[X86] Remove __builtin_ia32_pabs intrinsics and use generic __builtin_elementwise_abs

[RKSimon]

Summary

D111986 added the generic __builtin_elementwise_abs() intrinsic with the same integer absolute behaviour as the SSE/AVX instructions (abs(INT__MIN) == INT_MIN)

This patch removes the __builtin_ia32_pabs* intrinsics and just uses __builtin_elementwise_abs - the existing tests see no changes:

__m256i test_mm256_abs_epi8(__m256i a) {
  // CHECK-LABEL: test_mm256_abs_epi8
  // CHECK: [[ABS:%.*]] = call <32 x i8> @llvm.abs.v32i8(<32 x i8> %{{.*}}, i1 false)
  return _mm256_abs_epi8(a);
}

This requires us to add a __v64qs explicitly signed char vector type (we already have __v16qs and __v32qs).

[craig.topper]

Just to confirm since I can't see the test CHECK lines. builtin_elementwise_abs is not undefined for INT_MIN?

[RKSimon]

In D117791#3258602, @craig.topper wrote:

Just to confirm since I can't see the test CHECK lines. builtin_elementwise_abs is not undefined for INT_MIN?

That's why I added a copy of the check (from llvm-project\clang\test\CodeGen\X86\avx2-builtins.c) in the summary - the 'is_int_min_poison' flag in the abs intrinsic call is still false

[craig.topper]

In D117791#3258785, @RKSimon wrote:

In D117791#3258602, @craig.topper wrote:

Just to confirm since I can't see the test CHECK lines. builtin_elementwise_abs is not undefined for INT_MIN?

That's why I added a copy of the check (from llvm-project\clang\test\CodeGen\X86\avx2-builtins.c) in the summary - the 'is_int_min_poison' flag in the abs intrinsic call is still false

I shouldn't try to read patches too early in the morning I guess. Thanks.

[craig.topper]

LGTM

[pengfei]

clang/lib/Headers/avx512fintrin.h

31

Do we need to declare explicit signed?

[craig.topper]

clang/lib/Headers/avx512fintrin.h

31

The signedness of char is platform specific.

[pengfei]

clang/lib/Headers/avx512fintrin.h

31

Got it, thank you!

[RKSimon]

Quick update: we're hitting an issue with the libc-x86_64-debian-fullbuild-dbg and asan build bots which is using the wrong headers and breaks when the __builtin_ia32_pabs* builtins are removed. Once that has been addressed I'll recommit.

clang/include/clang/Basic/BuiltinsX86.def

 TARGET_BUILTIN(__builtin_ia32_phsubd128, "V4iV4iV4i", "ncV:128:", "ssse3")
 TARGET_BUILTIN(__builtin_ia32_phsubsw128, "V8sV8sV8s", "ncV:128:", "ssse3")
 TARGET_BUILTIN(__builtin_ia32_pmaddubsw128, "V8sV16cV16c", "ncV:128:", "ssse3")
 TARGET_BUILTIN(__builtin_ia32_pmulhrsw128, "V8sV8sV8s", "ncV:128:", "ssse3")
 TARGET_BUILTIN(__builtin_ia32_pshufb128, "V16cV16cV16c", "ncV:128:", "ssse3")
 TARGET_BUILTIN(__builtin_ia32_psignb128, "V16cV16cV16c", "ncV:128:", "ssse3")
 TARGET_BUILTIN(__builtin_ia32_psignw128, "V8sV8sV8s", "ncV:128:", "ssse3")
 TARGET_BUILTIN(__builtin_ia32_psignd128, "V4iV4iV4i", "ncV:128:", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pabsb128, "V16cV16c", "ncV:128:", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pabsw128, "V8sV8s", "ncV:128:", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pabsd128, "V4iV4i", "ncV:128:", "ssse3")
 
 TARGET_BUILTIN(__builtin_ia32_ldmxcsr, "vUi", "n", "sse")
 TARGET_HEADER_BUILTIN(_mm_setcsr, "vUi", "nh","xmmintrin.h", ALL_LANGUAGES, "sse")
 TARGET_BUILTIN(__builtin_ia32_stmxcsr, "Ui", "n", "sse")
 TARGET_HEADER_BUILTIN(_mm_getcsr, "Ui", "nh", "xmmintrin.h", ALL_LANGUAGES, "sse")
 TARGET_BUILTIN(__builtin_ia32_cvtss2si, "iV4f", "ncV:128:", "sse")
 TARGET_BUILTIN(__builtin_ia32_cvttss2si, "iV4f", "ncV:128:", "sse")
 TARGET_BUILTIN(__builtin_ia32_movmskps, "iV4f", "nV:128:", "sse")

