通过

Vectorizer and Parallelizer Messages

  • 项目

可以使用 Visual C++ 编译器选项 /Qpar-report 和 /Qvec-report 设置 自动并行化错误和警告 到输出原因代码和信息性消息有关其事件。 本文说明原因代码和消息。

信息性消息

根据您指定的报告级别，下面的信息性消息之一为每个循环显示。

有关原因代码的信息，请参见本文的下一节。

信息性消息

描述

5001

向量化的循环。

5002

循环不向量化由于原因“declaration”。

5011

并行化的循环。

5012

循环不并行化的由于原因“declaration”。

5021

无法将循环用杂注。

原因代码

以下部分列出自动 parallelizer 的可能原因代码和自动 vectorizer。

JJ658585.collapse_all(zh-cn,VS.110).gif50x

50x 原因代码应用于自动 parallelizer 和自动 vectorizer。

原因代码

说明

500

这是包含多个用例 (例如的一般消息，该循环包含多个退出点，、或循环标头不会增加归纳变量结束。

501

归纳变量不本地；或上限不是循环固定。

502

除了简单 +1. 外，归纳变量通过单步执行。

503

循环包含异常处理或 switch 语句。

504

循环体可能会引发需要 c. c++ 对象析构的异常。

 
void code_500(int *A)
{
    // Code 500 is emitted if the loop has non-vectorizable flow.
    // This can include "if", "break", "continue", the conditional 
    // operator "?", or function calls.
    // It also encompasses correct definition and use of the induction
    // variable "i", in that the increment "++i" or "i++" must be the last
    // statement in the loop.

    int i = 0;
    while (i<1000)
    {
        if (i == 4) 
        {
            break;
        }

        ++i;

        A[i] = A[i] + 1;
    }
    // To resolve code 500, use a 'for' loop with single increment of 
    // induction variable.

    for (int i=0; i<1000; ++i)
    {       
        A[i] = A[i] + 1;
    }    
}

int bound();
void code_501_example1(int *A)
{
    // Code 501 is emitted if the compiler cannot discern the
    // induction variable of this loop. In this case, when it checks
    // the upperbound of 'i', the compiler cannot prove that the 
    // function call "bound()" returns the same value each time.
    // Also, the compiler cannot prove that the call to "bound()"
    // does not modify the values of array A.

    for (int i=0; i<bound(); ++i)
    {
        A[i] = A[i] + 1;
    }

    // To resolve code 501, ensure that the induction variable is 
    // a local variable, and ensure that the upperbound is a
    // provably loop invariant value.

    for (int i=0, imax = bound(); i<imax; ++i)
    {
        A[i] = A[i] + 1;
    }
}

int i;
void code_501_example2(int *A)
{
    // Code 501 is emitted if the compiler cannot discern the
    // induction variable of this loop. In this case, 'i' is
    // a global.

    for (i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }

    // To resolve code 501, ensure that the induction variable is 
    // a local variable, and ensure that the upperbound is a
    // provably loop invariant value.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

void code_502(int *A)
{
    // Code 502 is emitted if the compiler cannot discern
    // the induction variable of the loop. In this case,
    // there are three increments to "i", one of which
    // is conditional.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
        ++i;

        if (i < 100) 
        {
            ++i;
        }
    }

    // To resolve code 502, ensure that there is just one 
    // increment of the induction variable, placed in the usual
    // spot in the "for" loop.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}


// compile with /EHsc
void code_503(int *A, int x)
{
    // Code 503 is emitted if there are inadmissible
    // operations in the loop - for example, exception handling and
    // switch statements.

    for (int i=0; i<1000; ++i)
    {
        try 
        {
            A[i] = A[i] + 1;
        }
        catch (...)
        {

        }

        switch (x)
        {
        case 1: A[i] = A[i] + 1;
        case 2: A[i] = A[i] + 2;
        case 3: A[i] = A[i] + 3;
            break;
        }
    }

