[X86] Lowering integer truncation intrinsics to native IR

[mike.dvoretsky]

Summary

This patch lowers the _mm[256|512]_cvtepi{64|32|16}_epi{32|16|8} intrinsics to native IR in cases where the result's length is less than 128 bits.

The resulting IR for 256-bit inputs is folded into VPMOV instructions in D46957, while for 128-bit inputs the vpshufb (or, in the 64-to-32-bit case, vinsertps) instructions are generated instead. D48822 adds fast-isel tests that demonstrate generated instructions.

[mike.dvoretsky]

Uploaded the correct diff.

[craig.topper]

clang/lib/Headers/avx512vlintrin.h

7421

If you need to add more zeroes than the width of the input, can you just repeat the zero vector in order as many times as necessary. It should look more like a concatentation.

so 0, 1, 2, 3, 4, 5, 6, 7, 4, 5, 6, 7, 4, 5, 6, 7

[craig.topper]

clang/lib/Headers/avx512vlintrin.h

33

Can you just do a local typedef in the functions that need them? That way we don't expose them to all users of the header.

[mike.dvoretsky]

Updated per comments. Typedefs for intermediate short vectors moved into the bodies of the functions using them.

[RKSimon]

clang/lib/Headers/avx512vlbwintrin.h

1501

Are we happy with using illegal types like this? What about flipping the shuffle and convert?

return (__m128i)__builtin_convertvector(
                  __builtin_shufflevector((__v8hi)__A,
                                          (__v8hi){0, 0, 0, 0, 0, 0, 0, 0},
                                          0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15), __v16qi);

[mike.dvoretsky]

clang/lib/Headers/avx512vlbwintrin.h

1501

This would bring its own issues, since in the cvtepi64_epi8 cases the inner shuffle would produce vectors of 16 64-bit values. There would be no extra typedef, but in the back-end these would be split in type legalization, making it harder to fold them into VPMOV instructions.

[RKSimon]

Please can you create a llvm side parallel patch that updates the relevant fast-isel tests

clang/lib/Headers/avx512vlbwintrin.h

1501

Yeah, neither solution is particularly clean. Please keep it as is.

[RKSimon]

LGTM - @craig.topper any comments?

[craig.topper]

LGTM

clang/lib/Headers/avx512vlbwintrin.h

 _mm256_maskz_cvtusepi16_epi8 (__mmask16 __M, __m256i __A) {
   return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A,
                 (__v16qi) _mm_setzero_si128(),
                 __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_cvtepi16_epi8 (__m128i __A) {
-
-  return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A,
-               (__v16qi) _mm_setzero_si128(),
-               (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v8hi)__A, __v8qi),
+      (__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
+      12, 13, 14, 15);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_cvtepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) {
   return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A,
                (__v16qi) __O,
                __M);
 }

clang/lib/Headers/avx512vlintrin.h

 #error "Never use <avx512vlintrin.h> directly; include <immintrin.h> instead."
 #endif
 
 #ifndef __AVX512VLINTRIN_H
 #define __AVX512VLINTRIN_H
 
 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl")))
 
+typedef short __v2hi __attribute__((__vector_size__(4)));
+typedef char __v4qi __attribute__((__vector_size__(4)));
+typedef char __v2qi __attribute__((__vector_size__(2)));
+
 /* Integer compare */
 
 #define _mm_cmpeq_epi32_mask(A, B) \
     _mm_cmp_epi32_mask((A), (B), _MM_CMPINT_EQ)
 #define _mm_mask_cmpeq_epi32_mask(k, A, B) \
     _mm_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_EQ)
 #define _mm_cmpge_epi32_mask(A, B) \
     _mm_cmp_epi32_mask((A), (B), _MM_CMPINT_GE)

 _mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
 {
   __builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_cvtepi32_epi8 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
-              (__v16qi)_mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v4si)__A, __v4qi), (__v4qi){0, 0, 0, 0}, 0, 1,
+      2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
 {
   return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
               (__v16qi) __O, __M);
 }

