GCC 5和更高版本中对AVX2的支持 [英] AVX2 support in GCC 5 and later
本文介绍了GCC 5和更高版本中对AVX2的支持的处理方法，对大家解决问题具有一定的参考价值，需要的朋友们下面随着小编来一起学习吧！
问题描述
我写了下面的"T"类来加速对 使用AVX2的字符集".然后我发现它在 gcc 5和更高版本，当我使用"-O3"时. 谁能帮我追溯到一些编程结构， 已知在最新的编译器/系统上不起作用?
此代码的工作方式:基础结构("_bits")是一个256字节的块(已为AVX2对齐并分配)，可以将其作为char [256]或AVX2元素进行访问，具体取决于元素是否为访问或整个操作都在矢量操作中使用.似乎它在AVX2平台上应该可以很好地工作.不?
这真的很难调试，因为"valgrind"说它很干净， 而且我不能使用调试器(由于问题在 我删除"-O3").但是我不满意仅使用"| =" 解决方法，因为如果此代码确实是错误的，那么我可能 在其他地方犯同样的错误，搞砸了一切 我开发！
有趣的是，"|"操作员有问题，但是 "| ="不.问题是否可能与从中返回结构有关 一个功能?但我认为自1990年以来返回结构已奏效 之类的.
// g++ -std=c++11 -mavx2 -O3 gcc_fail.cpp
#include "assert.h"
#include "immintrin.h" // AVX
class T {
public:
__m256i _bits[8];
inline bool& operator[](unsigned char c) {return ((bool*)_bits)[c];}
inline bool operator[](unsigned char c) const {return ((bool*)_bits)[c];}
inline T() {}
inline explicit T(char const*);
inline T operator| (T const& b) const;
inline T & operator|=(T const& b);
inline bool operator! () const;
};
T::T(char const* s)
{
_bits[0] = _bits[1] = _bits[2] = _bits[3] = _mm256_set1_epi32(0);
_bits[4] = _bits[5] = _bits[6] = _bits[7] = _mm256_set1_epi32(0);
char c;
while ((c = *s++))
(*this)[c] = true;
}
T T::operator| (T const& b) const
{
T res;
for (int i = 0; i < 8; i++)
res._bits[i] = _mm256_or_si256(_bits[i], b._bits[i]);
// FIXME why does the above code fail with -O3 in new gcc?
for (int i=0; i<256; i++)
assert(res[i] == ((*this)[i] || b[i]));
// gcc 4.7.0 - PASS
// gcc 4.7.2 - PASS
// gcc 4.8.0 - PASS
// gcc 4.9.2 - PASS
// gcc 5.2.0 - FAIL
// gcc 5.3.0 - FAIL
// gcc 5.3.1 - FAIL
// gcc 6.1.0 - FAIL
return res;
}
T & T::operator|=(T const& b)
{
for (int i = 0; i < 8; i++)
_bits[i] = _mm256_or_si256(_bits[i], b._bits[i]);
return *this;
}
bool T::operator! () const
{
for (int i = 0; i < 8; i++)
if (!_mm256_testz_si256(_bits[i], _bits[i]))
return false;
return true;
}
int Main()
{
T sep (" ,\t\n");
T end ("");
return !(sep|end);
}
int main()
{
return Main();
}
解决方案
您的代码的问题是，当您应该一直使用unsigned char*时使用了bool*，这使GCC 5可以进行指针别名优化. /p>
由GCC 4.8.5和5.3.1生成的功能Main()的两个机器代码转储位于附录的此答案的末尾，以供参考.
查看代码:
反编译
序言之后，T sep的_bits初始化为零...
_bits[0] = _bits[1] = _bits[2] = _bits[3] = _mm256_set1_epi32(0);
_bits[4] = _bits[5] = _bits[6] = _bits[7] = _mm256_set1_epi32(0);
40063d: c5 fd 7f 44 24 60 vmovdqa %ymm0,0x60(%rsp)
400643: c5 fd 7f 44 24 40 vmovdqa %ymm0,0x40(%rsp)
400649: c5 fd 7f 44 24 20 vmovdqa %ymm0,0x20(%rsp)
40064f: c5 fd 7f 04 24 vmovdqa %ymm0,(%rsp)
400654: c5 fd 7f 84 24 e0 00 00 00 vmovdqa %ymm0,0xe0(%rsp)
40065d: c5 fd 7f 84 24 c0 00 00 00 vmovdqa %ymm0,0xc0(%rsp)
400666: c5 fd 7f 84 24 a0 00 00 00 vmovdqa %ymm0,0xa0(%rsp)
40066f: c5 fd 7f 84 24 80 00 00 00 vmovdqa %ymm0,0x80(%rsp)
，然后基于char* s循环写入.