    // To resolve code 503, try to remove as many switch statements
    // and exception handling constructs as possible.
}

// compile with /EHsc

int code_504_helper();
class C504
{
public:
    C504();
    ~C504();
};

void code_504() {
    // Code 504 is emitted if a C++ object was created and
    // that object requires EH unwind tracking information under
    // /EHs or /EHsc.

    for(int i = 0; i < 1000; ++i)
    {
        C504 c;
        A[i] = code_504_helper();
    }

}

JJ658585.collapse_all(zh-cn,VS.110).gif100x

100x 原因代码应用于自动 parallelizer。

原因代码

说明

1000

编译器检测到在循环体中的数据依赖项。

1001

编译器检测到存储到在循环体的标量变量，因此，该标量在循环外的使用。

1002

编译器尝试并行化具有内部循环已经并行化的循环。

1003

循环体包含可读取或写入内存的内部调用。

1004

对循环体的标量降低。 如果循环向量化，标量减少会发生此错误。

1005

no_parallel 说明指定的。

1006

此功能包含 openmp。 通过移除此函数的所有 openmp 解决此问题。

1007

循环变量的归纳或循环边界不是有符号的 32 位数字 (int 或 long)。 通过更改归纳变量的类型解决此问题。

1008

编译器检测到此循环不执行足够的工作来保证自动并行化。

1009

编译器检测出尝试并行化“，则将在”循环。 仅自动 parallelizer 目标 for 循环。

 
void func();
void code_1000()
{
    // Code 1000 is emitted if the compiler detects a 
    // data dependence in the loop body. 

    // You can resolve this by using the ivdep pragma.
    // CAUTION -- the compiler will trust your
    // assertion that there are no data dependencies
    // in the loop body. If there are, you are generating
    // code that may have race conditions.

#pragma loop(hint_parallel(0))
    //#pragma loop(ivdep) // ivdep will force this through.
    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i-1] + 1;  // data dependence here
        func();             // data dependence here
    }
}

int code_1001()
{
    // Code 1001 is emitted if the compiler detects
    // a store to a scalar variable in the loop
    // body, and that scalar has a use beyond the loop.

    // Resolve this by rewriting your code so
    // that the scalar is not needed.

    int s = 0;
#pragma loop(hint_parallel(0))
    for (int i=0; i<1000; ++i)
    {
        s = A[i];
    }
    return s;
}

void code_1002()
{
    // Code 1002 is emitted when the compiler tries to
    // parallelize a loop that has an inner loop that
    // has already been parallelized.

#pragma loop(hint_parallel(0))
    for (int i=0; i<1000; ++i) // emit code 1002 for this loop
    {
#pragma loop(hint_parallel(0))
        for (int j=0; j<1000; ++j) // this loop gets parallelized
        {
            A[j] = A[j] + 1;
        }
    }
}

extern "C" void __stosb(unsigned char*, unsigned char, size_t);
void code_1003(unsigned char *dst)
{
    // Code 1003 is emitted when the loop body contains an intrinsic
    // call that may read or write to memory.

    // This can be resolved by using the ivdep pragma.
    // CAUTION -- the compiler will trust your
    // assertion that there are no data dependencies
    // in the loop body. If there are, you are generating
    // code that may have race conditions.

#pragma loop(hint_parallel(0))
    //#pragma loop(ivdep) // ivdep will force this through.
    for (int i=0; i<1000; ++i)
    {
        __stosb(dst, 'c', 10);
        A[i] = A[i] + 1;
    }
}

int code_1004()
{
    // Code 1004 is emitted when there is a scalar reduction
    // in the loop body, which can occur if the loop has been
    // vectorized.

    // You can resolve this by rewriting your code so that it
    // does not have a scalar reduction.

    int s = 0;
#pragma loop(hint_parallel(0))
    for (int i=0; i<1000; ++i)
    {
        s += A[i];
    }
    return s;
}

void code_1005()
{
    // Code 1005 is emitted when the 
    // no_parallel pragma is specified.