 TARGET_BUILTIN(__builtin_ia32_vec_ext_v16hi, "sV16sIi", "ncV:256:", "avx")
 TARGET_BUILTIN(__builtin_ia32_vec_ext_v8si, "iV8iIi", "ncV:256:", "avx")
 TARGET_BUILTIN(__builtin_ia32_vec_set_v32qi, "V32cV32ccIi", "ncV:256:", "avx")
 TARGET_BUILTIN(__builtin_ia32_vec_set_v16hi, "V16sV16ssIi", "ncV:256:", "avx")
 TARGET_BUILTIN(__builtin_ia32_vec_set_v8si, "V8iV8iiIi", "ncV:256:", "avx")
 
 // AVX2
 TARGET_BUILTIN(__builtin_ia32_mpsadbw256, "V32cV32cV32cIc", "ncV:256:", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pabsb256, "V32cV32c", "ncV:256:", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pabsw256, "V16sV16s", "ncV:256:", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pabsd256, "V8iV8i", "ncV:256:", "avx2")
 TARGET_BUILTIN(__builtin_ia32_packsswb256, "V32cV16sV16s", "ncV:256:", "avx2")
 TARGET_BUILTIN(__builtin_ia32_packssdw256, "V16sV8iV8i", "ncV:256:", "avx2")
 TARGET_BUILTIN(__builtin_ia32_packuswb256, "V32cV16sV16s", "ncV:256:", "avx2")
 TARGET_BUILTIN(__builtin_ia32_packusdw256, "V16sV8iV8i", "ncV:256:", "avx2")
 TARGET_BUILTIN(__builtin_ia32_paddsb256, "V32cV32cV32c", "ncV:256:", "avx2")
 TARGET_BUILTIN(__builtin_ia32_paddsw256, "V16sV16sV16s", "ncV:256:", "avx2")
 TARGET_BUILTIN(__builtin_ia32_psubsb256, "V32cV32cV32c", "ncV:256:", "avx2")
 TARGET_BUILTIN(__builtin_ia32_psubsw256, "V16sV16sV16s", "ncV:256:", "avx2")

 TARGET_BUILTIN(__builtin_ia32_minpd512, "V8dV8dV8dIi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_maxps512, "V16fV16fV16fIi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_maxpd512, "V8dV8dV8dIi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtdq2ps512_mask, "V16fV16iV16fUsIi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtudq2ps512_mask, "V16fV16iV16fUsIi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvtpd2ps512_mask, "V8fV8dV8fUcIi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vcvtps2ph512_mask, "V16sV16fIiV16sUs", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vcvtph2ps512_mask, "V16fV16sV16fUsIi", "ncV:512:", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pabsd512, "V16iV16i", "ncV:512:", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pabsq512, "V8OiV8Oi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmaxsd512, "V16iV16iV16i", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmaxsq512, "V8OiV8OiV8Oi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmaxud512, "V16iV16iV16i", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmaxuq512, "V8OiV8OiV8Oi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pminsd512, "V16iV16iV16i", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pminsq512, "V8OiV8OiV8Oi", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pminud512, "V16iV16iV16i", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pminuq512, "V8OiV8OiV8Oi", "ncV:512:", "avx512f")

 TARGET_BUILTIN(__builtin_ia32_ucmpd256_mask, "UcV8iV8iIiUc", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_ucmpq256_mask, "UcV4OiV4OiIiUc", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_ucmpw256_mask, "UsV16sV16sIiUs", "ncV:256:", "avx512vl,avx512bw")
 TARGET_BUILTIN(__builtin_ia32_ucmpb512_mask, "UOiV64cV64cIiUOi", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_ucmpd512_mask, "UsV16iV16iIiUs", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_ucmpq512_mask, "UcV8OiV8OiIiUc", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_ucmpw512_mask, "UiV32sV32sIiUi", "ncV:512:", "avx512bw")
 
-TARGET_BUILTIN(__builtin_ia32_pabsb512, "V64cV64c", "ncV:512:", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pabsw512, "V32sV32s", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_packssdw512, "V32sV16iV16i", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_packsswb512, "V64cV32sV32s", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_packusdw512, "V32sV16iV16i", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_packuswb512, "V64cV32sV32s", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_paddsb512, "V64cV64cV64c", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_paddsw512, "V32sV32sV32s", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_paddusb512, "V64cV64cV64c", "ncV:512:", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_paddusw512, "V32sV32sV32s", "ncV:512:", "avx512bw")

