From 2b77bd95e57926e506e776ccc46b773180725a9d Mon Sep 17 00:00:00 2001
From: Asaf Badouh <asaf.badouh@intel.com>
Date: Thu, 23 Jul 2015 07:07:25 +0000
Subject: [PATCH] [X86][AVX512BW] add madd and maddubs intrinsics

Differential Revision: http://reviews.llvm.org/D11420

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@242986 91177308-0d34-0410-b5e6-96231b3b80d8
---
 include/clang/Basic/BuiltinsX86.def |  3 ++
 lib/Headers/avx512bwintrin.h        | 54 +++++++++++++++++++++++++++++
 test/CodeGen/avx512bw-builtins.c    | 30 ++++++++++++++++
 3 files changed, 87 insertions(+)

diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index bfd10da50f..84cd669a48 100644
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -1228,4 +1228,7 @@ BUILTIN(__builtin_ia32_mulpd512_mask, "V8dV8dV8dV8dUcIi", "")
 BUILTIN(__builtin_ia32_mulps512_mask, "V16fV16fV16fV16fUsIi", "")
 BUILTIN(__builtin_ia32_subpd512_mask, "V8dV8dV8dV8dUcIi", "")
 BUILTIN(__builtin_ia32_subps512_mask, "V16fV16fV16fV16fUsIi", "")
+BUILTIN(__builtin_ia32_pmaddubsw512_mask, "V32sV64cV64cV32sUi", "")
+BUILTIN(__builtin_ia32_pmaddwd512_mask, "V16iV32sV32sV16iUs", "")
+
 #undef BUILTIN
diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h
index 5e17cbb50e..ec2e163ffd 100644
--- a/lib/Headers/avx512bwintrin.h
+++ b/lib/Headers/avx512bwintrin.h
@@ -1294,6 +1294,60 @@ _mm512_maskz_mulhi_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
                (__mmask32) __U);
 }
 
+static __inline__ __m512i __attribute__((__always_inline__, __nodebug__))
+_mm512_maddubs_epi16 (__m512i __X, __m512i __Y) {
+  return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X,
+                 (__v64qi) __Y,
+                 (__v32hi)
+                 _mm512_setzero_hi (),
+                 (__mmask32) -1);
+}
+
+static __inline__ __m512i __attribute__((__always_inline__, __nodebug__))
+_mm512_mask_maddubs_epi16 (__m512i __W, __mmask32 __U, __m512i __X,
+         __m512i __Y) {
+  return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X,
+                 (__v64qi) __Y,
+                 (__v32hi) __W,
+                 (__mmask32) __U);
+}
+
+static __inline__ __m512i __attribute__((__always_inline__, __nodebug__))
+_mm512_maskz_maddubs_epi16 (__mmask32 __U, __m512i __X, __m512i __Y) {
+  return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X,
+                 (__v64qi) __Y,
+                 (__v32hi)
+                 _mm512_setzero_hi (),
+                 (__mmask32) __U);
+}
+
+static __inline__ __m512i __attribute__((__always_inline__, __nodebug__))
+_mm512_madd_epi16 (__m512i __A, __m512i __B) {
+  return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A,
+               (__v32hi) __B,
+               (__v16si)
+               _mm512_setzero_si512 (),
+               (__mmask16) -1);
+}
+
+static __inline__ __m512i __attribute__((__always_inline__, __nodebug__))
+_mm512_mask_madd_epi16 (__m512i __W, __mmask16 __U, __m512i __A,
+      __m512i __B) {
+  return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A,
+               (__v32hi) __B,
+               (__v16si) __W,
+               (__mmask16) __U);
+}
+
+static __inline__ __m512i __attribute__((__always_inline__, __nodebug__))
+_mm512_maskz_madd_epi16 (__mmask16 __U, __m512i __A, __m512i __B) {
+  return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A,
+               (__v32hi) __B,
+               (__v16si)
+               _mm512_setzero_si512 (),
+               (__mmask16) __U);
+}
+
 #define _mm512_cmp_epi8_mask(a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \
                                          (__v64qi)(__m512i)(b), \
diff --git a/test/CodeGen/avx512bw-builtins.c b/test/CodeGen/avx512bw-builtins.c
index 9c57b56ec8..7109449636 100644
--- a/test/CodeGen/avx512bw-builtins.c
+++ b/test/CodeGen/avx512bw-builtins.c
@@ -880,3 +880,33 @@ __m512i test_mm512_maskz_mulhi_epu16(__mmask32 __U, __m512i __A, __m512i __B) {
   return _mm512_maskz_mulhi_epu16(__U,__A,__B); 
 }
 
+__m512i test_mm512_maddubs_epi16(__m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_maddubs_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmaddubs.w.512
+  return _mm512_maddubs_epi16(__X,__Y); 
+}
+__m512i test_mm512_mask_maddubs_epi16(__m512i __W, __mmask32 __U, __m512i __X,         __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_mask_maddubs_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmaddubs.w.512
+  return _mm512_mask_maddubs_epi16(__W,__U,__X,__Y); 
+}
+__m512i test_mm512_maskz_maddubs_epi16(__mmask32 __U, __m512i __X, __m512i __Y) {
+  // CHECK-LABEL: @test_mm512_maskz_maddubs_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmaddubs.w.512
+  return _mm512_maskz_maddubs_epi16(__U,__X,__Y); 
+}
+__m512i test_mm512_madd_epi16(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_madd_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmaddw.d.512
+  return _mm512_madd_epi16(__A,__B); 
+}
+__m512i test_mm512_mask_madd_epi16(__m512i __W, __mmask16 __U, __m512i __A,      __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_madd_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmaddw.d.512
+  return _mm512_mask_madd_epi16(__W,__U,__A,__B); 
+}
+__m512i test_mm512_maskz_madd_epi16(__mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_madd_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmaddw.d.512
+  return _mm512_maskz_madd_epi16(__U,__A,__B); 
+}
-- 
2.40.0