char c;
while ((c = *s++))
(*this)[c] = true;
400680: 48 83 c2 01 add $0x1,%rdx
400684: c6 04 04 01 movb $0x1,(%rsp,%rax,1)
400688: 0f b6 42 ff movzbl -0x1(%rdx),%eax
40068c: 84 c0 test %al,%al
40068e: 75 f0 jne 400680 <_Z4Mainv+0x60>
两个编译器然后将T end初始化为0:
400690: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400694: 31 c0 xor %eax,%eax
400696: c5 fd 7f 84 24 60 01 00 00 vmovdqa %ymm0,0x160(%rsp)
40069f: c5 fd 7f 84 24 40 01 00 00 vmovdqa %ymm0,0x140(%rsp)
4006a8: c5 fd 7f 84 24 20 01 00 00 vmovdqa %ymm0,0x120(%rsp)
4006b1: c5 fd 7f 84 24 00 01 00 00 vmovdqa %ymm0,0x100(%rsp)
4006ba: c5 fd 7f 84 24 e0 01 00 00 vmovdqa %ymm0,0x1e0(%rsp)
4006c3: c5 fd 7f 84 24 c0 01 00 00 vmovdqa %ymm0,0x1c0(%rsp)
4006cc: c5 fd 7f 84 24 a0 01 00 00 vmovdqa %ymm0,0x1a0(%rsp)
4006d5: c5 fd 7f 84 24 80 01 00 00 vmovdqa %ymm0,0x180(%rsp)
然后两个编译器都会优化_mm256_or_si256()操作，因为已知T end是0.但是，然后，GCC 4.8.5 从T sep复制到T res (这是计算结果，当您将任何值或零变量或为零时，会发生这种情况)，而GCC 5.3.1 会初始化0 operator []方法中，将类型为__m256i*的指针强制转换为bool*，并且允许编译器假定这些指针没有别名.因此，在GCC 4.8.5中，您会看到
4006de: c5 fd 6f 04 24 vmovdqa (%rsp),%ymm0
4006e3: c5 fd 7f 84 24 00 02 00 00 vmovdqa %ymm0,0x200(%rsp)
4006ec: c5 fd 6f 44 24 20 vmovdqa 0x20(%rsp),%ymm0
4006f2: c5 fd 7f 84 24 20 02 00 00 vmovdqa %ymm0,0x220(%rsp)
4006fb: c5 fd 6f 44 24 40 vmovdqa 0x40(%rsp),%ymm0
400701: c5 fd 7f 84 24 40 02 00 00 vmovdqa %ymm0,0x240(%rsp)
40070a: c5 fd 6f 44 24 60 vmovdqa 0x60(%rsp),%ymm0
400710: c5 fd 7f 84 24 60 02 00 00 vmovdqa %ymm0,0x260(%rsp)
400719: c5 fd 6f 84 24 80 00 00 00 vmovdqa 0x80(%rsp),%ymm0
400722: c5 fd 7f 84 24 80 02 00 00 vmovdqa %ymm0,0x280(%rsp)
40072b: c5 fd 6f 84 24 a0 00 00 00 vmovdqa 0xa0(%rsp),%ymm0
400734: c5 fd 7f 84 24 a0 02 00 00 vmovdqa %ymm0,0x2a0(%rsp)
40073d: c5 fd 6f 84 24 c0 00 00 00 vmovdqa 0xc0(%rsp),%ymm0
400746: c5 fd 7f 84 24 c0 02 00 00 vmovdqa %ymm0,0x2c0(%rsp)
40074f: c5 fd 6f 84 24 e0 00 00 00 vmovdqa 0xe0(%rsp),%ymm0
400758: c5 fd 7f 84 24 e0 02 00 00 vmovdqa %ymm0,0x2e0(%rsp)
在GCC 5.3.1中，您会看到
4006fa: c5 fd 7f 85 f0 fe ff ff vmovdqa %ymm0,-0x110(%rbp)
400702: c5 fd 7f 85 10 ff ff ff vmovdqa %ymm0,-0xf0(%rbp)
40070a: c5 fd 7f 85 30 ff ff ff vmovdqa %ymm0,-0xd0(%rbp)
400712: c5 fd 7f 85 50 ff ff ff vmovdqa %ymm0,-0xb0(%rbp)
40071a: c5 fd 7f 85 70 ff ff ff vmovdqa %ymm0,-0x90(%rbp)
400722: c5 fd 7f 45 90 vmovdqa %ymm0,-0x70(%rbp)
400727: c5 fd 7f 45 b0 vmovdqa %ymm0,-0x50(%rbp)
40072c: c5 fd 7f 45 d0 vmovdqa %ymm0,-0x30(%rbp)
因此，对assert()的读取将失败.
标准关于指针别名的规定:
ISO C ++ 11在以下部分中引用了别名，这明确说明类型__m256i*的变量无法使用bool*进行访问，但可以使用char*/unsigned char*进行访问:
§3.10左值和右值[basic.lval]
[...]
如果程序尝试通过以下类型之一以外的glvalue访问对象的存储值，则行为未定义:[52]
- 对象的动态类型，
- 对象的动态类型的cv限定版本，
- 类似于对象的动态类型的类型(定义见4.4)
- 一种类型，它是与对象的动态类型相对应的有符号或无符号类型，
- 一种类型，它是与对象的动态类型的cv限定版本相对应的有符号或无符号类型，
- 在其元素或非静态数据成员(包括递归地包括子聚合或所包含的并集的元素或非静态数据成员)中包括上述类型之一的集合或联合类型，
- 一种类型，它是对象动态类型的(可能是cv限定的)基类类型，
char或unsigned char类型.
52)此列表的目的是指定对象可能会别名也可能不会别名的那些情况.
