How to instruct compiler to generate unaligned loads for __m128

Question

I've got some code that works with __m128 values. I'm using x86-64 SSE intrinsics on these values and I find that if the values are unaligned in memory I get a crash. This is due to my compiler (clang in this instance) generating only aligned load instructions.

Can I instruct my compiler to generate unaligned loads instead, either globally or for certain values (perhaps with an annotation of some kind)?

---

The reason I have unaligned values in the first place is that I'm trying to save memory. I have a struct roughly as follows:

#pragma pack(push, 4)
struct Foobar {
    __m128 a;
    __m128 b;
    int c;
};
#pragma pack(pop)

I am then creating an array of these structs. The 2nd element in the array starts at 36 bytes, which is not a multiple of 16.

I know I could switch to a structure of arrays representation, or remove the packing pragma (at the cost of increasing the size of the struct from 36 to 48 bytes); but I also know that unaligned loads aren't that expensive these days and would like to try that first.

---

Update to answer some of the comments below:

My actual code was closer to this:

struct Vector4 {
    __m128 data;
    Vector4(__m128 v) : data(v) {}
};
struct Foobar {
    Vector4 a;
    Vector4 b;
    int c;
}

I then have some utility functions such as:

inline Vector4 add( const Vector4& a, const Vector4 &b ) {
    return Vector4(_mm_add_ps(a.data, b.data));
}

inline Vector4 subtract( const Vector4& a, const Vector4& b ) {
    return Vector4(_mm_sub_ps(a.data, b.data));
}

// etc..

I use these utilities often in combination. Fake example:

Foobar myArray[1000];
myArray[i+1].b = sub(add(myArray[i].a, myArray[i].b), myArray[i+1].a);

When looking at "Z Bozon"'s answer my code effectively changed into:

struct Vector4 {
    float data[4];
};

inline Vector4 add( const Vector4& a, const Vector4 &b ) {
    Vector4 result;
    _mm_storeu_ps(result.data, _mm_add_ps(_mm_loadu_ps(a.data), _mm_loadu_ps(b.data)));
    return result;
}

My concern was that when the utility functions were used in combination as above, that the generated code might have redundant load/store instructions. It turns out this was not a problem. I tested my compiler (clang), and it had removed them all. I'll accept Z Bozon's answer.

Answer 1

Clang has -fmax-type-align. If you set -fmax-type-align=8 then no 16-byte aligned instruction will be generated.

Answer 2

In my opinion you should write your data structures using standard C++ constructions (of which __m128i is not). When you want to use intrinsics which are are not standard C++ you "enter SSE world" through intrinsics such as _mm_loadu_ps and you "leave SSE world" back to standard C++ with an intrinsic such as _mm_storeu_ps. Don't rely on implicit SSE loads and stores. I have seen too many mistakes on SO doing this.

In this case you should use

struct Foobar {
    float a[4];
    float b[4];
    int c;
};

then you can do

Foobar foo[16];

In this case foo[1] won't be 16 byte aligned but when you want to use SSE and leave standard C++ do

__m128 a4 = _mm_loadu_ps(foo[1].a);
__m128 b4 = _mm_loadu_ps(foo[1].b);
__m128 max = _mm_max_ps(a4,b4);
_mm_storeu_ps(array, max);

then go back to standard C++.

Another thing you can consider is this

struct Foobar {
    float a[16];
    float b[16];
    int c[4];
};

then to get an array of 16 of the original struct do

Foobar foo[4];

In this case as long the first element is aligned so are all the other elements.

---

If you want utility functions which act on SSE registers then don't use explicit or implicit load/stores in the utility function. Pass const references to __m128 and return __m128 if you need to.

//SSE utility function
static inline __m128 mulk_SSE(__m128 const &a, float k)
{
    return _mm_mul_ps(_mm_set1_ps(k),a);
}

//main function
void foo(float *x, float *y n) 
{
    for(int i=0; i<n; i+=4)
        __m128 t1 = _mm_loadu_ps(x[i]);
        __m128 t2 = mulk_SSE(x4,3.14159f);
        _mm_store_ps(&y[i], t2);
    }
}

The reason to use a const reference is that MSVC cannot pass __m128 by value. Without a const reference you get an error

error C2719: formal parameter with __declspec(align('16')) won't be aligned.

__m128 for MSVC is really a union anyway.

typedef union __declspec(intrin_type) _CRT_ALIGN(16) __m128 {
     float               m128_f32[4];
     unsigned __int64    m128_u64[2];
     __int8              m128_i8[16];
     __int16             m128_i16[8];
     __int32             m128_i32[4];
     __int64             m128_i64[2];
     unsigned __int8     m128_u8[16];
     unsigned __int16    m128_u16[8];
     unsigned __int32    m128_u32[4];
 } __m128;

presumably MSVC should not have to load the union when the SSE utility functions are inlined.

---

Based on the OPs latest code update here is what I would suggest

#include <x86intrin.h>
struct Vector4 {
    __m128 data;
    Vector4() {
    }
    Vector4(__m128 const &v) {
        data = v;
    }
    Vector4 & load(float const *x) {
        data = _mm_loadu_ps(x);
        return *this;
    }
    void store(float *x) const {
        _mm_storeu_ps(x, data);
    }
    operator __m128() const {
        return data;
    }
};

static inline Vector4 operator + (Vector4 const & a, Vector4 const & b) {
    return _mm_add_ps(a, b);
}

static inline Vector4 operator - (Vector4 const & a, Vector4 const & b) {
    return _mm_sub_ps(a, b);
}

struct Foobar {
    float a[4];
    float b[4];
    int c;
};

int main(void)
{
    Foobar myArray[10];
    // note that myArray[0].a, myArray[0].b, and myArray[1].b should be      // initialized before doing the following 
    Vector4 a0 = Vector4().load(myArray[0].a);
    Vector4 b0 = Vector4().load(myArray[0].b);
    Vector4 a1 = Vector4().load(myArray[1].a);        
    (a0 + b0 - a1).store(myArray[1].b);
}

This code was based on ideas from Agner Fog's Vector Class Library.

Answer 3

If you use auto-vectorization or explicit OpenMP4/Cilk/pragmas-driven vectorization, then you can enforce compiler to use unaligned loads for vectorized loop by using:

#pragma vector unaligned //for C/C++ 

CDEC$ vector unaligned ; for Fortran

This is primarily intended to control trade-offs between "aligned but peeled" vs. "not peeled, but unaligned". Read more details at https://software.intel.com/en-us/articles/utilizing-full-vectors

This only works for Intel Compilers as far as I know. Intel Compilers also have internal compilation switch -mP2OPT_vec_alignment=6 to do the same for the whole compilation unit.

I didn't check if it could be effectively applied to implementations where intrinsics/assembly is used together with OpenMP/Cilk.

Answer 4

You could try changing your struct to:

#pragma pack(push, 4)
struct Foobar {
    int c;
    __m128 a;
    __m128 b;
};
#pragma pack(pop)

That would have the same size of course, and should in theory force clang to generate unaligned loads/stores.

---

Alternatively you could use explicit unaligned loads/stores, e.g. change:

v = _mm_max_ps(myArray[300].a, myArray[301].a)

to:

__m128i v1 = _mm_loadu_ps((float *)&myArray[300].a);
__m128i v2 = _mm_loadu_ps((float *)&myArray[301].a);
v = _mm_max_ps(v1, v2);