r331814 - [X86] Mark builtins 'const' where possible

Craig Topper via cfe-commits cfe-commits at lists.llvm.org
Tue May 8 15:01:43 PDT 2018


Author: ctopper
Date: Tue May  8 15:01:43 2018
New Revision: 331814

URL: http://llvm.org/viewvc/llvm-project?rev=331814&view=rev
Log:
[X86] Mark builtins 'const' where possible

I attempted to go through all the builtins and marked them const if they didn't touch memory or other hidden state.

I don't know how to test this or if it really matters.

Differential Revision: https://reviews.llvm.org/D46349

Modified:
    cfe/trunk/include/clang/Basic/BuiltinsX86.def
    cfe/trunk/include/clang/Basic/BuiltinsX86_64.def

Modified: cfe/trunk/include/clang/Basic/BuiltinsX86.def
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/include/clang/Basic/BuiltinsX86.def?rev=331814&r1=331813&r2=331814&view=diff
==============================================================================
--- cfe/trunk/include/clang/Basic/BuiltinsX86.def (original)
+++ cfe/trunk/include/clang/Basic/BuiltinsX86.def Tue May  8 15:01:43 2018
@@ -27,8 +27,6 @@
 #  define TARGET_HEADER_BUILTIN(ID, TYPE, ATTRS, HEADER, LANG, FEATURE) BUILTIN(ID, TYPE, ATTRS)
 #endif
 
-// FIXME: Are these const?
-
 // Miscellaneous builtin for checking x86 cpu features.
 // TODO: Make this somewhat generic so that other backends
 // can use it?
@@ -87,250 +85,250 @@ TARGET_BUILTIN(__builtin_ia32_pswapdsi,
 // doesn't work in the presence of re-declaration of _mm_prefetch for windows.
 TARGET_BUILTIN(_mm_prefetch, "vcC*i", "nc", "mmx")
 TARGET_BUILTIN(__builtin_ia32_emms, "v", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_paddb, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_paddw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_paddd, "V2iV2iV2i", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_paddsb, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_paddsw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_paddusb, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_paddusw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psubb, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psubw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psubd, "V2iV2iV2i", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psubsb, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psubsw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psubusb, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psubusw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pmulhw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pmullw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pmaddwd, "V2iV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pand, "V1LLiV1LLiV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pandn, "V1LLiV1LLiV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_por, "V1LLiV1LLiV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pxor, "V1LLiV1LLiV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psllw, "V4sV4sV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pslld, "V2iV2iV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psllq, "V1LLiV1LLiV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psrlw, "V4sV4sV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psrld, "V2iV2iV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psrlq, "V1LLiV1LLiV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psraw, "V4sV4sV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psrad, "V2iV2iV1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psllwi, "V4sV4si", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pslldi, "V2iV2ii", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psllqi, "V1LLiV1LLii", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psrlwi, "V4sV4si", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psrldi, "V2iV2ii", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psrlqi, "V1LLiV1LLii", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psrawi, "V4sV4si", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_psradi, "V2iV2ii", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_packsswb, "V8cV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_packssdw, "V4sV2iV2i", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_packuswb, "V8cV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_punpckhbw, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_punpckhwd, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_punpckhdq, "V2iV2iV2i", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_punpcklbw, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_punpcklwd, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_punpckldq, "V2iV2iV2i", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pcmpeqb, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pcmpeqw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pcmpeqd, "V2iV2iV2i", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pcmpgtb, "V8cV8cV8c", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pcmpgtw, "V4sV4sV4s", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_pcmpgtd, "V2iV2iV2i", "n", "mmx")
+TARGET_BUILTIN(__builtin_ia32_paddb, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_paddw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_paddd, "V2iV2iV2i", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_paddsb, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_paddsw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_paddusb, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_paddusw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psubb, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psubw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psubd, "V2iV2iV2i", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psubsb, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psubsw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psubusb, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psubusw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pmulhw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pmullw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pmaddwd, "V2iV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pand, "V1LLiV1LLiV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pandn, "V1LLiV1LLiV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_por, "V1LLiV1LLiV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pxor, "V1LLiV1LLiV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psllw, "V4sV4sV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pslld, "V2iV2iV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psllq, "V1LLiV1LLiV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psrlw, "V4sV4sV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psrld, "V2iV2iV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psrlq, "V1LLiV1LLiV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psraw, "V4sV4sV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psrad, "V2iV2iV1LLi", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psllwi, "V4sV4si", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pslldi, "V2iV2ii", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psllqi, "V1LLiV1LLii", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psrlwi, "V4sV4si", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psrldi, "V2iV2ii", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psrlqi, "V1LLiV1LLii", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psrawi, "V4sV4si", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_psradi, "V2iV2ii", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_packsswb, "V8cV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_packssdw, "V4sV2iV2i", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_packuswb, "V8cV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_punpckhbw, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_punpckhwd, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_punpckhdq, "V2iV2iV2i", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_punpcklbw, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_punpcklwd, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_punpckldq, "V2iV2iV2i", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pcmpeqb, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pcmpeqw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pcmpeqd, "V2iV2iV2i", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pcmpgtb, "V8cV8cV8c", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pcmpgtw, "V4sV4sV4s", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_pcmpgtd, "V2iV2iV2i", "nc", "mmx")
 TARGET_BUILTIN(__builtin_ia32_maskmovq, "vV8cV8cc*", "n", "mmx")
 TARGET_BUILTIN(__builtin_ia32_movntq, "vV1LLi*V1LLi", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_vec_init_v2si, "V2iii", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_vec_init_v4hi, "V4sssss", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_vec_init_v8qi, "V8ccccccccc", "n", "mmx")
-TARGET_BUILTIN(__builtin_ia32_vec_ext_v2si, "iV2ii", "n", "mmx")
+TARGET_BUILTIN(__builtin_ia32_vec_init_v2si, "V2iii", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_vec_init_v4hi, "V4sssss", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_vec_init_v8qi, "V8ccccccccc", "nc", "mmx")
+TARGET_BUILTIN(__builtin_ia32_vec_ext_v2si, "iV2ii", "nc", "mmx")
 
 // MMX2 (MMX+SSE) intrinsics
-TARGET_BUILTIN(__builtin_ia32_cvtpi2ps, "V4fV4fV2i", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cvtps2pi, "V2iV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cvttps2pi, "V2iV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pavgb, "V8cV8cV8c", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pavgw, "V4sV4sV4s", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pmaxsw, "V4sV4sV4s", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pmaxub, "V8cV8cV8c", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pminsw, "V4sV4sV4s", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pminub, "V8cV8cV8c", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pmovmskb, "iV8c", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pmulhuw, "V4sV4sV4s", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_psadbw, "V4sV8cV8c", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_pshufw, "V4sV4sIc", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_vec_ext_v4hi, "iV4sIi", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_vec_set_v4hi, "V4sV4siIi", "n", "sse")
+TARGET_BUILTIN(__builtin_ia32_cvtpi2ps, "V4fV4fV2i", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cvtps2pi, "V2iV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cvttps2pi, "V2iV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pavgb, "V8cV8cV8c", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pavgw, "V4sV4sV4s", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pmaxsw, "V4sV4sV4s", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pmaxub, "V8cV8cV8c", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pminsw, "V4sV4sV4s", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pminub, "V8cV8cV8c", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pmovmskb, "iV8c", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pmulhuw, "V4sV4sV4s", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_psadbw, "V4sV8cV8c", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_pshufw, "V4sV4sIc", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_vec_ext_v4hi, "iV4sIi", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_vec_set_v4hi, "V4sV4siIi", "nc", "sse")
 
 // MMX+SSE2
-TARGET_BUILTIN(__builtin_ia32_cvtpd2pi, "V2iV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvtpi2pd, "V2dV2i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2pi, "V2iV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_paddq, "V1LLiV1LLiV1LLi", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pmuludq, "V1LLiV2iV2i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psubq, "V1LLiV1LLiV1LLi", "n", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2pi, "V2iV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtpi2pd, "V2dV2i", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2pi, "V2iV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_paddq, "V1LLiV1LLiV1LLi", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pmuludq, "V1LLiV2iV2i", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psubq, "V1LLiV1LLiV1LLi", "nc", "sse2")
 
 // MMX+SSSE3
-TARGET_BUILTIN(__builtin_ia32_pabsb, "V8cV8c", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pabsd, "V2iV2i", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pabsw, "V4sV4s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_palignr, "V8cV8cV8cIc", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phaddd, "V2iV2iV2i", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phaddsw, "V4sV4sV4s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phaddw, "V4sV4sV4s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phsubd, "V2iV2iV2i", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phsubsw, "V4sV4sV4s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phsubw, "V4sV4sV4s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pmaddubsw, "V8cV8cV8c", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pmulhrsw, "V4sV4sV4s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pshufb, "V8cV8cV8c", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_psignw, "V4sV4sV4s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_psignb, "V8cV8cV8c", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_psignd, "V2iV2iV2i", "n", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pabsb, "V8cV8c", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pabsd, "V2iV2i", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pabsw, "V4sV4s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_palignr, "V8cV8cV8cIc", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phaddd, "V2iV2iV2i", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phaddsw, "V4sV4sV4s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phaddw, "V4sV4sV4s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phsubd, "V2iV2iV2i", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phsubsw, "V4sV4sV4s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phsubw, "V4sV4sV4s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pmaddubsw, "V8cV8cV8c", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pmulhrsw, "V4sV4sV4s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pshufb, "V8cV8cV8c", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_psignw, "V4sV4sV4s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_psignb, "V8cV8cV8c", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_psignd, "V2iV2iV2i", "nc", "ssse3")
 
 // SSE intrinsics.
-TARGET_BUILTIN(__builtin_ia32_comieq, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_comilt, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_comile, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_comigt, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_comige, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_comineq, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_ucomieq, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_ucomilt, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_ucomile, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_ucomigt, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_ucomige, "iV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_ucomineq, "iV4fV4f", "n", "sse")
-
-TARGET_BUILTIN(__builtin_ia32_comisdeq, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_comisdlt, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_comisdle, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_comisdgt, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_comisdge, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_comisdneq, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_ucomisdeq, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_ucomisdlt, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_ucomisdle, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_ucomisdgt, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_ucomisdge, "iV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_ucomisdneq, "iV2dV2d", "n", "sse2")
-
-TARGET_BUILTIN(__builtin_ia32_cmpeqps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpltps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpleps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpunordps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpneqps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpnltps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpnleps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpordps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpeqss, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpltss, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpless, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpunordss, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpneqss, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpnltss, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpnless, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cmpordss, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_minps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_maxps, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_minss, "V4fV4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_maxss, "V4fV4fV4f", "n", "sse")
-
-TARGET_BUILTIN(__builtin_ia32_cmpeqpd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpltpd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmplepd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpunordpd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpneqpd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpnltpd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpnlepd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpordpd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpeqsd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpltsd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmplesd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpunordsd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpneqsd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpnltsd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpnlesd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cmpordsd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_minpd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_maxpd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_minsd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_maxsd, "V2dV2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_paddsb128, "V16cV16cV16c", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_paddsw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psubsb128, "V16cV16cV16c", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psubsw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_paddusb128, "V16cV16cV16c", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_paddusw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psubusb128, "V16cV16cV16c", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psubusw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pmulhw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pmaxub128, "V16cV16cV16c", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pmaxsw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pminub128, "V16cV16cV16c", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pminsw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_packsswb128, "V16cV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_packssdw128, "V8sV4iV4i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_packuswb128, "V16cV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pmulhuw128, "V8sV8sV8s", "n", "sse2")
-
-TARGET_BUILTIN(__builtin_ia32_addsubps, "V4fV4fV4f", "n", "sse3")
-TARGET_BUILTIN(__builtin_ia32_addsubpd, "V2dV2dV2d", "n", "sse3")
-TARGET_BUILTIN(__builtin_ia32_haddps, "V4fV4fV4f", "n", "sse3")
-TARGET_BUILTIN(__builtin_ia32_haddpd, "V2dV2dV2d", "n", "sse3")
-TARGET_BUILTIN(__builtin_ia32_hsubps, "V4fV4fV4f", "n", "sse3")
-TARGET_BUILTIN(__builtin_ia32_hsubpd, "V2dV2dV2d", "n", "sse3")
-TARGET_BUILTIN(__builtin_ia32_phaddw128, "V8sV8sV8s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phaddd128, "V4iV4iV4i", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phaddsw128, "V8sV8sV8s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phsubw128, "V8sV8sV8s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phsubd128, "V4iV4iV4i", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_phsubsw128, "V8sV8sV8s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pmaddubsw128, "V8sV16cV16c", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pmulhrsw128, "V8sV8sV8s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pshufb128, "V16cV16cV16c", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_psignb128, "V16cV16cV16c", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_psignw128, "V8sV8sV8s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_psignd128, "V4iV4iV4i", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pabsb128, "V16cV16c", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pabsw128, "V8sV8s", "n", "ssse3")
-TARGET_BUILTIN(__builtin_ia32_pabsd128, "V4iV4i", "n", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_comieq, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_comilt, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_comile, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_comigt, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_comige, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_comineq, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_ucomieq, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_ucomilt, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_ucomile, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_ucomigt, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_ucomige, "iV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_ucomineq, "iV4fV4f", "nc", "sse")
+
+TARGET_BUILTIN(__builtin_ia32_comisdeq, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_comisdlt, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_comisdle, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_comisdgt, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_comisdge, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_comisdneq, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_ucomisdeq, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_ucomisdlt, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_ucomisdle, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_ucomisdgt, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_ucomisdge, "iV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_ucomisdneq, "iV2dV2d", "nc", "sse2")
+
+TARGET_BUILTIN(__builtin_ia32_cmpeqps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpltps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpleps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpunordps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpneqps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpnltps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpnleps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpordps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpeqss, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpltss, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpless, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpunordss, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpneqss, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpnltss, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpnless, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cmpordss, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_minps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_maxps, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_minss, "V4fV4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_maxss, "V4fV4fV4f", "nc", "sse")
+
+TARGET_BUILTIN(__builtin_ia32_cmpeqpd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpltpd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmplepd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpunordpd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpneqpd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpnltpd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpnlepd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpordpd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpeqsd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpltsd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmplesd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpunordsd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpneqsd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpnltsd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpnlesd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cmpordsd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_minpd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_maxpd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_minsd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_maxsd, "V2dV2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_paddsb128, "V16cV16cV16c", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_paddsw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psubsb128, "V16cV16cV16c", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psubsw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_paddusb128, "V16cV16cV16c", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_paddusw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psubusb128, "V16cV16cV16c", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psubusw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pmulhw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pmaxub128, "V16cV16cV16c", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pmaxsw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pminub128, "V16cV16cV16c", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pminsw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_packsswb128, "V16cV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_packssdw128, "V8sV4iV4i", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_packuswb128, "V16cV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pmulhuw128, "V8sV8sV8s", "nc", "sse2")
+
+TARGET_BUILTIN(__builtin_ia32_addsubps, "V4fV4fV4f", "nc", "sse3")
+TARGET_BUILTIN(__builtin_ia32_addsubpd, "V2dV2dV2d", "nc", "sse3")
+TARGET_BUILTIN(__builtin_ia32_haddps, "V4fV4fV4f", "nc", "sse3")
+TARGET_BUILTIN(__builtin_ia32_haddpd, "V2dV2dV2d", "nc", "sse3")
+TARGET_BUILTIN(__builtin_ia32_hsubps, "V4fV4fV4f", "nc", "sse3")
+TARGET_BUILTIN(__builtin_ia32_hsubpd, "V2dV2dV2d", "nc", "sse3")
+TARGET_BUILTIN(__builtin_ia32_phaddw128, "V8sV8sV8s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phaddd128, "V4iV4iV4i", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phaddsw128, "V8sV8sV8s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phsubw128, "V8sV8sV8s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phsubd128, "V4iV4iV4i", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_phsubsw128, "V8sV8sV8s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pmaddubsw128, "V8sV16cV16c", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pmulhrsw128, "V8sV8sV8s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pshufb128, "V16cV16cV16c", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_psignb128, "V16cV16cV16c", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_psignw128, "V8sV8sV8s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_psignd128, "V4iV4iV4i", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pabsb128, "V16cV16c", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pabsw128, "V8sV8s", "nc", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_pabsd128, "V4iV4i", "nc", "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", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cvttss2si, "iV4f", "n", "sse")
+TARGET_BUILTIN(__builtin_ia32_cvtss2si, "iV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cvttss2si, "iV4f", "nc", "sse")
 TARGET_BUILTIN(__builtin_ia32_storehps, "vV2i*V4f", "n", "sse")
 TARGET_BUILTIN(__builtin_ia32_storelps, "vV2i*V4f", "n", "sse")
 TARGET_BUILTIN(__builtin_ia32_movmskps, "iV4f", "n", "sse")
 TARGET_BUILTIN(__builtin_ia32_sfence, "v", "n", "sse")
 TARGET_HEADER_BUILTIN(_mm_sfence, "v", "nh", "xmmintrin.h", ALL_LANGUAGES, "sse")
-TARGET_BUILTIN(__builtin_ia32_rcpps, "V4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_rcpss, "V4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_rsqrtps, "V4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_rsqrtss, "V4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_sqrtps, "V4fV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_sqrtss, "V4fV4f", "n", "sse")
+TARGET_BUILTIN(__builtin_ia32_rcpps, "V4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_rcpss, "V4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_rsqrtps, "V4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_rsqrtss, "V4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_sqrtps, "V4fV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_sqrtss, "V4fV4f", "nc", "sse")
 