#pragma loop(no_parallel)
    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

#include <omp.h>

// Compile with /openmp
void code_1006()
{
    // Code 1006 is emitted when this function contains
    // openmp. Resolve this by removing any openmp in this
    // function.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }

#pragma omp parallel num_threads(4)
    {
        int i = omp_get_thread_num();
        A[i] = A[i] + 1;
    }
}

void code_1007()
{
    // Code 1007 is emitted when the loop induction variable
    // or the loop bounds are not signed 32-bit numbers (int 
    // or long). Resolve this by changing the type of the 
    // induction variable.

#pragma loop(hint_parallel(0))
    for (unsigned int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

void code_1008()
{
    // Code 1008 is emitted when the compiler detects that
    // this loop does not perform enough work to warrant 
    // auto-parallelization.

    // You can resolve this by specifying the hint_parallel
    // pragma. CAUTION -- if the loop does not perform
    // enough work, parallelizing might cause a potentially 
    // large performance penalty.

    // #pragma loop(hint_parallel(0)) //  hint_parallel will force this through
    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

int A[1000];
void code_1009()
{
    // Code 1009 is emitted when the compiler tries to parallelize a 
    // "do-while" loop. The auto-parallelizer only targets "for" loops.

    int i = 0;
#pragma loop(hint_parallel(0))
    do
    {
        A[i] = A[i] + 1;
    } 
    while (++i < 1000);
}

JJ658585.collapse_all(zh-cn,VS.110).gif110x

110x 原因代码应用于自动 vectorizer。

原因代码

说明

1100

循环包含控制流 (例如，“，则”或“?”。

1101

循环包含数据类型转换可能隐式不能向量化。

1102

循环包含非算术或其他非 vectorizable 操作。

1103

循环体包括大小可能在循环内更改移位操作数。

1104

循环体包含标量变量。

1105

循环包含无法识别的减少操作。

1106

内部循环已经向量化。 不还可以 vectorize 外层循环

 
void code_1100(int *A, int x) 
{
    // Code 1100 is emitted when the compiler detects control flow
    // in the loop - for example, "if", the ternary operator "?", and
    // the like. Resolve this by flattening or removing control
    // flow in the loop body.

    // Not all control flow causes 1100; some is indeed  
    // vectorized.

    for (int i=0; i<1000; ++i)
    {
        // straightline code is more amenable to vectorization
        if (x)
        {
            A[i] = A[i] + 1;
        }
    }
}

void code_1101(int *A, char *B)
{
    // Code 1101 is emitted when the compiler is unable to vectorize
    // convert operations in the loop body. Many C/C++ conversions
    // are implicit. 

    // In this example, the 1-byte load of "B[i]" is converted to 4 bytes
    // prior to the addition by "1". The compiler, although able to 
    // vectorize most conversions, does not vectorize char -> int 
    // conversions.

    // Resolve this by eliminating conversions where possible.

    for (int i=0; i<1000; ++i)
    {
        A[i] = B[i] + 1;
    }
}

extern "C"  long _InterlockedExchange(long * Target, long Value);
void code_1102(int *A, long *x)
{
    // Code 1102 is emitted when the compiler is unable to vectorize
    // an operation in the loop body. For example, intrinsics and other
    // non-arithmetic, non-logical, and non-memory operations are not
    // vectorizable.

    // Resolve this by removing as many non-vectorizable operations
    // as possible from the loop body.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + _InterlockedExchange(x, 1);
    }
}

void code_1103(int *A, int *B)
{
    // Code 1103 is emitted when the compiler is unable to vectorize
    // a "shift" operation. In this example, there are two shifts
    // that cannot be vectorized.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] >> B[i]; // not vectorizable

        int x = B[i];
        A[i] = A[i] >> x; // not vectorizable
    }

    // To resolve this, ensure that your shift amounts are loop 
    // invariant. If the shift amounts cannot be loop invariant,
    // it may not be possible to vectorize this loop.

    int x = B[0];
    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] >> x; // vectorizable
    }
}

int code_1104(int *A, int *B)
{
    // When it vectorizes a loop, the compiler must 'expand' scalar
    // variables to a vector size such that they can fit in
    // vector registers. Code 1104 is emitted when the compiler
    // cannot 'expand' such scalars.