 TARGET_BUILTIN(__builtin_ia32_expandsf128_mask, "V4fV4fV4fUc", "ncV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_expandsf256_mask, "V8fV8fV8fUc", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_expandsi128_mask, "V4iV4iV4iUc", "ncV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_expandsi256_mask, "V8iV8iV8iUc", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_getexppd128_mask, "V2dV2dV2dUc", "ncV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_getexppd256_mask, "V4dV4dV4dUc", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_getexpps128_mask, "V4fV4fV4fUc", "ncV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_getexpps256_mask, "V8fV8fV8fUc", "ncV:256:", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pabsq128, "V2OiV2Oi", "ncV:128:", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pabsq256, "V4OiV4Oi", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmaxsq128, "V2OiV2OiV2Oi", "ncV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmaxsq256, "V4OiV4OiV4Oi", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmaxuq128, "V2OiV2OiV2Oi", "ncV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmaxuq256, "V4OiV4OiV4Oi", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pminsq128, "V2OiV2OiV2Oi", "ncV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pminsq256, "V4OiV4OiV4Oi", "ncV:256:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pminuq128, "V2OiV2OiV2Oi", "ncV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pminuq256, "V4OiV4OiV4Oi", "ncV:256:", "avx512vl")

clang/lib/CodeGen/CGBuiltin.cpp

if (Builder.getIsFPConstrained()) {
       Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt,
                                      Ops[0]->getType());
       return Builder.CreateConstrainedFPCall(F, Ops[0]);
     } else {
       Function *F = CGM.getIntrinsic(Intrinsic::sqrt, Ops[0]->getType());
       return Builder.CreateCall(F, Ops[0]);
     }
   }
-  case X86::BI__builtin_ia32_pabsb128:
-  case X86::BI__builtin_ia32_pabsw128:
-  case X86::BI__builtin_ia32_pabsd128:
-  case X86::BI__builtin_ia32_pabsb256:
-  case X86::BI__builtin_ia32_pabsw256:
-  case X86::BI__builtin_ia32_pabsd256:
-  case X86::BI__builtin_ia32_pabsq128:
-  case X86::BI__builtin_ia32_pabsq256:
-  case X86::BI__builtin_ia32_pabsb512:
-  case X86::BI__builtin_ia32_pabsw512:
-  case X86::BI__builtin_ia32_pabsd512:
-  case X86::BI__builtin_ia32_pabsq512: {
-    Function *F = CGM.getIntrinsic(Intrinsic::abs, Ops[0]->getType());
-    return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
-  }
   case X86::BI__builtin_ia32_pmaxsb128:
   case X86::BI__builtin_ia32_pmaxsw128:
   case X86::BI__builtin_ia32_pmaxsd128:
   case X86::BI__builtin_ia32_pmaxsq128:
   case X86::BI__builtin_ia32_pmaxsb256:
   case X86::BI__builtin_ia32_pmaxsw256:
   case X86::BI__builtin_ia32_pmaxsd256:
   case X86::BI__builtin_ia32_pmaxsq256:

clang/lib/Headers/avx2intrin.h

 /* SSE4 Multiple Packed Sums of Absolute Difference.  */
 #define _mm256_mpsadbw_epu8(X, Y, M) \
   ((__m256i)__builtin_ia32_mpsadbw256((__v32qi)(__m256i)(X), \
                                       (__v32qi)(__m256i)(Y), (int)(M)))
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_abs_epi8(__m256i __a)
 {
-    return (__m256i)__builtin_ia32_pabsb256((__v32qi)__a);
+    return (__m256i)__builtin_elementwise_abs((__v32qs)__a);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_abs_epi16(__m256i __a)
 {
-    return (__m256i)__builtin_ia32_pabsw256((__v16hi)__a);
+    return (__m256i)__builtin_elementwise_abs((__v16hi)__a);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_abs_epi32(__m256i __a)
 {
-    return (__m256i)__builtin_ia32_pabsd256((__v8si)__a);
+    return (__m256i)__builtin_elementwise_abs((__v8si)__a);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_packs_epi16(__m256i __a, __m256i __b)
 {
   return (__m256i)__builtin_ia32_packsswb256((__v16hi)__a, (__v16hi)__b);
 }
 