附录
GCC 4.8.5:
0000000000400620 <_Z4Mainv>:
400620: 55 push %rbp
400621: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400625: ba e5 08 40 00 mov $0x4008e5,%edx
40062a: b8 20 00 00 00 mov $0x20,%eax
40062f: 48 89 e5 mov %rsp,%rbp
400632: 48 83 e4 e0 and $0xffffffffffffffe0,%rsp
400636: 48 81 ec 00 03 00 00 sub $0x300,%rsp
40063d: c5 fd 7f 44 24 60 vmovdqa %ymm0,0x60(%rsp)
400643: c5 fd 7f 44 24 40 vmovdqa %ymm0,0x40(%rsp)
400649: c5 fd 7f 44 24 20 vmovdqa %ymm0,0x20(%rsp)
40064f: c5 fd 7f 04 24 vmovdqa %ymm0,(%rsp)
400654: c5 fd 7f 84 24 e0 00 00 00 vmovdqa %ymm0,0xe0(%rsp)
40065d: c5 fd 7f 84 24 c0 00 00 00 vmovdqa %ymm0,0xc0(%rsp)
400666: c5 fd 7f 84 24 a0 00 00 00 vmovdqa %ymm0,0xa0(%rsp)
40066f: c5 fd 7f 84 24 80 00 00 00 vmovdqa %ymm0,0x80(%rsp)
400678: 0f 1f 84 00 00 00 00 00 nopl 0x0(%rax,%rax,1)
400680: 48 83 c2 01 add $0x1,%rdx
400684: c6 04 04 01 movb $0x1,(%rsp,%rax,1)
400688: 0f b6 42 ff movzbl -0x1(%rdx),%eax
40068c: 84 c0 test %al,%al
40068e: 75 f0 jne 400680 <_Z4Mainv+0x60>
400690: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400694: 31 c0 xor %eax,%eax
400696: c5 fd 7f 84 24 60 01 00 00 vmovdqa %ymm0,0x160(%rsp)
40069f: c5 fd 7f 84 24 40 01 00 00 vmovdqa %ymm0,0x140(%rsp)
4006a8: c5 fd 7f 84 24 20 01 00 00 vmovdqa %ymm0,0x120(%rsp)
4006b1: c5 fd 7f 84 24 00 01 00 00 vmovdqa %ymm0,0x100(%rsp)
4006ba: c5 fd 7f 84 24 e0 01 00 00 vmovdqa %ymm0,0x1e0(%rsp)
4006c3: c5 fd 7f 84 24 c0 01 00 00 vmovdqa %ymm0,0x1c0(%rsp)
4006cc: c5 fd 7f 84 24 a0 01 00 00 vmovdqa %ymm0,0x1a0(%rsp)
4006d5: c5 fd 7f 84 24 80 01 00 00 vmovdqa %ymm0,0x180(%rsp)
4006de: c5 fd 6f 04 24 vmovdqa (%rsp),%ymm0
4006e3: c5 fd 7f 84 24 00 02 00 00 vmovdqa %ymm0,0x200(%rsp)
4006ec: c5 fd 6f 44 24 20 vmovdqa 0x20(%rsp),%ymm0
4006f2: c5 fd 7f 84 24 20 02 00 00 vmovdqa %ymm0,0x220(%rsp)
4006fb: c5 fd 6f 44 24 40 vmovdqa 0x40(%rsp),%ymm0
400701: c5 fd 7f 84 24 40 02 00 00 vmovdqa %ymm0,0x240(%rsp)
40070a: c5 fd 6f 44 24 60 vmovdqa 0x60(%rsp),%ymm0
400710: c5 fd 7f 84 24 60 02 00 00 vmovdqa %ymm0,0x260(%rsp)
400719: c5 fd 6f 84 24 80 00 00 00 vmovdqa 0x80(%rsp),%ymm0
400722: c5 fd 7f 84 24 80 02 00 00 vmovdqa %ymm0,0x280(%rsp)
40072b: c5 fd 6f 84 24 a0 00 00 00 vmovdqa 0xa0(%rsp),%ymm0
400734: c5 fd 7f 84 24 a0 02 00 00 vmovdqa %ymm0,0x2a0(%rsp)
40073d: c5 fd 6f 84 24 c0 00 00 00 vmovdqa 0xc0(%rsp),%ymm0
400746: c5 fd 7f 84 24 c0 02 00 00 vmovdqa %ymm0,0x2c0(%rsp)
40074f: c5 fd 6f 84 24 e0 00 00 00 vmovdqa 0xe0(%rsp),%ymm0
400758: c5 fd 7f 84 24 e0 02 00 00 vmovdqa %ymm0,0x2e0(%rsp)
400761: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
400768: 80 3c 04 00 cmpb $0x0,(%rsp,%rax,1)
40076c: 0f b6 8c 04 00 02 00 00 movzbl 0x200(%rsp,%rax,1),%ecx
400774: ba 01 00 00 00 mov $0x1,%edx
400779: 75 08 jne 400783 <_Z4Mainv+0x163>
40077b: 0f b6 94 04 00 01 00 00 movzbl 0x100(%rsp,%rax,1),%edx
400783: 38 d1 cmp %dl,%cl
400785: 0f 85 b2 00 00 00 jne 40083d <_Z4Mainv+0x21d>
40078b: 48 83 c0 01 add $0x1,%rax
40078f: 48 3d 00 01 00 00 cmp $0x100,%rax
400795: 75 d1 jne 400768 <_Z4Mainv+0x148>
400797: c5 fd 6f 8c 24 00 02 00 00 vmovdqa 0x200(%rsp),%ymm1
4007a0: 31 c0 xor %eax,%eax
4007a2: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007a7: 0f 94 c0 sete %al
4007aa: 0f 85 88 00 00 00 jne 400838 <_Z4Mainv+0x218>
4007b0: c5 fd 6f 8c 24 20 02 00 00 vmovdqa 0x220(%rsp),%ymm1
4007b9: 31 c0 xor %eax,%eax
4007bb: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007c0: 0f 94 c0 sete %al
4007c3: 75 73 jne 400838 <_Z4Mainv+0x218>
4007c5: c5 fd 6f 8c 24 40 02 00 00 vmovdqa 0x240(%rsp),%ymm1
4007ce: 31 c0 xor %eax,%eax
4007d0: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007d5: 0f 94 c0 sete %al
4007d8: 75 5e jne 400838 <_Z4Mainv+0x218>
4007da: c5 fd 6f 8c 24 60 02 00 00 vmovdqa 0x260(%rsp),%ymm1
4007e3: 31 c0 xor %eax,%eax
4007e5: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007ea: 0f 94 c0 sete %al
4007ed: 75 49 jne 400838 <_Z4Mainv+0x218>
4007ef: c5 fd 6f 8c 24 80 02 00 00 vmovdqa 0x280(%rsp),%ymm1
4007f8: 31 c0 xor %eax,%eax
4007fa: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007ff: 0f 94 c0 sete %al
400802: 75 34 jne 400838 <_Z4Mainv+0x218>
400804: c5 fd 6f 8c 24 a0 02 00 00 vmovdqa 0x2a0(%rsp),%ymm1
40080d: 31 c0 xor %eax,%eax
40080f: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
400814: 0f 94 c0 sete %al