    // In this example, we try to 'expand' x to be used in the 
    // vectorized loop. However, there is a use of 'x' 
    // beyond the loop body, which prohibits this expansion.

    // To resolve this, try to limit scalars to be used only in
    // the loop body and not beyond, and try to keep their types
    // consistent with the loop types.

    int x;
    for (int i=0; i<1000; ++i)
    {
        x = B[i];
        A[i] = A[i] + x;
    }

    return x;
}

int code_1105(int *A)
{
    // The compiler performs an optimization that's known as "reduction"
    // when it operates on each element of an array and computes
    // a resulting scalar value - for example, in this piece of code, which
    // computes the sum of each element in the array:

    int s = 0;
    for (int i=0; i<1000; ++i)
    {
        s += A[i]; // vectorizable
    }

    // The reduction pattern must resemble the loop in the example. The
    // compiler emits code 1105 if it cannot deduce the reduction
    // pattern, as shown in this example:

    for (int i=0; i<1000; ++i)
    {
        s += A[i] + s;  // code 1105
    }

    // Similarly, reductions of "float" or "double" types require
    // that the /fp:fast switch is thrown. Strictly speaking,
    // the reduction optimization that the compiler performs uses
    // "floating point reassociation". Reassociation is only
    // allowed when /fp:fast is thrown.

    return s;    
}

void code_1106(int *A)
{
    // Code 1106 is emitted when the compiler tries to vectorize
    // an outer loop.

    for (int i=0; i<1000; ++i) // this loop is not vectorized
    {
        for (int j=0; j<1000; ++j) // this loop is vectorized
        {
            A[j] = A[j] + 1;
        }
    }
}

JJ658585.collapse_all(zh-cn,VS.110).gif120x

120x 原因代码应用于自动 vectorizer。

原因代码

说明

1200

循环包含阻止 vectorization 的循环包含的数据依赖项。 循环的其他迭代相互影响这样 vectorizing 循环将错误导致答案，并且，自动 vectorizer 不能证明到自身没有此类数据依赖项。

1201

数组在循环中的基础更改。

1202

在结构的字段不宽 32 或 64 位。

1203

循环体包含非连续的访问数组。

 
void fn();
void code_1200(int *A)
{
    // Code 1200 is emitted when data dependence is prohibiting
    // vectorization. This can only be resolved by rewriting the
    // loop, and considering the marking of loop function calls as 
    // __forceinline.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i-1] + 1; // vectorization-prohibiting
        fn();               // vectorization-prohibiting
    }
}

void code_1201(int *A)
{
    // Code 1201 is emitted when an array base changes
    // in the loop body. Resolve this by rewriting your
    // code so that varying the array base is not necessary.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
        A++;
    }
}

typedef struct S_1202
{
    short a;
    short b;
} S_1202;

void code_1202(S_1202 *s)
{
    // Code 1202 is emitted when non-vectorizable struct accesses
    // are present in the loop body. Only struct accesses 
    // that are 32 or 64 bits are vectorized.

    for (int i=0; i<1000; ++i)
    {        
        s[i].a = s[i].b + 1; // this 16 bit struct access is not vectorizable
    }
}

void code_1203(int *A)
{
    // Code 1203 is emitted when non-vectorizable memory references
    // are present in the loop body. Vectorization of some non-contiguous 
    // memory access is supported - for example, the gather/scatter pattern.

    for (int i=0; i<1000; ++i)
    {
        A[i] += A[0] + 1;       // constant memory access not vectorized
        A[i] += A[i*2+2] + 2;  // non-contiguous memory access not vectorized
    }
}

