[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR

Differential Revision: https://reviews.llvm.org/D37562

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@313011 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 42 insertions, 28 deletions

diff --git a/lib/Headers/avx2intrin.h b/lib/Headers/avx2intrin.h
index 576f761b25..caf4ced920 100644
--- a/lib/Headers/avx2intrin.h
+++ b/lib/Headers/avx2intrin.h
@@ -145,13 +145,21 @@ _mm256_andnot_si256(__m256i __a, __m256i __b)
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_avg_epu8(__m256i __a, __m256i __b)
 {
-  return (__m256i)__builtin_ia32_pavgb256((__v32qi)__a, (__v32qi)__b);
+  typedef unsigned short __v32hu __attribute__((__vector_size__(64)));
+  return (__m256i)__builtin_convertvector(
+               ((__builtin_convertvector((__v32qu)__a, __v32hu) +
+                 __builtin_convertvector((__v32qu)__b, __v32hu)) + 1)
+                 >> 1, __v32qu);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_avg_epu16(__m256i __a, __m256i __b)
 {
-  return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b);
+  typedef unsigned int __v16su __attribute__((__vector_size__(64)));
+  return (__m256i)__builtin_convertvector(
+               ((__builtin_convertvector((__v16hu)__a, __v16su) +
+                 __builtin_convertvector((__v16hu)__b, __v16su)) + 1)
+                 >> 1, __v16hu);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h
index 41958b7214..fa4fb20432 100644
--- a/lib/Headers/avx512bwintrin.h
+++ b/lib/Headers/avx512bwintrin.h
@@ -706,57 +706,55 @@ _mm512_maskz_adds_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_avg_epu8 (__m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
-              (__v64qi) __B,
-              (__v64qi) _mm512_setzero_qi(),
-              (__mmask64) -1);
+  typedef unsigned short __v64hu __attribute__((__vector_size__(128)));
+  return (__m512i)__builtin_convertvector(
+              ((__builtin_convertvector((__v64qu) __A, __v64hu) +
+                __builtin_convertvector((__v64qu) __B, __v64hu)) + 1)
+                >> 1, __v64qu);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_mask_avg_epu8 (__m512i __W, __mmask64 __U, __m512i __A,
           __m512i __B)
 {
-  return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
-              (__v64qi) __B,
-              (__v64qi) __W,
-              (__mmask64) __U);
+  return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U,
+              (__v64qi)_mm512_avg_epu8(__A, __B),
+              (__v64qi)__W);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_maskz_avg_epu8 (__mmask64 __U, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
-              (__v64qi) __B,
-              (__v64qi) _mm512_setzero_qi(),
-              (__mmask64) __U);
+  return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U,
+              (__v64qi)_mm512_avg_epu8(__A, __B),
+              (__v64qi)_mm512_setzero_qi());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_avg_epu16 (__m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v32hi) _mm512_setzero_hi(),
-              (__mmask32) -1);
+  typedef unsigned int __v32su __attribute__((__vector_size__(128)));
+  return (__m512i)__builtin_convertvector(
+              ((__builtin_convertvector((__v32hu) __A, __v32su) +
+                __builtin_convertvector((__v32hu) __B, __v32su)) + 1)
+                >> 1, __v32hu);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_mask_avg_epu16 (__m512i __W, __mmask32 __U, __m512i __A,
            __m512i __B)
 {
-  return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v32hi) __W,
-              (__mmask32) __U);
+  return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U,
+              (__v32hi)_mm512_avg_epu16(__A, __B),
+              (__v32hi)__W);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v32hi) _mm512_setzero_hi(),
-              (__mmask32) __U);
+  return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U,
+              (__v32hi)_mm512_avg_epu16(__A, __B),
+              (__v32hi) _mm512_setzero_hi());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
diff --git a/lib/Headers/emmintrin.h b/lib/Headers/emmintrin.h
index 709815cbb4..3372508a7f 100644
--- a/lib/Headers/emmintrin.h
+++ b/lib/Headers/emmintrin.h
@@ -2258,7 +2258,11 @@ _mm_adds_epu16(__m128i __a, __m128i __b)
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_avg_epu8(__m128i __a, __m128i __b)
 {
-  return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b);
+  typedef unsigned short __v16hu __attribute__ ((__vector_size__ (32)));
+  return (__m128i)__builtin_convertvector(
+               ((__builtin_convertvector((__v16qu)__a, __v16hu) +
+                 __builtin_convertvector((__v16qu)__b, __v16hu)) + 1)
+                 >> 1, __v16qu);
 }
 
 /// \brief Computes the rounded avarages of corresponding elements of two
@@ -2278,7 +2282,11 @@ _mm_avg_epu8(__m128i __a, __m128i __b)
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_avg_epu16(__m128i __a, __m128i __b)
 {
-  return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b);
+  typedef unsigned int __v8su __attribute__ ((__vector_size__ (32)));
+  return (__m128i)__builtin_convertvector(
+               ((__builtin_convertvector((__v8hu)__a, __v8su) +
+                 __builtin_convertvector((__v8hu)__b, __v8su)) + 1)
+                 >> 1, __v8hu);
 }
 
 /// \brief Multiplies the corresponding elements of two 128-bit signed [8 x i16]