[clang] [Clang][LoongArch] Support builtin functions for LSX and LASX (PR #70404)

via cfe-commits cfe-commits at lists.llvm.org
Fri Oct 27 00:42:14 PDT 2023


https://github.com/yjijd updated https://github.com/llvm/llvm-project/pull/70404

>From 298b4d0d51026b7e35945e743528c95f198fd98d Mon Sep 17 00:00:00 2001
From: licongtian <licongtian at loongson.cn>
Date: Wed, 20 Sep 2023 11:21:56 +0800
Subject: [PATCH 1/4] Support compiler options -mlsx/-mlasx for clang

This patch adds compiler options -mlsx/-mlasx which enables the
instruction sets of LSX and LASX, and sets related predefined macros
acording to the options.
---
 .../clang/Basic/DiagnosticDriverKinds.td      |  6 +++
 clang/include/clang/Driver/Options.td         | 10 +++++
 clang/lib/Basic/Targets/LoongArch.cpp         | 12 +++++-
 clang/lib/Basic/Targets/LoongArch.h           |  4 ++
 .../lib/Driver/ToolChains/Arch/LoongArch.cpp  | 32 +++++++++++++++
 clang/test/Driver/loongarch-mlasx-error.c     | 15 +++++++
 clang/test/Driver/loongarch-mlasx.c           | 37 +++++++++++++++++
 clang/test/Driver/loongarch-mlsx-error.c      | 12 ++++++
 clang/test/Driver/loongarch-mlsx.c            | 41 +++++++++++++++++++
 clang/test/Preprocessor/init-loongarch.c      | 35 ++++++++++++++++
 10 files changed, 203 insertions(+), 1 deletion(-)
 create mode 100644 clang/test/Driver/loongarch-mlasx-error.c
 create mode 100644 clang/test/Driver/loongarch-mlasx.c
 create mode 100644 clang/test/Driver/loongarch-mlsx-error.c
 create mode 100644 clang/test/Driver/loongarch-mlsx.c

diff --git a/clang/include/clang/Basic/DiagnosticDriverKinds.td b/clang/include/clang/Basic/DiagnosticDriverKinds.td
index c0ccd64a2a7b82e..676f1a62b49dd0d 100644
--- a/clang/include/clang/Basic/DiagnosticDriverKinds.td
+++ b/clang/include/clang/Basic/DiagnosticDriverKinds.td
@@ -765,6 +765,12 @@ def warn_drv_loongarch_conflicting_implied_val : Warning<
   InGroup<OptionIgnored>;
 def err_drv_loongarch_invalid_mfpu_EQ : Error<
   "invalid argument '%0' to -mfpu=; must be one of: 64, 32, none, 0 (alias for none)">;
+def err_drv_loongarch_wrong_fpu_width_for_lsx : Error<
+  "wrong fpu width; LSX depends on 64-bit FPU.">;
+def err_drv_loongarch_wrong_fpu_width_for_lasx : Error<
+  "wrong fpu width; LASX depends on 64-bit FPU.">;
+def err_drv_loongarch_invalid_simd_option_combination : Error<
+  "invalid option combination; LASX depends on LSX.">;
 
 def err_drv_expand_response_file : Error<
   "failed to expand response file: %0">;
diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
index 7f3f5125d42e7a9..c8b730e0f7ecd84 100644
--- a/clang/include/clang/Driver/Options.td
+++ b/clang/include/clang/Driver/Options.td
@@ -208,6 +208,8 @@ def m_riscv_Features_Group : OptionGroup<"<riscv features group>">,
                              Group<m_Group>, DocName<"RISC-V">;
 def m_ve_Features_Group : OptionGroup<"<ve features group>">,
                           Group<m_Group>, DocName<"VE">;
+def m_loongarch_Features_Group : OptionGroup<"<loongarch features group>">,
+                                 Group<m_Group>, DocName<"LoongArch">;
 
