[Clang][constexpr] Subbexpression not valid in a constant expression for pimmintrin.h functions.

[mahesh-attarde]

Hi
Latest changes related to pimmintrin.h broke compiler for simple valid code.
here is repro.
https://godbolt.org/z/GfoT9Ydff

#include <pmmintrin.h>
#include <stdio.h>

#define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)

constexpr bool match_m128(__m128 _v, __m128 _p) {
  __v4su v = (__v4su)_v;
    __v4su p = (__v4su)_p;
  return v[0] == __builtin_bit_cast(unsigned, p[0]) && v[1] == __builtin_bit_cast(unsigned, p[1]) && v[2] == __builtin_bit_cast(unsigned, p[2]) && v[3] == __builtin_bit_cast(unsigned, p[3]);
}

int main() {

  TEST_CONSTEXPR(match_m128(_mm_addsub_pd((__m128d){+2.0, +2.0}, (__m128d){+1.0, +2.0}), (__m128d){+1.0, +4.0}));
  return 0;
 }

Error Reported is

x86-64 clang (assertions trunk)
-msse3 
123
<Compilation failed>

# For more information see the output window
x86-64 clang (assertions trunk) - 500ms
<source>:15:18: error: static assertion expression is not an integral constant expression
   15 |   TEST_CONSTEXPR(match_m128(_mm_addsub_pd((__m128d){+2.0, +2.0}, (__m128d){+1.0, +2.0}), (__m128d){+1.0, +4.0}));
      |                  ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<source>:4:43: note: expanded from macro 'TEST_CONSTEXPR'
    4 | #define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)
      |                                           ^~~~~~~~~~~
/cefs/96/96a61a354c7fcaba5936d879_clang-assertions-trunk-20251106/lib/clang/22/include/pmmintrin.h:171:10: note: subexpression not valid in a constant expression
  171 |   return __builtin_ia32_addsubpd((__v2df)__a, (__v2df)__b);
      |          ^~~~~~~~~~~~~~~~~~~~~~~
<source>:15:29: note: in call to '_mm_addsub_pd({2.000000e+00, 2.000000e+00}, {1.000000e+00, 2.000000e+00})'
   15 |   TEST_CONSTEXPR(match_m128(_mm_addsub_pd((__m128d){+2.0, +2.0}, (__m128d){+1.0, +2.0}), (__m128d){+1.0, +4.0}));
      |                             ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<source>:4:43: note: expanded from macro 'TEST_CONSTEXPR'
    4 | #define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)
      |                                           ^~~~~~~~~~~
1 error generated.
Compiler returned: 1

Can you help identify root cause and possible WA?

[mahesh-attarde]

We have seen similar case with mmx intrinsics as well. #159905
cc @tbaederr

[AaronBallman]

CC @whytolearn @RKSimon seems potentially related to 53ddeb4

[RKSimon]

Yes, this seems to be something to do with bitcasts, and it's showing up because we're stressing them more with all the new intrinsics support.

Not sure if _mm_addsub_pd supports constexpr yet though?

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Mahesh-Attarde (mahesh-attarde)

Hi Latest changes related to `pimmintrin.h` broke compiler for simple valid code. here is repro. https://godbolt.org/z/GfoT9Ydff

#include <pmmintrin.h>
#include <stdio.h>

#define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)

constexpr bool match_m128(__m128 _v, __m128 _p) {
  __v4su v = (__v4su)_v;
    __v4su p = (__v4su)_p;
  return v[0] == __builtin_bit_cast(unsigned, p[0]) && v[1] == __builtin_bit_cast(unsigned, p[1]) && v[2] == __builtin_bit_cast(unsigned, p[2]) && v[3] == __builtin_bit_cast(unsigned, p[3]);
}

int main() {

  TEST_CONSTEXPR(match_m128(_mm_addsub_pd((__m128d){+2.0, +2.0}, (__m128d){+1.0, +2.0}), (__m128d){+1.0, +4.0}));
  return 0;
 }

