[X86] Replace default _mm512_sqrt_pd/s/h implementations with generic __builtin_elementwise_sqrt

[RKSimon]

Followup to #165682

[llvmbot]

@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-clang

Author: Simon Pilgrim (RKSimon)

Changes

Followup to #165682

---

Full diff: https://github.com/llvm/llvm-project/pull/168057.diff

3 Files Affected:

• (modified) clang/lib/Headers/avx512fintrin.h (+14-28)
• (modified) clang/lib/Headers/avx512fp16intrin.h (+5-9)
• (modified) clang/test/CodeGen/X86/avx512fp16-builtins-constrained.c (+4-4)

diff --git a/clang/lib/Headers/avx512fintrin.h b/clang/lib/Headers/avx512fintrin.h
index 79c37173ac838..976eba816b8bf 100644
--- a/clang/lib/Headers/avx512fintrin.h
+++ b/clang/lib/Headers/avx512fintrin.h
@@ -1450,26 +1450,19 @@ _mm512_mask_mullox_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) {
                                        (__v8df)_mm512_sqrt_round_pd((A), (R)), \
                                        (__v8df)_mm512_setzero_pd()))
 
-static  __inline__ __m512d __DEFAULT_FN_ATTRS512
-_mm512_sqrt_pd(__m512d __A)
-{
-  return (__m512d)__builtin_ia32_sqrtpd512((__v8df)__A,
-                                           _MM_FROUND_CUR_DIRECTION);
+static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_sqrt_pd(__m512d __A) {
+  return (__m512d)__builtin_elementwise_sqrt((__v8df)__A);
 }
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS512
-_mm512_mask_sqrt_pd (__m512d __W, __mmask8 __U, __m512d __A)
-{
-  return (__m512d)__builtin_ia32_selectpd_512(__U,
-                                              (__v8df)_mm512_sqrt_pd(__A),
+_mm512_mask_sqrt_pd(__m512d __W, __mmask8 __U, __m512d __A) {
+  return (__m512d)__builtin_ia32_selectpd_512(__U, (__v8df)_mm512_sqrt_pd(__A),
                                               (__v8df)__W);
 }
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS512
-_mm512_maskz_sqrt_pd (__mmask8 __U, __m512d __A)
-{
-  return (__m512d)__builtin_ia32_selectpd_512(__U,
-                                              (__v8df)_mm512_sqrt_pd(__A),
+_mm512_maskz_sqrt_pd(__mmask8 __U, __m512d __A) {
+  return (__m512d)__builtin_ia32_selectpd_512(__U, (__v8df)_mm512_sqrt_pd(__A),
                                               (__v8df)_mm512_setzero_pd());
 }
 
@@ -1486,26 +1479,19 @@ _mm512_maskz_sqrt_pd (__mmask8 __U, __m512d __A)
                                       (__v16sf)_mm512_sqrt_round_ps((A), (R)), \
                                       (__v16sf)_mm512_setzero_ps()))
 
-static  __inline__ __m512 __DEFAULT_FN_ATTRS512
-_mm512_sqrt_ps(__m512 __A)
-{
-  return (__m512)__builtin_ia32_sqrtps512((__v16sf)__A,
-                                          _MM_FROUND_CUR_DIRECTION);
+static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_sqrt_ps(__m512 __A) {
+  return (__m512)__builtin_elementwise_sqrt((__v16sf)__A);
 }
 
-static  __inline__ __m512 __DEFAULT_FN_ATTRS512
-_mm512_mask_sqrt_ps(__m512 __W, __mmask16 __U, __m512 __A)
-{
-  return (__m512)__builtin_ia32_selectps_512(__U,
-                                             (__v16sf)_mm512_sqrt_ps(__A),
+static __inline__ __m512 __DEFAULT_FN_ATTRS512
+_mm512_mask_sqrt_ps(__m512 __W, __mmask16 __U, __m512 __A) {
+  return (__m512)__builtin_ia32_selectps_512(__U, (__v16sf)_mm512_sqrt_ps(__A),
                                              (__v16sf)__W);
 }
 
-static  __inline__ __m512 __DEFAULT_FN_ATTRS512
-_mm512_maskz_sqrt_ps( __mmask16 __U, __m512 __A)
-{
-  return (__m512)__builtin_ia32_selectps_512(__U,
-                                             (__v16sf)_mm512_sqrt_ps(__A),
+static __inline__ __m512 __DEFAULT_FN_ATTRS512
+_mm512_maskz_sqrt_ps(__mmask16 __U, __m512 __A) {
+  return (__m512)__builtin_ia32_selectps_512(__U, (__v16sf)_mm512_sqrt_ps(__A),
                                              (__v16sf)_mm512_setzero_ps());
 }
 
diff --git a/clang/lib/Headers/avx512fp16intrin.h b/clang/lib/Headers/avx512fp16intrin.h
index 25051228f3e0a..2776450387480 100644
--- a/clang/lib/Headers/avx512fp16intrin.h
+++ b/clang/lib/Headers/avx512fp16intrin.h
@@ -1401,24 +1401,20 @@ _mm_maskz_scalef_sh(__mmask8 __U, __m128h __A, __m128h __B) {
       (__v32hf)_mm512_setzero_ph()))
 