 def m_libc_Group : OptionGroup<"<m libc group>">, Group<m_mips_Features_Group>,
                    Flags<[HelpHidden]>;
@@ -4886,6 +4888,14 @@ def mstack_protector_guard_reg_EQ : Joined<["-"], "mstack-protector-guard-reg=">
 def mfentry : Flag<["-"], "mfentry">, HelpText<"Insert calls to fentry at function entry (x86/SystemZ only)">,
   Visibility<[ClangOption, CC1Option]>, Group<m_Group>,
   MarshallingInfoFlag<CodeGenOpts<"CallFEntry">>;
+def mlsx : Flag<["-"], "mlsx">, Group<m_loongarch_Features_Group>,
+  HelpText<"Enable Loongson SIMD Extension (LSX).">;
+def mno_lsx : Flag<["-"], "mno-lsx">, Group<m_loongarch_Features_Group>,
+  HelpText<"Disable Loongson SIMD Extension (LSX).">;
+def mlasx : Flag<["-"], "mlasx">, Group<m_loongarch_Features_Group>,
+  HelpText<"Enable Loongson Advanced SIMD Extension (LASX).">;
+def mno_lasx : Flag<["-"], "mno-lasx">, Group<m_loongarch_Features_Group>,
+  HelpText<"Disable Loongson Advanced SIMD Extension (LASX).">;
 def mnop_mcount : Flag<["-"], "mnop-mcount">, HelpText<"Generate mcount/__fentry__ calls as nops. To activate they need to be patched in.">,
   Visibility<[ClangOption, CC1Option]>, Group<m_Group>,
   MarshallingInfoFlag<CodeGenOpts<"MNopMCount">>;
diff --git a/clang/lib/Basic/Targets/LoongArch.cpp b/clang/lib/Basic/Targets/LoongArch.cpp
index 4448a2ae10a1725..88537989a05129f 100644
--- a/clang/lib/Basic/Targets/LoongArch.cpp
+++ b/clang/lib/Basic/Targets/LoongArch.cpp
@@ -208,6 +208,11 @@ void LoongArchTargetInfo::getTargetDefines(const LangOptions &Opts,
     TuneCPU = ArchName;
   Builder.defineMacro("__loongarch_tune", Twine('"') + TuneCPU + Twine('"'));
 
+  if (HasFeatureLSX)
+    Builder.defineMacro("__loongarch_sx", Twine(1));
+  if (HasFeatureLASX)
+    Builder.defineMacro("__loongarch_asx", Twine(1));
+
   StringRef ABI = getABI();
   if (ABI == "lp64d" || ABI == "lp64f" || ABI == "lp64s")
     Builder.defineMacro("__loongarch_lp64");
@@ -257,6 +262,8 @@ bool LoongArchTargetInfo::hasFeature(StringRef Feature) const {
       .Case("loongarch64", Is64Bit)
       .Case("32bit", !Is64Bit)
       .Case("64bit", Is64Bit)
+      .Case("lsx", HasFeatureLSX)
+      .Case("lasx", HasFeatureLASX)
       .Default(false);
 }
 
@@ -274,7 +281,10 @@ bool LoongArchTargetInfo::handleTargetFeatures(
       if (Feature == "+d") {
         HasFeatureD = true;
       }
-    }
+    } else if (Feature == "+lsx")
+      HasFeatureLSX = true;
+    else if (Feature == "+lasx")
+      HasFeatureLASX = true;
   }
   return true;
 }
diff --git a/clang/lib/Basic/Targets/LoongArch.h b/clang/lib/Basic/Targets/LoongArch.h
index ba7fb78ab94cd23..3313102492cb8dc 100644
--- a/clang/lib/Basic/Targets/LoongArch.h
+++ b/clang/lib/Basic/Targets/LoongArch.h
@@ -27,12 +27,16 @@ class LLVM_LIBRARY_VISIBILITY LoongArchTargetInfo : public TargetInfo {
   std::string CPU;
   bool HasFeatureD;
   bool HasFeatureF;
+  bool HasFeatureLSX;
+  bool HasFeatureLASX;
 