 TARGET_BUILTIN(__builtin_ia32_maskmovdqu, "vV16cV16cc*", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_movmskpd, "iV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pmovmskb128, "iV16c", "n", "sse2")
+TARGET_BUILTIN(__builtin_ia32_movmskpd, "iV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pmovmskb128, "iV16c", "nc", "sse2")
 TARGET_BUILTIN(__builtin_ia32_movnti, "vi*i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psadbw128, "V2LLiV16cV16c", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_sqrtpd, "V2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_sqrtsd, "V2dV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvtdq2ps, "V4fV4i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2dq, "V2LLiV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2ps, "V4fV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2dq, "V4iV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvtsd2si, "iV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvttsd2si, "iV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvtsd2ss, "V4fV4fV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvtps2dq, "V4iV4f", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvttps2dq, "V4iV4f", "n", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psadbw128, "V2LLiV16cV16c", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_sqrtpd, "V2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_sqrtsd, "V2dV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtdq2ps, "V4fV4i", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2dq, "V2LLiV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2ps, "V4fV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2dq, "V4iV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtsd2si, "iV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvttsd2si, "iV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtsd2ss, "V4fV4fV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtps2dq, "V4iV4f", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvttps2dq, "V4iV4f", "nc", "sse2")
 TARGET_BUILTIN(__builtin_ia32_clflush, "vvC*", "n", "sse2")
 TARGET_HEADER_BUILTIN(_mm_clflush, "vvC*", "nh", "emmintrin.h", ALL_LANGUAGES, "sse2")
 TARGET_BUILTIN(__builtin_ia32_lfence, "v", "n", "sse2")
@@ -339,179 +337,179 @@ TARGET_BUILTIN(__builtin_ia32_mfence, "v
 TARGET_HEADER_BUILTIN(_mm_mfence, "v", "nh", "emmintrin.h", ALL_LANGUAGES, "sse2")
 TARGET_BUILTIN(__builtin_ia32_pause, "v", "n", "")
 TARGET_HEADER_BUILTIN(_mm_pause, "v", "nh", "emmintrin.h", ALL_LANGUAGES, "")
-TARGET_BUILTIN(__builtin_ia32_pmuludq128, "V2LLiV4iV4i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psraw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psrad128, "V4iV4iV4i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psrlw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psrld128, "V4iV4iV4i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psrlq128, "V2LLiV2LLiV2LLi", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psllw128, "V8sV8sV8s", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pslld128, "V4iV4iV4i", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psllq128, "V2LLiV2LLiV2LLi", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psllwi128, "V8sV8si", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pslldi128, "V4iV4ii", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psllqi128, "V2LLiV2LLii", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psrlwi128, "V8sV8si", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psrldi128, "V4iV4ii", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psrlqi128, "V2LLiV2LLii", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psrawi128, "V8sV8si", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_psradi128, "V4iV4ii", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_pmaddwd128, "V4iV8sV8s", "n", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pmuludq128, "V2LLiV4iV4i", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psraw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psrad128, "V4iV4iV4i", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psrlw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psrld128, "V4iV4iV4i", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psrlq128, "V2LLiV2LLiV2LLi", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psllw128, "V8sV8sV8s", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pslld128, "V4iV4iV4i", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psllq128, "V2LLiV2LLiV2LLi", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psllwi128, "V8sV8si", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pslldi128, "V4iV4ii", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psllqi128, "V2LLiV2LLii", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psrlwi128, "V8sV8si", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psrldi128, "V4iV4ii", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psrlqi128, "V2LLiV2LLii", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psrawi128, "V8sV8si", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_psradi128, "V4iV4ii", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_pmaddwd128, "V4iV8sV8s", "nc", "sse2")
 
 TARGET_BUILTIN(__builtin_ia32_monitor, "vv*UiUi", "n", "sse3")
 TARGET_BUILTIN(__builtin_ia32_mwait, "vUiUi", "n", "sse3")
 TARGET_BUILTIN(__builtin_ia32_lddqu, "V16ccC*", "n", "sse3")
 
-TARGET_BUILTIN(__builtin_ia32_palignr128, "V16cV16cV16cIi", "n", "ssse3")
+TARGET_BUILTIN(__builtin_ia32_palignr128, "V16cV16cV16cIi", "nc", "ssse3")
 
-TARGET_BUILTIN(__builtin_ia32_insertps128, "V4fV4fV4fIc", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pblendvb128, "V16cV16cV16cV16c", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_blendvpd, "V2dV2dV2dV2d", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_blendvps, "V4fV4fV4fV4f", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_packusdw128, "V8sV4iV4i", "n", "sse4.1")
-
-TARGET_BUILTIN(__builtin_ia32_pmaxsb128, "V16cV16cV16c", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pmaxsd128, "V4iV4iV4i", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pmaxud128, "V4iV4iV4i", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pmaxuw128, "V8sV8sV8s", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pminsb128, "V16cV16cV16c", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pminsd128, "V4iV4iV4i", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pminud128, "V4iV4iV4i", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pminuw128, "V8sV8sV8s", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_pmuldq128, "V2LLiV4iV4i", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_roundps, "V4fV4fIi", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_roundss, "V4fV4fV4fIi", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_roundsd, "V2dV2dV2dIi", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_roundpd, "V2dV2dIi", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_dpps, "V4fV4fV4fIc", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_dppd, "V2dV2dV2dIc", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_ptestz128, "iV2LLiV2LLi", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_ptestc128, "iV2LLiV2LLi", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_ptestnzc128, "iV2LLiV2LLi", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_mpsadbw128, "V16cV16cV16cIc", "n", "sse4.1")
-TARGET_BUILTIN(__builtin_ia32_phminposuw128, "V8sV8s", "n", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_insertps128, "V4fV4fV4fIc", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pblendvb128, "V16cV16cV16cV16c", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_blendvpd, "V2dV2dV2dV2d", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_blendvps, "V4fV4fV4fV4f", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_packusdw128, "V8sV4iV4i", "nc", "sse4.1")
+
+TARGET_BUILTIN(__builtin_ia32_pmaxsb128, "V16cV16cV16c", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pmaxsd128, "V4iV4iV4i", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pmaxud128, "V4iV4iV4i", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pmaxuw128, "V8sV8sV8s", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pminsb128, "V16cV16cV16c", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pminsd128, "V4iV4iV4i", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pminud128, "V4iV4iV4i", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pminuw128, "V8sV8sV8s", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_pmuldq128, "V2LLiV4iV4i", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_roundps, "V4fV4fIi", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_roundss, "V4fV4fV4fIi", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_roundsd, "V2dV2dV2dIi", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_roundpd, "V2dV2dIi", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_dpps, "V4fV4fV4fIc", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_dppd, "V2dV2dV2dIc", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_ptestz128, "iV2LLiV2LLi", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_ptestc128, "iV2LLiV2LLi", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_ptestnzc128, "iV2LLiV2LLi", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_mpsadbw128, "V16cV16cV16cIc", "nc", "sse4.1")
+TARGET_BUILTIN(__builtin_ia32_phminposuw128, "V8sV8s", "nc", "sse4.1")
 
 // SSE 4.2
-TARGET_BUILTIN(__builtin_ia32_pcmpistrm128, "V16cV16cV16cIc", "n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpistri128, "iV16cV16cIc", "n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpestrm128, "V16cV16ciV16ciIc", "n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpestri128, "iV16ciV16ciIc","n", "sse4.2")
-
-TARGET_BUILTIN(__builtin_ia32_pcmpistria128, "iV16cV16cIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpistric128, "iV16cV16cIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpistrio128, "iV16cV16cIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpistris128, "iV16cV16cIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpistriz128, "iV16cV16cIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpestria128, "iV16ciV16ciIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpestric128, "iV16ciV16ciIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpestrio128, "iV16ciV16ciIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpestris128, "iV16ciV16ciIc","n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_pcmpestriz128, "iV16ciV16ciIc","n", "sse4.2")
-
-TARGET_BUILTIN(__builtin_ia32_crc32qi, "UiUiUc", "n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_crc32hi, "UiUiUs", "n", "sse4.2")
-TARGET_BUILTIN(__builtin_ia32_crc32si, "UiUiUi", "n", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpistrm128, "V16cV16cV16cIc", "nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpistri128, "iV16cV16cIc", "nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpestrm128, "V16cV16ciV16ciIc", "nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpestri128, "iV16ciV16ciIc","nc", "sse4.2")
+
+TARGET_BUILTIN(__builtin_ia32_pcmpistria128, "iV16cV16cIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpistric128, "iV16cV16cIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpistrio128, "iV16cV16cIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpistris128, "iV16cV16cIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpistriz128, "iV16cV16cIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpestria128, "iV16ciV16ciIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpestric128, "iV16ciV16ciIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpestrio128, "iV16ciV16ciIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpestris128, "iV16ciV16ciIc","nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_pcmpestriz128, "iV16ciV16ciIc","nc", "sse4.2")
+
+TARGET_BUILTIN(__builtin_ia32_crc32qi, "UiUiUc", "nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_crc32hi, "UiUiUs", "nc", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_crc32si, "UiUiUi", "nc", "sse4.2")
 
 // SSE4a
-TARGET_BUILTIN(__builtin_ia32_extrqi, "V2LLiV2LLiIcIc", "n", "sse4a")
-TARGET_BUILTIN(__builtin_ia32_extrq, "V2LLiV2LLiV16c", "n", "sse4a")
-TARGET_BUILTIN(__builtin_ia32_insertqi, "V2LLiV2LLiV2LLiIcIc", "n", "sse4a")
-TARGET_BUILTIN(__builtin_ia32_insertq, "V2LLiV2LLiV2LLi", "n", "sse4a")
+TARGET_BUILTIN(__builtin_ia32_extrqi, "V2LLiV2LLiIcIc", "nc", "sse4a")
+TARGET_BUILTIN(__builtin_ia32_extrq, "V2LLiV2LLiV16c", "nc", "sse4a")
+TARGET_BUILTIN(__builtin_ia32_insertqi, "V2LLiV2LLiV2LLiIcIc", "nc", "sse4a")
+TARGET_BUILTIN(__builtin_ia32_insertq, "V2LLiV2LLiV2LLi", "nc", "sse4a")
 TARGET_BUILTIN(__builtin_ia32_movntsd, "vd*V2d", "n", "sse4a")
 TARGET_BUILTIN(__builtin_ia32_movntss, "vf*V4f", "n", "sse4a")
 
 // AES
-TARGET_BUILTIN(__builtin_ia32_aesenc128, "V2LLiV2LLiV2LLi", "n", "aes")
-TARGET_BUILTIN(__builtin_ia32_aesenclast128, "V2LLiV2LLiV2LLi", "n", "aes")
-TARGET_BUILTIN(__builtin_ia32_aesdec128, "V2LLiV2LLiV2LLi", "n", "aes")
-TARGET_BUILTIN(__builtin_ia32_aesdeclast128, "V2LLiV2LLiV2LLi", "n", "aes")
-TARGET_BUILTIN(__builtin_ia32_aesimc128, "V2LLiV2LLi", "n", "aes")
-TARGET_BUILTIN(__builtin_ia32_aeskeygenassist128, "V2LLiV2LLiIc", "n", "aes")
+TARGET_BUILTIN(__builtin_ia32_aesenc128, "V2LLiV2LLiV2LLi", "nc", "aes")
+TARGET_BUILTIN(__builtin_ia32_aesenclast128, "V2LLiV2LLiV2LLi", "nc", "aes")
+TARGET_BUILTIN(__builtin_ia32_aesdec128, "V2LLiV2LLiV2LLi", "nc", "aes")
+TARGET_BUILTIN(__builtin_ia32_aesdeclast128, "V2LLiV2LLiV2LLi", "nc", "aes")
+TARGET_BUILTIN(__builtin_ia32_aesimc128, "V2LLiV2LLi", "nc", "aes")
+TARGET_BUILTIN(__builtin_ia32_aeskeygenassist128, "V2LLiV2LLiIc", "nc", "aes")
 
 // VAES
-TARGET_BUILTIN(__builtin_ia32_aesenc256, "V4LLiV4LLiV4LLi", "n", "vaes")
-TARGET_BUILTIN(__builtin_ia32_aesenc512, "V8LLiV8LLiV8LLi", "n", "avx512f,vaes")
-TARGET_BUILTIN(__builtin_ia32_aesenclast256, "V4LLiV4LLiV4LLi", "n", "vaes")
-TARGET_BUILTIN(__builtin_ia32_aesenclast512, "V8LLiV8LLiV8LLi", "n", "avx512f,vaes")
-TARGET_BUILTIN(__builtin_ia32_aesdec256, "V4LLiV4LLiV4LLi", "n", "vaes")
-TARGET_BUILTIN(__builtin_ia32_aesdec512, "V8LLiV8LLiV8LLi", "n", "avx512f,vaes")
-TARGET_BUILTIN(__builtin_ia32_aesdeclast256, "V4LLiV4LLiV4LLi", "n", "vaes")
-TARGET_BUILTIN(__builtin_ia32_aesdeclast512, "V8LLiV8LLiV8LLi", "n", "avx512f,vaes")
+TARGET_BUILTIN(__builtin_ia32_aesenc256, "V4LLiV4LLiV4LLi", "nc", "vaes")
+TARGET_BUILTIN(__builtin_ia32_aesenc512, "V8LLiV8LLiV8LLi", "nc", "avx512f,vaes")
+TARGET_BUILTIN(__builtin_ia32_aesenclast256, "V4LLiV4LLiV4LLi", "nc", "vaes")
+TARGET_BUILTIN(__builtin_ia32_aesenclast512, "V8LLiV8LLiV8LLi", "nc", "avx512f,vaes")
+TARGET_BUILTIN(__builtin_ia32_aesdec256, "V4LLiV4LLiV4LLi", "nc", "vaes")
+TARGET_BUILTIN(__builtin_ia32_aesdec512, "V8LLiV8LLiV8LLi", "nc", "avx512f,vaes")
+TARGET_BUILTIN(__builtin_ia32_aesdeclast256, "V4LLiV4LLiV4LLi", "nc", "vaes")
+TARGET_BUILTIN(__builtin_ia32_aesdeclast512, "V8LLiV8LLiV8LLi", "nc", "avx512f,vaes")
 
 // GFNI
-TARGET_BUILTIN(__builtin_ia32_vgf2p8affineinvqb_v16qi, "V16cV16cV16cIc", "n", "gfni")
-TARGET_BUILTIN(__builtin_ia32_vgf2p8affineinvqb_v32qi, "V32cV32cV32cIc", "n", "avx,gfni")
-TARGET_BUILTIN(__builtin_ia32_vgf2p8affineinvqb_v64qi, "V64cV64cV64cIc", "n", "avx512bw,gfni")
-TARGET_BUILTIN(__builtin_ia32_vgf2p8affineqb_v16qi, "V16cV16cV16cIc", "n", "gfni")
-TARGET_BUILTIN(__builtin_ia32_vgf2p8affineqb_v32qi, "V32cV32cV32cIc", "n", "avx,gfni")
-TARGET_BUILTIN(__builtin_ia32_vgf2p8affineqb_v64qi, "V64cV64cV64cIc", "n", "avx512bw,gfni")
-TARGET_BUILTIN(__builtin_ia32_vgf2p8mulb_v16qi, "V16cV16cV16c", "n", "gfni")
-TARGET_BUILTIN(__builtin_ia32_vgf2p8mulb_v32qi, "V32cV32cV32c", "n", "avx,gfni")
-TARGET_BUILTIN(__builtin_ia32_vgf2p8mulb_v64qi, "V64cV64cV64c", "n", "avx512bw,gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8affineinvqb_v16qi, "V16cV16cV16cIc", "nc", "gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8affineinvqb_v32qi, "V32cV32cV32cIc", "nc", "avx,gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8affineinvqb_v64qi, "V64cV64cV64cIc", "nc", "avx512bw,gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8affineqb_v16qi, "V16cV16cV16cIc", "nc", "gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8affineqb_v32qi, "V32cV32cV32cIc", "nc", "avx,gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8affineqb_v64qi, "V64cV64cV64cIc", "nc", "avx512bw,gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8mulb_v16qi, "V16cV16cV16c", "nc", "gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8mulb_v32qi, "V32cV32cV32c", "nc", "avx,gfni")
+TARGET_BUILTIN(__builtin_ia32_vgf2p8mulb_v64qi, "V64cV64cV64c", "nc", "avx512bw,gfni")
 
 // CLMUL
-TARGET_BUILTIN(__builtin_ia32_pclmulqdq128, "V2LLiV2LLiV2LLiIc", "n", "pclmul")
+TARGET_BUILTIN(__builtin_ia32_pclmulqdq128, "V2LLiV2LLiV2LLiIc", "nc", "pclmul")
 
 // VPCLMULQDQ
-TARGET_BUILTIN(__builtin_ia32_pclmulqdq256, "V4LLiV4LLiV4LLiIc", "n", "vpclmulqdq")
-TARGET_BUILTIN(__builtin_ia32_pclmulqdq512, "V8LLiV8LLiV8LLiIc", "n", "avx512f,vpclmulqdq")
+TARGET_BUILTIN(__builtin_ia32_pclmulqdq256, "V4LLiV4LLiV4LLiIc", "nc", "vpclmulqdq")
+TARGET_BUILTIN(__builtin_ia32_pclmulqdq512, "V8LLiV8LLiV8LLiIc", "nc", "avx512f,vpclmulqdq")
 
 // AVX
-TARGET_BUILTIN(__builtin_ia32_addsubpd256, "V4dV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_addsubps256, "V8fV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_haddpd256, "V4dV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_hsubps256, "V8fV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_hsubpd256, "V4dV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_haddps256, "V8fV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_maxpd256, "V4dV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_maxps256, "V8fV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_minpd256, "V4dV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_minps256, "V8fV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vpermilvarpd, "V2dV2dV2LLi", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vpermilvarps, "V4fV4fV4i", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vpermilvarpd256, "V4dV4dV4LLi", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vpermilvarps256, "V8fV8fV8i", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_blendvpd256, "V4dV4dV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_blendvps256, "V8fV8fV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_dpps256, "V8fV8fV8fIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cmppd, "V2dV2dV2dIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cmppd256, "V4dV4dV4dIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cmpps, "V4fV4fV4fIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cmpps256, "V8fV8fV8fIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cmpsd, "V2dV2dV2dIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cmpss, "V4fV4fV4fIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cvtdq2ps256, "V8fV8i", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2ps256, "V4fV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cvtps2dq256, "V8iV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2dq256, "V4iV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2dq256, "V4iV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_cvttps2dq256, "V8iV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vperm2f128_pd256, "V4dV4dV4dIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vperm2f128_ps256, "V8fV8fV8fIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vperm2f128_si256, "V8iV8iV8iIc", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_sqrtpd256, "V4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_sqrtps256, "V8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_rsqrtps256, "V8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_rcpps256, "V8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_roundpd256, "V4dV4dIi", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_roundps256, "V8fV8fIi", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestzpd, "iV2dV2d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestcpd, "iV2dV2d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestnzcpd, "iV2dV2d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestzps, "iV4fV4f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestcps, "iV4fV4f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestnzcps, "iV4fV4f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestzpd256, "iV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestcpd256, "iV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestnzcpd256, "iV4dV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestzps256, "iV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestcps256, "iV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_vtestnzcps256, "iV8fV8f", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_ptestz256, "iV4LLiV4LLi", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_ptestc256, "iV4LLiV4LLi", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_ptestnzc256, "iV4LLiV4LLi", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_movmskpd256, "iV4d", "n", "avx")
-TARGET_BUILTIN(__builtin_ia32_movmskps256, "iV8f", "n", "avx")
+TARGET_BUILTIN(__builtin_ia32_addsubpd256, "V4dV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_addsubps256, "V8fV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_haddpd256, "V4dV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_hsubps256, "V8fV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_hsubpd256, "V4dV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_haddps256, "V8fV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_maxpd256, "V4dV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_maxps256, "V8fV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_minpd256, "V4dV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_minps256, "V8fV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarpd, "V2dV2dV2LLi", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarps, "V4fV4fV4i", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarpd256, "V4dV4dV4LLi", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarps256, "V8fV8fV8i", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_blendvpd256, "V4dV4dV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_blendvps256, "V8fV8fV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_dpps256, "V8fV8fV8fIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cmppd, "V2dV2dV2dIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cmppd256, "V4dV4dV4dIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cmpps, "V4fV4fV4fIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cmpps256, "V8fV8fV8fIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cmpsd, "V2dV2dV2dIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cmpss, "V4fV4fV4fIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cvtdq2ps256, "V8fV8i", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2ps256, "V4fV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cvtps2dq256, "V8iV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2dq256, "V4iV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2dq256, "V4iV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_cvttps2dq256, "V8iV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vperm2f128_pd256, "V4dV4dV4dIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vperm2f128_ps256, "V8fV8fV8fIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vperm2f128_si256, "V8iV8iV8iIc", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_sqrtpd256, "V4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_sqrtps256, "V8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_rsqrtps256, "V8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_rcpps256, "V8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_roundpd256, "V4dV4dIi", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_roundps256, "V8fV8fIi", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestzpd, "iV2dV2d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestcpd, "iV2dV2d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestnzcpd, "iV2dV2d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestzps, "iV4fV4f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestcps, "iV4fV4f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestnzcps, "iV4fV4f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestzpd256, "iV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestcpd256, "iV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestnzcpd256, "iV4dV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestzps256, "iV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestcps256, "iV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_vtestnzcps256, "iV8fV8f", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_ptestz256, "iV4LLiV4LLi", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_ptestc256, "iV4LLiV4LLi", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_ptestnzc256, "iV4LLiV4LLi", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_movmskpd256, "iV4d", "nc", "avx")
+TARGET_BUILTIN(__builtin_ia32_movmskps256, "iV8f", "nc", "avx")
 TARGET_BUILTIN(__builtin_ia32_vzeroall, "v", "n", "avx")
 TARGET_BUILTIN(__builtin_ia32_vzeroupper, "v", "n", "avx")
 TARGET_BUILTIN(__builtin_ia32_vbroadcastf128_pd256, "V4dV2dC*", "n", "avx")
@@ -527,74 +525,74 @@ TARGET_BUILTIN(__builtin_ia32_maskstorep
 TARGET_BUILTIN(__builtin_ia32_maskstoreps256, "vV8f*V8iV8f", "n", "avx")
 