 _mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
 {
   __builtin_ia32_pmovdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm256_cvtepi32_epi8 (__m256i __A)
 {
-  return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
-              (__v16qi)_mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v8si)__A, __v8qi),
+      (__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
+      12, 13, 14, 15);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm256_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
 {
   return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
               (__v16qi) __O, __M);
 }

 _mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
 {
   __builtin_ia32_pmovdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_cvtepi32_epi16 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
-              (__v8hi) _mm_setzero_si128 (),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v4si)__A, __v4hi), (__v4hi){0, 0, 0, 0}, 0, 1,
+      2, 3, 4, 5, 6, 7);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
 {
   return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
               (__v8hi) __O, __M);
 }

 _mm256_mask_cvtepi32_storeu_epi16 (void *  __P, __mmask8 __M, __m256i __A)
 {
   __builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_cvtepi64_epi8 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
-              (__v16qi) _mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v2di)__A, __v2qi), (__v2qi){0, 0}, 0, 1, 2, 3,
+      3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
 {
   return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
               (__v16qi) __O, __M);
 }

 _mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
 {
   __builtin_ia32_pmovqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm256_cvtepi64_epi8 (__m256i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
-              (__v16qi) _mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v4di)__A, __v4qi), (__v4qi){0, 0, 0, 0}, 0, 1,
+      2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm256_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
 {
   return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
               (__v16qi) __O, __M);
 }

 _mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
 {
   __builtin_ia32_pmovqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_cvtepi64_epi32 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
-              (__v4si)_mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v2di)__A, __v2si), (__v2si){0, 0}, 0, 1, 2, 3);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
 {
   return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
               (__v4si) __O, __M);
 }

 _mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
 {
   __builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_cvtepi64_epi16 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
-              (__v8hi) _mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v2di)__A, __v2hi), (__v2hi){0, 0}, 0, 1, 2, 3,
+      3, 3, 3, 3);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
 {
   return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
               (__v8hi)__O,
               __M);

 _mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
 {
   __builtin_ia32_pmovqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm256_cvtepi64_epi16 (__m256i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
-              (__v8hi)_mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v4di)__A, __v4hi), (__v4hi){0, 0, 0, 0}, 0, 1,
+      2, 3, 4, 5, 6, 7);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm256_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
 {
   return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
               (__v8hi) __O, __M);
 }

clang/test/CodeGen/avx512vl-builtins.c

 void test_mm256_mask_cvtusepi64_storeu_epi16(void * __P, __mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtusepi64_storeu_epi16
   // CHECK: @llvm.x86.avx512.mask.pmovus.qw.mem.256
   return _mm256_mask_cvtusepi64_storeu_epi16(__P, __M, __A); 
 }
 
 __m128i test_mm_cvtepi32_epi8(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi32_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.db.128
+  // CHECK: trunc <4 x i32> %{{.*}} to <4 x i8>
+  // CHECK: shufflevector <4 x i8> %{{.*}}, <4 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>
   return _mm_cvtepi32_epi8(__A); 
 }
 
 __m128i test_mm_mask_cvtepi32_epi8(__m128i __O, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi32_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.db.128
   return _mm_mask_cvtepi32_epi8(__O, __M, __A); 
 }
 
 __m128i test_mm_maskz_cvtepi32_epi8(__mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_maskz_cvtepi32_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.db.128
   return _mm_maskz_cvtepi32_epi8(__M, __A); 
 }
 
 void test_mm_mask_cvtepi32_storeu_epi8(void * __P, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi32_storeu_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.db.mem.128
   return _mm_mask_cvtepi32_storeu_epi8(__P, __M, __A); 
 }
 
 __m128i test_mm256_cvtepi32_epi8(__m256i __A) {
   // CHECK-LABEL: @test_mm256_cvtepi32_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.db.256
+  // CHECK: trunc <8 x i32> %{{.*}} to <8 x i8>
+  // CHECK: shufflevector <8 x i8> %{{.*}}, <8 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
   return _mm256_cvtepi32_epi8(__A); 
 }
 
