[Clang][AVX512] completing missing intrinsics [roundscale, ceil, floor]

Differential Revision: http://reviews.llvm.org/D20070

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@269022 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h
index 5a09b0141d915490c7ccb4338ce9ba901d77ddae..6107790852a1cab7bb2af79d411099188dea727f 100644 (file)
--- a/lib/Headers/avx512fintrin.h
+++ b/lib/Headers/avx512fintrin.h
@@ -1307,6 +1307,16 @@ _mm512_floor_ps(__m512 __A)
                                                   _MM_FROUND_CUR_DIRECTION);
 }
 
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask_floor_ps (__m512 __W, __mmask16 __U, __m512 __A)
+{
+  return (__m512) __builtin_ia32_rndscaleps_mask ((__v16sf) __A,
+                   _MM_FROUND_FLOOR,
+                   (__v16sf) __W, __U,
+                   _MM_FROUND_CUR_DIRECTION);
+}
+
+
 static __inline __m512d __DEFAULT_FN_ATTRS
 _mm512_floor_pd(__m512d __A)
 {
@@ -1316,6 +1326,25 @@ _mm512_floor_pd(__m512d __A)
                                                    _MM_FROUND_CUR_DIRECTION);
 }
 
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask_floor_pd (__m512d __W, __mmask8 __U, __m512d __A)
+{
+  return (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __A,
+                _MM_FROUND_FLOOR,
+                (__v8df) __W, __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask_ceil_ps (__m512 __W, __mmask16 __U, __m512 __A)
+{
+  return (__m512) __builtin_ia32_rndscaleps_mask ((__v16sf) __A,
+                   _MM_FROUND_CEIL,
+                   (__v16sf) __W, __U,
+                   _MM_FROUND_CUR_DIRECTION);
+}
+
 static __inline __m512 __DEFAULT_FN_ATTRS
 _mm512_ceil_ps(__m512 __A)
 {
@@ -1334,6 +1363,15 @@ _mm512_ceil_pd(__m512d __A)
                                                    _MM_FROUND_CUR_DIRECTION);
 }
 
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask_ceil_pd (__m512d __W, __mmask8 __U, __m512d __A)
+{
+  return (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __A,
+                _MM_FROUND_CEIL,
+                (__v8df) __W, __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
 static __inline __m512i __DEFAULT_FN_ATTRS
 _mm512_abs_epi64(__m512i __A)
 {
@@ -1839,10 +1877,34 @@ _mm512_maskz_div_ps(__mmask16 __U, __m512 __A, __m512 __B) {
   (__m512)__builtin_ia32_rndscaleps_mask((__v16sf)(A), (B), (__v16sf)(A), \
                                          -1, _MM_FROUND_CUR_DIRECTION); })
 
+#define _mm512_mask_roundscale_ps( __A, __B, __C, __imm) __extension__ ({\
+  (__m512) __builtin_ia32_rndscaleps_mask ((__v16sf) __C, __imm,\
+                      (__v16sf) __A,(__mmask16) __B,\
+                     _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm512_maskz_roundscale_ps( __A, __B, __imm) __extension__ ({\
+  (__m512) __builtin_ia32_rndscaleps_mask ((__v16sf) __B, __imm,\
+                      (__v16sf) _mm512_setzero_ps (),\
+                      (__mmask16) __A, _MM_FROUND_CUR_DIRECTION);\
+})
+                    
 #define _mm512_roundscale_pd(A, B) __extension__ ({ \
   (__m512d)__builtin_ia32_rndscalepd_mask((__v8df)(A), (B), (__v8df)(A), \
                                           -1, _MM_FROUND_CUR_DIRECTION); })
 