 // AVX2
-TARGET_BUILTIN(__builtin_ia32_mpsadbw256, "V32cV32cV32cIc", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pabsb256, "V32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pabsw256, "V16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pabsd256, "V8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_packsswb256, "V32cV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_packssdw256, "V16sV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_packuswb256, "V32cV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_packusdw256, "V16sV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_paddsb256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_paddsw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psubsb256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psubsw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_paddusb256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_paddusw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psubusb256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psubusw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_palignr256, "V32cV32cV32cIi", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pblendvb256, "V32cV32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_phaddw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_phaddd256, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_phaddsw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_phsubw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_phsubd256, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_phsubsw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmaddubsw256, "V16sV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmaddwd256, "V8iV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmaxub256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmaxuw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmaxud256, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmaxsb256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmaxsw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmaxsd256, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pminub256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pminuw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pminud256, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pminsb256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pminsw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pminsd256, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmovmskb256, "iV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmuldq256, "V4LLiV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmulhrsw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmulhuw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmulhw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pmuludq256, "V4LLiV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psadbw256, "V4LLiV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pshufb256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psignb256, "V32cV32cV32c", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psignw256, "V16sV16sV16s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psignd256, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psllwi256, "V16sV16si", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psllw256, "V16sV16sV8s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pslldi256, "V8iV8ii", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_pslld256, "V8iV8iV4i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psllqi256, "V4LLiV4LLii", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psllq256, "V4LLiV4LLiV2LLi", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrawi256, "V16sV16si", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psraw256, "V16sV16sV8s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psradi256, "V8iV8ii", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrad256, "V8iV8iV4i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrlwi256, "V16sV16si", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrlw256, "V16sV16sV8s", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrldi256, "V8iV8ii", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrld256, "V8iV8iV4i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrlqi256, "V4LLiV4LLii", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrlq256, "V4LLiV4LLiV2LLi", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_permvarsi256, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_permvarsf256, "V8fV8fV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_permti256, "V4LLiV4LLiV4LLiIc", "n", "avx2")
+TARGET_BUILTIN(__builtin_ia32_mpsadbw256, "V32cV32cV32cIc", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pabsb256, "V32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pabsw256, "V16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pabsd256, "V8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_packsswb256, "V32cV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_packssdw256, "V16sV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_packuswb256, "V32cV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_packusdw256, "V16sV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_paddsb256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_paddsw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psubsb256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psubsw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_paddusb256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_paddusw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psubusb256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psubusw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_palignr256, "V32cV32cV32cIi", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pblendvb256, "V32cV32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_phaddw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_phaddd256, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_phaddsw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_phsubw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_phsubd256, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_phsubsw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmaddubsw256, "V16sV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmaddwd256, "V8iV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmaxub256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmaxuw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmaxud256, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmaxsb256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmaxsw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmaxsd256, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pminub256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pminuw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pminud256, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pminsb256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pminsw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pminsd256, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmovmskb256, "iV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmuldq256, "V4LLiV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmulhrsw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmulhuw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmulhw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pmuludq256, "V4LLiV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psadbw256, "V4LLiV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pshufb256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psignb256, "V32cV32cV32c", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psignw256, "V16sV16sV16s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psignd256, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psllwi256, "V16sV16si", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psllw256, "V16sV16sV8s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pslldi256, "V8iV8ii", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_pslld256, "V8iV8iV4i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psllqi256, "V4LLiV4LLii", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psllq256, "V4LLiV4LLiV2LLi", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrawi256, "V16sV16si", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psraw256, "V16sV16sV8s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psradi256, "V8iV8ii", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrad256, "V8iV8iV4i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrlwi256, "V16sV16si", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrlw256, "V16sV16sV8s", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrldi256, "V8iV8ii", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrld256, "V8iV8iV4i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrlqi256, "V4LLiV4LLii", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrlq256, "V4LLiV4LLiV2LLi", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_permvarsi256, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_permvarsf256, "V8fV8fV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_permti256, "V4LLiV4LLiV4LLiIc", "nc", "avx2")
 TARGET_BUILTIN(__builtin_ia32_maskloadd256, "V8iV8iC*V8i", "n", "avx2")
 TARGET_BUILTIN(__builtin_ia32_maskloadq256, "V4LLiV4LLiC*V4LLi", "n", "avx2")
 TARGET_BUILTIN(__builtin_ia32_maskloadd, "V4iV4iC*V4i", "n", "avx2")
@@ -603,16 +601,16 @@ TARGET_BUILTIN(__builtin_ia32_maskstored
 TARGET_BUILTIN(__builtin_ia32_maskstoreq256, "vV4LLi*V4LLiV4LLi", "n", "avx2")
 TARGET_BUILTIN(__builtin_ia32_maskstored, "vV4i*V4iV4i", "n", "avx2")
 TARGET_BUILTIN(__builtin_ia32_maskstoreq, "vV2LLi*V2LLiV2LLi", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psllv8si, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psllv4si, "V4iV4iV4i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psllv4di, "V4LLiV4LLiV4LLi", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psllv2di, "V2LLiV2LLiV2LLi", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrav8si, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrav4si, "V4iV4iV4i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrlv8si, "V8iV8iV8i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrlv4si, "V4iV4iV4i", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrlv4di, "V4LLiV4LLiV4LLi", "n", "avx2")
-TARGET_BUILTIN(__builtin_ia32_psrlv2di, "V2LLiV2LLiV2LLi", "n", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psllv8si, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psllv4si, "V4iV4iV4i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psllv4di, "V4LLiV4LLiV4LLi", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psllv2di, "V2LLiV2LLiV2LLi", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrav8si, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrav4si, "V4iV4iV4i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrlv8si, "V8iV8iV8i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrlv4si, "V4iV4iV4i", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrlv4di, "V4LLiV4LLiV4LLi", "nc", "avx2")
+TARGET_BUILTIN(__builtin_ia32_psrlv2di, "V2LLiV2LLiV2LLi", "nc", "avx2")
 
 // GATHER
 TARGET_BUILTIN(__builtin_ia32_gatherd_pd, "V2dV2ddC*V4iV2dIc", "n", "avx2")
@@ -634,10 +632,10 @@ TARGET_BUILTIN(__builtin_ia32_gatherq_d,
 TARGET_BUILTIN(__builtin_ia32_gatherq_d256, "V4iV4iiC*V4LLiV4iIc", "n", "avx2")
 
 // F16C
-TARGET_BUILTIN(__builtin_ia32_vcvtps2ph, "V8sV4fIi", "n", "f16c")
-TARGET_BUILTIN(__builtin_ia32_vcvtps2ph256, "V8sV8fIi", "n", "f16c")
-TARGET_BUILTIN(__builtin_ia32_vcvtph2ps, "V4fV8s", "n", "f16c")
-TARGET_BUILTIN(__builtin_ia32_vcvtph2ps256, "V8fV8s", "n", "f16c")
+TARGET_BUILTIN(__builtin_ia32_vcvtps2ph, "V8sV4fIi", "nc", "f16c")
+TARGET_BUILTIN(__builtin_ia32_vcvtps2ph256, "V8sV8fIi", "nc", "f16c")
+TARGET_BUILTIN(__builtin_ia32_vcvtph2ps, "V4fV8s", "nc", "f16c")
+TARGET_BUILTIN(__builtin_ia32_vcvtph2ps256, "V8fV8s", "nc", "f16c")
 
 // RDRAND
 TARGET_BUILTIN(__builtin_ia32_rdrand16_step, "UiUs*", "n", "rdrnd")
@@ -685,15 +683,15 @@ TARGET_BUILTIN(__builtin_ia32_rdseed16_s
 TARGET_BUILTIN(__builtin_ia32_rdseed32_step, "UiUi*", "n", "rdseed")
 
 // BMI
-TARGET_BUILTIN(__builtin_ia32_bextr_u32, "UiUiUi", "n", "bmi")
+TARGET_BUILTIN(__builtin_ia32_bextr_u32, "UiUiUi", "nc", "bmi")
 
 // BMI2
-TARGET_BUILTIN(__builtin_ia32_bzhi_si, "UiUiUi", "n", "bmi2")
-TARGET_BUILTIN(__builtin_ia32_pdep_si, "UiUiUi", "n", "bmi2")
-TARGET_BUILTIN(__builtin_ia32_pext_si, "UiUiUi", "n", "bmi2")
+TARGET_BUILTIN(__builtin_ia32_bzhi_si, "UiUiUi", "nc", "bmi2")
+TARGET_BUILTIN(__builtin_ia32_pdep_si, "UiUiUi", "nc", "bmi2")
+TARGET_BUILTIN(__builtin_ia32_pext_si, "UiUiUi", "nc", "bmi2")
 
 // TBM
-TARGET_BUILTIN(__builtin_ia32_bextri_u32, "UiUiIUi", "n", "tbm")
+TARGET_BUILTIN(__builtin_ia32_bextri_u32, "UiUiIUi", "nc", "tbm")
 
 // LWP
 TARGET_BUILTIN(__builtin_ia32_llwpcb, "vv*", "n", "lwp")
@@ -702,184 +700,184 @@ TARGET_BUILTIN(__builtin_ia32_lwpins32,
 TARGET_BUILTIN(__builtin_ia32_lwpval32, "vUiUiUi", "n", "lwp")
 
 // SHA
-TARGET_BUILTIN(__builtin_ia32_sha1rnds4, "V4iV4iV4iIc", "n", "sha")
-TARGET_BUILTIN(__builtin_ia32_sha1nexte, "V4iV4iV4i", "n", "sha")
-TARGET_BUILTIN(__builtin_ia32_sha1msg1, "V4iV4iV4i", "n", "sha")
-TARGET_BUILTIN(__builtin_ia32_sha1msg2, "V4iV4iV4i", "n", "sha")
-TARGET_BUILTIN(__builtin_ia32_sha256rnds2, "V4iV4iV4iV4i", "n", "sha")
-TARGET_BUILTIN(__builtin_ia32_sha256msg1, "V4iV4iV4i", "n", "sha")
-TARGET_BUILTIN(__builtin_ia32_sha256msg2, "V4iV4iV4i", "n", "sha")
+TARGET_BUILTIN(__builtin_ia32_sha1rnds4, "V4iV4iV4iIc", "nc", "sha")
+TARGET_BUILTIN(__builtin_ia32_sha1nexte, "V4iV4iV4i", "nc", "sha")
+TARGET_BUILTIN(__builtin_ia32_sha1msg1, "V4iV4iV4i", "nc", "sha")
+TARGET_BUILTIN(__builtin_ia32_sha1msg2, "V4iV4iV4i", "nc", "sha")
+TARGET_BUILTIN(__builtin_ia32_sha256rnds2, "V4iV4iV4iV4i", "nc", "sha")
+TARGET_BUILTIN(__builtin_ia32_sha256msg1, "V4iV4iV4i", "nc", "sha")
+TARGET_BUILTIN(__builtin_ia32_sha256msg2, "V4iV4iV4i", "nc", "sha")
 
 // FMA
-TARGET_BUILTIN(__builtin_ia32_vfmaddps, "V4fV4fV4fV4f", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd, "V2dV2dV2dV2d", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddss3, "V4fV4fV4fV4f", "n", "fma")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsd3, "V2dV2dV2dV2d", "n", "fma")
-TARGET_BUILTIN(__builtin_ia32_vfmaddss, "V4fV4fV4fV4f", "n", "fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsd, "V2dV2dV2dV2d", "n", "fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps, "V4fV4fV4fV4f", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd, "V2dV2dV2dV2d", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddps256, "V8fV8fV8fV8f", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd256, "V4dV4dV4dV4d", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfnmaddps256, "V8fV8fV8fV8f", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfnmaddpd256, "V4dV4dV4dV4d", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256, "V8fV8fV8fV8f", "n", "fma|fma4")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256, "V4dV4dV4dV4d", "n", "fma|fma4")
-
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd128_mask, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd128_mask3, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd128_maskz, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_mask, "V4dV4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_mask3, "V4dV4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_maskz, "V4dV4dV4dV4dUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd512_mask3, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddpd512_maskz, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfmaddps128_mask, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddps128_mask3, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddps128_maskz, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddps256_mask, "V8fV8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddps256_mask3, "V8fV8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddps256_maskz, "V8fV8fV8fV8fUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfmaddps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddps512_mask3, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddps512_maskz, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd128_mask, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd128_mask3, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd128_maskz, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_mask, "V4dV4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_mask3, "V4dV4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_maskz, "V4dV4dV4dV4dUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd512_mask3, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd512_maskz, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps128_mask, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps128_mask3, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps128_maskz, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_mask, "V8fV8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_mask3, "V8fV8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_maskz, "V8fV8fV8fV8fUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps512_mask3, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsubps512_maskz, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfmsubpd128_mask3, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmsubpd256_mask3, "V4dV4dV4dV4dUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfmsubpd512_mask3, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfmsubps128_mask3, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmsubps256_mask3, "V8fV8fV8fV8fUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfmsubps512_mask3, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfmsubaddpd128_mask3, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmsubaddpd256_mask3, "V4dV4dV4dV4dUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfmsubaddpd512_mask3, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfmsubaddps128_mask3, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfmsubaddps256_mask3, "V8fV8fV8fV8fUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfmsubaddps512_mask3, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfnmaddpd128_mask, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfnmaddpd256_mask, "V4dV4dV4dV4dUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfnmaddpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfnmaddps128_mask, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfnmaddps256_mask, "V8fV8fV8fV8fUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfnmaddps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfnmsubpd128_mask, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubpd128_mask3, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubpd256_mask, "V4dV4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubpd256_mask3, "V4dV4dV4dV4dUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vfnmsubpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubpd512_mask3, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vfnmsubps128_mask, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubps128_mask3, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubps256_mask, "V8fV8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubps256_mask3, "V8fV8fV8fV8fUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps, "V4fV4fV4fV4f", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd, "V2dV2dV2dV2d", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddss3, "V4fV4fV4fV4f", "nc", "fma")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsd3, "V2dV2dV2dV2d", "nc", "fma")
+TARGET_BUILTIN(__builtin_ia32_vfmaddss, "V4fV4fV4fV4f", "nc", "fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsd, "V2dV2dV2dV2d", "nc", "fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps, "V4fV4fV4fV4f", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd, "V2dV2dV2dV2d", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps256, "V8fV8fV8fV8f", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd256, "V4dV4dV4dV4d", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfnmaddps256, "V8fV8fV8fV8f", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfnmaddpd256, "V4dV4dV4dV4d", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256, "V8fV8fV8fV8f", "nc", "fma|fma4")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256, "V4dV4dV4dV4d", "nc", "fma|fma4")
+
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd128_mask, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd128_mask3, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd128_maskz, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_mask, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_mask3, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_maskz, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd512_mask3, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddpd512_maskz, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfmaddps128_mask, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps128_mask3, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps128_maskz, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps256_mask, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps256_mask3, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps256_maskz, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfmaddps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps512_mask3, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddps512_maskz, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd128_mask, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd128_mask3, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd128_maskz, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_mask, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_mask3, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_maskz, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd512_mask3, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd512_maskz, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps128_mask, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps128_mask3, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps128_maskz, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_mask, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_mask3, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_maskz, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps512_mask3, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsubps512_maskz, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfmsubpd128_mask3, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmsubpd256_mask3, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfmsubpd512_mask3, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfmsubps128_mask3, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmsubps256_mask3, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfmsubps512_mask3, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfmsubaddpd128_mask3, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmsubaddpd256_mask3, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfmsubaddpd512_mask3, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfmsubaddps128_mask3, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfmsubaddps256_mask3, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfmsubaddps512_mask3, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfnmaddpd128_mask, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfnmaddpd256_mask, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfnmaddpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfnmaddps128_mask, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfnmaddps256_mask, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfnmaddps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfnmsubpd128_mask, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubpd128_mask3, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubpd256_mask, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubpd256_mask3, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vfnmsubpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubpd512_mask3, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vfnmsubps128_mask, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubps128_mask3, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubps256_mask, "V8fV8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubps256_mask3, "V8fV8fV8fV8fUc", "nc", "avx512vl")
 
-TARGET_BUILTIN(__builtin_ia32_vfnmsubps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubps512_mask3, "V16fV16fV16fV16fUsIi", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubps512_mask3, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
 
