypto_pmulldi_ppp (a, b);——静态 __inline poly128_tvmull_high_p64 (poly64x2_t a, poly64x2_t b){返回 __builtin_aarch64_crypto_pmullv2di_ppp (a, b);

I'm learning GCC inline assembler under under ARM on a 64-bit Aarch64 device. I'm seeing an error message I don't quite understand. The error message in from GCC's inline assembler:

$ gcc -DNDEBUG -g3 -O1 -march=armv8-a+crc+crypto test.cc -o test.exe
/tmp/ccCHOWrn.s: Assembler messages:
/tmp/ccCHOWrn.s:19: Error: invalid use of vector register at operand 1 -- `pmull v0,v0,v0'

The sample program simply tries to exercise the polynomial multiply:

$ cat test.cc 
#include <arm_neon.h>
int main(int argc, char* argv[])
{
    uint64x2_t r = {0,0}, a = {2,4};
    __asm__ __volatile__
    (
        "pmull %0, %1, %1;"
        : "=w" (r)
        : "w" (a), "w" (a)
        : "cc"
    );

    return (int)r[0];
}

The "w" is an Aarch64 machine constraint. In this case, its described as "Floating point or SIMD vector register", which seems to be what I want.

The uint64x2_t type is typically used with ARM intrinsics. But its a 128-bit type and aligned for SIMD coprocessor, so it seemed like a good choice for the sample.

The device is a LeMaker HiKey with a Linaro image and GCC 4.9.2 compiler. It looks like this bug was fixed a couple of years ago, but I'm not sure if its related: fbb ftbfs on arm64.

I have two questions:

• What does the error mean, and how can I fix it?
• Is there an intrinsic for pmull and pmull2?

---

I tried adding the arrangement specifiers, but I'm not surprised it did not work since I don't know the syntax:

$ gcc -DNDEBUG -g3 -O1 -march=armv8-a+crc+crypto test.cc -o test.exe
test.cc: In function ‘int main(int, char**)’:
test.cc:8:15: error: expected ‘)’ before numeric constant
      : "=w" (r.1q)
               ^
test.cc:8:15: error: expected ‘)’ before numeric constant
test.cc:9:6: error: expected ‘;’ before ‘:’ token
      : "w" (a.1d), "w" (a.1d)
      ^
test.cc:9:6: error: expected primary-expression before ‘:’ token

---

I also tried adding double percent signs (i.e., %%0 and %%1) since the assembler was having trouble with .att_stntax and .intel_syntax:

$ gcc -DNDEBUG -g3 -O1 -march=armv8-a+crc+crypto test.cc -o test.exe
/tmp/ccPpnvUP.s: Assembler messages:
/tmp/ccPpnvUP.s:19: Error: operand 1 should be a SIMD vector register -- `pmull %0,%1,%1'

解决方案

What does the error mean, and how can I fix it?

I guess the error means a constraint violation occurred. It looks like the following does the trick:

$ cat test.cc
#include <arm_neon.h>
#include <stdio.h>
int main(int argc, char* argv[])
{
  uint64x2_t r = {0,0}, a = {2,4};
  __asm__ __volatile__
    (
     "pmull %0.1q, %1.1d, %1.1d;"
     : "=w" (r)
     : "w" (a[0]), "w" (a[1])
     : "cc"
     );

  fprintf(stdout, "%d, %d\n", r[0], r[1]);
  return 0;
}

And:

$ gcc -march=armv8-a+crc+crypto test.cc -o test.exe 
$ ./test.exe 
4, 0

And:

$ gdb -batch -ex 'disassemble main' ./test.exe 
Dump of assembler code for function main:
   0x00000000004005f0 <+0>: stp x29, x30, [sp,#-64]!
   0x00000000004005f4 <+4>: mov x29, sp
   0x00000000004005f8 <+8>: str w0, [x29,#28]
   0x00000000004005fc <+12>:    str x1, [x29,#16]
   0x0000000000400600 <+16>:    movi    v0.4s, #0x0
   0x0000000000400604 <+20>:    mov x0, v0.d[0]
   0x0000000000400608 <+24>:    mov x1, v0.d[1]
   0x000000000040060c <+28>:    fmov    d0, x0
   0x0000000000400610 <+32>:    mov v0.d[1], x1
   0x0000000000400614 <+36>:    str q0, [x29,#48]
   0x0000000000400618 <+40>:    ldr q0, 0x4006a0
   0x000000000040061c <+44>:    mov x0, v0.d[0]
   0x0000000000400620 <+48>:    mov x1, v0.d[1]
   0x0000000000400624 <+52>:    fmov    d0, x0
   0x0000000000400628 <+56>:    mov v0.d[1], x1
   0x000000000040062c <+60>:    str q0, [x29,#32]
   0x0000000000400630 <+64>:    ldr x0, [x29,#32]
   0x0000000000400634 <+68>:    ldr x1, [x29,#40]
   0x0000000000400638 <+72>:    fmov    d0, x0
   0x000000000040063c <+76>:    fmov    d1, x1
   0x0000000000400640 <+80>:    pmull   v0.1q, v0.1d, v0.1d
   0x0000000000400644 <+84>:    mov x0, v0.d[0]
   0x0000000000400648 <+88>:    mov x1, v0.d[1]
   0x000000000040064c <+92>:    fmov    d0, x0
   0x0000000000400650 <+96>:    mov v0.d[1], x1
   0x0000000000400654 <+100>:   str q0, [x29,#48]
   0x0000000000400658 <+104>:   adrp    x0, 0x410000
   0x000000000040065c <+108>:   add x0, x0, #0x9f0
   0x0000000000400660 <+112>:   ldr x4, [x0]
   0x0000000000400664 <+116>:   ldr x1, [x29,#48]
   0x0000000000400668 <+120>:   ldr x2, [x29,#56]
   0x000000000040066c <+124>:   adrp    x0, 0x400000
   0x0000000000400670 <+128>:   add x0, x0, #0x748
   0x0000000000400674 <+132>:   mov x3, x2
   0x0000000000400678 <+136>:   mov x2, x1
   0x000000000040067c <+140>:   mov x1, x0
   0x0000000000400680 <+144>:   mov x0, x4
   0x0000000000400684 <+148>:   bl  0x4004a0 <fprintf@plt>
   0x0000000000400688 <+152>:   mov w0, #0x0                    // #0
   0x000000000040068c <+156>:   ldp x29, x30, [sp],#64
   0x0000000000400690 <+160>:   ret
End of assembler dump.

---

Is there an intrinsic for pmull and pmull2?

It looks like there are some intrinsics:

$ gcc -march=armv8-a+crc+crypto -E test.cc | grep -B 4 pmull
__extension__ static __inline poly16x8_t __attribute__ ((__always_inline__))
vmull_high_p8 (poly8x16_t a, poly8x16_t b)
{
  poly16x8_t result;
  __asm__ ("pmull2 %0.8h,%1.16b,%2.16b"
--
__extension__ static __inline poly16x8_t __attribute__ ((__always_inline__))
vmull_p8 (poly8x8_t a, poly8x8_t b)
{
  poly16x8_t result;
  __asm__ ("pmull %0.8h, %1.8b, %2.8b"
--
static __inline poly128_t
vmull_p64 (poly64_t a, poly64_t b)
{
  return
    __builtin_aarch64_crypto_pmulldi_ppp (a, b);
--

static __inline poly128_t
vmull_high_p64 (poly64x2_t a, poly64x2_t b)
{
  return __builtin_aarch64_crypto_pmullv2di_ppp (a, b);

这篇关于错误:在操作数 1 处无效使用向量寄存器的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！