clang/lib/Headers/avx512bwintrin.h

_mm512_mask_blend_epi16 (__mmask32 __U, __m512i __A, __m512i __W)
   return (__m512i) __builtin_ia32_selectw_512 ((__mmask32) __U,
               (__v32hi) __W,
               (__v32hi) __A);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_abs_epi8 (__m512i __A)
 {
-  return (__m512i)__builtin_ia32_pabsb512((__v64qi)__A);
+  return (__m512i)__builtin_elementwise_abs((__v64qs)__A);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_mask_abs_epi8 (__m512i __W, __mmask64 __U, __m512i __A)
 {
   return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U,
                                              (__v64qi)_mm512_abs_epi8(__A),
                                              (__v64qi)__W);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_maskz_abs_epi8 (__mmask64 __U, __m512i __A)
 {
   return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U,
                                              (__v64qi)_mm512_abs_epi8(__A),
                                              (__v64qi)_mm512_setzero_si512());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_abs_epi16 (__m512i __A)
 {
-  return (__m512i)__builtin_ia32_pabsw512((__v32hi)__A);
+  return (__m512i)__builtin_elementwise_abs((__v32hi)__A);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_mask_abs_epi16 (__m512i __W, __mmask32 __U, __m512i __A)
 {
   return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U,
                                              (__v32hi)_mm512_abs_epi16(__A),
                                              (__v32hi)__W);

clang/lib/Headers/avx512fintrin.h

 typedef int __v16si __attribute__((__vector_size__(64)));
 
 /* Unsigned types */
 typedef unsigned char __v64qu __attribute__((__vector_size__(64)));
 typedef unsigned short __v32hu __attribute__((__vector_size__(64)));
 typedef unsigned long long __v8du __attribute__((__vector_size__(64)));
 typedef unsigned int __v16su __attribute__((__vector_size__(64)));
 
+/* We need an explicitly signed variant for char. Note that this shouldn't
+ * appear in the interface though. */
+typedef signed char __v64qs __attribute__((__vector_size__(64)));
+
 typedef float __m512 __attribute__((__vector_size__(64), __aligned__(64)));
 typedef double __m512d __attribute__((__vector_size__(64), __aligned__(64)));
 typedef long long __m512i __attribute__((__vector_size__(64), __aligned__(64)));
 
 typedef float __m512_u __attribute__((__vector_size__(64), __aligned__(1)));
 typedef double __m512d_u __attribute__((__vector_size__(64), __aligned__(1)));
 typedef long long __m512i_u __attribute__((__vector_size__(64), __aligned__(1)));
 
return (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __A,
                 _MM_FROUND_CEIL,
                 (__v8df) __W, __U,
                 _MM_FROUND_CUR_DIRECTION);
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_abs_epi64(__m512i __A)
 {
-  return (__m512i)__builtin_ia32_pabsq512((__v8di)__A);
+  return (__m512i)__builtin_elementwise_abs((__v8di)__A);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_mask_abs_epi64 (__m512i __W, __mmask8 __U, __m512i __A)
 {
   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U,
                                              (__v8di)_mm512_abs_epi64(__A),
                                              (__v8di)__W);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_maskz_abs_epi64 (__mmask8 __U, __m512i __A)
 {
   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U,
                                              (__v8di)_mm512_abs_epi64(__A),
                                              (__v8di)_mm512_setzero_si512());
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_abs_epi32(__m512i __A)
 {
-  return (__m512i)__builtin_ia32_pabsd512((__v16si) __A);
+  return (__m512i)__builtin_elementwise_abs((__v16si) __A);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_mask_abs_epi32 (__m512i __W, __mmask16 __U, __m512i __A)
 {
   return (__m512i)__builtin_ia32_selectd_512(__U,
                                              (__v16si)_mm512_abs_epi32(__A),
                                              (__v16si)__W);

clang/lib/Headers/avx512vlintrin.h

 _mm256_maskz_abs_epi32(__mmask8 __U, __m256i __A) {
   return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
                                              (__v8si)_mm256_abs_epi32(__A),
                                              (__v8si)_mm256_setzero_si256());
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_abs_epi64 (__m128i __A) {
-  return (__m128i)__builtin_ia32_pabsq128((__v2di)__A);
+  return (__m128i)__builtin_elementwise_abs((__v2di)__A);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A) {
   return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
                                              (__v2di)_mm_abs_epi64(__A),
                                              (__v2di)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A) {
   return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
                                              (__v2di)_mm_abs_epi64(__A),
                                              (__v2di)_mm_setzero_si128());
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_abs_epi64 (__m256i __A) {
-  return (__m256i)__builtin_ia32_pabsq256 ((__v4di)__A);
+  return (__m256i)__builtin_elementwise_abs((__v4di)__A);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A) {
   return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
                                              (__v4di)_mm256_abs_epi64(__A),
                                              (__v4di)__W);
 }