 // XOP
-TARGET_BUILTIN(__builtin_ia32_vpmacssww, "V8sV8sV8sV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacsww, "V8sV8sV8sV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacsswd, "V4iV8sV8sV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacswd, "V4iV8sV8sV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacssdd, "V4iV4iV4iV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacsdd, "V4iV4iV4iV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacssdql, "V2LLiV4iV4iV2LLi", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacsdql, "V2LLiV4iV4iV2LLi", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacssdqh, "V2LLiV4iV4iV2LLi", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmacsdqh, "V2LLiV4iV4iV2LLi", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmadcsswd, "V4iV8sV8sV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpmadcswd, "V4iV8sV8sV4i", "n", "xop")
-
-TARGET_BUILTIN(__builtin_ia32_vphaddbw, "V8sV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphaddbd, "V4iV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphaddbq, "V2LLiV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphaddwd, "V4iV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphaddwq, "V2LLiV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphadddq, "V2LLiV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphaddubw, "V8sV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphaddubd, "V4iV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphaddubq, "V2LLiV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphadduwd, "V4iV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphadduwq, "V2LLiV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphaddudq, "V2LLiV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphsubbw, "V8sV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphsubwd, "V4iV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vphsubdq, "V2LLiV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpperm, "V16cV16cV16cV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vprotb, "V16cV16cV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vprotw, "V8sV8sV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vprotd, "V4iV4iV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vprotq, "V2LLiV2LLiV2LLi", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vprotbi, "V16cV16cIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vprotwi, "V8sV8sIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vprotdi, "V4iV4iIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vprotqi, "V2LLiV2LLiIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpshlb, "V16cV16cV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpshlw, "V8sV8sV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpshld, "V4iV4iV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpshlq, "V2LLiV2LLiV2LLi", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpshab, "V16cV16cV16c", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpshaw, "V8sV8sV8s", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpshad, "V4iV4iV4i", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpshaq, "V2LLiV2LLiV2LLi", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpcomub, "V16cV16cV16cIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpcomuw, "V8sV8sV8sIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpcomud, "V4iV4iV4iIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpcomuq, "V2LLiV2LLiV2LLiIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpcomb, "V16cV16cV16cIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpcomw, "V8sV8sV8sIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpcomd, "V4iV4iV4iIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpcomq, "V2LLiV2LLiV2LLiIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpermil2pd, "V2dV2dV2dV2LLiIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpermil2pd256, "V4dV4dV4dV4LLiIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpermil2ps, "V4fV4fV4fV4iIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vpermil2ps256, "V8fV8fV8fV8iIc", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vfrczss, "V4fV4f", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vfrczsd, "V2dV2d", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vfrczps, "V4fV4f", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vfrczpd, "V2dV2d", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vfrczps256, "V8fV8f", "n", "xop")
-TARGET_BUILTIN(__builtin_ia32_vfrczpd256, "V4dV4d", "n", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacssww, "V8sV8sV8sV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacsww, "V8sV8sV8sV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacsswd, "V4iV8sV8sV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacswd, "V4iV8sV8sV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacssdd, "V4iV4iV4iV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacsdd, "V4iV4iV4iV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacssdql, "V2LLiV4iV4iV2LLi", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacsdql, "V2LLiV4iV4iV2LLi", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacssdqh, "V2LLiV4iV4iV2LLi", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmacsdqh, "V2LLiV4iV4iV2LLi", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmadcsswd, "V4iV8sV8sV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpmadcswd, "V4iV8sV8sV4i", "nc", "xop")
+
+TARGET_BUILTIN(__builtin_ia32_vphaddbw, "V8sV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphaddbd, "V4iV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphaddbq, "V2LLiV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphaddwd, "V4iV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphaddwq, "V2LLiV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphadddq, "V2LLiV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphaddubw, "V8sV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphaddubd, "V4iV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphaddubq, "V2LLiV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphadduwd, "V4iV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphadduwq, "V2LLiV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphaddudq, "V2LLiV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphsubbw, "V8sV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphsubwd, "V4iV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vphsubdq, "V2LLiV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpperm, "V16cV16cV16cV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vprotb, "V16cV16cV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vprotw, "V8sV8sV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vprotd, "V4iV4iV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vprotq, "V2LLiV2LLiV2LLi", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vprotbi, "V16cV16cIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vprotwi, "V8sV8sIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vprotdi, "V4iV4iIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vprotqi, "V2LLiV2LLiIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpshlb, "V16cV16cV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpshlw, "V8sV8sV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpshld, "V4iV4iV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpshlq, "V2LLiV2LLiV2LLi", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpshab, "V16cV16cV16c", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpshaw, "V8sV8sV8s", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpshad, "V4iV4iV4i", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpshaq, "V2LLiV2LLiV2LLi", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpcomub, "V16cV16cV16cIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpcomuw, "V8sV8sV8sIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpcomud, "V4iV4iV4iIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpcomuq, "V2LLiV2LLiV2LLiIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpcomb, "V16cV16cV16cIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpcomw, "V8sV8sV8sIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpcomd, "V4iV4iV4iIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpcomq, "V2LLiV2LLiV2LLiIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpermil2pd, "V2dV2dV2dV2LLiIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpermil2pd256, "V4dV4dV4dV4LLiIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpermil2ps, "V4fV4fV4fV4iIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vpermil2ps256, "V8fV8fV8fV8iIc", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vfrczss, "V4fV4f", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vfrczsd, "V2dV2d", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vfrczps, "V4fV4f", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vfrczpd, "V2dV2d", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vfrczps256, "V8fV8f", "nc", "xop")
+TARGET_BUILTIN(__builtin_ia32_vfrczpd256, "V4dV4d", "nc", "xop")
 
 TARGET_BUILTIN(__builtin_ia32_xbegin, "i", "n", "rtm")
 TARGET_BUILTIN(__builtin_ia32_xend, "v", "n", "rtm")
@@ -898,69 +896,69 @@ TARGET_BUILTIN(__builtin_ia32_rdpkru, "U
 TARGET_BUILTIN(__builtin_ia32_wrpkru, "vUi", "n", "pku")
 
 // AVX-512
-TARGET_BUILTIN(__builtin_ia32_sqrtpd512_mask, "V8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_sqrtps512_mask, "V16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14sd_mask, "V2dV2dV2dV2dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14ss_mask, "V4fV4fV4fV4fUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14pd512_mask, "V8dV8dV8dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14ps512_mask, "V16fV16fV16fUs", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_rsqrt28sd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_rsqrt28ss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_rsqrt28pd_mask, "V8dV8dV8dUcIi", "n", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_rsqrt28ps_mask, "V16fV16fV16fUsIi", "n", "avx512er")
-
-TARGET_BUILTIN(__builtin_ia32_rcp14sd_mask, "V2dV2dV2dV2dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rcp14ss_mask, "V4fV4fV4fV4fUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rcp14pd512_mask, "V8dV8dV8dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rcp14ps512_mask, "V16fV16fV16fUs", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_rcp28sd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_rcp28ss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_rcp28pd_mask, "V8dV8dV8dUcIi", "n", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_rcp28ps_mask, "V16fV16fV16fUsIi", "n", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_exp2pd_mask, "V8dV8dV8dUcIi", "n", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_exp2ps_mask, "V16fV16fV16fUsIi", "n", "avx512er")
-
-TARGET_BUILTIN(__builtin_ia32_cvttps2dq512_mask, "V16iV16fV16iUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvttps2udq512_mask, "V16iV16fV16iUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2dq512_mask, "V8iV8dV8iUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2udq512_mask, "V8iV8dV8iUcIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_cmpps512_mask,   "UsV16fV16fIiUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cmpps256_mask,   "UcV8fV8fIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cmpps128_mask,   "UcV4fV4fIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cmppd512_mask, "UcV8dV8dIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cmppd256_mask, "UcV4dV4dIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cmppd128_mask, "UcV2dV2dIiUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_rndscaleps_mask, "V16fV16fIiV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rndscalepd_mask, "V8dV8dIiV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtps2dq512_mask, "V16iV16fV16iUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2dq512_mask, "V8iV8dV8iUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtps2udq512_mask, "V16iV16fV16iUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2udq512_mask, "V8iV8dV8iUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_minps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_minpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_maxps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_maxpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtdq2ps512_mask, "V16fV16iV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtudq2ps512_mask, "V16fV16iV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2ps512_mask, "V8fV8dV8fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtps2ph512_mask, "V16sV16fIiV16sUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtph2ps512_mask, "V16fV16sV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pabsd512_mask, "V16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pabsq512_mask, "V8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmaxsd512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmaxsq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmaxud512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmaxuq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pminsd512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pminsq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pminud512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pminuq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmuldq512, "V8LLiV16iV16i", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmuludq512, "V8LLiV16iV16i", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_sqrtpd512_mask, "V8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_sqrtps512_mask, "V16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14sd_mask, "V2dV2dV2dV2dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14ss_mask, "V4fV4fV4fV4fUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14pd512_mask, "V8dV8dV8dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14ps512_mask, "V16fV16fV16fUs", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_rsqrt28sd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rsqrt28ss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rsqrt28pd_mask, "V8dV8dV8dUcIi", "nc", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rsqrt28ps_mask, "V16fV16fV16fUsIi", "nc", "avx512er")
+
+TARGET_BUILTIN(__builtin_ia32_rcp14sd_mask, "V2dV2dV2dV2dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rcp14ss_mask, "V4fV4fV4fV4fUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rcp14pd512_mask, "V8dV8dV8dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rcp14ps512_mask, "V16fV16fV16fUs", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_rcp28sd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rcp28ss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rcp28pd_mask, "V8dV8dV8dUcIi", "nc", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rcp28ps_mask, "V16fV16fV16fUsIi", "nc", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_exp2pd_mask, "V8dV8dV8dUcIi", "nc", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_exp2ps_mask, "V16fV16fV16fUsIi", "nc", "avx512er")
+
+TARGET_BUILTIN(__builtin_ia32_cvttps2dq512_mask, "V16iV16fV16iUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvttps2udq512_mask, "V16iV16fV16iUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2dq512_mask, "V8iV8dV8iUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2udq512_mask, "V8iV8dV8iUcIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_cmpps512_mask,   "UsV16fV16fIiUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cmpps256_mask,   "UcV8fV8fIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cmpps128_mask,   "UcV4fV4fIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cmppd512_mask, "UcV8dV8dIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cmppd256_mask, "UcV4dV4dIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cmppd128_mask, "UcV2dV2dIiUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_rndscaleps_mask, "V16fV16fIiV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rndscalepd_mask, "V8dV8dIiV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtps2dq512_mask, "V16iV16fV16iUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2dq512_mask, "V8iV8dV8iUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtps2udq512_mask, "V16iV16fV16iUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2udq512_mask, "V8iV8dV8iUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_minps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_minpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_maxps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_maxpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtdq2ps512_mask, "V16fV16iV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtudq2ps512_mask, "V16fV16iV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2ps512_mask, "V8fV8dV8fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtps2ph512_mask, "V16sV16fIiV16sUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtph2ps512_mask, "V16fV16sV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pabsd512_mask, "V16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pabsq512_mask, "V8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmaxsd512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmaxsq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmaxud512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmaxuq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pminsd512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pminsq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pminud512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pminuq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmuldq512, "V8LLiV16iV16i", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmuludq512, "V8LLiV16iV16i", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_loaddqusi512_mask, "V16iiC*V16iUs", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_loaddqudi512_mask, "V8LLiLLiC*V8LLiUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_loadups512_mask, "V16ffC*V16fUs", "n", "avx512f")
@@ -973,35 +971,35 @@ TARGET_BUILTIN(__builtin_ia32_storeupd51
 TARGET_BUILTIN(__builtin_ia32_storeapd512_mask, "vV8d*V8dUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_storeups512_mask, "vf*V16fUs", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_storeaps512_mask, "vV16f*V16fUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermt2vard512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varps512_mask, "V16fV16iV16fV16fUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varpd512_mask, "V8dV8LLiV8dV8dUc", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_vpdpbusd128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusd256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusd512_mask, "V16iV16iV16iV16iUs", "n", "avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusds128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusds256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusds512_mask, "V16iV16iV16iV16iUs", "n", "avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssd128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssd256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssd512_mask, "V16iV16iV16iV16iUs", "n", "avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssds128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssds256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssds512_mask, "V16iV16iV16iV16iUs", "n", "avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusd128_maskz, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusd256_maskz, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusd512_maskz, "V16iV16iV16iV16iUs", "n", "avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusds128_maskz, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusds256_maskz, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpbusds512_maskz, "V16iV16iV16iV16iUs", "n", "avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssd128_maskz, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssd256_maskz, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssd512_maskz, "V16iV16iV16iV16iUs", "n", "avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssds128_maskz, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssds256_maskz, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vnni")
-TARGET_BUILTIN(__builtin_ia32_vpdpwssds512_maskz, "V16iV16iV16iV16iUs", "n", "avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpermt2vard512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varps512_mask, "V16fV16iV16fV16fUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varpd512_mask, "V8dV8LLiV8dV8dUc", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_vpdpbusd128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusd256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusd512_mask, "V16iV16iV16iV16iUs", "nc", "avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusds128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusds256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusds512_mask, "V16iV16iV16iV16iUs", "nc", "avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssd128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssd256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssd512_mask, "V16iV16iV16iV16iUs", "nc", "avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssds128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssds256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssds512_mask, "V16iV16iV16iV16iUs", "nc", "avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusd128_maskz, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusd256_maskz, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusd512_maskz, "V16iV16iV16iV16iUs", "nc", "avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusds128_maskz, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusds256_maskz, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpbusds512_maskz, "V16iV16iV16iV16iUs", "nc", "avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssd128_maskz, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssd256_maskz, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssd512_maskz, "V16iV16iV16iV16iUs", "nc", "avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssds128_maskz, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssds256_maskz, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vnni")
+TARGET_BUILTIN(__builtin_ia32_vpdpwssds512_maskz, "V16iV16iV16iV16iUs", "nc", "avx512vnni")
 
 TARGET_BUILTIN(__builtin_ia32_gather3div2df, "V2dV2ddC*V2LLiUcIi", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_gather3div2di, "V2LLiV2LLiLLiC*V2LLiUcIi", "n", "avx512vl")
@@ -1045,138 +1043,138 @@ TARGET_BUILTIN(__builtin_ia32_scatterpfd
 TARGET_BUILTIN(__builtin_ia32_scatterpfqpd, "vUcV8LLiLLi*IiIi", "n", "avx512pf")
 TARGET_BUILTIN(__builtin_ia32_scatterpfqps, "vUcV8LLii*IiIi", "n", "avx512pf")
 
-TARGET_BUILTIN(__builtin_ia32_knothi, "UsUs", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_knothi, "UsUs", "nc", "avx512f")
 