400817: 75 1f jne 400838 <_Z4Mainv+0x218>
400819: c5 fd 6f 8c 24 c0 02 00 00 vmovdqa 0x2c0(%rsp),%ymm1
400822: 31 c0 xor %eax,%eax
400824: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
400829: 0f 94 c0 sete %al
40082c: 75 0a jne 400838 <_Z4Mainv+0x218>
40082e: 31 c0 xor %eax,%eax
400830: c4 e2 7d 17 c0 vptest %ymm0,%ymm0
400835: 0f 94 c0 sete %al
400838: c5 f8 77 vzeroupper
40083b: c9 leaveq
40083c: c3 retq
40083d: b9 20 09 40 00 mov $0x400920,%ecx
400842: ba 26 00 00 00 mov $0x26,%edx
400847: be e9 08 40 00 mov $0x4008e9,%esi
40084c: bf f8 08 40 00 mov $0x4008f8,%edi
400851: c5 f8 77 vzeroupper
400854: e8 97 fc ff ff callq 4004f0 <__assert_fail@plt>
400859: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
海湾合作委员会5:
0000000000400630 <_Z4Mainv>:
400630: 4c 8d 54 24 08 lea 0x8(%rsp),%r10
400635: 48 83 e4 e0 and $0xffffffffffffffe0,%rsp
400639: b8 20 00 00 00 mov $0x20,%eax
40063e: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400642: ba 25 08 40 00 mov $0x400825,%edx
400647: 41 ff 72 f8 pushq -0x8(%r10)
40064b: 55 push %rbp
40064c: 48 89 e5 mov %rsp,%rbp
40064f: 41 52 push %r10
400651: 48 81 ec 08 03 00 00 sub $0x308,%rsp
400658: c5 fd 7f 85 50 fd ff ff vmovdqa %ymm0,-0x2b0(%rbp)
400660: c5 fd 7f 85 30 fd ff ff vmovdqa %ymm0,-0x2d0(%rbp)
400668: c5 fd 7f 85 10 fd ff ff vmovdqa %ymm0,-0x2f0(%rbp)
400670: c5 fd 7f 85 f0 fc ff ff vmovdqa %ymm0,-0x310(%rbp)
400678: c5 fd 7f 85 d0 fd ff ff vmovdqa %ymm0,-0x230(%rbp)
400680: c5 fd 7f 85 b0 fd ff ff vmovdqa %ymm0,-0x250(%rbp)
400688: c5 fd 7f 85 90 fd ff ff vmovdqa %ymm0,-0x270(%rbp)
400690: c5 fd 7f 85 70 fd ff ff vmovdqa %ymm0,-0x290(%rbp)
400698: 0f 1f 84 00 00 00 00 00 nopl 0x0(%rax,%rax,1)
4006a0: 48 83 c2 01 add $0x1,%rdx
4006a4: c6 84 05 f0 fc ff ff 01 movb $0x1,-0x310(%rbp,%rax,1)
4006ac: 0f b6 42 ff movzbl -0x1(%rdx),%eax
4006b0: 84 c0 test %al,%al
4006b2: 75 ec jne 4006a0 <_Z4Mainv+0x70>
4006b4: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
4006b8: 31 c0 xor %eax,%eax
4006ba: c5 fd 7f 85 50 fe ff ff vmovdqa %ymm0,-0x1b0(%rbp)
4006c2: c5 fd 7f 85 30 fe ff ff vmovdqa %ymm0,-0x1d0(%rbp)
4006ca: c5 fd 7f 85 10 fe ff ff vmovdqa %ymm0,-0x1f0(%rbp)
4006d2: c5 fd 7f 85 f0 fd ff ff vmovdqa %ymm0,-0x210(%rbp)
4006da: c5 fd 7f 85 d0 fe ff ff vmovdqa %ymm0,-0x130(%rbp)
4006e2: c5 fd 7f 85 b0 fe ff ff vmovdqa %ymm0,-0x150(%rbp)
4006ea: c5 fd 7f 85 90 fe ff ff vmovdqa %ymm0,-0x170(%rbp)
4006f2: c5 fd 7f 85 70 fe ff ff vmovdqa %ymm0,-0x190(%rbp)
4006fa: c5 fd 7f 85 f0 fe ff ff vmovdqa %ymm0,-0x110(%rbp)
400702: c5 fd 7f 85 10 ff ff ff vmovdqa %ymm0,-0xf0(%rbp)
40070a: c5 fd 7f 85 30 ff ff ff vmovdqa %ymm0,-0xd0(%rbp)
400712: c5 fd 7f 85 50 ff ff ff vmovdqa %ymm0,-0xb0(%rbp)
40071a: c5 fd 7f 85 70 ff ff ff vmovdqa %ymm0,-0x90(%rbp)
400722: c5 fd 7f 45 90 vmovdqa %ymm0,-0x70(%rbp)
400727: c5 fd 7f 45 b0 vmovdqa %ymm0,-0x50(%rbp)
40072c: c5 fd 7f 45 d0 vmovdqa %ymm0,-0x30(%rbp)
400731: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
400738: 0f b6 94 05 f0 fc ff ff movzbl -0x310(%rbp,%rax,1),%edx
400740: 0f b6 8c 05 f0 fe ff ff movzbl -0x110(%rbp,%rax,1),%ecx
400748: 84 d2 test %dl,%dl
40074a: 75 08 jne 400754 <_Z4Mainv+0x124>
40074c: 0f b6 94 05 f0 fd ff ff movzbl -0x210(%rbp,%rax,1),%edx
400754: 38 d1 cmp %dl,%cl
400756: 75 2c jne 400784 <_Z4Mainv+0x154>
400758: 48 83 c0 01 add $0x1,%rax
40075c: 48 3d 00 01 00 00 cmp $0x100,%rax
400762: 75 d4 jne 400738 <_Z4Mainv+0x108>
400764: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400768: 31 c0 xor %eax,%eax
40076a: c4 e2 7d 17 c0 vptest %ymm0,%ymm0
40076f: 0f 94 c0 sete %al
400772: c5 f8 77 vzeroupper
400775: 48 81 c4 08 03 00 00 add $0x308,%rsp
40077c: 41 5a pop %r10
40077e: 5d pop %rbp
40077f: 49 8d 62 f8 lea -0x8(%r10),%rsp
400783: c3 retq
400784: b9 60 08 40 00 mov $0x400860,%ecx
400789: ba 26 00 00 00 mov $0x26,%edx
40078e: be 29 08 40 00 mov $0x400829,%esi
400793: bf 38 08 40 00 mov $0x400838,%edi
400798: c5 f8 77 vzeroupper
40079b: e8 50 fd ff ff callq 4004f0 <__assert_fail@plt>
I wrote the following class "T" to accelerate manipulations of "sets of characters" using AVX2. Then I found that it doesn't work in gcc 5 and later when I use "-O3". Can anyone help me trace this down to some programming construct that is known not to work on the latest compilers/systems?