JJ658585.collapse_all(zh-cn,VS.110).gif130x

130x 原因代码应用于自动 vectorizer。

原因代码

说明

1300

循环体包含 NO-或非常小计算。

1301

循环、不是 +1。

1302

循环是“，则将在”。

1303

vectorization 的很少循环迭代可以提供值。

1304

循环包含不同大小的分配。

1305

没有足够的信息。

 
void code_1300(int *A, int *B)
{
    // Code 1300 is emitted when the compiler detects that there is
    // no computation in the loop body.

    for (int i=0; i<1000; ++i)
    {
        A[i] = B[i]; // Do not vectorize, instead emit memcpy
    }
}

void code_1301(int *A)
{
    // Code 1301 is emitted when the stride of a loop is not positive 1.
    // Only loops that have a stride of positive 1 are vectorized;
    // rewriting your loop may be required.

    for (int i=0; i<1000; i += 2)
    {
        A[i] = A[i] + 1;
    }
}

void code_1302(int *A)
{
    // Code 1302 is emitted for "do-while" loops. Only "while"  
    // and "for" loops are vectorized.

    int i = 0;
    do
    {
        A[i] = A[i] + 1;
    } while (++i < 1000);
}

int code_1303(int *A, int *B)
{
    // Code 1303 is emitted when the compiler detects that
    // the number of iterations of the loop is too small to
    // make vectorization profitable.

    // If the loop computation fits perfectly in 
    // vector registers - for example, the upperbound is 4, or 8 in 
    // this case - then the loop _may_ be vectorized.

    // This loop is not vectorized because there are 5 iterations

    for (int i=0; i<5; ++i)
    {
        A[i] = A[i] + 1;
    }

    // This loop is vectorized

    for (int i=0; i<4; ++i)
    {
        A[i] = A[i] + 1;
    }

    // This loop is not vectorized because runtime pointer checks
    // are required to check that A and B don't overlap. It is not
    // worth it to vectorize this loop.

    for (int i=0; i<4; ++i)
    {
        A[i] = B[i] + 1;
    }

    // This loop is not vectorized because of the scalar reduction.

    int s = 0;
    for (int i=0; i<4; ++i)
    {
        s += A[i];
    }
    return s;
}

void code_1304(int *A, short *B)
{
    // Code 1304 is emitted when the compiler detects
    // different sized statements in the loop body.
    // In this case, there is an 32-bit statement and a
    // 16-bit statement.

    // In cases like this consider splitting the loop into loops to 
    // maximize vector register utilization.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
        B[i] = B[i] + 1;
    }
}

typedef struct S_1305
{
    int a;
    int b;
} S_1305;

void code_1305( S_1305 *s, S_1305 x)
{
    // Code 1305 is emitted when the compiler can't discern
    // proper vectorizable type information for this loop.
    // This includes non-scalar loop types such as struct 
    // assignments, as in this example.

    // Resolve this by ensuring that your loops have statements
    // that operate on integers or floating point types.

    for (int i=0; i<1000; ++i)
    {
        s[i] = x;
    }
}

JJ658585.collapse_all(zh-cn,VS.110).gif140x

140x 原因代码时，会发生的自动 vectorization 不兼容的某些选项指定。

原因代码

说明

1400

#pragma loop(no_vector) 指定。

1401

/kernel 开关指定。

1402

/arch:IA32 开关指定。

1403

/arch:ATOM 开关指定，并且该循环包含在二进制文件的操作。

1404

/O1 或 /Os 开关指定。

 
void code_1400(int *A)
{
    // Code 1400 is emitted when the no_vector pragma 
    // is specified. 