clang/lib/Headers/tmmintrin.h

 ///
 /// \param __a
 ///    A 128-bit vector of [16 x i8].
 /// \returns A 128-bit integer vector containing the absolute values of the
 ///    elements in the operand.
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_abs_epi8(__m128i __a)
 {
-    return (__m128i)__builtin_ia32_pabsb128((__v16qi)__a);
+    return (__m128i)__builtin_elementwise_abs((__v16qs)__a);
 }
 
 /// Computes the absolute value of each of the packed 16-bit signed
 ///    integers in the source operand and stores the 16-bit unsigned integer
 ///    results in the destination.
 ///
 /// \headerfile <x86intrin.h>
 ///

 ///
 /// \param __a
 ///    A 128-bit vector of [8 x i16].
 /// \returns A 128-bit integer vector containing the absolute values of the
 ///    elements in the operand.
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_abs_epi16(__m128i __a)
 {
-    return (__m128i)__builtin_ia32_pabsw128((__v8hi)__a);
+    return (__m128i)__builtin_elementwise_abs((__v8hi)__a);
 }
 
 /// Computes the absolute value of each of the packed 32-bit signed
 ///    integers in the source operand and stores the 32-bit unsigned integer
 ///    results in the destination.
 ///
 /// \headerfile <x86intrin.h>
 ///

 ///
 /// \param __a
 ///    A 128-bit vector of [4 x i32].
 /// \returns A 128-bit integer vector containing the absolute values of the
 ///    elements in the operand.
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_abs_epi32(__m128i __a)
 {
-    return (__m128i)__builtin_ia32_pabsd128((__v4si)__a);
+    return (__m128i)__builtin_elementwise_abs((__v4si)__a);
 }
 
 /// Concatenates the two 128-bit integer vector operands, and
 ///    right-shifts the result by the number of bytes specified in the immediate
 ///    operand.
 ///
 /// \headerfile <x86intrin.h>
 ///

clang/test/CodeGen/builtins-x86.c

#endif
   tmp_V16c = __builtin_ia32_pshufb128(tmp_V16c, tmp_V16c);
   tmp_V8c = __builtin_ia32_pshufb(tmp_V8c, tmp_V8c);
   tmp_V16c = __builtin_ia32_psignb128(tmp_V16c, tmp_V16c);
   tmp_V8c = __builtin_ia32_psignb(tmp_V8c, tmp_V8c);
   tmp_V8s = __builtin_ia32_psignw128(tmp_V8s, tmp_V8s);
   tmp_V4s = __builtin_ia32_psignw(tmp_V4s, tmp_V4s);
   tmp_V4i = __builtin_ia32_psignd128(tmp_V4i, tmp_V4i);
   tmp_V2i = __builtin_ia32_psignd(tmp_V2i, tmp_V2i);
-  tmp_V16c = __builtin_ia32_pabsb128(tmp_V16c);
   tmp_V8c = __builtin_ia32_pabsb(tmp_V8c);
-  tmp_V8s = __builtin_ia32_pabsw128(tmp_V8s);
   tmp_V4s = __builtin_ia32_pabsw(tmp_V4s);
-  tmp_V4i = __builtin_ia32_pabsd128(tmp_V4i);
   tmp_V2i = __builtin_ia32_pabsd(tmp_V2i);
   tmp_V4s = __builtin_ia32_psllw(tmp_V4s, tmp_V1LLi);
   tmp_V2i = __builtin_ia32_pslld(tmp_V2i, tmp_V1LLi);
   tmp_V1LLi = __builtin_ia32_psllq(tmp_V1LLi, tmp_V1LLi);
   tmp_V4s = __builtin_ia32_psrlw(tmp_V4s, tmp_V1LLi);
   tmp_V2i = __builtin_ia32_psrld(tmp_V2i, tmp_V1LLi);
   tmp_V1LLi = __builtin_ia32_psrlq(tmp_V1LLi, tmp_V1LLi);
   tmp_V4s = __builtin_ia32_psraw(tmp_V4s, tmp_V1LLi);