How this code works: The underlying structure ("_bits") is a block of 256 bytes (aligned and allocated for AVX2), which can be accessed either as char[256] or AVX2 elements, depending on whether an element is accessed or the whole thing is used in a vector operation. Seems like it should work perfectly well on the AVX2 platform. No?
This is really hard to debug, because "valgrind" says it's clean, and I can't use a debugger (due to the problem disappearing when I remove "-O3"). But I am not happy with just going with the "|=" workaround because if this code is really wrong, then I'm probably making the same mistake in other places and screwing up everything I develop!
It is interesting to note that the "|" operator has the problem but the "|=" does not. Could the problem be related to returning a struct from a function? But I thought that returning a struct has worked since 1990 or something.
// g++ -std=c++11 -mavx2 -O3 gcc_fail.cpp
#include "assert.h"
#include "immintrin.h" // AVX
class T {
public:
__m256i _bits[8];
inline bool& operator[](unsigned char c) {return ((bool*)_bits)[c];}
inline bool operator[](unsigned char c) const {return ((bool*)_bits)[c];}
inline T() {}
inline explicit T(char const*);
inline T operator| (T const& b) const;
inline T & operator|=(T const& b);
inline bool operator! () const;
};
T::T(char const* s)
{
_bits[0] = _bits[1] = _bits[2] = _bits[3] = _mm256_set1_epi32(0);
_bits[4] = _bits[5] = _bits[6] = _bits[7] = _mm256_set1_epi32(0);
char c;
while ((c = *s++))
(*this)[c] = true;
}
T T::operator| (T const& b) const
{
T res;
for (int i = 0; i < 8; i++)
res._bits[i] = _mm256_or_si256(_bits[i], b._bits[i]);
// FIXME why does the above code fail with -O3 in new gcc?
for (int i=0; i<256; i++)
assert(res[i] == ((*this)[i] || b[i]));
// gcc 4.7.0 - PASS
// gcc 4.7.2 - PASS
// gcc 4.8.0 - PASS
// gcc 4.9.2 - PASS
// gcc 5.2.0 - FAIL
// gcc 5.3.0 - FAIL
// gcc 5.3.1 - FAIL
// gcc 6.1.0 - FAIL
return res;
}
T & T::operator|=(T const& b)
{
for (int i = 0; i < 8; i++)
_bits[i] = _mm256_or_si256(_bits[i], b._bits[i]);
return *this;
}
bool T::operator! () const
{
for (int i = 0; i < 8; i++)
if (!_mm256_testz_si256(_bits[i], _bits[i]))
return false;
return true;
}
int Main()
{
T sep (" ,\t\n");
T end ("");
return !(sep|end);
}
int main()
{
return Main();
}
解决方案
Your code's problem is the use of bool* when you should have been using unsigned char*, which allowed GCC 5 to proceed with a pointer alias optimization.
The two dumps of the machine code for function Main(), produced both by GCC 4.8.5 and 5.3.1, are at the end of this answer in appendix for reference.
Looking at the code:
Decompilation
After the prologue, T sep's _bits are initialized to zero...
_bits[0] = _bits[1] = _bits[2] = _bits[3] = _mm256_set1_epi32(0);
_bits[4] = _bits[5] = _bits[6] = _bits[7] = _mm256_set1_epi32(0);
40063d: c5 fd 7f 44 24 60 vmovdqa %ymm0,0x60(%rsp)
400643: c5 fd 7f 44 24 40 vmovdqa %ymm0,0x40(%rsp)
400649: c5 fd 7f 44 24 20 vmovdqa %ymm0,0x20(%rsp)
40064f: c5 fd 7f 04 24 vmovdqa %ymm0,(%rsp)
400654: c5 fd 7f 84 24 e0 00 00 00 vmovdqa %ymm0,0xe0(%rsp)
40065d: c5 fd 7f 84 24 c0 00 00 00 vmovdqa %ymm0,0xc0(%rsp)
400666: c5 fd 7f 84 24 a0 00 00 00 vmovdqa %ymm0,0xa0(%rsp)
40066f: c5 fd 7f 84 24 80 00 00 00 vmovdqa %ymm0,0x80(%rsp)
and then written to in a loop based on char* s.