Error Reported is

x86-64 clang (assertions trunk)
-msse3 
123
<Compilation failed>

# For more information see the output window
x86-64 clang (assertions trunk) - 500ms
<source>:15:18: error: static assertion expression is not an integral constant expression
   15 |   TEST_CONSTEXPR(match_m128(_mm_addsub_pd((__m128d){+2.0, +2.0}, (__m128d){+1.0, +2.0}), (__m128d){+1.0, +4.0}));
      |                  ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<source>:4:43: note: expanded from macro 'TEST_CONSTEXPR'
    4 | #define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)
      |                                           ^~~~~~~~~~~
/cefs/96/96a61a354c7fcaba5936d879_clang-assertions-trunk-20251106/lib/clang/22/include/pmmintrin.h:171:10: note: subexpression not valid in a constant expression
  171 |   return __builtin_ia32_addsubpd((__v2df)__a, (__v2df)__b);
      |          ^~~~~~~~~~~~~~~~~~~~~~~
<source>:15:29: note: in call to '_mm_addsub_pd({2.000000e+00, 2.000000e+00}, {1.000000e+00, 2.000000e+00})'
   15 |   TEST_CONSTEXPR(match_m128(_mm_addsub_pd((__m128d){+2.0, +2.0}, (__m128d){+1.0, +2.0}), (__m128d){+1.0, +4.0}));
      |                             ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<source>:4:43: note: expanded from macro 'TEST_CONSTEXPR'
    4 | #define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)
      |                                           ^~~~~~~~~~~
1 error generated.
Compiler returned: 1

Can you help identify root cause and possible WA?

[frederick-vs-ja]

I guess this is already tracked in some subissue(s) of #30794.

[tbaederr]

Reduced:

typedef float __m128 __attribute__((__vector_size__(16), __aligned__(16)));
typedef double __m128d __attribute__((__vector_size__(16), __aligned__(16)));
typedef unsigned  __v4su __attribute__((__vector_size__(16)));
typedef float __v4sf __attribute__((__vector_size__(16)));
typedef double __v2df __attribute__((__vector_size__(16)));

static __m128d constexpr
_mm_addsub_pd(__m128d __a, __m128d __b) {
  return __builtin_ia32_addsubpd((__v2df)__a, (__v2df)__b);
}

static_assert(_mm_addsub_pd((__m128d){+2.0, +2.0}, (__m128d){+1.0, +2.0})[0] == 10);

this is simply __builtin_ia32_addsubpd not being handled in ExprConstant.cpp(or rather, not being marked as constexpr in Builtins.td), but marked constexpr in the header.

[mahesh-attarde]

Reduced:

typedef float __m128 attribute((vector_size(16), aligned(16)));
typedef double __m128d attribute((vector_size(16), aligned(16)));
typedef unsigned __v4su attribute((vector_size(16)));
typedef float __v4sf attribute((vector_size(16)));
typedef double __v2df attribute((vector_size(16)));

static __m128d constexpr
_mm_addsub_pd(__m128d __a, __m128d __b) {
return __builtin_ia32_addsubpd((__v2df)__a, (__v2df)__b);
}

static_assert(_mm_addsub_pd((__m128d){+2.0, +2.0}, (__m128d){+1.0, +2.0})[0] == 10);
this is simply __builtin_ia32_addsubpd not being handled in ExprConstant.cpp(or rather, not being marked as constexpr in Builtins.td), but marked constexpr in the header.

Thanks @tbaederr for debugging.
cc @ykhatav

[ykhatav]

Thanks @tbaederr .
FYI, the same error occurs for _mm_mask_move_ss, _mm_maskz_move_ss, _mm_mask_move_sd and _mm_maskz_move_sd which were all marked constexpr in: 7fe0691
@ahmednoursphinx

`#include <immintrin.h>
#include <stdio.h>

#define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)

constexpr bool match_m128(__m128 _v, __m128 _p) {
  __v4su v = (__v4su)_v;
  __v4su p = (__v4su)_p;
  return v[0] == __builtin_bit_cast(unsigned, p[0]) && 
         v[1] == __builtin_bit_cast(unsigned, p[1]) && 
         v[2] == __builtin_bit_cast(unsigned, p[2]) && 
         v[3] == __builtin_bit_cast(unsigned, p[3]);
}

constexpr bool match_m128d(__m128d _v, __m128d _p) {
  __v2du v = (__v2du)_v;
  __v2du p = (__v2du)_p;
  return v[0] == __builtin_bit_cast(unsigned long long, p[0]) && 
         v[1] == __builtin_bit_cast(unsigned long long, p[1]);
}

int main() {
  
 
  TEST_CONSTEXPR(match_m128(
    _mm_mask_move_ss((__m128){9.0f, 10.0f, 11.0f, 12.0f}, (__mmask8)1, (__m128){1.0f, 2.0f, 3.0f, 4.0f}, (__m128){5.0f, 6.0f, 7.0f, 8.0f}),
    (__m128){5.0f, 2.0f, 3.0f, 4.0f}
  ));
  
 
  TEST_CONSTEXPR(match_m128(
    _mm_maskz_move_ss((__mmask8)1, (__m128){1.0f, 2.0f, 3.0f, 4.0f}, (__m128){5.0f, 6.0f, 7.0f, 8.0f}),
    (__m128){5.0f, 2.0f, 3.0f, 4.0f}
  ));
  