 public:
   LoongArchTargetInfo(const llvm::Triple &Triple, const TargetOptions &)
       : TargetInfo(Triple) {
     HasFeatureD = false;
     HasFeatureF = false;
+    HasFeatureLSX = false;
+    HasFeatureLASX = false;
     LongDoubleWidth = 128;
     LongDoubleAlign = 128;
     LongDoubleFormat = &llvm::APFloat::IEEEquad();
diff --git a/clang/lib/Driver/ToolChains/Arch/LoongArch.cpp b/clang/lib/Driver/ToolChains/Arch/LoongArch.cpp
index 65925e9ed610107..31153a67ad28403 100644
--- a/clang/lib/Driver/ToolChains/Arch/LoongArch.cpp
+++ b/clang/lib/Driver/ToolChains/Arch/LoongArch.cpp
@@ -175,6 +175,38 @@ void loongarch::getLoongArchTargetFeatures(const Driver &D,
     A->ignoreTargetSpecific();
   if (Arg *A = Args.getLastArgNoClaim(options::OPT_mfpu_EQ))
     A->ignoreTargetSpecific();
+
+  // Select lsx feature determined by -m[no-]lsx.
+  if (const Arg *A = Args.getLastArg(options::OPT_mlsx, options::OPT_mno_lsx)) {
+    // LSX depends on 64-bit FPU.
+    // -m*-float and -mfpu=none/0/32 conflict with -mlsx.
+    if (A->getOption().matches(options::OPT_mlsx)) {
+      if (llvm::find(Features, "-d") != Features.end())
+        D.Diag(diag::err_drv_loongarch_wrong_fpu_width_for_lsx);
+      else /*-mlsx*/
+        Features.push_back("+lsx");
+    } else /*-mno-lsx*/ {
+      Features.push_back("-lsx");
+    }
+  }
+
+  // Select lasx feature determined by -m[no-]lasx.
+  if (const Arg *A =
+          Args.getLastArg(options::OPT_mlasx, options::OPT_mno_lasx)) {
+    // LASX depends on 64-bit FPU and LSX.
+    // -mno-lsx conflicts with -mlasx.
+    if (A->getOption().matches(options::OPT_mlasx)) {
+      if (llvm::find(Features, "-d") != Features.end())
+        D.Diag(diag::err_drv_loongarch_wrong_fpu_width_for_lasx);
+      else if (llvm::find(Features, "-lsx") != Features.end())
+        D.Diag(diag::err_drv_loongarch_invalid_simd_option_combination);
+      else { /*-mlasx*/
+        Features.push_back("+lsx");
+        Features.push_back("+lasx");
+      }
+    } else /*-mno-lasx*/
+      Features.push_back("-lasx");
+  }
 }
 
 std::string loongarch::postProcessTargetCPUString(const std::string &CPU,
diff --git a/clang/test/Driver/loongarch-mlasx-error.c b/clang/test/Driver/loongarch-mlasx-error.c
new file mode 100644
index 000000000000000..e66f277f7c292f1
--- /dev/null
+++ b/clang/test/Driver/loongarch-mlasx-error.c
@@ -0,0 +1,15 @@
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlasx -msingle-float 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LASX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlasx -msoft-float 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LASX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlasx -mfpu=32 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LASX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlasx -mfpu=0 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LASX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlasx -mfpu=none 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LASX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlasx -mno-lsx 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LASX_FPU128 %s
+
+// ERROR_LASX_FPU64: error: wrong fpu width; LASX depends on 64-bit FPU.
+// ERROR_LASX_FPU128: error: invalid option combination; LASX depends on LSX.
diff --git a/clang/test/Driver/loongarch-mlasx.c b/clang/test/Driver/loongarch-mlasx.c
new file mode 100644
index 000000000000000..0b934f125c9e462
--- /dev/null
+++ b/clang/test/Driver/loongarch-mlasx.c
@@ -0,0 +1,37 @@
+/// Test -m[no-]lasx options.
+
+// RUN: %clang --target=loongarch64 -mlasx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-LASX
+// RUN: %clang --target=loongarch64 -mno-lasx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-NOLASX
+// RUN: %clang --target=loongarch64 -mlasx -mno-lasx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-NOLASX
+// RUN: %clang --target=loongarch64 -mno-lasx -mlasx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-LASX
+// RUN: %clang --target=loongarch64 -mlsx -mlasx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-LASX
+// RUN: %clang --target=loongarch64 -mlasx -mlsx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-LASX
+
+// RUN: %clang --target=loongarch64 -mlasx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-LASX
+// RUN: %clang --target=loongarch64 -mno-lasx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-NOLASX
+// RUN: %clang --target=loongarch64 -mlasx -mno-lasx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-NOLASX
+// RUN: %clang --target=loongarch64 -mno-lasx -mlasx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-LASX
+// RUN: %clang --target=loongarch64 -mlsx -mlasx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-LASX
+// RUN: %clang --target=loongarch64 -mlasx -mlsx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-LASX
+
+// CC1-LASX: "-target-feature" "+lsx" "-target-feature" "+lasx"
+// CC1-NOLASX: "-target-feature" "-lasx"
+
+// IR-LASX: attributes #[[#]] ={{.*}}"target-features"="{{(.*,)?}}+lasx{{(,.*)?}}"
+// IR-NOLASX: attributes #[[#]] ={{.*}}"target-features"="{{(.*,)?}}-lasx{{(,.*)?}}"
+
+int foo(void){
+  return 3;
+}
diff --git a/clang/test/Driver/loongarch-mlsx-error.c b/clang/test/Driver/loongarch-mlsx-error.c
new file mode 100644
index 000000000000000..bd6b8e2718bf608
--- /dev/null
+++ b/clang/test/Driver/loongarch-mlsx-error.c
@@ -0,0 +1,12 @@
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlsx -msingle-float 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LSX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlsx -msoft-float 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LSX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlsx -mfpu=32 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LSX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlsx -mfpu=0 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LSX_FPU64 %s
+// RUN: not %clang --target=loongarch64 %s -fsyntax-only -mlsx -mfpu=none 2>&1 \
+// RUN:   FileCheck --check-prefix=ERROR_LSX_FPU64 %s
+
+// ERROR_LSX_FPU64: error: wrong fpu width; LSX depends on 64-bit FPU.
diff --git a/clang/test/Driver/loongarch-mlsx.c b/clang/test/Driver/loongarch-mlsx.c
new file mode 100644
index 000000000000000..7d4307b078e1a06
--- /dev/null
+++ b/clang/test/Driver/loongarch-mlsx.c
@@ -0,0 +1,41 @@
+/// Test -m[no-]lsx options.
+
+// RUN: %clang --target=loongarch64 -mlsx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-LSX
+// RUN: %clang --target=loongarch64 -mno-lsx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-NOLSX
+// RUN: %clang --target=loongarch64 -mlsx -mno-lsx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-NOLSX
+// RUN: %clang --target=loongarch64 -mno-lsx -mlsx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-LSX
+// RUN: %clang --target=loongarch64 -mlsx -mno-lasx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-LSX
+// RUN: %clang --target=loongarch64 -mno-lasx -mlsx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-LSX
+// RUN: %clang --target=loongarch64 -mno-lsx -mno-lasx -fsyntax-only %s -### 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CC1-NOLSX
+
+// RUN: %clang --target=loongarch64 -mlsx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-LSX
+// RUN: %clang --target=loongarch64 -mno-lsx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-NOLSX
+// RUN: %clang --target=loongarch64 -mlsx -mno-lsx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-NOLSX
+// RUN: %clang --target=loongarch64 -mno-lsx -mlsx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-LSX
+// RUN: %clang --target=loongarch64 -mlsx -mno-lasx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-LSX
+// RUN: %clang --target=loongarch64 -mno-lasx -mlsx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-LSX
+// RUN: %clang --target=loongarch64 -mno-lsx -mno-lasx -S -emit-llvm %s -o - | \
+// RUN:   FileCheck %s --check-prefix=IR-NOLSX
+
+// CC1-LSX: "-target-feature" "+lsx"
+// CC1-NOLSX: "-target-feature" "-lsx"
+
+// IR-LSX: attributes #[[#]] ={{.*}}"target-features"="{{(.*,)?}}+lsx{{(,.*)?}}"
+// IR-NOLSX: attributes #[[#]] ={{.*}}"target-features"="{{(.*,)?}}-lsx{{(,.*)?}}"
+
+int foo(void){
+  return 3;
+}
diff --git a/clang/test/Preprocessor/init-loongarch.c b/clang/test/Preprocessor/init-loongarch.c
index 4ef42a921ec0339..e235a728302153a 100644
--- a/clang/test/Preprocessor/init-loongarch.c
+++ b/clang/test/Preprocessor/init-loongarch.c
@@ -807,3 +807,38 @@
 