char c;
while ((c = *s++))
(*this)[c] = true;
400680: 48 83 c2 01 add $0x1,%rdx
400684: c6 04 04 01 movb $0x1,(%rsp,%rax,1)
400688: 0f b6 42 ff movzbl -0x1(%rdx),%eax
40068c: 84 c0 test %al,%al
40068e: 75 f0 jne 400680 <_Z4Mainv+0x60>
Both compilers then initialize T end to 0s:
400690: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400694: 31 c0 xor %eax,%eax
400696: c5 fd 7f 84 24 60 01 00 00 vmovdqa %ymm0,0x160(%rsp)
40069f: c5 fd 7f 84 24 40 01 00 00 vmovdqa %ymm0,0x140(%rsp)
4006a8: c5 fd 7f 84 24 20 01 00 00 vmovdqa %ymm0,0x120(%rsp)
4006b1: c5 fd 7f 84 24 00 01 00 00 vmovdqa %ymm0,0x100(%rsp)
4006ba: c5 fd 7f 84 24 e0 01 00 00 vmovdqa %ymm0,0x1e0(%rsp)
4006c3: c5 fd 7f 84 24 c0 01 00 00 vmovdqa %ymm0,0x1c0(%rsp)
4006cc: c5 fd 7f 84 24 a0 01 00 00 vmovdqa %ymm0,0x1a0(%rsp)
4006d5: c5 fd 7f 84 24 80 01 00 00 vmovdqa %ymm0,0x180(%rsp)
Both compilers then optimize out the _mm256_or_si256() operations because T end is known to be 0. But then, GCC 4.8.5 copies from T sep to T res (which is computationally what happens when you OR anything into a zero variable), while GCC 5.3.1 initializes T res to 0. It's entitled to do that because in your operator [] method you cast a pointer of type __m256i* to bool*, and the compiler is permitted to assume the pointers do not alias. Thus in GCC 4.8.5 you see
4006de: c5 fd 6f 04 24 vmovdqa (%rsp),%ymm0
4006e3: c5 fd 7f 84 24 00 02 00 00 vmovdqa %ymm0,0x200(%rsp)
4006ec: c5 fd 6f 44 24 20 vmovdqa 0x20(%rsp),%ymm0
4006f2: c5 fd 7f 84 24 20 02 00 00 vmovdqa %ymm0,0x220(%rsp)
4006fb: c5 fd 6f 44 24 40 vmovdqa 0x40(%rsp),%ymm0
400701: c5 fd 7f 84 24 40 02 00 00 vmovdqa %ymm0,0x240(%rsp)
40070a: c5 fd 6f 44 24 60 vmovdqa 0x60(%rsp),%ymm0
400710: c5 fd 7f 84 24 60 02 00 00 vmovdqa %ymm0,0x260(%rsp)
400719: c5 fd 6f 84 24 80 00 00 00 vmovdqa 0x80(%rsp),%ymm0
400722: c5 fd 7f 84 24 80 02 00 00 vmovdqa %ymm0,0x280(%rsp)
40072b: c5 fd 6f 84 24 a0 00 00 00 vmovdqa 0xa0(%rsp),%ymm0
400734: c5 fd 7f 84 24 a0 02 00 00 vmovdqa %ymm0,0x2a0(%rsp)
40073d: c5 fd 6f 84 24 c0 00 00 00 vmovdqa 0xc0(%rsp),%ymm0
400746: c5 fd 7f 84 24 c0 02 00 00 vmovdqa %ymm0,0x2c0(%rsp)
40074f: c5 fd 6f 84 24 e0 00 00 00 vmovdqa 0xe0(%rsp),%ymm0
400758: c5 fd 7f 84 24 e0 02 00 00 vmovdqa %ymm0,0x2e0(%rsp)
while in GCC 5.3.1 you see
4006fa: c5 fd 7f 85 f0 fe ff ff vmovdqa %ymm0,-0x110(%rbp)
400702: c5 fd 7f 85 10 ff ff ff vmovdqa %ymm0,-0xf0(%rbp)
40070a: c5 fd 7f 85 30 ff ff ff vmovdqa %ymm0,-0xd0(%rbp)
400712: c5 fd 7f 85 50 ff ff ff vmovdqa %ymm0,-0xb0(%rbp)
40071a: c5 fd 7f 85 70 ff ff ff vmovdqa %ymm0,-0x90(%rbp)
400722: c5 fd 7f 45 90 vmovdqa %ymm0,-0x70(%rbp)
400727: c5 fd 7f 45 b0 vmovdqa %ymm0,-0x50(%rbp)
40072c: c5 fd 7f 45 d0 vmovdqa %ymm0,-0x30(%rbp)
Whereupon the reads for the assert() then fail.
The Standard's Ruling on Pointer Aliasing:
ISO C++11 refers to aliasing under the following section, which makes clear that variables of type __m256i* cannot be accessed using bool*, but may be accessed with a char*/unsigned char*:
§ 3.10 Lvalues and rvalues [basic.lval]
[...]
If a program attempts to access the stored value of an object through a glvalue of other than one of the following types the behavior is undefined: [52]
- the dynamic type of the object,
- a cv-qualified version of the dynamic type of the object,
- a type similar (as defined in 4.4) to the dynamic type of the object,
- a type that is the signed or unsigned type corresponding to the dynamic type of the object,
- a type that is the signed or unsigned type corresponding to a cv-qualified version of the dynamic type of the object,
- an aggregate or union type that includes one of the aforementioned types among its elements or non-static data members (including, recursively, an element or non-static data member of a subaggregate or contained union),
- a type that is a (possibly cv-qualified) base class type of the dynamic type of the object,
- a
charorunsigned chartype.
52) The intent of this list is to specify those circumstances in which an object may or may not be aliased.
Appendix
GCC 4.8.5:
0000000000400620 <_Z4Mainv>:
400620: 55 push %rbp