-TARGET_BUILTIN(__builtin_ia32_cmpb128_mask, "UsV16cV16cIiUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cmpd128_mask, "UcV4iV4iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cmpq128_mask, "UcV2LLiV2LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cmpw128_mask, "UcV8sV8sIiUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cmpb256_mask, "UiV32cV32cIiUi", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cmpd256_mask, "UcV8iV8iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cmpq256_mask, "UcV4LLiV4LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cmpw256_mask, "UsV16sV16sIiUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cmpb512_mask, "ULLiV64cV64cIiULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cmpd512_mask, "UsV16iV16iIiUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cmpq512_mask, "UcV8LLiV8LLiIiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cmpw512_mask, "UiV32sV32sIiUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_ucmpb128_mask, "UsV16cV16cIiUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_ucmpd128_mask, "UcV4iV4iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_ucmpq128_mask, "UcV2LLiV2LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_ucmpw128_mask, "UcV8sV8sIiUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_ucmpb256_mask, "UiV32cV32cIiUi", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_ucmpd256_mask, "UcV8iV8iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_ucmpq256_mask, "UcV4LLiV4LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_ucmpw256_mask, "UsV16sV16sIiUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_ucmpb512_mask, "ULLiV64cV64cIiULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_ucmpd512_mask, "UsV16iV16iIiUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_ucmpq512_mask, "UcV8LLiV8LLiIiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_ucmpw512_mask, "UiV32sV32sIiUi", "n", "avx512bw")
-
-TARGET_BUILTIN(__builtin_ia32_pabsb512_mask, "V64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pabsw512_mask, "V32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_packssdw512, "V32sV16iV16i", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_packsswb512, "V64cV32sV32s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_packusdw512, "V32sV16iV16i", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_packuswb512, "V64cV32sV32s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_paddsb512_mask, "V64cV64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_paddsw512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_paddusb512_mask, "V64cV64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_paddusw512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmaxsb512_mask, "V64cV64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmaxsw512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmaxub512_mask, "V64cV64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmaxuw512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pminsb512_mask, "V64cV64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pminsw512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pminub512_mask, "V64cV64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pminuw512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pshufb512, "V64cV64cV64c", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psubsb512_mask, "V64cV64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psubsw512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psubusb512_mask, "V64cV64cV64cV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psubusw512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-
-TARGET_BUILTIN(__builtin_ia32_vpermi2varhi512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varhi512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varhi512_maskz, "V32sV32sV32sV32sUi", "n", "avx512bw")
-
-TARGET_BUILTIN(__builtin_ia32_vpconflictdi_128_mask, "V2LLiV2LLiV2LLiUc", "n", "avx512cd,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpconflictdi_256_mask, "V4LLiV4LLiV4LLiUc", "n", "avx512cd,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpconflictsi_128_mask, "V4iV4iV4iUc", "n", "avx512cd,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpconflictsi_256_mask, "V8iV8iV8iUc", "n", "avx512cd,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpconflictdi_512_mask, "V8LLiV8LLiV8LLiUc", "n", "avx512cd")
-TARGET_BUILTIN(__builtin_ia32_vpconflictsi_512_mask, "V16iV16iV16iUs", "n", "avx512cd")
-TARGET_BUILTIN(__builtin_ia32_vplzcntd_512_mask, "V16iV16iV16iUs", "n", "avx512cd")
-TARGET_BUILTIN(__builtin_ia32_vplzcntq_512_mask, "V8LLiV8LLiV8LLiUc", "n", "avx512cd")
-
-TARGET_BUILTIN(__builtin_ia32_vpopcntd_128, "V4iV4i", "n", "avx512vpopcntdq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpopcntq_128, "V2LLiV2LLi", "n", "avx512vpopcntdq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpopcntd_256, "V8iV8i", "n", "avx512vpopcntdq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpopcntq_256, "V4LLiV4LLi", "n", "avx512vpopcntdq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpopcntd_512, "V16iV16i", "n", "avx512vpopcntdq")
-TARGET_BUILTIN(__builtin_ia32_vpopcntq_512, "V8LLiV8LLi", "n", "avx512vpopcntdq")
-
-TARGET_BUILTIN(__builtin_ia32_vpopcntb_128, "V16cV16c", "n", "avx512vl,avx512bitalg")
-TARGET_BUILTIN(__builtin_ia32_vpopcntw_128, "V8sV8s", "n", "avx512vl,avx512bitalg")
-TARGET_BUILTIN(__builtin_ia32_vpopcntb_256, "V32cV32c", "n", "avx512vl,avx512bitalg")
-TARGET_BUILTIN(__builtin_ia32_vpopcntw_256, "V16sV16s", "n", "avx512vl,avx512bitalg")
-TARGET_BUILTIN(__builtin_ia32_vpopcntb_512, "V64cV64c", "n", "avx512bitalg")
-TARGET_BUILTIN(__builtin_ia32_vpopcntw_512, "V32sV32s", "n", "avx512bitalg")
-
-TARGET_BUILTIN(__builtin_ia32_vpshufbitqmb128_mask, "UsV16cV16cUs", "n", "avx512vl,avx512bitalg")
-TARGET_BUILTIN(__builtin_ia32_vpshufbitqmb256_mask, "UiV32cV32cUi", "n", "avx512vl,avx512bitalg")
-TARGET_BUILTIN(__builtin_ia32_vpshufbitqmb512_mask, "ULLiV64cV64cULLi", "n", "avx512bitalg")
-
-TARGET_BUILTIN(__builtin_ia32_vpermi2varhi128_mask, "V8sV8sV8sV8sUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varhi256_mask, "V16sV16sV16sV16sUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varhi128_mask, "V8sV8sV8sV8sUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varhi128_maskz, "V8sV8sV8sV8sUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varhi256_mask, "V16sV16sV16sV16sUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varhi256_maskz, "V16sV16sV16sV16sUs", "n", "avx512vl,avx512bw")
-
-TARGET_BUILTIN(__builtin_ia32_pmulhrsw512, "V32sV32sV32s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmulhuw512, "V32sV32sV32s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmulhw512, "V32sV32sV32s", "n", "avx512bw")
-
-TARGET_BUILTIN(__builtin_ia32_addpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_addps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_divpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_divps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_mulpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_mulps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_subpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_subps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_pmaddubsw512, "V32sV64cV64c", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmaddwd512, "V16iV32sV32s", "n", "avx512bw")
-
-TARGET_BUILTIN(__builtin_ia32_addss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_divss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_mulss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_subss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_maxss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_minss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_addsd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_divsd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_mulsd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_subsd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_maxsd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_minsd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-
-TARGET_BUILTIN(__builtin_ia32_compressdf128_mask, "V2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_compressdf256_mask, "V4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_compressdi128_mask, "V2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_compressdi256_mask, "V4LLiV4LLiV4LLiUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_compresshi128_mask, "V8sV8sV8sUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_compresshi256_mask, "V16sV16sV16sUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_compressqi128_mask, "V16cV16cV16cUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_compressqi256_mask, "V32cV32cV32cUi", "n", "avx512vl,avx512vbmi2")
-
-TARGET_BUILTIN(__builtin_ia32_compresssf128_mask, "V4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_compresssf256_mask, "V8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_compresssi128_mask, "V4iV4iV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_compresssi256_mask, "V8iV8iV8iUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cmpb128_mask, "UsV16cV16cIiUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cmpd128_mask, "UcV4iV4iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cmpq128_mask, "UcV2LLiV2LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cmpw128_mask, "UcV8sV8sIiUc", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cmpb256_mask, "UiV32cV32cIiUi", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cmpd256_mask, "UcV8iV8iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cmpq256_mask, "UcV4LLiV4LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cmpw256_mask, "UsV16sV16sIiUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cmpb512_mask, "ULLiV64cV64cIiULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cmpd512_mask, "UsV16iV16iIiUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cmpq512_mask, "UcV8LLiV8LLiIiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cmpw512_mask, "UiV32sV32sIiUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ucmpb128_mask, "UsV16cV16cIiUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ucmpd128_mask, "UcV4iV4iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ucmpq128_mask, "UcV2LLiV2LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ucmpw128_mask, "UcV8sV8sIiUc", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ucmpb256_mask, "UiV32cV32cIiUi", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ucmpd256_mask, "UcV8iV8iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ucmpq256_mask, "UcV4LLiV4LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_ucmpw256_mask, "UsV16sV16sIiUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ucmpb512_mask, "ULLiV64cV64cIiULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_ucmpd512_mask, "UsV16iV16iIiUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_ucmpq512_mask, "UcV8LLiV8LLiIiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_ucmpw512_mask, "UiV32sV32sIiUi", "nc", "avx512bw")
+
+TARGET_BUILTIN(__builtin_ia32_pabsb512_mask, "V64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pabsw512_mask, "V32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_packssdw512, "V32sV16iV16i", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_packsswb512, "V64cV32sV32s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_packusdw512, "V32sV16iV16i", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_packuswb512, "V64cV32sV32s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_paddsb512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_paddsw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_paddusb512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_paddusw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmaxsb512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmaxsw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmaxub512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmaxuw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pminsb512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pminsw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pminub512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pminuw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pshufb512, "V64cV64cV64c", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psubsb512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psubsw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psubusb512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psubusw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+
+TARGET_BUILTIN(__builtin_ia32_vpermi2varhi512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varhi512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varhi512_maskz, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+
+TARGET_BUILTIN(__builtin_ia32_vpconflictdi_128_mask, "V2LLiV2LLiV2LLiUc", "nc", "avx512cd,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpconflictdi_256_mask, "V4LLiV4LLiV4LLiUc", "nc", "avx512cd,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpconflictsi_128_mask, "V4iV4iV4iUc", "nc", "avx512cd,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpconflictsi_256_mask, "V8iV8iV8iUc", "nc", "avx512cd,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpconflictdi_512_mask, "V8LLiV8LLiV8LLiUc", "nc", "avx512cd")
+TARGET_BUILTIN(__builtin_ia32_vpconflictsi_512_mask, "V16iV16iV16iUs", "nc", "avx512cd")
+TARGET_BUILTIN(__builtin_ia32_vplzcntd_512_mask, "V16iV16iV16iUs", "nc", "avx512cd")
+TARGET_BUILTIN(__builtin_ia32_vplzcntq_512_mask, "V8LLiV8LLiV8LLiUc", "nc", "avx512cd")
+
+TARGET_BUILTIN(__builtin_ia32_vpopcntd_128, "V4iV4i", "nc", "avx512vpopcntdq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpopcntq_128, "V2LLiV2LLi", "nc", "avx512vpopcntdq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpopcntd_256, "V8iV8i", "nc", "avx512vpopcntdq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpopcntq_256, "V4LLiV4LLi", "nc", "avx512vpopcntdq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpopcntd_512, "V16iV16i", "nc", "avx512vpopcntdq")
+TARGET_BUILTIN(__builtin_ia32_vpopcntq_512, "V8LLiV8LLi", "nc", "avx512vpopcntdq")
+
+TARGET_BUILTIN(__builtin_ia32_vpopcntb_128, "V16cV16c", "nc", "avx512vl,avx512bitalg")
+TARGET_BUILTIN(__builtin_ia32_vpopcntw_128, "V8sV8s", "nc", "avx512vl,avx512bitalg")
+TARGET_BUILTIN(__builtin_ia32_vpopcntb_256, "V32cV32c", "nc", "avx512vl,avx512bitalg")
+TARGET_BUILTIN(__builtin_ia32_vpopcntw_256, "V16sV16s", "nc", "avx512vl,avx512bitalg")
+TARGET_BUILTIN(__builtin_ia32_vpopcntb_512, "V64cV64c", "nc", "avx512bitalg")
+TARGET_BUILTIN(__builtin_ia32_vpopcntw_512, "V32sV32s", "nc", "avx512bitalg")
+
+TARGET_BUILTIN(__builtin_ia32_vpshufbitqmb128_mask, "UsV16cV16cUs", "nc", "avx512vl,avx512bitalg")
+TARGET_BUILTIN(__builtin_ia32_vpshufbitqmb256_mask, "UiV32cV32cUi", "nc", "avx512vl,avx512bitalg")
+TARGET_BUILTIN(__builtin_ia32_vpshufbitqmb512_mask, "ULLiV64cV64cULLi", "nc", "avx512bitalg")
+
+TARGET_BUILTIN(__builtin_ia32_vpermi2varhi128_mask, "V8sV8sV8sV8sUc", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varhi256_mask, "V16sV16sV16sV16sUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varhi128_mask, "V8sV8sV8sV8sUc", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varhi128_maskz, "V8sV8sV8sV8sUc", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varhi256_mask, "V16sV16sV16sV16sUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varhi256_maskz, "V16sV16sV16sV16sUs", "nc", "avx512vl,avx512bw")
+
+TARGET_BUILTIN(__builtin_ia32_pmulhrsw512, "V32sV32sV32s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmulhuw512, "V32sV32sV32s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmulhw512, "V32sV32sV32s", "nc", "avx512bw")
+
+TARGET_BUILTIN(__builtin_ia32_addpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_addps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_divpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_divps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_mulpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_mulps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_subpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_subps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_pmaddubsw512, "V32sV64cV64c", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmaddwd512, "V16iV32sV32s", "nc", "avx512bw")
+
+TARGET_BUILTIN(__builtin_ia32_addss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_divss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_mulss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_subss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_maxss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_minss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_addsd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_divsd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_mulsd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_subsd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_maxsd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_minsd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+
+TARGET_BUILTIN(__builtin_ia32_compressdf128_mask, "V2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_compressdf256_mask, "V4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_compressdi128_mask, "V2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_compressdi256_mask, "V4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_compresshi128_mask, "V8sV8sV8sUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_compresshi256_mask, "V16sV16sV16sUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_compressqi128_mask, "V16cV16cV16cUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_compressqi256_mask, "V32cV32cV32cUi", "nc", "avx512vl,avx512vbmi2")
+
+TARGET_BUILTIN(__builtin_ia32_compresssf128_mask, "V4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_compresssf256_mask, "V8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_compresssi128_mask, "V4iV4iV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_compresssi256_mask, "V8iV8iV8iUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_compressstoredf128_mask, "vV2d*V2dUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_compressstoredf256_mask, "vV4d*V4dUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_compressstoredi128_mask, "vV2LLi*V2LLiUc", "n", "avx512vl")
@@ -1191,28 +1189,28 @@ TARGET_BUILTIN(__builtin_ia32_compressst
 TARGET_BUILTIN(__builtin_ia32_compressstoresf256_mask, "vV8f*V8fUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_compressstoresi128_mask, "vV4i*V4iUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_compressstoresi256_mask, "vV8i*V8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2dq128_mask, "V4iV2dV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2ps_mask, "V4fV2dV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2udq128_mask, "V4iV2dV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2udq256_mask, "V4iV4dV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtps2udq128_mask, "V4iV4fV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtps2udq256_mask, "V8iV8fV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2dq128_mask, "V4iV2dV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2udq128_mask, "V4iV2dV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2udq256_mask, "V4iV4dV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvttps2udq128_mask, "V4iV4fV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvttps2udq256_mask, "V8iV8fV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtudq2ps128_mask, "V4fV4iV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtudq2ps256_mask, "V8fV8iV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_expanddf128_mask, "V2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_expanddf256_mask, "V4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_expanddi128_mask, "V2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_expanddi256_mask, "V4LLiV4LLiV4LLiUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_expandhi128_mask, "V8sV8sV8sUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_expandhi256_mask, "V16sV16sV16sUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_expandqi128_mask, "V16cV16cV16cUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_expandqi256_mask, "V32cV32cV32cUi", "n", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2dq128_mask, "V4iV2dV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2ps_mask, "V4fV2dV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2udq128_mask, "V4iV2dV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2udq256_mask, "V4iV4dV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtps2udq128_mask, "V4iV4fV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtps2udq256_mask, "V8iV8fV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2dq128_mask, "V4iV2dV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2udq128_mask, "V4iV2dV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2udq256_mask, "V4iV4dV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvttps2udq128_mask, "V4iV4fV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvttps2udq256_mask, "V8iV8fV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtudq2ps128_mask, "V4fV4iV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtudq2ps256_mask, "V8fV8iV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_expanddf128_mask, "V2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_expanddf256_mask, "V4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_expanddi128_mask, "V2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_expanddi256_mask, "V4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_expandhi128_mask, "V8sV8sV8sUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_expandhi256_mask, "V16sV16sV16sUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_expandqi128_mask, "V16cV16cV16cUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_expandqi256_mask, "V32cV32cV32cUi", "nc", "avx512vl,avx512vbmi2")
 
 TARGET_BUILTIN(__builtin_ia32_expandloaddf128_mask, "V2dV2d*V2dUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_expandloaddf256_mask, "V4dV4d*V4dUc", "n", "avx512vl")
@@ -1228,32 +1226,32 @@ TARGET_BUILTIN(__builtin_ia32_expandload
 TARGET_BUILTIN(__builtin_ia32_expandloadsf256_mask, "V8fV8f*V8fUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_expandloadsi128_mask, "V4iV4i*V4iUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_expandloadsi256_mask, "V8iV8i*V8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_expandsf128_mask, "V4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_expandsf256_mask, "V8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_expandsi128_mask, "V4iV4iV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_expandsi256_mask, "V8iV8iV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getexppd128_mask, "V2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getexppd256_mask, "V4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getexpps128_mask, "V4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getexpps256_mask, "V8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pabsq128_mask, "V2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pabsq256_mask, "V4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmaxsq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmaxsq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmaxuq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmaxuq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pminsq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_expandsf128_mask, "V4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_expandsf256_mask, "V8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_expandsi128_mask, "V4iV4iV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_expandsi256_mask, "V8iV8iV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_getexppd128_mask, "V2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_getexppd256_mask, "V4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_getexpps128_mask, "V4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_getexpps256_mask, "V8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pabsq128_mask, "V2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pabsq256_mask, "V4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmaxsq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmaxsq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmaxuq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmaxuq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pminsq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pminsq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pminuq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pminuq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rndscalepd_128_mask, "V2dV2dIiV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rndscalepd_256_mask, "V4dV4dIiV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rndscaleps_128_mask, "V4fV4fIiV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rndscaleps_256_mask, "V8fV8fIiV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_scalefpd128_mask, "V2dV2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_scalefpd256_mask, "V4dV4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_scalefps128_mask, "V4fV4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_scalefps256_mask, "V8fV8fV8fV8fUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rndscalepd_128_mask, "V2dV2dIiV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rndscalepd_256_mask, "V4dV4dIiV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rndscaleps_128_mask, "V4fV4fIiV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rndscaleps_256_mask, "V8fV8fIiV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_scalefpd128_mask, "V2dV2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_scalefpd256_mask, "V4dV4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_scalefps128_mask, "V4fV4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_scalefps256_mask, "V8fV8fV8fV8fUc", "nc", "avx512vl")
 
 TARGET_BUILTIN(__builtin_ia32_scatterdiv2df, "vd*UcV2LLiV2dIi", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_scatterdiv2di, "vLLi*UcV2LLiV2LLiIi", "n", "avx512vl")
@@ -1272,195 +1270,195 @@ TARGET_BUILTIN(__builtin_ia32_scattersiv
 TARGET_BUILTIN(__builtin_ia32_scattersiv8sf, "vf*UcV8iV8fIi", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_scattersiv8si, "vi*UcV8iV8iIi", "n", "avx512vl")
 
