[AVX512][Clang] Adding missing instructions' variations

[coby]

Summary

The following patch introduces the following alternations:

1. correcting the 'vpmulshiftqb' instruction, along with the respective intrinsics
2. Adding AVX512 missing instructions variants.

llvm part: https://reviews.llvm.org/D25759

[craig.topper]

lib/Headers/avx512vbmivlintrin.h

225

Intel docs show that this intrinsic should accept __mask32 so I don't think this is correct.

[coby]

Amending vpmultishiftqb is currently being maintained at the following patches:
https://reviews.llvm.org/D26258
https://reviews.llvm.org/D26257
All other changes are discarded

include/clang/Basic/BuiltinsX86.def

 TARGET_BUILTIN(__builtin_ia32_cvtsi2sd64, "V2dV2dLLiIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtsi2ss32, "V4fV4fiIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtsi2ss64, "V4fV4fLLiIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtss2sd_round_mask, "V2dV2dV4fV2dUcIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtusi2sd32, "V2dV2dUi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtusi2sd64, "V2dV2dULLiIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtusi2ss32, "V4fV4fUiIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtusi2ss64, "V4fV4fULLiIi","","avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb512_mask, "V64cV64cV64cV64cULLi","","avx512vbmi")
-TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb128_mask, "V16cV16cV16cV16cUs","","avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb256_mask, "V32cV32cV32cV32cUi","","avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb512_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512vbmi")
+TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb128_mask, "V2LLiV2LLiV2LLiV2LLiUc","","avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb256_mask, "V4LLiV4LLiV4LLiV4LLiUc","","avx512vbmi,avx512vl")
 
 // generic select intrinsics
 TARGET_BUILTIN(__builtin_ia32_selectb_128, "V16cUsV16cV16c", "", "")
 TARGET_BUILTIN(__builtin_ia32_selectb_256, "V32cUiV32cV32c", "", "")
 TARGET_BUILTIN(__builtin_ia32_selectb_512, "V64cULLiV64cV64c", "", "")
 TARGET_BUILTIN(__builtin_ia32_selectw_128, "V8sUcV8sV8s", "", "")
 TARGET_BUILTIN(__builtin_ia32_selectw_256, "V16sUsV16sV16s", "", "")
 TARGET_BUILTIN(__builtin_ia32_selectw_512, "V32sUiV32sV32s", "", "")

lib/Headers/avx512vbmiintrin.h

 {
   return (__m512i) __builtin_ia32_permvarqi512_mask ((__v64qi) __B,
                  (__v64qi) __A,
                  (__v64qi) __W,
                  (__mmask64) __M);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_multishift_epi64_epi8 (__m512i __W, __mmask64 __M, __m512i __X, __m512i __Y)
+_mm512_mask_multishift_epi64_epi8 (__m512i __W, __mmask8 __M, __m512i __X, __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_vpmultishiftqb512_mask ((__v64qi) __X,
-                (__v64qi) __Y,
-                (__v64qi) __W,
-                (__mmask64) __M);
+  return (__m512i) __builtin_ia32_vpmultishiftqb512_mask ((__v8di) __X,
+                (__v8di) __Y,
+                (__v8di) __W,
+                (__mmask8) __M);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_multishift_epi64_epi8 (__mmask64 __M, __m512i __X, __m512i __Y)
+_mm512_maskz_multishift_epi64_epi8 (__mmask8 __M, __m512i __X, __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_vpmultishiftqb512_mask ((__v64qi) __X,
-                (__v64qi) __Y,
-                (__v64qi) _mm512_setzero_si512 (),
-                (__mmask64) __M);
+  return (__m512i) __builtin_ia32_vpmultishiftqb512_mask ((__v8di) __X,
+                (__v8di) __Y,
+                (__v8di) _mm512_setzero_si512 (),
+                (__mmask8) __M);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_multishift_epi64_epi8 (__m512i __X, __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_vpmultishiftqb512_mask ((__v64qi) __X,
-                (__v64qi) __Y,
-                (__v64qi) _mm512_undefined_epi32 (),
-                (__mmask64) -1);
+  return (__m512i) __builtin_ia32_vpmultishiftqb512_mask ((__v8di) __X,
+                (__v8di) __Y,
+                (__v8di) _mm512_undefined_epi32 (),
+                (__mmask8) -1);
 }
 