 __m128i test_mm256_mask_cvtepi32_epi8(__m128i __O, __mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtepi32_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.db.256
   return _mm256_mask_cvtepi32_epi8(__O, __M, __A); 
 }
 
 __m128i test_mm256_maskz_cvtepi32_epi8(__mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_maskz_cvtepi32_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.db.256
   return _mm256_maskz_cvtepi32_epi8(__M, __A); 
 }
 
 void test_mm256_mask_cvtepi32_storeu_epi8(void * __P, __mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtepi32_storeu_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.db.mem.256
   return _mm256_mask_cvtepi32_storeu_epi8(__P, __M, __A); 
 }
 
 __m128i test_mm_cvtepi32_epi16(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi32_epi16
-  // CHECK: @llvm.x86.avx512.mask.pmov.dw.128
+  // CHECK: trunc <4 x i32> %{{.*}} to <4 x i16>
+  // CHECK: shufflevector <4 x i16> %{{.*}}, <4 x i16> %{{.*}}, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   return _mm_cvtepi32_epi16(__A); 
 }
 
 __m128i test_mm_mask_cvtepi32_epi16(__m128i __O, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi32_epi16
   // CHECK: @llvm.x86.avx512.mask.pmov.dw.128
   return _mm_mask_cvtepi32_epi16(__O, __M, __A); 
 }

 void test_mm256_mask_cvtepi32_storeu_epi16(void *  __P, __mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtepi32_storeu_epi16
   // CHECK: @llvm.x86.avx512.mask.pmov.dw.mem.256
   return _mm256_mask_cvtepi32_storeu_epi16(__P, __M, __A); 
 }
 
 __m128i test_mm_cvtepi64_epi8(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi64_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.qb.128
+  // CHECK: trunc <2 x i64> %{{.*}} to <2 x i8>
+  // CHECK: shufflevector <2 x i8> %{{.*}}, <2 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>
   return _mm_cvtepi64_epi8(__A); 
 }
 
 __m128i test_mm_mask_cvtepi64_epi8(__m128i __O, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi64_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.qb.128
   return _mm_mask_cvtepi64_epi8(__O, __M, __A); 
 }
 
 __m128i test_mm_maskz_cvtepi64_epi8(__mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_maskz_cvtepi64_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.qb.128
   return _mm_maskz_cvtepi64_epi8(__M, __A); 
 }
 
 void test_mm_mask_cvtepi64_storeu_epi8(void * __P, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi64_storeu_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.qb.mem.128
   return _mm_mask_cvtepi64_storeu_epi8(__P, __M, __A); 
 }
 
 __m128i test_mm256_cvtepi64_epi8(__m256i __A) {
   // CHECK-LABEL: @test_mm256_cvtepi64_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.qb.256
+  // CHECK: trunc <4 x i64> %{{.*}} to <4 x i8>
+  // CHECK: shufflevector <4 x i8> %{{.*}}, <4 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>
   return _mm256_cvtepi64_epi8(__A); 
 }
 
 __m128i test_mm256_mask_cvtepi64_epi8(__m128i __O, __mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtepi64_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.qb.256
   return _mm256_mask_cvtepi64_epi8(__O, __M, __A); 
 }
 
 __m128i test_mm256_maskz_cvtepi64_epi8(__mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_maskz_cvtepi64_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.qb.256
   return _mm256_maskz_cvtepi64_epi8(__M, __A); 
 }
 
 void test_mm256_mask_cvtepi64_storeu_epi8(void * __P, __mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtepi64_storeu_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.qb.mem.256
   return _mm256_mask_cvtepi64_storeu_epi8(__P, __M, __A); 
 }
 
