[Bug c/108191] New: Add support to usage of *intrin.h without -mavx512f -mavx512cd

[Bug c/108191] New: Add support to usage of *intrin.h without -mavx512f -mavx512cd

[luoyonggang at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108191

            Bug ID: 108191
           Summary: Add support to usage of *intrin.h without -mavx512f
                    -mavx512cd
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: luoyonggang at gmail dot com
  Target Milestone: ---

This is for getting the following command to be works
```
gcc -fPIC -O2 -D__SSE3__=1 -D__SSSE3__=1 \
-D__SSE4_1__=1 -D__SSE4_2__=1 -D__SSE4A__=1 \
-D__POPCNT__=1 -D__XSAVE__=1 -D__CRC32__=1 \
-D__AVX__=1 -D__AVX2__=1 \
-D__FP_FAST_FMAF32=1 \
-D__FP_FAST_FMAF64=1 \
-D__FP_FAST_FMAF=1 \
-D__FP_FAST_FMAF32x=1 \
-D__AVX512F__=1 -D__AVX512CD__=1 test.c
```
That is generating code for SSE2 only, and we can using 
#include <x86intrin.h>
by using runtime flags.

Indeed, MSVC are aready can did that, if gcc can also support for that, we can
reduce the usage of inline assembly, because MSVC(x64) doesn't support for
inline assembly, so that we can reduce the code complex

The content of test.c is:
```
#if defined(_MSC_VER)
#include <intrin.h>
#else
#include <x86intrin.h>
#endif

#include <math.h>

static inline int
util_iround(float f)
{
   __m128 m = _mm_set_ss(f);
   return _mm_cvtss_i32(m);
}

int util_iround_outside(int x, float y) {
    return x + util_iround(y);
}
```

The compile error is something like:
```
In file included from
C:/CI-Tools/msys64/mingw64/lib/gcc/x86_64-w64-mingw32/12.2.0/include/immintrin.h:35,
                 from
C:/CI-Tools/msys64/mingw64/lib/gcc/x86_64-w64-mingw32/12.2.0/include/x86intrin.h:32,
                 from test.c:4:
C:/CI-Tools/msys64/mingw64/lib/gcc/x86_64-w64-mingw32/12.2.0/include/pmmintrin.h:
In function '_mm_addsub_ps':
C:/CI-Tools/msys64/mingw64/lib/gcc/x86_64-w64-mingw32/12.2.0/include/pmmintrin.h:53:3:
error: cannot convert a value of type 'int' to vector type '__vector(4) float'
which has different size
   53 |   return (__m128) __builtin_ia32_addsubps ((__v4sf)__X, (__v4sf)__Y);
      |   ^~~~~~
C:/CI-Tools/msys64/mingw64/lib/gcc/x86_64-w64-mingw32/12.2.0/include/pmmintrin.h:
In function '_mm_hadd_ps':
C:/CI-Tools/msys64/mingw64/lib/gcc/x86_64-w64-mingw32/12.2.0/include/pmmintrin.h:59:3:
error: cannot convert a value of type 'int' to vector type '__vector(4) float'
which has different size
   59 |   return (__m128) __builtin_ia32_haddps ((__v4sf)__X, (__v4sf)__Y);
      |   ^~~~~~
C:/CI-Tools/msys64/mingw64/lib/gcc/x86_64-w64-mingw32/12.2.0/include/pmmintrin.h:
In function '_mm_hsub_ps':
C:/CI-Tools/msys64/mingw64/lib/gcc/x86_64-w64-mingw32/12.2.0/include/pmmintrin.h:65:3:
error: cannot convert a value of type 'int' to vector type '__vector(4) float'
which has different size
   65 |   return (__m128) __builtin_ia32_hsubps ((__v4sf)__X, (__v4sf)__Y);
```

[Bug target/108191] Add support to usage of *intrin.h without -mavx512f -mavx512cd

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108191

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
You are lying to the compiler, don't.  In GCC you can #include <x86intrin.h>
with SSE2 only and later in say __attribute__((target ("avx512cd"))) function
use avx512f/avx512cd intrinsics, no need to do the what you show above.

[Bug target/108191] Add support to usage of *intrin.h without -mavx512f -mavx512cd

[luoyonggang at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108191

--- Comment #2 from 罗勇刚(Yonggang Luo) <luoyonggang at gmail dot com> ---
(In reply to Jakub Jelinek from comment #1)
> You are lying to the compiler, don't.  In GCC you can #include <x86intrin.h>
> with SSE2 only and later in say __attribute__((target ("avx512cd")))
> function use avx512f/avx512cd intrinsics, no need to do the what you show
> above.

Can you be more specific, show me the code, thanks:)

[Bug target/108191] Add support to usage of *intrin.h without -mavx512f -mavx512cd

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108191

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
         Resolution|---                         |WONTFIX
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
I suppose the issue will be that __attribute__((target)) isn't supported by
MSVC?  But indeed this isn't something we are going to support.  Note another
way is to put the functions into different translation units.

[Bug target/108191] Add support to usage of *intrin.h without -mavx512f -mavx512cd

[luoyonggang at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108191

--- Comment #4 from 罗勇刚(Yonggang Luo) <luoyonggang at gmail dot com> ---
(In reply to Richard Biener from comment #3)
> I suppose the issue will be that __attribute__((target)) isn't supported by
> MSVC?  But indeed this isn't something we are going to support.  Note
> another way is to put the functions into different translation units.

gcc is enough, no need care about msvc, msvc can support without attribute, we
can  use macro to deal with that.