400621: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400625: ba e5 08 40 00 mov $0x4008e5,%edx
40062a: b8 20 00 00 00 mov $0x20,%eax
40062f: 48 89 e5 mov %rsp,%rbp
400632: 48 83 e4 e0 and $0xffffffffffffffe0,%rsp
400636: 48 81 ec 00 03 00 00 sub $0x300,%rsp
40063d: c5 fd 7f 44 24 60 vmovdqa %ymm0,0x60(%rsp)
400643: c5 fd 7f 44 24 40 vmovdqa %ymm0,0x40(%rsp)
400649: c5 fd 7f 44 24 20 vmovdqa %ymm0,0x20(%rsp)
40064f: c5 fd 7f 04 24 vmovdqa %ymm0,(%rsp)
400654: c5 fd 7f 84 24 e0 00 00 00 vmovdqa %ymm0,0xe0(%rsp)
40065d: c5 fd 7f 84 24 c0 00 00 00 vmovdqa %ymm0,0xc0(%rsp)
400666: c5 fd 7f 84 24 a0 00 00 00 vmovdqa %ymm0,0xa0(%rsp)
40066f: c5 fd 7f 84 24 80 00 00 00 vmovdqa %ymm0,0x80(%rsp)
400678: 0f 1f 84 00 00 00 00 00 nopl 0x0(%rax,%rax,1)
400680: 48 83 c2 01 add $0x1,%rdx
400684: c6 04 04 01 movb $0x1,(%rsp,%rax,1)
400688: 0f b6 42 ff movzbl -0x1(%rdx),%eax
40068c: 84 c0 test %al,%al
40068e: 75 f0 jne 400680 <_Z4Mainv+0x60>
400690: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400694: 31 c0 xor %eax,%eax
400696: c5 fd 7f 84 24 60 01 00 00 vmovdqa %ymm0,0x160(%rsp)
40069f: c5 fd 7f 84 24 40 01 00 00 vmovdqa %ymm0,0x140(%rsp)
4006a8: c5 fd 7f 84 24 20 01 00 00 vmovdqa %ymm0,0x120(%rsp)
4006b1: c5 fd 7f 84 24 00 01 00 00 vmovdqa %ymm0,0x100(%rsp)
4006ba: c5 fd 7f 84 24 e0 01 00 00 vmovdqa %ymm0,0x1e0(%rsp)
4006c3: c5 fd 7f 84 24 c0 01 00 00 vmovdqa %ymm0,0x1c0(%rsp)
4006cc: c5 fd 7f 84 24 a0 01 00 00 vmovdqa %ymm0,0x1a0(%rsp)
4006d5: c5 fd 7f 84 24 80 01 00 00 vmovdqa %ymm0,0x180(%rsp)
4006de: c5 fd 6f 04 24 vmovdqa (%rsp),%ymm0
4006e3: c5 fd 7f 84 24 00 02 00 00 vmovdqa %ymm0,0x200(%rsp)
4006ec: c5 fd 6f 44 24 20 vmovdqa 0x20(%rsp),%ymm0
4006f2: c5 fd 7f 84 24 20 02 00 00 vmovdqa %ymm0,0x220(%rsp)
4006fb: c5 fd 6f 44 24 40 vmovdqa 0x40(%rsp),%ymm0
400701: c5 fd 7f 84 24 40 02 00 00 vmovdqa %ymm0,0x240(%rsp)
40070a: c5 fd 6f 44 24 60 vmovdqa 0x60(%rsp),%ymm0
400710: c5 fd 7f 84 24 60 02 00 00 vmovdqa %ymm0,0x260(%rsp)
400719: c5 fd 6f 84 24 80 00 00 00 vmovdqa 0x80(%rsp),%ymm0
400722: c5 fd 7f 84 24 80 02 00 00 vmovdqa %ymm0,0x280(%rsp)
40072b: c5 fd 6f 84 24 a0 00 00 00 vmovdqa 0xa0(%rsp),%ymm0
400734: c5 fd 7f 84 24 a0 02 00 00 vmovdqa %ymm0,0x2a0(%rsp)
40073d: c5 fd 6f 84 24 c0 00 00 00 vmovdqa 0xc0(%rsp),%ymm0
400746: c5 fd 7f 84 24 c0 02 00 00 vmovdqa %ymm0,0x2c0(%rsp)
40074f: c5 fd 6f 84 24 e0 00 00 00 vmovdqa 0xe0(%rsp),%ymm0
400758: c5 fd 7f 84 24 e0 02 00 00 vmovdqa %ymm0,0x2e0(%rsp)
400761: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
400768: 80 3c 04 00 cmpb $0x0,(%rsp,%rax,1)
40076c: 0f b6 8c 04 00 02 00 00 movzbl 0x200(%rsp,%rax,1),%ecx
400774: ba 01 00 00 00 mov $0x1,%edx
400779: 75 08 jne 400783 <_Z4Mainv+0x163>
40077b: 0f b6 94 04 00 01 00 00 movzbl 0x100(%rsp,%rax,1),%edx
400783: 38 d1 cmp %dl,%cl
400785: 0f 85 b2 00 00 00 jne 40083d <_Z4Mainv+0x21d>
40078b: 48 83 c0 01 add $0x1,%rax
40078f: 48 3d 00 01 00 00 cmp $0x100,%rax
400795: 75 d1 jne 400768 <_Z4Mainv+0x148>
400797: c5 fd 6f 8c 24 00 02 00 00 vmovdqa 0x200(%rsp),%ymm1
4007a0: 31 c0 xor %eax,%eax
4007a2: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007a7: 0f 94 c0 sete %al
4007aa: 0f 85 88 00 00 00 jne 400838 <_Z4Mainv+0x218>
4007b0: c5 fd 6f 8c 24 20 02 00 00 vmovdqa 0x220(%rsp),%ymm1
4007b9: 31 c0 xor %eax,%eax
4007bb: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007c0: 0f 94 c0 sete %al
4007c3: 75 73 jne 400838 <_Z4Mainv+0x218>
4007c5: c5 fd 6f 8c 24 40 02 00 00 vmovdqa 0x240(%rsp),%ymm1
4007ce: 31 c0 xor %eax,%eax
4007d0: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007d5: 0f 94 c0 sete %al
4007d8: 75 5e jne 400838 <_Z4Mainv+0x218>
4007da: c5 fd 6f 8c 24 60 02 00 00 vmovdqa 0x260(%rsp),%ymm1
4007e3: 31 c0 xor %eax,%eax
4007e5: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007ea: 0f 94 c0 sete %al
4007ed: 75 49 jne 400838 <_Z4Mainv+0x218>
4007ef: c5 fd 6f 8c 24 80 02 00 00 vmovdqa 0x280(%rsp),%ymm1
4007f8: 31 c0 xor %eax,%eax
4007fa: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
4007ff: 0f 94 c0 sete %al
400802: 75 34 jne 400838 <_Z4Mainv+0x218>
400804: c5 fd 6f 8c 24 a0 02 00 00 vmovdqa 0x2a0(%rsp),%ymm1
40080d: 31 c0 xor %eax,%eax