-TARGET_BUILTIN(__builtin_ia32_vpermi2vard128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2vard256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varpd128_mask, "V2dV2dV2LLiV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varpd256_mask, "V4dV4dV4LLiV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varps128_mask, "V4fV4fV4iV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varps256_mask, "V8fV8fV8iV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2vard128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2vard128_maskz, "V4iV4iV4iV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2vard256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2vard256_maskz, "V8iV8iV8iV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varpd128_mask, "V2dV2LLiV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varpd128_maskz, "V2dV2LLiV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varpd256_mask, "V4dV4LLiV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varpd256_maskz, "V4dV4LLiV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varps128_mask, "V4fV4iV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varps128_maskz, "V4fV4iV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varps256_mask, "V8fV8iV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varps256_maskz, "V8fV8iV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-
-TARGET_BUILTIN(__builtin_ia32_vpshldd128_mask, "V4iV4iV4iIiV4iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldd256_mask, "V8iV8iV8iIiV8iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldd512_mask, "V16iV16iV16iIiV16iUs", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldq128_mask, "V2LLiV2LLiV2LLiIiV2LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldq256_mask, "V4LLiV4LLiV4LLiIiV4LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldq512_mask, "V8LLiV8LLiV8LLiIiV8LLiUc", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldw128_mask, "V8sV8sV8sIiV8sUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldw256_mask, "V16sV16sV16sIiV16sUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldw512_mask, "V32sV32sV32sIiV32sUi", "n", "avx512vbmi2")
-
-TARGET_BUILTIN(__builtin_ia32_vpshldvd128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvd256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvd512_mask, "V16iV16iV16iV16iUs", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvw128_mask, "V8sV8sV8sV8sUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvw256_mask, "V16sV16sV16sV16sUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvw512_mask, "V32sV32sV32sV32sUi", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvd128_maskz, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvd256_maskz, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvd512_maskz, "V16iV16iV16iV16iUs", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvw128_maskz, "V8sV8sV8sV8sUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvw256_maskz, "V16sV16sV16sV16sUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshldvw512_maskz, "V32sV32sV32sV32sUi", "n", "avx512vbmi2")
-
-TARGET_BUILTIN(__builtin_ia32_vpshrdvd128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvd256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvd512_mask, "V16iV16iV16iV16iUs", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvw128_mask, "V8sV8sV8sV8sUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvw256_mask, "V16sV16sV16sV16sUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvw512_mask, "V32sV32sV32sV32sUi", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvd128_maskz, "V4iV4iV4iV4iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvd256_maskz, "V8iV8iV8iV8iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvd512_maskz, "V16iV16iV16iV16iUs", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvw128_maskz, "V8sV8sV8sV8sUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvw256_maskz, "V16sV16sV16sV16sUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdvw512_maskz, "V32sV32sV32sV32sUi", "n", "avx512vbmi2")
-
-TARGET_BUILTIN(__builtin_ia32_vpshrdd128_mask, "V4iV4iV4iIiV4iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdd256_mask, "V8iV8iV8iIiV8iUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdd512_mask, "V16iV16iV16iIiV16iUs", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdq128_mask, "V2LLiV2LLiV2LLiIiV2LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdq256_mask, "V4LLiV4LLiV4LLiIiV4LLiUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdq512_mask, "V8LLiV8LLiV8LLiIiV8LLiUc", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdw128_mask, "V8sV8sV8sIiV8sUc", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdw256_mask, "V16sV16sV16sIiV16sUs", "n", "avx512vl,avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_vpshrdw512_mask, "V32sV32sV32sIiV32sUi", "n", "avx512vbmi2")
-
-TARGET_BUILTIN(__builtin_ia32_pmovswb512_mask, "V32cV32sV32cUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovuswb512_mask, "V32cV32sV32cUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovwb512_mask, "V32cV32sV32cUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2qq128_mask, "V2LLiV2dV2LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2qq256_mask, "V4LLiV4dV4LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2uqq128_mask, "V2LLiV2dV2LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2uqq256_mask, "V4LLiV4dV4LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtps2qq128_mask, "V2LLiV4fV2LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtps2qq256_mask, "V4LLiV4fV4LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtps2uqq128_mask, "V2LLiV4fV2LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtps2uqq256_mask, "V4LLiV4fV4LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtqq2pd128_mask, "V2dV2LLiV2dUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtqq2pd256_mask, "V4dV4LLiV4dUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtqq2ps128_mask, "V4fV2LLiV4fUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtqq2ps256_mask, "V4fV4LLiV4fUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2qq128_mask, "V2LLiV2dV2LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2qq256_mask, "V4LLiV4dV4LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2uqq128_mask, "V2LLiV2dV2LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2uqq256_mask, "V4LLiV4dV4LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttps2qq128_mask, "V2LLiV4fV2LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttps2qq256_mask, "V4LLiV4fV4LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttps2uqq128_mask, "V2LLiV4fV2LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttps2uqq256_mask, "V4LLiV4fV4LLiUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtuqq2pd128_mask, "V2dV2LLiV2dUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtuqq2pd256_mask, "V4dV4LLiV4dUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtuqq2ps128_mask, "V4fV2LLiV4fUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtuqq2ps256_mask, "V4fV4LLiV4fUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_rangepd128_mask, "V2dV2dV2dIiV2dUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_rangepd256_mask, "V4dV4dV4dIiV4dUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_rangeps128_mask, "V4fV4fV4fIiV4fUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_rangeps256_mask, "V8fV8fV8fIiV8fUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_rangesd128_round_mask, "V2dV2dV2dV2dUcIiIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_rangess128_round_mask, "V4fV4fV4fV4fUcIiIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_reducepd128_mask, "V2dV2dIiV2dUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_reducepd256_mask, "V4dV4dIiV4dUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_reduceps128_mask, "V4fV4fIiV4fUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_reduceps256_mask, "V8fV8fIiV8fUc", "n", "avx512vl,avx512dq")
-TARGET_BUILTIN(__builtin_ia32_reducesd_mask, "V2dV2dV2dV2dUcIiIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_reducess_mask, "V4fV4fV4fV4fUcIiIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_pmovswb128_mask, "V16cV8sV16cUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovswb256_mask, "V16cV16sV16cUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovuswb128_mask, "V16cV8sV16cUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovuswb256_mask, "V16cV16sV16cUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovwb128_mask, "V16cV8sV16cUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovwb256_mask, "V16cV16sV16cUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2qq512_mask, "V8LLiV8dV8LLiUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtpd2uqq512_mask, "V8LLiV8dV8LLiUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtps2qq512_mask, "V8LLiV8fV8LLiUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtps2uqq512_mask, "V8LLiV8fV8LLiUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtqq2pd512_mask, "V8dV8LLiV8dUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtqq2ps512_mask, "V8fV8LLiV8fUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2qq512_mask, "V8LLiV8dV8LLiUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttpd2uqq512_mask, "V8LLiV8dV8LLiUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttps2qq512_mask, "V8LLiV8fV8LLiUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvttps2uqq512_mask, "V8LLiV8fV8LLiUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtuqq2pd512_mask, "V8dV8LLiV8dUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtuqq2ps512_mask, "V8fV8LLiV8fUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_rangepd512_mask, "V8dV8dV8dIiV8dUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_rangeps512_mask, "V16fV16fV16fIiV16fUsIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_reducepd512_mask, "V8dV8dIiV8dUcIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_reduceps512_mask, "V16fV16fIiV16fUsIi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_prold512_mask, "V16iV16iIiV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_prolq512_mask, "V8LLiV8LLiIiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_prold128_mask, "V4iV4iIiV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prold256_mask, "V8iV8iIiV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prolq128_mask, "V2LLiV2LLiIiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prolq256_mask, "V4LLiV4LLiIiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prolvd512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_prolvq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_prord512_mask, "V16iV16iiV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_prorq512_mask, "V8LLiV8LLiiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_prolvd128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prolvd256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prolvq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prolvq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prord128_mask, "V4iV4iIiV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prord256_mask, "V8iV8iIiV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prorq128_mask, "V2LLiV2LLiIiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prorq256_mask, "V4LLiV4LLiIiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prorvd512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_prorvq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_prorvd128_mask, "V4iV4iV4iV4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prorvd256_mask, "V8iV8iV8iV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prorvq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_prorvq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psllv32hi, "V32sV32sV32s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psllw512, "V32sV32sV8s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psllwi512, "V32sV32si", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psllv16hi, "V16sV16sV16s", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psllv8hi, "V8sV8sV8s", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pslldi512, "V16iV16ii", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psllqi512, "V8LLiV8LLii", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrlv32hi, "V32sV32sV32s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psrlv16hi, "V16sV16sV16s", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psrlv8hi, "V8sV8sV8s", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psrldi512, "V16iV16ii", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrlqi512, "V8LLiV8LLii", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrav32hi, "V32sV32sV32s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psrav16hi, "V16sV16sV16s", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psrav8hi, "V8sV8sV8s", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psravq128, "V2LLiV2LLiV2LLi", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psravq256, "V4LLiV4LLiV4LLi", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psraw512, "V32sV32sV8s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psrawi512, "V32sV32si", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psrlw512, "V32sV32sV8s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psrlwi512, "V32sV32si", "n", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpermi2vard128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2vard256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varpd128_mask, "V2dV2dV2LLiV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varpd256_mask, "V4dV4dV4LLiV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varps128_mask, "V4fV4fV4iV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varps256_mask, "V8fV8fV8iV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2vard128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2vard128_maskz, "V4iV4iV4iV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2vard256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2vard256_maskz, "V8iV8iV8iV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varpd128_mask, "V2dV2LLiV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varpd128_maskz, "V2dV2LLiV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varpd256_mask, "V4dV4LLiV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varpd256_maskz, "V4dV4LLiV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varps128_mask, "V4fV4iV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varps128_maskz, "V4fV4iV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varps256_mask, "V8fV8iV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varps256_maskz, "V8fV8iV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+
+TARGET_BUILTIN(__builtin_ia32_vpshldd128_mask, "V4iV4iV4iIiV4iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldd256_mask, "V8iV8iV8iIiV8iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldd512_mask, "V16iV16iV16iIiV16iUs", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldq128_mask, "V2LLiV2LLiV2LLiIiV2LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldq256_mask, "V4LLiV4LLiV4LLiIiV4LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldq512_mask, "V8LLiV8LLiV8LLiIiV8LLiUc", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldw128_mask, "V8sV8sV8sIiV8sUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldw256_mask, "V16sV16sV16sIiV16sUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldw512_mask, "V32sV32sV32sIiV32sUi", "nc", "avx512vbmi2")
+
+TARGET_BUILTIN(__builtin_ia32_vpshldvd128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvd256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvd512_mask, "V16iV16iV16iV16iUs", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvw128_mask, "V8sV8sV8sV8sUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvw256_mask, "V16sV16sV16sV16sUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvd128_maskz, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvd256_maskz, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvd512_maskz, "V16iV16iV16iV16iUs", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvw128_maskz, "V8sV8sV8sV8sUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvw256_maskz, "V16sV16sV16sV16sUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshldvw512_maskz, "V32sV32sV32sV32sUi", "nc", "avx512vbmi2")
+
+TARGET_BUILTIN(__builtin_ia32_vpshrdvd128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvd256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvd512_mask, "V16iV16iV16iV16iUs", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvw128_mask, "V8sV8sV8sV8sUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvw256_mask, "V16sV16sV16sV16sUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvw512_mask, "V32sV32sV32sV32sUi", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvd128_maskz, "V4iV4iV4iV4iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvd256_maskz, "V8iV8iV8iV8iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvd512_maskz, "V16iV16iV16iV16iUs", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvw128_maskz, "V8sV8sV8sV8sUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvw256_maskz, "V16sV16sV16sV16sUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdvw512_maskz, "V32sV32sV32sV32sUi", "nc", "avx512vbmi2")
+
+TARGET_BUILTIN(__builtin_ia32_vpshrdd128_mask, "V4iV4iV4iIiV4iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdd256_mask, "V8iV8iV8iIiV8iUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdd512_mask, "V16iV16iV16iIiV16iUs", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdq128_mask, "V2LLiV2LLiV2LLiIiV2LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdq256_mask, "V4LLiV4LLiV4LLiIiV4LLiUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdq512_mask, "V8LLiV8LLiV8LLiIiV8LLiUc", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdw128_mask, "V8sV8sV8sIiV8sUc", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdw256_mask, "V16sV16sV16sIiV16sUs", "nc", "avx512vl,avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_vpshrdw512_mask, "V32sV32sV32sIiV32sUi", "nc", "avx512vbmi2")
+
+TARGET_BUILTIN(__builtin_ia32_pmovswb512_mask, "V32cV32sV32cUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmovuswb512_mask, "V32cV32sV32cUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmovwb512_mask, "V32cV32sV32cUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2qq128_mask, "V2LLiV2dV2LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2qq256_mask, "V4LLiV4dV4LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2uqq128_mask, "V2LLiV2dV2LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2uqq256_mask, "V4LLiV4dV4LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtps2qq128_mask, "V2LLiV4fV2LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtps2qq256_mask, "V4LLiV4fV4LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtps2uqq128_mask, "V2LLiV4fV2LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtps2uqq256_mask, "V4LLiV4fV4LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtqq2pd128_mask, "V2dV2LLiV2dUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtqq2pd256_mask, "V4dV4LLiV4dUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtqq2ps128_mask, "V4fV2LLiV4fUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtqq2ps256_mask, "V4fV4LLiV4fUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2qq128_mask, "V2LLiV2dV2LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2qq256_mask, "V4LLiV4dV4LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2uqq128_mask, "V2LLiV2dV2LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2uqq256_mask, "V4LLiV4dV4LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttps2qq128_mask, "V2LLiV4fV2LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttps2qq256_mask, "V4LLiV4fV4LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttps2uqq128_mask, "V2LLiV4fV2LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttps2uqq256_mask, "V4LLiV4fV4LLiUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtuqq2pd128_mask, "V2dV2LLiV2dUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtuqq2pd256_mask, "V4dV4LLiV4dUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtuqq2ps128_mask, "V4fV2LLiV4fUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtuqq2ps256_mask, "V4fV4LLiV4fUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_rangepd128_mask, "V2dV2dV2dIiV2dUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_rangepd256_mask, "V4dV4dV4dIiV4dUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_rangeps128_mask, "V4fV4fV4fIiV4fUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_rangeps256_mask, "V8fV8fV8fIiV8fUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_rangesd128_round_mask, "V2dV2dV2dV2dUcIiIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_rangess128_round_mask, "V4fV4fV4fV4fUcIiIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_reducepd128_mask, "V2dV2dIiV2dUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_reducepd256_mask, "V4dV4dIiV4dUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_reduceps128_mask, "V4fV4fIiV4fUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_reduceps256_mask, "V8fV8fIiV8fUc", "nc", "avx512vl,avx512dq")
+TARGET_BUILTIN(__builtin_ia32_reducesd_mask, "V2dV2dV2dV2dUcIiIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_reducess_mask, "V4fV4fV4fV4fUcIiIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_pmovswb128_mask, "V16cV8sV16cUc", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmovswb256_mask, "V16cV16sV16cUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmovuswb128_mask, "V16cV8sV16cUc", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmovuswb256_mask, "V16cV16sV16cUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmovwb128_mask, "V16cV8sV16cUc", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_pmovwb256_mask, "V16cV16sV16cUs", "nc", "avx512vl,avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2qq512_mask, "V8LLiV8dV8LLiUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtpd2uqq512_mask, "V8LLiV8dV8LLiUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtps2qq512_mask, "V8LLiV8fV8LLiUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtps2uqq512_mask, "V8LLiV8fV8LLiUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtqq2pd512_mask, "V8dV8LLiV8dUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtqq2ps512_mask, "V8fV8LLiV8fUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2qq512_mask, "V8LLiV8dV8LLiUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttpd2uqq512_mask, "V8LLiV8dV8LLiUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttps2qq512_mask, "V8LLiV8fV8LLiUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvttps2uqq512_mask, "V8LLiV8fV8LLiUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtuqq2pd512_mask, "V8dV8LLiV8dUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtuqq2ps512_mask, "V8fV8LLiV8fUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_rangepd512_mask, "V8dV8dV8dIiV8dUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_rangeps512_mask, "V16fV16fV16fIiV16fUsIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_reducepd512_mask, "V8dV8dIiV8dUcIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_reduceps512_mask, "V16fV16fIiV16fUsIi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_prold512_mask, "V16iV16iIiV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_prolq512_mask, "V8LLiV8LLiIiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_prold128_mask, "V4iV4iIiV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prold256_mask, "V8iV8iIiV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prolq128_mask, "V2LLiV2LLiIiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prolq256_mask, "V4LLiV4LLiIiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prolvd512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_prolvq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_prord512_mask, "V16iV16iiV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_prorq512_mask, "V8LLiV8LLiiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_prolvd128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prolvd256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prolvq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prolvq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prord128_mask, "V4iV4iIiV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prord256_mask, "V8iV8iIiV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prorq128_mask, "V2LLiV2LLiIiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prorq256_mask, "V4LLiV4LLiIiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prorvd512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_prorvq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_prorvd128_mask, "V4iV4iV4iV4iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prorvd256_mask, "V8iV8iV8iV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prorvq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_prorvq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psllv32hi, "V32sV32sV32s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psllw512, "V32sV32sV8s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psllwi512, "V32sV32si", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psllv16hi, "V16sV16sV16s", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psllv8hi, "V8sV8sV8s", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pslldi512, "V16iV16ii", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psllqi512, "V8LLiV8LLii", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrlv32hi, "V32sV32sV32s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psrlv16hi, "V16sV16sV16s", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psrlv8hi, "V8sV8sV8s", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psrldi512, "V16iV16ii", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrlqi512, "V8LLiV8LLii", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrav32hi, "V32sV32sV32s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psrav16hi, "V16sV16sV16s", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psrav8hi, "V8sV8sV8s", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psravq128, "V2LLiV2LLiV2LLi", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psravq256, "V4LLiV4LLiV4LLi", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psraw512, "V32sV32sV8s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psrawi512, "V32sV32si", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psrlw512, "V32sV32sV8s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psrlwi512, "V32sV32si", "nc", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_movdqa32load128_mask, "V4iV4i*V4iUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_movdqa32load256_mask, "V8iV8i*V8iUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_movdqa32load512_mask, "V16iV16iC*V16iUs", "n", "avx512f")
@@ -1473,57 +1471,57 @@ TARGET_BUILTIN(__builtin_ia32_movdqa64lo
 TARGET_BUILTIN(__builtin_ia32_movdqa64load256_mask, "V4LLiV4LLiC*V4LLiUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_movdqa64store128_mask, "vV2LLi*V2LLiUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_movdqa64store256_mask, "vV4LLi*V4LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52huq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512ifma")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52huq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512ifma")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52luq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512ifma")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52luq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512ifma")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52huq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512ifma,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52huq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512ifma,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52huq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512ifma,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52huq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512ifma,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52luq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512ifma,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52luq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "n", "avx512ifma,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52luq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512ifma,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmadd52luq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512ifma,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varqi512_mask, "V64cV64cV64cV64cULLi", "n", "avx512vbmi")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varqi512_mask, "V64cV64cV64cV64cULLi", "n", "avx512vbmi")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varqi512_maskz, "V64cV64cV64cV64cULLi", "n", "avx512vbmi")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varqi128_mask, "V16cV16cV16cV16cUs", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varqi256_mask, "V32cV32cV32cV32cUi", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varqi128_mask, "V16cV16cV16cV16cUs", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varqi128_maskz, "V16cV16cV16cV16cUs", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varqi256_mask, "V32cV32cV32cV32cUi", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varqi256_maskz, "V32cV32cV32cV32cUi", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vcomisd, "iV2dV2dIiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcomiss, "iV4fV4fIiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_kunpckdi, "ULLiULLiULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_kunpcksi, "UiUiUi", "n", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52huq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512ifma")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52huq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512ifma")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52luq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512ifma")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52luq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512ifma")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52huq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512ifma,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52huq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512ifma,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52huq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512ifma,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52huq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512ifma,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52luq128_mask, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512ifma,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52luq128_maskz, "V2LLiV2LLiV2LLiV2LLiUc", "nc", "avx512ifma,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52luq256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512ifma,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmadd52luq256_maskz, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512ifma,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varqi512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512vbmi")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varqi512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512vbmi")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varqi512_maskz, "V64cV64cV64cV64cULLi", "nc", "avx512vbmi")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varqi128_mask, "V16cV16cV16cV16cUs", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varqi256_mask, "V32cV32cV32cV32cUi", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varqi128_mask, "V16cV16cV16cV16cUs", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varqi128_maskz, "V16cV16cV16cV16cUs", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varqi256_mask, "V32cV32cV32cV32cUi", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varqi256_maskz, "V32cV32cV32cV32cUi", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vcomisd, "iV2dV2dIiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcomiss, "iV4fV4fIiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_kunpckdi, "ULLiULLiULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_kunpcksi, "UiUiUi", "nc", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_loaddquhi512_mask, "V32sV32s*V32sUi", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_loaddquqi512_mask, "V64cV64c*V64cULLi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_fixupimmpd512_mask, "V8dV8dV8dV8LLiIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_fixupimmpd512_maskz, "V8dV8dV8dV8LLiIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_fixupimmps512_mask, "V16fV16fV16fV16iIiUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_fixupimmps512_maskz, "V16fV16fV16fV16iIiUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_fixupimmsd_mask, "V2dV2dV2dV2LLiIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_fixupimmsd_maskz, "V2dV2dV2dV2LLiIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_fixupimmss_mask, "V4fV4fV4fV4iIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_fixupimmss_maskz, "V4fV4fV4fV4iIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_getexpsd128_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_getexpss128_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_getmantsd_round_mask, "V2dV2dV2dIiV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_getmantss_round_mask, "V4fV4fV4fIiV4fUcIi", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_fixupimmpd512_mask, "V8dV8dV8dV8LLiIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_fixupimmpd512_maskz, "V8dV8dV8dV8LLiIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_fixupimmps512_mask, "V16fV16fV16fV16iIiUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_fixupimmps512_maskz, "V16fV16fV16fV16iIiUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_fixupimmsd_mask, "V2dV2dV2dV2LLiIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_fixupimmsd_maskz, "V2dV2dV2dV2LLiIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_fixupimmss_mask, "V4fV4fV4fV4iIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_fixupimmss_maskz, "V4fV4fV4fV4iIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_getexpsd128_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_getexpss128_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_getmantsd_round_mask, "V2dV2dV2dIiV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_getmantss_round_mask, "V4fV4fV4fIiV4fUcIi", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_loaddquhi128_mask, "V8sV8s*V8sUc", "n", "avx512bw,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loaddquhi256_mask, "V16sV16s*V16sUs", "n", "avx512bw,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loaddquqi128_mask, "V16cV16c*V16cUs", "n", "avx512bw,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loaddquqi256_mask, "V32cV32c*V32cUi", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fixupimmpd128_mask, "V2dV2dV2dV2LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fixupimmpd128_maskz, "V2dV2dV2dV2LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fixupimmpd256_mask, "V4dV4dV4dV4LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fixupimmpd256_maskz, "V4dV4dV4dV4LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fixupimmps128_mask, "V4fV4fV4fV4iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fixupimmps128_maskz, "V4fV4fV4fV4iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fixupimmps256_mask, "V8fV8fV8fV8iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fixupimmps256_maskz, "V8fV8fV8fV8iIiUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fixupimmpd128_mask, "V2dV2dV2dV2LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fixupimmpd128_maskz, "V2dV2dV2dV2LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fixupimmpd256_mask, "V4dV4dV4dV4LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fixupimmpd256_maskz, "V4dV4dV4dV4LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fixupimmps128_mask, "V4fV4fV4fV4iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fixupimmps128_maskz, "V4fV4fV4fV4iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fixupimmps256_mask, "V8fV8fV8fV8iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fixupimmps256_maskz, "V8fV8fV8fV8iIiUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loadapd128_mask, "V2dV2d*V2dUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loadsd128_mask, "V8dV8d*V8dUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_loadapd256_mask, "V4dV4d*V4dUc", "n", "avx512vl")
@@ -1558,315 +1556,315 @@ TARGET_BUILTIN(__builtin_ia32_storeupd12
 TARGET_BUILTIN(__builtin_ia32_storeupd256_mask, "vV4d*V4dUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_storeups128_mask, "vV4f*V4fUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_storeups256_mask, "vV8f*V8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rcp14pd128_mask, "V2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rcp14pd256_mask, "V4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rcp14ps128_mask, "V4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rcp14ps256_mask, "V8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vplzcntd_128_mask, "V4iV4iV4iUc", "n", "avx512cd,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vplzcntd_256_mask, "V8iV8iV8iUc", "n", "avx512cd,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vplzcntq_128_mask, "V2LLiV2LLiV2LLiUc", "n", "avx512cd,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vplzcntq_256_mask, "V4LLiV4LLiV4LLiUc", "n", "avx512cd,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vcvtsd2si32, "iV2dIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi32, "UiV2dIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtss2si32, "iV4fIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtss2usi32, "UiV4fIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvttsd2si32, "iV2dIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi32, "UiV2dIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvttss2si32, "iV4fIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvttss2usi32, "UiV4fIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermi2vard512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varpd512_mask, "V8dV8dV8LLiV8dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varps512_mask, "V16fV16fV16iV16fUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermi2varq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermilvarpd512, "V8dV8dV8LLi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermilvarps512, "V16fV16fV16i", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermt2vard512_maskz, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varpd512_maskz, "V8dV8LLiV8dV8dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varps512_maskz, "V16fV16iV16fV16fUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpermt2varq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rndscalesd_round_mask, "V2dV2dV2dV2dUcIiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rndscaless_round_mask, "V4fV4fV4fV4fUcIiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_scalefpd512_mask, "V8dV8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_scalefps512_mask, "V16fV16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_scalefsd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_scalefss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psradi512, "V16iV16ii", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psraqi512, "V8LLiV8LLii", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psraq128, "V2LLiV2LLiV2LLi", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psraq256, "V4LLiV4LLiV2LLi", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psraqi128, "V2LLiV2LLii", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_psraqi256, "V4LLiV4LLii", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pslld512, "V16iV16iV4i", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psllq512, "V8LLiV8LLiV2LLi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psllv16si, "V16iV16iV16i", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psllv8di, "V8LLiV8LLiV8LLi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrad512, "V16iV16iV4i", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psraq512, "V8LLiV8LLiV2LLi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrav16si, "V16iV16iV16i", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrav8di, "V8LLiV8LLiV8LLi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrld512, "V16iV16iV4i", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrlq512, "V8LLiV8LLiV2LLi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrlv16si, "V16iV16iV16i", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_psrlv8di, "V8LLiV8LLiV8LLi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pternlogd512_mask, "V16iV16iV16iV16iIiUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pternlogd512_maskz, "V16iV16iV16iV16iIiUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pternlogq512_mask, "V8LLiV8LLiV8LLiV8LLiIiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pternlogq512_maskz, "V8LLiV8LLiV8LLiV8LLiIiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pternlogd128_mask, "V4iV4iV4iV4iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pternlogd128_maskz, "V4iV4iV4iV4iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pternlogd256_mask, "V8iV8iV8iV8iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pternlogd256_maskz, "V8iV8iV8iV8iIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pternlogq128_mask, "V2LLiV2LLiV2LLiV2LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pternlogq128_maskz, "V2LLiV2LLiV2LLiV2LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pternlogq256_mask, "V4LLiV4LLiV4LLiV4LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pternlogq256_maskz, "V4LLiV4LLiV4LLiV4LLiIiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_shuf_f32x4_mask, "V16fV16fV16fIiV16fUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_shuf_f64x2_mask, "V8dV8dV8dIiV8dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_shuf_i32x4_mask, "V16iV16iV16iIiV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_shuf_i64x2_mask, "V8LLiV8LLiV8LLiIiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_shuf_f32x4_256_mask, "V8fV8fV8fIiV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_shuf_f64x2_256_mask, "V4dV4dV4dIiV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_shuf_i32x4_256_mask, "V8iV8iV8iIiV8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_shuf_i64x2_256_mask, "V4LLiV4LLiV4LLiIiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_sqrtsd_round_mask, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_sqrtss_round_mask, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14pd128_mask, "V2dV2dV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14pd256_mask, "V4dV4dV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14ps128_mask, "V4fV4fV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14ps256_mask, "V8fV8fV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtb2mask512, "ULLiV64c", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2b512, "V64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2w512, "V32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cvtd2mask512, "UsV16i", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2d512, "V16iUs", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2q512, "V8LLiUc", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtq2mask512, "UcV8LLi", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_cvtb2mask128, "UsV16c", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtb2mask256, "UiV32c", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2b128, "V16cUs", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2b256, "V32cUi", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2w128, "V8sUc", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2w256, "V16sUs", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtd2mask128, "UcV4i", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtd2mask256, "UcV8i", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2d128, "V4iUc", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2d256, "V8iUc", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2q128, "V2LLiUc", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtmask2q256, "V4LLiUc", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtq2mask128, "UcV2LLi", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtq2mask256, "UcV4LLi", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovsdb512_mask, "V16cV16iV16cUs", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rcp14pd128_mask, "V2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rcp14pd256_mask, "V4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rcp14ps128_mask, "V4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rcp14ps256_mask, "V8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vplzcntd_128_mask, "V4iV4iV4iUc", "nc", "avx512cd,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vplzcntd_256_mask, "V8iV8iV8iUc", "nc", "avx512cd,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vplzcntq_128_mask, "V2LLiV2LLiV2LLiUc", "nc", "avx512cd,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vplzcntq_256_mask, "V4LLiV4LLiV4LLiUc", "nc", "avx512cd,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2si32, "iV2dIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi32, "UiV2dIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2si32, "iV4fIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2usi32, "UiV4fIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2si32, "iV2dIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi32, "UiV2dIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2si32, "iV4fIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2usi32, "UiV4fIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermi2vard512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varpd512_mask, "V8dV8dV8LLiV8dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varps512_mask, "V16fV16fV16iV16fUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermi2varq512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarpd512, "V8dV8dV8LLi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermilvarps512, "V16fV16fV16i", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2vard512_maskz, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varpd512_maskz, "V8dV8LLiV8dV8dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varps512_maskz, "V16fV16iV16fV16fUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpermt2varq512_maskz, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rndscalesd_round_mask, "V2dV2dV2dV2dUcIiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rndscaless_round_mask, "V4fV4fV4fV4fUcIiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_scalefpd512_mask, "V8dV8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_scalefps512_mask, "V16fV16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_scalefsd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_scalefss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psradi512, "V16iV16ii", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psraqi512, "V8LLiV8LLii", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psraq128, "V2LLiV2LLiV2LLi", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psraq256, "V4LLiV4LLiV2LLi", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psraqi128, "V2LLiV2LLii", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_psraqi256, "V4LLiV4LLii", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pslld512, "V16iV16iV4i", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psllq512, "V8LLiV8LLiV2LLi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psllv16si, "V16iV16iV16i", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psllv8di, "V8LLiV8LLiV8LLi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrad512, "V16iV16iV4i", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psraq512, "V8LLiV8LLiV2LLi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrav16si, "V16iV16iV16i", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrav8di, "V8LLiV8LLiV8LLi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrld512, "V16iV16iV4i", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrlq512, "V8LLiV8LLiV2LLi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrlv16si, "V16iV16iV16i", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_psrlv8di, "V8LLiV8LLiV8LLi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogd512_mask, "V16iV16iV16iV16iIiUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogd512_maskz, "V16iV16iV16iV16iIiUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogq512_mask, "V8LLiV8LLiV8LLiV8LLiIiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogq512_maskz, "V8LLiV8LLiV8LLiV8LLiIiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pternlogd128_mask, "V4iV4iV4iV4iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogd128_maskz, "V4iV4iV4iV4iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogd256_mask, "V8iV8iV8iV8iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogd256_maskz, "V8iV8iV8iV8iIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogq128_mask, "V2LLiV2LLiV2LLiV2LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogq128_maskz, "V2LLiV2LLiV2LLiV2LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogq256_mask, "V4LLiV4LLiV4LLiV4LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pternlogq256_maskz, "V4LLiV4LLiV4LLiV4LLiIiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_f32x4_mask, "V16fV16fV16fIiV16fUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_f64x2_mask, "V8dV8dV8dIiV8dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_i32x4_mask, "V16iV16iV16iIiV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_i64x2_mask, "V8LLiV8LLiV8LLiIiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_f32x4_256_mask, "V8fV8fV8fIiV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_f64x2_256_mask, "V4dV4dV4dIiV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_i32x4_256_mask, "V8iV8iV8iIiV8iUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_i64x2_256_mask, "V4LLiV4LLiV4LLiIiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_sqrtsd_round_mask, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_sqrtss_round_mask, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14pd128_mask, "V2dV2dV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14pd256_mask, "V4dV4dV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14ps128_mask, "V4fV4fV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14ps256_mask, "V8fV8fV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtb2mask512, "ULLiV64c", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2b512, "V64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2w512, "V32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtd2mask512, "UsV16i", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2d512, "V16iUs", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2q512, "V8LLiUc", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtq2mask512, "UcV8LLi", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtb2mask128, "UsV16c", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtb2mask256, "UiV32c", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2b128, "V16cUs", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2b256, "V32cUi", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2w128, "V8sUc", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2w256, "V16sUs", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtd2mask128, "UcV4i", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtd2mask256, "UcV8i", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2d128, "V4iUc", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2d256, "V8iUc", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2q128, "V2LLiUc", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2q256, "V4LLiUc", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtq2mask128, "UcV2LLi", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtq2mask256, "UcV4LLi", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsdb512_mask, "V16cV16iV16cUs", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovsdb512mem_mask, "vV16c*V16iUs", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovswb512mem_mask, "vV32c*V32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovsdw512_mask, "V16sV16iV16sUs", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovsdw512_mask, "V16sV16iV16sUs", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovsdw512mem_mask, "vV16s*V16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovsqb512_mask, "V16cV8LLiV16cUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovsqb512_mask, "V16cV8LLiV16cUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovsqb512mem_mask, "vV16c*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovsqd512_mask, "V8iV8LLiV8iUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovsqd512_mask, "V8iV8LLiV8iUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovsqd512mem_mask, "vV8i*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovsqw512_mask, "V8sV8LLiV8sUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovsqw512_mask, "V8sV8LLiV8sUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovsqw512mem_mask, "vV8s*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovsdb128_mask, "V16cV4iV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsdb128_mask, "V16cV4iV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsdb128mem_mask, "vV16c*V4iUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovswb128mem_mask, "vV16c*V8sUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovsdb256_mask, "V16cV8iV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsdb256_mask, "V16cV8iV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsdb256mem_mask, "vV16c*V8iUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovswb256mem_mask, "vV16c*V16sUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovsdw128_mask, "V8sV4iV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsdw128_mask, "V8sV4iV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsdw128mem_mask, "vV8s*V4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovsdw256_mask, "V8sV8iV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsdw256_mask, "V8sV8iV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsdw256mem_mask, "vV8s*V8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovsqb128_mask, "V16cV2LLiV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsqb128_mask, "V16cV2LLiV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsqb128mem_mask, "vV16c*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovsqb256_mask, "V16cV4LLiV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsqb256_mask, "V16cV4LLiV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsqb256mem_mask, "vV16c*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovsqd128_mask, "V4iV2LLiV4iUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsqd128_mask, "V4iV2LLiV4iUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsqd128mem_mask, "vV4i*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovsqd256_mask, "V4iV4LLiV4iUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsqd256_mask, "V4iV4LLiV4iUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsqd256mem_mask, "vV4i*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovsqw128_mask, "V8sV2LLiV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsqw128_mask, "V8sV2LLiV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsqw128mem_mask, "vV8s*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovsqw256_mask, "V8sV4LLiV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovsqw256_mask, "V8sV4LLiV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovsqw256mem_mask, "vV8s*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovusdb512_mask, "V16cV16iV16cUs", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb512_mask, "V16cV16iV16cUs", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovusdb512mem_mask, "vV16c*V16iUs", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovuswb512mem_mask, "vV32c*V32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovusdw512_mask, "V16sV16iV16sUs", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw512_mask, "V16sV16iV16sUs", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovusdw512mem_mask, "vV16s*V16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovusqb512_mask, "V16cV8LLiV16cUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqb512_mask, "V16cV8LLiV16cUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovusqb512mem_mask, "vV16c*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovusqd512_mask, "V8iV8LLiV8iUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd512_mask, "V8iV8LLiV8iUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovusqd512mem_mask, "vV8i*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovusqw512_mask, "V8sV8LLiV8sUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw512_mask, "V8sV8LLiV8sUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovusqw512mem_mask, "vV8s*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovusdb128_mask, "V16cV4iV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb128_mask, "V16cV4iV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusdb128mem_mask, "vV16c*V4iUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovuswb128mem_mask, "vV16c*V8sUc", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovusdb256_mask, "V16cV8iV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb256_mask, "V16cV8iV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusdb256mem_mask, "vV16c*V8iUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovuswb256mem_mask, "vV16c*V16sUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovusdw128_mask, "V8sV4iV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw128_mask, "V8sV4iV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusdw128mem_mask, "vV8s*V4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovusdw256_mask, "V8sV8iV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw256_mask, "V8sV8iV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusdw256mem_mask, "vV8s*V8iUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusqb128_mask, "V16cV2LLiV16cUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusqb128mem_mask, "vV16c*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovusqb256_mask, "V16cV4LLiV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqb256_mask, "V16cV4LLiV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusqb256mem_mask, "vV16c*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovusqd128_mask, "V4iV2LLiV4iUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd128_mask, "V4iV2LLiV4iUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusqd128mem_mask, "vV4i*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovusqd256_mask, "V4iV4LLiV4iUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd256_mask, "V4iV4LLiV4iUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusqd256mem_mask, "vV4i*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovusqw128_mask, "V8sV2LLiV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw128_mask, "V8sV2LLiV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusqw128mem_mask, "vV8s*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovusqw256_mask, "V8sV4LLiV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw256_mask, "V8sV4LLiV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovusqw256mem_mask, "vV8s*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovdb512_mask, "V16cV16iV16cUs", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovdb512_mask, "V16cV16iV16cUs", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovdb512mem_mask, "vV16c*V16iUs", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovwb512mem_mask, "vV32c*V32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovdw512_mask, "V16sV16iV16sUs", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovdw512_mask, "V16sV16iV16sUs", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovdw512mem_mask, "vV16s*V16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovqb512_mask, "V16cV8LLiV16cUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovqb512_mask, "V16cV8LLiV16cUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovqb512mem_mask, "vV16c*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovqd512_mask, "V8iV8LLiV8iUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovqd512_mask, "V8iV8LLiV8iUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovqd512mem_mask, "vV8i*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovqw512_mask, "V8sV8LLiV8sUc", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovqw512_mask, "V8sV8LLiV8sUc", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_pmovqw512mem_mask, "vV8s*V8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_pmovdb128_mask, "V16cV4iV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovdb128_mask, "V16cV4iV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovwb128mem_mask, "vV16c*V8sUc", "n", "avx512vl,avx512bw")
 TARGET_BUILTIN(__builtin_ia32_pmovdb128mem_mask, "vV16c*V4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovdb256_mask, "V16cV8iV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovdb256_mask, "V16cV8iV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovdb256mem_mask, "vV16c*V8iUc", "n", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovwb256mem_mask, "vV16c*V16sUs", "n", "avx512vl,avx512bw")
-TARGET_BUILTIN(__builtin_ia32_pmovdw128_mask, "V8sV4iV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovdw128_mask, "V8sV4iV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovdw128mem_mask, "vV8s*V4iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovdw256_mask, "V8sV8iV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovdw256_mask, "V8sV8iV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovdw256mem_mask, "vV8s*V8iUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovqb128_mask, "V16cV2LLiV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovqb128_mask, "V16cV2LLiV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovqb128mem_mask, "vV16c*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovqb256_mask, "V16cV4LLiV16cUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovqb256_mask, "V16cV4LLiV16cUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovqb256mem_mask, "vV16c*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovqd128_mask, "V4iV2LLiV4iUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovqd128_mask, "V4iV2LLiV4iUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovqd128mem_mask, "vV4i*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovqd256_mask, "V4iV4LLiV4iUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovqd256_mask, "V4iV4LLiV4iUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovqd256mem_mask, "vV4i*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovqw128_mask, "V8sV2LLiV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovqw128_mask, "V8sV2LLiV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovqw128mem_mask, "vV8s*V2LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_pmovqw256_mask, "V8sV4LLiV8sUc", "n", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovqw256_mask, "V8sV4LLiV8sUc", "nc", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_pmovqw256mem_mask, "vV8s*V4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getmantpd128_mask, "V2dV2diV2dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getmantpd256_mask, "V4dV4diV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getmantps128_mask, "V4fV4fiV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getmantps256_mask, "V8fV8fiV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_getmantpd512_mask, "V8dV8diV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_getmantps512_mask, "V16fV16fiV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_getexppd512_mask, "V8dV8dV8dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_getexpps512_mask, "V16fV16fV16fUsIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddss3_mask,  "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddss3_maskz, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddss3_mask3, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_mask,  "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_maskz, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_mask3, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmsubsd3_mask3, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfmsubss3_mask3, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubsd3_mask3, "V2dV2dV2dV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vfnmsubss3_mask3, "V4fV4fV4fV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_permvarhi512_mask, "V32sV32sV32sV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_permvardf512_mask, "V8dV8dV8LLiV8dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_permvardi512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_permvarsf512_mask, "V16fV16fV16iV16fUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_permvarsi512_mask, "V16iV16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_permvarqi512_mask, "V64cV64cV64cV64cULLi", "n", "avx512vbmi")
-TARGET_BUILTIN(__builtin_ia32_permvarqi128_mask, "V16cV16cV16cV16cUs", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_permvarqi256_mask, "V32cV32cV32cV32cUi", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_permvarhi128_mask, "V8sV8sV8sV8sUc", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_permvarhi256_mask, "V16sV16sV16sV16sUs", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_permvardf256_mask, "V4dV4dV4LLiV4dUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_permvardi256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fpclasspd128_mask, "UcV2dIiUc", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fpclasspd256_mask, "UcV4dIiUc", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fpclassps128_mask, "UcV4fIiUc", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fpclassps256_mask, "UcV8fIiUc", "n", "avx512dq,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_fpclassps512_mask, "UsV16fIiUs", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_fpclasspd512_mask, "UcV8dIiUc", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_fpclasssd_mask, "UcV2dIiUc", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_fpclassss_mask, "UcV4fIiUc", "n", "avx512dq")
-TARGET_BUILTIN(__builtin_ia32_kandhi, "UsUsUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_kandnhi, "UsUsUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_korhi, "UsUsUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_kortestchi, "iUsUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_kortestzhi, "iUsUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_kunpckhi, "UsUsUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_kxnorhi, "UsUsUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_kxorhi, "UsUsUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_palignr512_mask, "V64cV64cV64cIiV64cULLi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_dbpsadbw128_mask, "V8sV16cV16cIiV8sUc", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_dbpsadbw256_mask, "V16sV32cV32cIiV16sUs", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_dbpsadbw512_mask, "V32sV64cV64cIiV32sUi", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_psadbw512, "V8LLiV64cV64c", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_compressdf512_mask, "V8dV8dV8dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_compressdi512_mask, "V8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_compresshi512_mask, "V32sV32sV32sUi", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_compressqi512_mask, "V64cV64cV64cULLi", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_compresssf512_mask, "V16fV16fV16fUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_compresssi512_mask, "V16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cmpsd_mask, "UcV2dV2dIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cmpss_mask, "UcV4fV4fIiUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_expanddf512_mask, "V8dV8dV8dUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_expanddi512_mask, "V8LLiV8LLiV8LLiUc", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_expandhi512_mask, "V32sV32sV32sUi", "n", "avx512vbmi2")
-TARGET_BUILTIN(__builtin_ia32_expandqi512_mask, "V64cV64cV64cULLi", "n", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_getmantpd128_mask, "V2dV2diV2dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_getmantpd256_mask, "V4dV4diV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_getmantps128_mask, "V4fV4fiV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_getmantps256_mask, "V8fV8fiV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_getmantpd512_mask, "V8dV8diV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_getmantps512_mask, "V16fV16fiV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_getexppd512_mask, "V8dV8dV8dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_getexpps512_mask, "V16fV16fV16fUsIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddss3_mask,  "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddss3_maskz, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddss3_mask3, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_mask,  "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_maskz, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_mask3, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmsubsd3_mask3, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmsubss3_mask3, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubsd3_mask3, "V2dV2dV2dV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubss3_mask3, "V4fV4fV4fV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvarhi512_mask, "V32sV32sV32sV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_permvardf512_mask, "V8dV8dV8LLiV8dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvardi512_mask, "V8LLiV8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvarsf512_mask, "V16fV16fV16iV16fUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvarsi512_mask, "V16iV16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvarqi512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512vbmi")
+TARGET_BUILTIN(__builtin_ia32_permvarqi128_mask, "V16cV16cV16cV16cUs", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvarqi256_mask, "V32cV32cV32cV32cUi", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvarhi128_mask, "V8sV8sV8sV8sUc", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvarhi256_mask, "V16sV16sV16sV16sUs", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvardf256_mask, "V4dV4dV4LLiV4dUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvardi256_mask, "V4LLiV4LLiV4LLiV4LLiUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fpclasspd128_mask, "UcV2dIiUc", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fpclasspd256_mask, "UcV4dIiUc", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fpclassps128_mask, "UcV4fIiUc", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fpclassps256_mask, "UcV8fIiUc", "nc", "avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_fpclassps512_mask, "UsV16fIiUs", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_fpclasspd512_mask, "UcV8dIiUc", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_fpclasssd_mask, "UcV2dIiUc", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_fpclassss_mask, "UcV4fIiUc", "nc", "avx512dq")
+TARGET_BUILTIN(__builtin_ia32_kandhi, "UsUsUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_kandnhi, "UsUsUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_korhi, "UsUsUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_kortestchi, "iUsUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_kortestzhi, "iUsUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_kunpckhi, "UsUsUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_kxnorhi, "UsUsUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_kxorhi, "UsUsUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_palignr512_mask, "V64cV64cV64cIiV64cULLi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_dbpsadbw128_mask, "V8sV16cV16cIiV8sUc", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_dbpsadbw256_mask, "V16sV32cV32cIiV16sUs", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_dbpsadbw512_mask, "V32sV64cV64cIiV32sUi", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_psadbw512, "V8LLiV64cV64c", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_compressdf512_mask, "V8dV8dV8dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_compressdi512_mask, "V8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_compresshi512_mask, "V32sV32sV32sUi", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_compressqi512_mask, "V64cV64cV64cULLi", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_compresssf512_mask, "V16fV16fV16fUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_compresssi512_mask, "V16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cmpsd_mask, "UcV2dV2dIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cmpss_mask, "UcV4fV4fIiUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_expanddf512_mask, "V8dV8dV8dUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_expanddi512_mask, "V8LLiV8LLiV8LLiUc", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_expandhi512_mask, "V32sV32sV32sUi", "nc", "avx512vbmi2")
+TARGET_BUILTIN(__builtin_ia32_expandqi512_mask, "V64cV64cV64cULLi", "nc", "avx512vbmi2")
 TARGET_BUILTIN(__builtin_ia32_expandloaddf512_mask, "V8dV8dC*V8dUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_expandloaddi512_mask, "V8LLiV8LLiC*V8LLiUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_expandloadhi512_mask, "V32sV32sC*V32sUi", "n", "avx512vbmi2")
 TARGET_BUILTIN(__builtin_ia32_expandloadqi512_mask, "V64cV64cC*V64cULLi", "n", "avx512vbmi2")
 TARGET_BUILTIN(__builtin_ia32_expandloadsf512_mask, "V16fV16fC*V16fUs", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_expandloadsi512_mask, "V16iV16iC*V16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_expandsf512_mask, "V16fV16fV16fUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_expandsi512_mask, "V16iV16iV16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtps2pd512_mask, "V8dV8fV8dUcIi", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_expandsf512_mask, "V16fV16fV16fUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_expandsi512_mask, "V16iV16iV16iUs", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtps2pd512_mask, "V8dV8fV8dUcIi", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_compressstoredf512_mask, "vV8d*V8dUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_compressstoredi512_mask, "vV8LLi*V8LLiUc", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_compressstorehi512_mask, "vV32s*V32sUi", "n", "avx512vbmi2")
 TARGET_BUILTIN(__builtin_ia32_compressstoreqi512_mask, "vV64c*V64cULLi", "n", "avx512vbmi2")
 TARGET_BUILTIN(__builtin_ia32_compressstoresf512_mask, "vV16f*V16fUs", "n", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_compressstoresi512_mask, "vV16i*V16iUs", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtph2ps_mask, "V4fV8sV4fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vcvtph2ps256_mask, "V8fV8sV8fUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vcvtps2ph_mask, "V8sV4fIiV8sUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vcvtps2ph256_mask, "V8sV8fIiV8sUc", "n", "avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtw2mask512, "UiV32s", "n", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_cvtw2mask128, "UcV8s", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtw2mask256, "UsV16s", "n", "avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_cvtsd2ss_round_mask, "V4fV4fV2dV4fUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtsi2ss32, "V4fV4fiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtss2sd_round_mask, "V2dV2dV4fV2dUcIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtusi2sd32, "V2dV2dUi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtusi2ss32, "V4fV4fUiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb512_mask, "V64cV64cV64cV64cULLi", "n", "avx512vbmi")
-TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb128_mask, "V16cV16cV16cV16cUs", "n", "avx512vbmi,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb256_mask, "V32cV32cV32cV32cUi", "n", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vcvtph2ps_mask, "V4fV8sV4fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vcvtph2ps256_mask, "V8fV8sV8fUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vcvtps2ph_mask, "V8sV4fIiV8sUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vcvtps2ph256_mask, "V8sV8fIiV8sUc", "nc", "avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtw2mask512, "UiV32s", "nc", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtw2mask128, "UcV8s", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtw2mask256, "UsV16s", "nc", "avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtsd2ss_round_mask, "V4fV4fV2dV4fUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtsi2ss32, "V4fV4fiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtss2sd_round_mask, "V2dV2dV4fV2dUcIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtusi2sd32, "V2dV2dUi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtusi2ss32, "V4fV4fUiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb512_mask, "V64cV64cV64cV64cULLi", "nc", "avx512vbmi")
+TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb128_mask, "V16cV16cV16cV16cUs", "nc", "avx512vbmi,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vpmultishiftqb256_mask, "V32cV32cV32cV32cUi", "nc", "avx512vbmi,avx512vl")
 