 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_sqrt_ph(__m512h __A) {
-  return (__m512h)__builtin_ia32_sqrtph512((__v32hf)__A,
-                                           _MM_FROUND_CUR_DIRECTION);
+  return (__m512h)__builtin_elementwise_sqrt((__v32hf)__A);
 }
 
 static __inline__ __m512h __DEFAULT_FN_ATTRS512
 _mm512_mask_sqrt_ph(__m512h __W, __mmask32 __U, __m512h __A) {
   return (__m512h)__builtin_ia32_selectph_512(
-      (__mmask32)(__U),
-      (__v32hf)__builtin_ia32_sqrtph512((__A), (_MM_FROUND_CUR_DIRECTION)),
-      (__v32hf)(__m512h)(__W));
+      (__mmask32)(__U), (__v32hf)_mm512_sqrt_ph(__A), (__v32hf)(__m512h)(__W));
 }
 
 static __inline__ __m512h __DEFAULT_FN_ATTRS512
 _mm512_maskz_sqrt_ph(__mmask32 __U, __m512h __A) {
-  return (__m512h)__builtin_ia32_selectph_512(
-      (__mmask32)(__U),
-      (__v32hf)__builtin_ia32_sqrtph512((__A), (_MM_FROUND_CUR_DIRECTION)),
-      (__v32hf)_mm512_setzero_ph());
+  return (__m512h)__builtin_ia32_selectph_512((__mmask32)(__U),
+                                              (__v32hf)_mm512_sqrt_ph(__A),
+                                              (__v32hf)_mm512_setzero_ph());
 }
 
 #define _mm_sqrt_round_sh(A, B, R)                                             \
diff --git a/clang/test/CodeGen/X86/avx512fp16-builtins-constrained.c b/clang/test/CodeGen/X86/avx512fp16-builtins-constrained.c
index cb029ccdcbf59..95403aecae9bd 100644
--- a/clang/test/CodeGen/X86/avx512fp16-builtins-constrained.c
+++ b/clang/test/CodeGen/X86/avx512fp16-builtins-constrained.c
@@ -65,8 +65,8 @@ __m512h test_mm512_sqrt_ph(__m512h x) {
 __m512h test_mm512_mask_sqrt_ph (__m512h __W, __mmask32 __U, __m512h __A)
 {
   // COMMON-LABEL: test_mm512_mask_sqrt_ph
-  // UNCONSTRAINED: call <32 x half> @llvm.sqrt.v32f16(<32 x half> %{{.*}})
-  // CONSTRAINED: call <32 x half> @llvm.experimental.constrained.sqrt.v32f16(<32 x half> %{{.*}}, metadata !{{.*}})
+  // UNCONSTRAINED: call {{.*}}<32 x half> @llvm.sqrt.v32f16(<32 x half> %{{.*}})
+  // CONSTRAINED: call {{.*}}<32 x half> @llvm.experimental.constrained.sqrt.v32f16(<32 x half> %{{.*}}, metadata !{{.*}})
   // CHECK-ASM: vsqrtph %zmm{{.*}},
   // COMMONIR: bitcast i32 %{{.*}} to <32 x i1>
   // COMMONIR: select <32 x i1> %{{.*}}, <32 x half> %{{.*}}, <32 x half> %{{.*}}
@@ -76,8 +76,8 @@ __m512h test_mm512_mask_sqrt_ph (__m512h __W, __mmask32 __U, __m512h __A)
 __m512h test_mm512_maskz_sqrt_ph (__mmask32 __U, __m512h __A)
 {
   // COMMON-LABEL: test_mm512_maskz_sqrt_ph
-  // UNCONSTRAINED: call <32 x half> @llvm.sqrt.v32f16(<32 x half> %{{.*}})
-  // CONSTRAINED: call <32 x half> @llvm.experimental.constrained.sqrt.v32f16(<32 x half> %{{.*}}, metadata !{{.*}})
+  // UNCONSTRAINED: call {{.*}}<32 x half> @llvm.sqrt.v32f16(<32 x half> %{{.*}})
+  // CONSTRAINED: call {{.*}}<32 x half> @llvm.experimental.constrained.sqrt.v32f16(<32 x half> %{{.*}}, metadata !{{.*}})
   // CHECK-ASM: vsqrtph %zmm{{.*}},
   // COMMONIR: bitcast i32 %{{.*}} to <32 x i1>
   // COMMONIR: select <32 x i1> %{{.*}}, <32 x half> %{{.*}}, <32 x half> {{.*}}

[RKSimon]

@philnik777 if you want to merge this into #165682 that's fine with me

[phoebewang]

LGTM.