#pragma loop(no_vector)
    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

// Compile with /kernel
void code_1401(int *A)
{
    // Code 1401 is emitted when /kernel is specified.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

// Compile with /arch:IA32
void code_1402(int *A)
{
    // Code 1401 is emitted when /arch:IA32 is specified.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

// Compile with /favor:ATOM
void code_1403(double *A)
{
    // Code 1401 is emitted when /favor:ATOM is specified, and
    // the loop contains operations on "double" arrays.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

// Compile with /O1 or /Os
void code_1404(int *A)
{
    // Code 1401 is emitted when compiling for size.

    for (int i=0; i<1000; ++i)
    {
        A[i] = A[i] + 1;
    }
}

JJ658585.collapse_all(zh-cn,VS.110).gif150x

150x 原因代码块适用于失真。 如果内存的位置可由两个不同的名称时，获取发生失真。

原因代码

说明

1500

在多维数组的失真。

1501

可能的失真数组结构。

1502

可能的失真和数组索引以外的 n + K。

1503

可能的失真和数组索引有多个偏移量。

1504

可能的锯齿化；需要许多运行时检查。

1505

可能的失真，但是，运行时检查太复杂。

 
void code_1500(int A[100][100], int B[100][100])
{
    // Code 1500 is emitted when runtime pointer
    // disambiguation checks are required, and 
    // there are multidimensional array references.

    for (int i=0; i<100; ++i)
    {
        for (int j=0; j<100; ++j)
        {
            A[i][j] = B[i][j] + 1;
        }
    }
}

typedef struct S_1501
{
    int a;
    int b;
} S_1501;

void code_1501(S_1501 *s1, S_1501 *s2)
{
    // Code 1501 is emitted when runtime pointer
    // disambiguation checks are required, and 
    // there are array-of-struct accesses in the 
    // loop body.

    for (int i=0; i<100; ++i)
    {
        s1[i].a = s2[i].b + 1;
    }
}

void code_1502(int *A, int *B)
{
    // Code 1502 is emitted when runtime pointer
    // disambiguation checks are required, and 
    // an array reference has an offset that varies 
    // in the loop.

    int x = 0;
    for (int i=0; i<100; ++i)
    {
        A[i] = B[i + x] + 1;
        ++x;                   // 'x' varies in the loop
    }
}

void code_1503(int *A, int *B, int x, int y)
{
    // Code 1503 is emitted when runtime pointer
    // disambiguation checks are required, and 
    // an array reference has multiple offsets.

    for (int i=0; i<100; ++i)
    {
        A[i] = B[i+x] + B[i+y] + 1;   // multiple offsets when addressing 'B': {x, y}
        A[i] = B[i+x] + B[i] + 1;     // multiple offsets when addressing 'B': {x, 0}
        A[i] = B[i+x] + B[i+x] + 1;   // this is vectorized
    }
}

void code_1504(int *A1, int *A2, int *A3, int *A4, 
               int *A5, int *A6, int *A7, int *A8,
               int *A9, int *A10, int *A11, int *A12,
               int *A13, int *A14, int *A15, int *A16)
{
    // Code 1504 is emitted when too many runtime 
    // pointer disambiguation checks are required.

    for (int i=0; i<100; ++i)
    {
        ++A1[i];
        ++A2[i];
        ++A3[i];
        ++A4[i];
        ++A5[i];
        ++A6[i];
        ++A7[i];
        ++A8[i];
        ++A9[i];
        ++A10[i];
        ++A11[i];
        ++A12[i];
        ++A13[i];
        ++A14[i];
        ++A15[i];
        ++A16[i];
    }
}

void code_1505(int *A, int *B)
{
    // Code 1505 is emitted when runtime pointer 
    // disambiguation checks are required, but are
    // too complex for the compiler to discern.

    for (int i=0; i<100; ++i)
    {
        for (int j=0; j<100; ++j)
        {
            for (int k=0; k<100; ++k)
            {
                A[i+j-k] = B[i-j+k] + 1;
            }
        }
    }
}

请参见

参考

自动并行化错误和警告

#pragma 循环 ()

/Q 选项（低级别操作）

/Qpar-report (Auto-Parallelizer Reporting Level)

/Qvec-report (Auto-Vectorizer Reporting Level)

其他资源

并行编程方式在本机代码