 // generic select intrinsics
-TARGET_BUILTIN(__builtin_ia32_selectb_128, "V16cUsV16cV16c", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectb_256, "V32cUiV32cV32c", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectb_512, "V64cULLiV64cV64c", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectw_128, "V8sUcV8sV8s", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectw_256, "V16sUsV16sV16s", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectw_512, "V32sUiV32sV32s", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectd_128, "V4iUcV4iV4i", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectd_256, "V8iUcV8iV8i", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectd_512, "V16iUsV16iV16i", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectq_128, "V2LLiUcV2LLiV2LLi", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectq_256, "V4LLiUcV4LLiV4LLi", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectq_512, "V8LLiUcV8LLiV8LLi", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectps_128, "V4fUcV4fV4f", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectps_256, "V8fUcV8fV8f", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectps_512, "V16fUsV16fV16f", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectpd_128, "V2dUcV2dV2d", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectpd_256, "V4dUcV4dV4d", "n", "")
-TARGET_BUILTIN(__builtin_ia32_selectpd_512, "V8dUcV8dV8d", "n", "")
+TARGET_BUILTIN(__builtin_ia32_selectb_128, "V16cUsV16cV16c", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectb_256, "V32cUiV32cV32c", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectb_512, "V64cULLiV64cV64c", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectw_128, "V8sUcV8sV8s", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectw_256, "V16sUsV16sV16s", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectw_512, "V32sUiV32sV32s", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectd_128, "V4iUcV4iV4i", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectd_256, "V8iUcV8iV8i", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectd_512, "V16iUsV16iV16i", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectq_128, "V2LLiUcV2LLiV2LLi", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectq_256, "V4LLiUcV4LLiV4LLi", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectq_512, "V8LLiUcV8LLiV8LLi", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectps_128, "V4fUcV4fV4f", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectps_256, "V8fUcV8fV8f", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectps_512, "V16fUsV16fV16f", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectpd_128, "V2dUcV2dV2d", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectpd_256, "V4dUcV4dV4d", "nc", "")
+TARGET_BUILTIN(__builtin_ia32_selectpd_512, "V8dUcV8dV8d", "nc", "")
 