[Bug target/108191] Add support to usage of *intrin.h without -mavx512f -mavx512cd

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108191

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to 罗勇刚(Yonggang Luo) from comment #2)
> (In reply to Jakub Jelinek from comment #1)
> > You are lying to the compiler, don't.  In GCC you can #include <x86intrin.h>
> > with SSE2 only and later in say __attribute__((target ("avx512cd")))
> > function use avx512f/avx512cd intrinsics, no need to do the what you show
> > above.
> 
> Can you be more specific, show me the code, thanks:)

#include <x86intrin.h>

int __attribute__((target("avx512f"))) foo(float f)
{
  __m128 m = _mm_set_ss(f);
  return _mm_cvtss_i32(m);
}

results in (with just SSE2):

foo:
.LFB6614:
        .cfi_startproc
        vcvtss2sil      %xmm0, %eax
        ret

[Bug target/108191] Add support to usage of *intrin.h without -mavx512f -mavx512cd

[luoyonggang at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108191

--- Comment #6 from 罗勇刚(Yonggang Luo) <luoyonggang at gmail dot com> ---
Is the following command are valid usage? It's compiled properly

```

// compile args:  -fPIC -O2 -D__SSE3__=1 -D__SSSE3__=1 -D__SSE4_1__=1
-D__SSE4_2__=1 -D__SSE4A__=1 -D__POPCNT__=1 -D__XSAVE__=1 -D__CRC32__=1
-D__AVX__=1 -D__AVX2__=1 -D__FP_FAST_FMAF32=1 -D__FP_FAST_FMAF64=1
-D__FP_FAST_FMAF=1 -D__FP_FAST_FMAF32x=1 -D__AVX512F__=1 -D__AVX512CD__=1
#include <math.h>

#pragma GCC push_options
#pragma GCC target("avx512f")
#pragma GCC target("avx512cd")
#pragma GCC target("sse4a")

#if defined(_MSC_VER)
#include <intrin.h>
#else
#include <x86intrin.h>
#endif

#pragma GCC pop_options


#pragma GCC push_options
#pragma GCC target("avx512f")
#pragma GCC target("avx512cd")
#pragma GCC target("sse4a")

void util_fadd_512(float *a, float *b, float *c) {
    /* a = b + c */
    __m512 av = _mm512_load_ps(a);
    __m512 bv = _mm512_load_ps(b);
    __m512 cv = _mm512_add_ps(av, bv);
    _mm512_store_ps(c, cv);
}
static inline int
util_iround(float f)
{
   __m128 m = _mm_set_ss(f);
   return _mm_cvtss_i32(m);
}

#pragma GCC pop_options

int util_iround_outside(int x, float y) {
    return x + util_iround(y);
}
float util_fadd(float a, float b) {
   return a + b;
}
```

[Bug target/108191] Add support to usage of *intrin.h without -mavx512f -mavx512cd

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108191

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to 罗勇刚(Yonggang Luo) from comment #6)
> Is the following command are valid usage? It's compiled properly

No, this is invalid.
> 
> ```
> 
> // compile args:  -fPIC -O2 -D__SSE3__=1 -D__SSSE3__=1 -D__SSE4_1__=1
> -D__SSE4_2__=1 -D__SSE4A__=1 -D__POPCNT__=1 -D__XSAVE__=1 -D__CRC32__=1
> -D__AVX__=1 -D__AVX2__=1 -D__FP_FAST_FMAF32=1 -D__FP_FAST_FMAF64=1
> -D__FP_FAST_FMAF=1 -D__FP_FAST_FMAF32x=1 -D__AVX512F__=1 -D__AVX512CD__=1

Only -fPIC -O2 here, none of the -D arguments, all of them are internal
GCC macros that shouldn't be redefined by users.
Plus it isn't needed.

> #include <math.h>
> 
> #pragma GCC push_options
> #pragma GCC target("avx512f")
> #pragma GCC target("avx512cd")
> #pragma GCC target("sse4a")
> 
> #if defined(_MSC_VER)
> #include <intrin.h>
> #else
> #include <x86intrin.h>
> #endif
> 
> #pragma GCC pop_options

You can do it, but for GCC it is completely useless, you can just
#include <x86intrin.h> without anything further.

> #pragma GCC push_options
> #pragma GCC target("avx512f")
> #pragma GCC target("avx512cd")
> #pragma GCC target("sse4a")

This is certainly fine, but avx512f in there isn't needed, that is implied by
avx512cd.
Though, I don't see anything avx512cd nor sse4a-ish in there.
> 
> void util_fadd_512(float *a, float *b, float *c) {
>     /* a = b + c */
>     __m512 av = _mm512_load_ps(a);
>     __m512 bv = _mm512_load_ps(b);
>     __m512 cv = _mm512_add_ps(av, bv);
>     _mm512_store_ps(c, cv);
> }
> static inline int
> util_iround(float f)
> {
>    __m128 m = _mm_set_ss(f);
>    return _mm_cvtss_i32(m);
> }
> 
> #pragma GCC pop_options
> 
> int util_iround_outside(int x, float y) {
>     return x + util_iround(y);
> }
> float util_fadd(float a, float b) {
>    return a + b;
> }
> ```

That said, code with avx512cd etc. target won't inline into code without it.