 __m128i test_mm_cvtepi64_epi32(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi64_epi32
-  // CHECK: @llvm.x86.avx512.mask.pmov.qd.128
+  // CHECK: trunc <2 x i64> %{{.*}} to <2 x i32>
+  // CHECK: shufflevector <2 x i32> %{{.*}}, <2 x i32> %{{.*}}, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   return _mm_cvtepi64_epi32(__A); 
 }
 
 __m128i test_mm_mask_cvtepi64_epi32(__m128i __O, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi64_epi32
   // CHECK: @llvm.x86.avx512.mask.pmov.qd.128
   return _mm_mask_cvtepi64_epi32(__O, __M, __A); 
 }

 void test_mm256_mask_cvtepi64_storeu_epi32(void * __P, __mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtepi64_storeu_epi32
   // CHECK: @llvm.x86.avx512.mask.pmov.qd.mem.256
   return _mm256_mask_cvtepi64_storeu_epi32(__P, __M, __A); 
 }
 
 __m128i test_mm_cvtepi64_epi16(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi64_epi16
-  // CHECK: @llvm.x86.avx512.mask.pmov.qw.128
+  // CHECK: trunc <2 x i64> %{{.*}} to <2 x i16>
+  // CHECK: shufflevector <2 x i16> %{{.*}}, <2 x i16> %{{.*}}, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 3, i32 3, i32 3, i32 3>
   return _mm_cvtepi64_epi16(__A); 
 }
 
 __m128i test_mm_mask_cvtepi64_epi16(__m128i __O, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi64_epi16
   // CHECK: @llvm.x86.avx512.mask.pmov.qw.128
   return _mm_mask_cvtepi64_epi16(__O, __M, __A); 
 }
 
 __m128i test_mm_maskz_cvtepi64_epi16(__mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_maskz_cvtepi64_epi16
   // CHECK: @llvm.x86.avx512.mask.pmov.qw.128
   return _mm_maskz_cvtepi64_epi16(__M, __A); 
 }
 
 void test_mm_mask_cvtepi64_storeu_epi16(void * __P, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi64_storeu_epi16
   // CHECK: @llvm.x86.avx512.mask.pmov.qw.mem.128
   return _mm_mask_cvtepi64_storeu_epi16(__P, __M, __A); 
 }
 
 __m128i test_mm256_cvtepi64_epi16(__m256i __A) {
   // CHECK-LABEL: @test_mm256_cvtepi64_epi16
-  // CHECK: @llvm.x86.avx512.mask.pmov.qw.256
+  // CHECK: trunc <4 x i64> %{{.*}} to <4 x i16>
+  // CHECK: shufflevector <4 x i16> %{{.*}}, <4 x i16> %{{.*}}, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   return _mm256_cvtepi64_epi16(__A); 
 }
 
 __m128i test_mm256_mask_cvtepi64_epi16(__m128i __O, __mmask8 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtepi64_epi16
   // CHECK: @llvm.x86.avx512.mask.pmov.qw.256
   return _mm256_mask_cvtepi64_epi16(__O, __M, __A); 
 }

clang/test/CodeGen/avx512vlbw-builtins.c

 __m128i test_mm256_maskz_cvtusepi16_epi8(__mmask16 __M, __m256i __A) {
   // CHECK-LABEL: @test_mm256_maskz_cvtusepi16_epi8
   // CHECK: @llvm.x86.avx512.mask.pmovus.wb.256
   return _mm256_maskz_cvtusepi16_epi8(__M, __A); 
 }
 
 __m128i test_mm_cvtepi16_epi8(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi16_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.wb.128
+  // CHECK: trunc <8 x i16> %{{.*}} to <8 x i8>
+  // CHECK: shufflevector <8 x i8> %{{.*}}, <8 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
   return _mm_cvtepi16_epi8(__A); 
 }
 
 __m128i test_mm_mask_cvtepi16_epi8(__m128i __O, __mmask8 __M, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtepi16_epi8
   // CHECK: @llvm.x86.avx512.mask.pmov.wb.128
   return _mm_mask_cvtepi16_epi8(__O, __M, __A); 
 }