 // MONITORX/MWAITX
 TARGET_BUILTIN(__builtin_ia32_monitorx, "vv*UiUi", "n", "mwaitx")
@@ -1895,8 +1893,8 @@ TARGET_HEADER_BUILTIN(_ReadWriteBarrier,
 TARGET_HEADER_BUILTIN(_ReadBarrier,      "v", "nh", "intrin.h", ALL_MS_LANGUAGES, "")
 TARGET_HEADER_BUILTIN(_WriteBarrier,     "v", "nh", "intrin.h", ALL_MS_LANGUAGES, "")
 
-TARGET_HEADER_BUILTIN(__emul,  "LLiii",    "nh", "intrin.h", ALL_MS_LANGUAGES, "")
-TARGET_HEADER_BUILTIN(__emulu, "ULLiUiUi", "nh", "intrin.h", ALL_MS_LANGUAGES, "")
+TARGET_HEADER_BUILTIN(__emul,  "LLiii",    "nch", "intrin.h", ALL_MS_LANGUAGES, "")
+TARGET_HEADER_BUILTIN(__emulu, "ULLiUiUi", "nch", "intrin.h", ALL_MS_LANGUAGES, "")
 
 TARGET_HEADER_BUILTIN(_AddressOfReturnAddress, "v*", "nh", "intrin.h", ALL_MS_LANGUAGES, "")
 

Modified: cfe/trunk/include/clang/Basic/BuiltinsX86_64.def
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/include/clang/Basic/BuiltinsX86_64.def?rev=331814&r1=331813&r2=331814&view=diff
==============================================================================
--- cfe/trunk/include/clang/Basic/BuiltinsX86_64.def (original)
+++ cfe/trunk/include/clang/Basic/BuiltinsX86_64.def Tue May  8 15:01:43 2018
@@ -27,8 +27,8 @@ TARGET_HEADER_BUILTIN(_BitScanReverse64,
 
 TARGET_HEADER_BUILTIN(__mulh,  "LLiLLiLLi",    "nch", "intrin.h", ALL_MS_LANGUAGES, "")
 TARGET_HEADER_BUILTIN(__umulh, "ULLiULLiULLi", "nch", "intrin.h", ALL_MS_LANGUAGES, "")
-TARGET_HEADER_BUILTIN(_mul128, "LLiLLiLLiLLi*",      "nh",   "intrin.h", ALL_MS_LANGUAGES, "")
-TARGET_HEADER_BUILTIN(_umul128, "ULLiULLiULLiULLi*", "nh",   "intrin.h", ALL_MS_LANGUAGES, "")
+TARGET_HEADER_BUILTIN(_mul128, "LLiLLiLLiLLi*",      "nch",   "intrin.h", ALL_MS_LANGUAGES, "")
+TARGET_HEADER_BUILTIN(_umul128, "ULLiULLiULLiULLi*", "nch",   "intrin.h", ALL_MS_LANGUAGES, "")
 
 TARGET_HEADER_BUILTIN(__faststorefence, "v", "nh", "intrin.h", ALL_MS_LANGUAGES, "")
 
@@ -44,12 +44,12 @@ TARGET_HEADER_BUILTIN(_InterlockedCompar
 
 TARGET_BUILTIN(__builtin_ia32_readeflags_u64, "ULLi", "n", "")
 TARGET_BUILTIN(__builtin_ia32_writeeflags_u64, "vULLi", "n", "")
-TARGET_BUILTIN(__builtin_ia32_cvtss2si64, "LLiV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cvttss2si64, "LLiV4f", "n", "sse")
-TARGET_BUILTIN(__builtin_ia32_cvtsd2si64, "LLiV2d", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_cvttsd2si64, "LLiV2d", "n", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvtss2si64, "LLiV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cvttss2si64, "LLiV4f", "nc", "sse")
+TARGET_BUILTIN(__builtin_ia32_cvtsd2si64, "LLiV2d", "nc", "sse2")
+TARGET_BUILTIN(__builtin_ia32_cvttsd2si64, "LLiV2d", "nc", "sse2")
 TARGET_BUILTIN(__builtin_ia32_movnti64, "vLLi*LLi", "n", "sse2")
-TARGET_BUILTIN(__builtin_ia32_crc32di, "ULLiULLiULLi", "n", "sse4.2")
+TARGET_BUILTIN(__builtin_ia32_crc32di, "ULLiULLiULLi", "nc", "sse4.2")
 TARGET_BUILTIN(__builtin_ia32_rdfsbase32, "Ui", "n", "fsgsbase")
 TARGET_BUILTIN(__builtin_ia32_rdfsbase64, "ULLi", "n", "fsgsbase")
 TARGET_BUILTIN(__builtin_ia32_rdgsbase32, "Ui", "n", "fsgsbase")
@@ -75,25 +75,25 @@ TARGET_BUILTIN(__builtin_ia32_addcarry_u
 TARGET_BUILTIN(__builtin_ia32_subborrow_u64, "UcUcULLiULLiULLi*", "n", "")
 TARGET_BUILTIN(__builtin_ia32_rdrand64_step, "UiULLi*", "n", "rdrnd")
 TARGET_BUILTIN(__builtin_ia32_rdseed64_step, "UiULLi*", "n", "rdseed")
-TARGET_BUILTIN(__builtin_ia32_bextr_u64, "ULLiULLiULLi", "n", "bmi")
-TARGET_BUILTIN(__builtin_ia32_bzhi_di, "ULLiULLiULLi", "n", "bmi2")
-TARGET_BUILTIN(__builtin_ia32_pdep_di, "ULLiULLiULLi", "n", "bmi2")
-TARGET_BUILTIN(__builtin_ia32_pext_di, "ULLiULLiULLi", "n", "bmi2")
-TARGET_BUILTIN(__builtin_ia32_bextri_u64, "ULLiULLiIULLi", "n", "tbm")
+TARGET_BUILTIN(__builtin_ia32_bextr_u64, "ULLiULLiULLi", "nc", "bmi")
+TARGET_BUILTIN(__builtin_ia32_bzhi_di, "ULLiULLiULLi", "nc", "bmi2")
+TARGET_BUILTIN(__builtin_ia32_pdep_di, "ULLiULLiULLi", "nc", "bmi2")
+TARGET_BUILTIN(__builtin_ia32_pext_di, "ULLiULLiULLi", "nc", "bmi2")
+TARGET_BUILTIN(__builtin_ia32_bextri_u64, "ULLiULLiIULLi", "nc", "tbm")
 TARGET_BUILTIN(__builtin_ia32_lwpins64, "UcULLiUiUi", "n", "lwp")
 TARGET_BUILTIN(__builtin_ia32_lwpval64, "vULLiUiUi", "n", "lwp")
-TARGET_BUILTIN(__builtin_ia32_vcvtsd2si64, "LLiV2dIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi64, "ULLiV2dIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtss2si64, "LLiV4fIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvtss2usi64, "ULLiV4fIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvttsd2si64, "LLiV2dIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi64, "ULLiV2dIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvttss2si64, "LLiV4fIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_vcvttss2usi64, "ULLiV4fIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtsi2sd64, "V2dV2dLLiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtsi2ss64, "V4fV4fLLiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtusi2sd64, "V2dV2dULLiIi", "n", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_cvtusi2ss64, "V4fV4fULLiIi", "n", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2si64, "LLiV2dIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi64, "ULLiV2dIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2si64, "LLiV4fIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2usi64, "ULLiV4fIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2si64, "LLiV2dIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi64, "ULLiV2dIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2si64, "LLiV4fIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2usi64, "ULLiV4fIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtsi2sd64, "V2dV2dLLiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtsi2ss64, "V4fV4fLLiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtusi2sd64, "V2dV2dULLiIi", "nc", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_cvtusi2ss64, "V4fV4fULLiIi", "nc", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_directstore_u64, "vULi*ULi", "n", "movdiri")
 
 #undef BUILTIN




More information about the cfe-commits mailing list