40080f: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
400814: 0f 94 c0 sete %al
400817: 75 1f jne 400838 <_Z4Mainv+0x218>
400819: c5 fd 6f 8c 24 c0 02 00 00 vmovdqa 0x2c0(%rsp),%ymm1
400822: 31 c0 xor %eax,%eax
400824: c4 e2 7d 17 c9 vptest %ymm1,%ymm1
400829: 0f 94 c0 sete %al
40082c: 75 0a jne 400838 <_Z4Mainv+0x218>
40082e: 31 c0 xor %eax,%eax
400830: c4 e2 7d 17 c0 vptest %ymm0,%ymm0
400835: 0f 94 c0 sete %al
400838: c5 f8 77 vzeroupper
40083b: c9 leaveq
40083c: c3 retq
40083d: b9 20 09 40 00 mov $0x400920,%ecx
400842: ba 26 00 00 00 mov $0x26,%edx
400847: be e9 08 40 00 mov $0x4008e9,%esi
40084c: bf f8 08 40 00 mov $0x4008f8,%edi
400851: c5 f8 77 vzeroupper
400854: e8 97 fc ff ff callq 4004f0 <__assert_fail@plt>
400859: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
GCC 5:
0000000000400630 <_Z4Mainv>:
400630: 4c 8d 54 24 08 lea 0x8(%rsp),%r10
400635: 48 83 e4 e0 and $0xffffffffffffffe0,%rsp
400639: b8 20 00 00 00 mov $0x20,%eax
40063e: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400642: ba 25 08 40 00 mov $0x400825,%edx
400647: 41 ff 72 f8 pushq -0x8(%r10)
40064b: 55 push %rbp
40064c: 48 89 e5 mov %rsp,%rbp
40064f: 41 52 push %r10
400651: 48 81 ec 08 03 00 00 sub $0x308,%rsp
400658: c5 fd 7f 85 50 fd ff ff vmovdqa %ymm0,-0x2b0(%rbp)
400660: c5 fd 7f 85 30 fd ff ff vmovdqa %ymm0,-0x2d0(%rbp)
400668: c5 fd 7f 85 10 fd ff ff vmovdqa %ymm0,-0x2f0(%rbp)
400670: c5 fd 7f 85 f0 fc ff ff vmovdqa %ymm0,-0x310(%rbp)
400678: c5 fd 7f 85 d0 fd ff ff vmovdqa %ymm0,-0x230(%rbp)
400680: c5 fd 7f 85 b0 fd ff ff vmovdqa %ymm0,-0x250(%rbp)
400688: c5 fd 7f 85 90 fd ff ff vmovdqa %ymm0,-0x270(%rbp)
400690: c5 fd 7f 85 70 fd ff ff vmovdqa %ymm0,-0x290(%rbp)
400698: 0f 1f 84 00 00 00 00 00 nopl 0x0(%rax,%rax,1)
4006a0: 48 83 c2 01 add $0x1,%rdx
4006a4: c6 84 05 f0 fc ff ff 01 movb $0x1,-0x310(%rbp,%rax,1)
4006ac: 0f b6 42 ff movzbl -0x1(%rdx),%eax
4006b0: 84 c0 test %al,%al
4006b2: 75 ec jne 4006a0 <_Z4Mainv+0x70>
4006b4: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
4006b8: 31 c0 xor %eax,%eax
4006ba: c5 fd 7f 85 50 fe ff ff vmovdqa %ymm0,-0x1b0(%rbp)
4006c2: c5 fd 7f 85 30 fe ff ff vmovdqa %ymm0,-0x1d0(%rbp)
4006ca: c5 fd 7f 85 10 fe ff ff vmovdqa %ymm0,-0x1f0(%rbp)
4006d2: c5 fd 7f 85 f0 fd ff ff vmovdqa %ymm0,-0x210(%rbp)
4006da: c5 fd 7f 85 d0 fe ff ff vmovdqa %ymm0,-0x130(%rbp)
4006e2: c5 fd 7f 85 b0 fe ff ff vmovdqa %ymm0,-0x150(%rbp)
4006ea: c5 fd 7f 85 90 fe ff ff vmovdqa %ymm0,-0x170(%rbp)
4006f2: c5 fd 7f 85 70 fe ff ff vmovdqa %ymm0,-0x190(%rbp)
4006fa: c5 fd 7f 85 f0 fe ff ff vmovdqa %ymm0,-0x110(%rbp)
400702: c5 fd 7f 85 10 ff ff ff vmovdqa %ymm0,-0xf0(%rbp)
40070a: c5 fd 7f 85 30 ff ff ff vmovdqa %ymm0,-0xd0(%rbp)
400712: c5 fd 7f 85 50 ff ff ff vmovdqa %ymm0,-0xb0(%rbp)
40071a: c5 fd 7f 85 70 ff ff ff vmovdqa %ymm0,-0x90(%rbp)
400722: c5 fd 7f 45 90 vmovdqa %ymm0,-0x70(%rbp)
400727: c5 fd 7f 45 b0 vmovdqa %ymm0,-0x50(%rbp)
40072c: c5 fd 7f 45 d0 vmovdqa %ymm0,-0x30(%rbp)
400731: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
400738: 0f b6 94 05 f0 fc ff ff movzbl -0x310(%rbp,%rax,1),%edx
400740: 0f b6 8c 05 f0 fe ff ff movzbl -0x110(%rbp,%rax,1),%ecx
400748: 84 d2 test %dl,%dl
40074a: 75 08 jne 400754 <_Z4Mainv+0x124>
40074c: 0f b6 94 05 f0 fd ff ff movzbl -0x210(%rbp,%rax,1),%edx
400754: 38 d1 cmp %dl,%cl
400756: 75 2c jne 400784 <_Z4Mainv+0x154>
400758: 48 83 c0 01 add $0x1,%rax
40075c: 48 3d 00 01 00 00 cmp $0x100,%rax
400762: 75 d4 jne 400738 <_Z4Mainv+0x108>
400764: c5 f9 ef c0 vpxor %xmm0,%xmm0,%xmm0
400768: 31 c0 xor %eax,%eax
40076a: c4 e2 7d 17 c0 vptest %ymm0,%ymm0
40076f: 0f 94 c0 sete %al
400772: c5 f8 77 vzeroupper
400775: 48 81 c4 08 03 00 00 add $0x308,%rsp
40077c: 41 5a pop %r10
40077e: 5d pop %rbp
40077f: 49 8d 62 f8 lea -0x8(%r10),%rsp
400783: c3 retq
400784: b9 60 08 40 00 mov $0x400860,%ecx
400789: ba 26 00 00 00 mov $0x26,%edx
40078e: be 29 08 40 00 mov $0x400829,%esi
400793: bf 38 08 40 00 mov $0x400838,%edi
400798: c5 f8 77 vzeroupper
40079b: e8 50 fd ff ff callq 4004f0 <__assert_fail@plt>
这篇关于GCC 5和更高版本中对AVX2的支持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！
查看全文