+#define _mm512_mask_roundscale_pd(__A, __B, __C, __imm) __extension__ ({\
+  (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __C, __imm,\
+                          (__v8df) __A, (__mmask8) __B,\
+                          _MM_FROUND_CUR_DIRECTION);\
+})
+
+#define _mm512_maskz_roundscale_pd(__A, __B, __imm) __extension__ ({\
+  (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __B, __imm,\
+                          (__v8df) _mm512_setzero_pd (),\
+                          (__mmask8) __A, _MM_FROUND_CUR_DIRECTION);\
+})
+                     
 #define _mm512_fmadd_round_pd(A, B, C, R) __extension__ ({ \
   (__m512d) __builtin_ia32_vfmaddpd512_mask ((__v8df) (A), \
                                              (__v8df) (B), (__v8df) (C), \

diff --git a/test/CodeGen/avx512f-builtins.c b/test/CodeGen/avx512f-builtins.c
index 7d88c2a682287889bb9f39d02330bb56a8d7ee21..01dd2165202ad287fa0ee79a0f412671e25f8da1 100644 (file)
--- a/test/CodeGen/avx512f-builtins.c
+++ b/test/CodeGen/avx512f-builtins.c
@@ -6023,4 +6023,58 @@ __m512 test_mm512_maskz_min_ps (__mmask16 __U, __m512 __A, __m512 __B)
   return _mm512_maskz_min_ps (__U,__A,__B);
 }
 
+__m512 test_mm512_mask_floor_ps (__m512 __W, __mmask16 __U, __m512 __A)
+{
+  // CHECK-LABEL: @test_mm512_mask_floor_ps 
+  // CHECK: @llvm.x86.avx512.mask.rndscale.ps.512
+  return _mm512_mask_floor_ps (__W,__U,__A);
+}
+
+__m512d test_mm512_mask_floor_pd (__m512d __W, __mmask8 __U, __m512d __A)
+{
+  // CHECK-LABEL: @test_mm512_mask_floor_pd 
+  // CHECK: @llvm.x86.avx512.mask.rndscale.pd.512
+  return _mm512_mask_floor_pd (__W,__U,__A);
+}
+
+__m512 test_mm512_mask_ceil_ps (__m512 __W, __mmask16 __U, __m512 __A)
+{
+  // CHECK-LABEL: @test_mm512_mask_ceil_ps 
+  // CHECK: @llvm.x86.avx512.mask.rndscale.ps.512
+  return _mm512_mask_ceil_ps (__W,__U,__A);
+}
+
+__m512d test_mm512_mask_ceil_pd (__m512d __W, __mmask8 __U, __m512d __A)
+{
+  // CHECK-LABEL: @test_mm512_mask_ceil_pd 
+  // CHECK: @llvm.x86.avx512.mask.rndscale.pd.512
+  return _mm512_mask_ceil_pd (__W,__U,__A);
+}
 
+__m512 test_mm512_mask_roundscale_ps(__m512d __W, __mmask8 __U, __m512d __A) 
+{
+  // CHECK-LABEL: @test_mm512_mask_roundscale_ps
+  // CHECK: @llvm.x86.avx512.mask.rndscale.ps.512
+  return _mm512_mask_roundscale_ps(__W,__U,__A, 1);
+}
+
+__m512 test_mm512_maskz_roundscale_ps(__mmask8 __U, __m512d __A) 
+{
+  // CHECK-LABEL: @test_mm512_maskz_roundscale_ps
+  // CHECK: @llvm.x86.avx512.mask.rndscale.ps.512
+  return _mm512_maskz_roundscale_ps(__U,__A, 1);
+}
+
+__m512d test_mm512_mask_roundscale_pd(__m512d __W, __mmask8 __U, __m512d __A) 
+{
+  // CHECK-LABEL: @test_mm512_mask_roundscale_pd
+  // CHECK: @llvm.x86.avx512.mask.rndscale.pd.512
+  return _mm512_mask_roundscale_pd(__W,__U,__A, 1);
+}
+
+__m512d test_mm512_maskz_roundscale_pd(__mmask8 __U, __m512d __A) 
+{
+  // CHECK-LABEL: @test_mm512_maskz_roundscale_pd
+  // CHECK: @llvm.x86.avx512.mask.rndscale.pd.512
+  return _mm512_maskz_roundscale_pd(__U,__A, 1);
+}