 
  TEST_CONSTEXPR(match_m128(
    _mm_maskz_move_ss((__mmask8)0, (__m128){1.0f, 2.0f, 3.0f, 4.0f}, (__m128){5.0f, 6.0f, 7.0f, 8.0f}),
    (__m128){0.0f, 2.0f, 3.0f, 4.0f}
  ));
  
 
  TEST_CONSTEXPR(match_m128d(
    _mm_mask_move_sd((__m128d){9.0, 10.0}, (__mmask8)1, (__m128d){1.0, 2.0}, (__m128d){5.0, 6.0}),
    (__m128d){5.0, 2.0}
  ));
  

  TEST_CONSTEXPR(match_m128d(
    _mm_maskz_move_sd((__mmask8)1, (__m128d){1.0, 2.0}, (__m128d){5.0, 6.0}),
    (__m128d){5.0, 2.0}
  ));
  
 
  TEST_CONSTEXPR(match_m128d(
    _mm_maskz_move_sd((__mmask8)0, (__m128d){1.0, 2.0}, (__m128d){5.0, 6.0}),
    (__m128d){0.0, 2.0}
  ));
  
  return 0;
}`

[ahmednoursphinx]

Thanks @tbaederr . FYI, the same error occurs for _mm_mask_move_ss, _mm_maskz_move_ss, _mm_mask_move_sd and _mm_maskz_move_sd which were all marked constexpr in: 7fe0691 @ahmednoursphinx

`#include <immintrin.h>
#include <stdio.h>

#define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)

constexpr bool match_m128(__m128 _v, __m128 _p) {
  __v4su v = (__v4su)_v;
  __v4su p = (__v4su)_p;
  return v[0] == __builtin_bit_cast(unsigned, p[0]) && 
         v[1] == __builtin_bit_cast(unsigned, p[1]) && 
         v[2] == __builtin_bit_cast(unsigned, p[2]) && 
         v[3] == __builtin_bit_cast(unsigned, p[3]);
}

constexpr bool match_m128d(__m128d _v, __m128d _p) {
  __v2du v = (__v2du)_v;
  __v2du p = (__v2du)_p;
  return v[0] == __builtin_bit_cast(unsigned long long, p[0]) && 
         v[1] == __builtin_bit_cast(unsigned long long, p[1]);
}

int main() {
  
 
  TEST_CONSTEXPR(match_m128(
    _mm_mask_move_ss((__m128){9.0f, 10.0f, 11.0f, 12.0f}, (__mmask8)1, (__m128){1.0f, 2.0f, 3.0f, 4.0f}, (__m128){5.0f, 6.0f, 7.0f, 8.0f}),
    (__m128){5.0f, 2.0f, 3.0f, 4.0f}
  ));
  
 
  TEST_CONSTEXPR(match_m128(
    _mm_maskz_move_ss((__mmask8)1, (__m128){1.0f, 2.0f, 3.0f, 4.0f}, (__m128){5.0f, 6.0f, 7.0f, 8.0f}),
    (__m128){5.0f, 2.0f, 3.0f, 4.0f}
  ));
  
 
  TEST_CONSTEXPR(match_m128(
    _mm_maskz_move_ss((__mmask8)0, (__m128){1.0f, 2.0f, 3.0f, 4.0f}, (__m128){5.0f, 6.0f, 7.0f, 8.0f}),
    (__m128){0.0f, 2.0f, 3.0f, 4.0f}
  ));
  
 
  TEST_CONSTEXPR(match_m128d(
    _mm_mask_move_sd((__m128d){9.0, 10.0}, (__mmask8)1, (__m128d){1.0, 2.0}, (__m128d){5.0, 6.0}),
    (__m128d){5.0, 2.0}
  ));
  

  TEST_CONSTEXPR(match_m128d(
    _mm_maskz_move_sd((__mmask8)1, (__m128d){1.0, 2.0}, (__m128d){5.0, 6.0}),
    (__m128d){5.0, 2.0}
  ));
  
 
  TEST_CONSTEXPR(match_m128d(
    _mm_maskz_move_sd((__mmask8)0, (__m128d){1.0, 2.0}, (__m128d){5.0, 6.0}),
    (__m128d){0.0, 2.0}
  ));
  
  return 0;
}`

Thanks @ykhatav should we open a seperate issue for that or it will be resolved within the scope of the current one

[ykhatav]

At a high level, the issues seem to stem from similar underlying causes and can be resolved as part of the same fix.