 
 #undef __DEFAULT_FN_ATTRS
 
 #endif

lib/Headers/avx512vbmivlintrin.h

 {
   return (__m256i) __builtin_ia32_permvarqi256_mask ((__v32qi) __B,
                  (__v32qi) __A,
                  (__v32qi) __W,
                  (__mmask32) __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_multishift_epi64_epi8 (__m128i __W, __mmask16 __M, __m128i __X, __m128i __Y)
+_mm_mask_multishift_epi64_epi8 (__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y)
 {
-  return (__m128i) __builtin_ia32_vpmultishiftqb128_mask ((__v16qi) __X,
-                (__v16qi) __Y,
-                (__v16qi) __W,
-                (__mmask16) __M);
+  return (__m128i) __builtin_ia32_vpmultishiftqb128_mask ((__v2di) __X,
+                (__v2di) __Y,
+                (__v2di) __W,
+                (__mmask8) __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_multishift_epi64_epi8 (__mmask16 __M, __m128i __X, __m128i __Y)
+_mm_maskz_multishift_epi64_epi8 (__mmask8 __M, __m128i __X, __m128i __Y)
 {
-  return (__m128i) __builtin_ia32_vpmultishiftqb128_mask ((__v16qi) __X,
-                (__v16qi) __Y,
-                (__v16qi)
+  return (__m128i) __builtin_ia32_vpmultishiftqb128_mask ((__v2di) __X,
+                (__v2di) __Y,
+                (__v2di)
                 _mm_setzero_si128 (),
-                (__mmask16) __M);
+                (__mmask8) __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_multishift_epi64_epi8 (__m128i __X, __m128i __Y)
 {
-  return (__m128i) __builtin_ia32_vpmultishiftqb128_mask ((__v16qi) __X,
-                (__v16qi) __Y,
-                (__v16qi)
+  return (__m128i) __builtin_ia32_vpmultishiftqb128_mask ((__v2di) __X,
+                (__v2di) __Y,
+                (__v2di)
                 _mm_undefined_si128 (),
-                (__mmask16) -1);
+                (__mmask8) -1);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_multishift_epi64_epi8 (__m256i __W, __mmask32 __M, __m256i __X, __m256i __Y)
+_mm256_mask_multishift_epi64_epi8 (__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y)
 {
-  return (__m256i) __builtin_ia32_vpmultishiftqb256_mask ((__v32qi) __X,
-                (__v32qi) __Y,
-                (__v32qi) __W,
-                (__mmask32) __M);
+  return (__m256i) __builtin_ia32_vpmultishiftqb256_mask ((__v4di) __X,
+                (__v4di) __Y,
+                (__v4di) __W,
+                (__mmask8) __M);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_multishift_epi64_epi8 (__mmask32 __M, __m256i __X, __m256i __Y)
+_mm256_maskz_multishift_epi64_epi8 (__mmask8 __M, __m256i __X, __m256i __Y)
 {
-  return (__m256i) __builtin_ia32_vpmultishiftqb256_mask ((__v32qi) __X,
-                (__v32qi) __Y,
-                (__v32qi)
+  return (__m256i) __builtin_ia32_vpmultishiftqb256_mask ((__v4di) __X,
+                (__v4di) __Y,
+                (__v4di)
                 _mm256_setzero_si256 (),
-                (__mmask32) __M);
+                (__mmask8) __M);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_multishift_epi64_epi8 (__m256i __X, __m256i __Y)
 {
-  return (__m256i) __builtin_ia32_vpmultishiftqb256_mask ((__v32qi) __X,
-                (__v32qi) __Y,
-                (__v32qi)
+  return (__m256i) __builtin_ia32_vpmultishiftqb256_mask ((__v4di) __X,
+                (__v4di) __Y,
+                (__v4di)
                 _mm256_undefined_si256 (),
-                (__mmask32) -1);
+                (__mmask8) -1);
 }
 
 
 #undef __DEFAULT_FN_ATTRS
 
 #endif

test/CodeGen/avx512vbmi-builtins.c

 }
 
 __m512i test_mm512_mask_permutexvar_epi8(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_mask_permutexvar_epi8
   // CHECK: @llvm.x86.avx512.mask.permvar.qi.512
   return _mm512_mask_permutexvar_epi8(__W, __M, __A, __B); 
 }
 
-__m512i test_mm512_mask_multishift_epi64_epi8(__m512i __W, __mmask64 __M, __m512i __X, __m512i __Y) {
+__m512i test_mm512_mask_multishift_epi64_epi8(__m512i __W, __mmask8 __M, __m512i __X, __m512i __Y) {
   // CHECK-LABEL: @test_mm512_mask_multishift_epi64_epi8
   // CHECK: @llvm.x86.avx512.mask.pmultishift.qb.512
   return _mm512_mask_multishift_epi64_epi8(__W, __M, __X, __Y); 
 }
 
-__m512i test_mm512_maskz_multishift_epi64_epi8(__mmask64 __M, __m512i __X, __m512i __Y) {
+__m512i test_mm512_maskz_multishift_epi64_epi8(__mmask8 __M, __m512i __X, __m512i __Y) {
   // CHECK-LABEL: @test_mm512_maskz_multishift_epi64_epi8
   // CHECK: @llvm.x86.avx512.mask.pmultishift.qb.512
   return _mm512_maskz_multishift_epi64_epi8(__M, __X, __Y); 
 }
 
 __m512i test_mm512_multishift_epi64_epi8(__m512i __X, __m512i __Y) {
   // CHECK-LABEL: @test_mm512_multishift_epi64_epi8
   // CHECK: @llvm.x86.avx512.mask.pmultishift.qb.512
   return _mm512_multishift_epi64_epi8(__X, __Y); 
 }