 // ARCH-TUNE: #define __loongarch_arch "[[ARCH]]"
 // ARCH-TUNE: #define __loongarch_tune "[[TUNE]]"
+
+// RUN: %clang --target=loongarch64 -mlsx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MLSX %s
+// RUN: %clang --target=loongarch64 -mno-lsx -mlsx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MLSX %s
+// RUN: %clang --target=loongarch64 -mlsx -mno-lasx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MLSX %s
+// RUN: %clang --target=loongarch64 -mno-lasx -mlsx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MLSX %s
+// MLSX-NOT: #define __loongarch_asx
+// MLSX: #define __loongarch_sx 1
+
+// RUN: %clang --target=loongarch64 -mlasx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MLASX %s
+// RUN: %clang --target=loongarch64 -mno-lasx -mlasx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MLASX %s
+// RUN: %clang --target=loongarch64 -mlsx -mlasx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MLASX %s
+// RUN: %clang --target=loongarch64 -mlasx -mlsx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MLASX %s
+// MLASX: #define __loongarch_asx 1
+// MLASX: #define __loongarch_sx 1
+
+// RUN: %clang --target=loongarch64 -mno-lsx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MNO-LSX %s
+// RUN: %clang --target=loongarch64 -mlsx -mno-lsx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MNO-LSX %s
+// RUN: %clang --target=loongarch64 -mno-lsx -mno-lasx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MNO-LSX %s
+// RUN: %clang --target=loongarch64 -mno-lasx -mno-lsx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MNO-LSX %s
+// RUN: %clang --target=loongarch64 -mno-lasx -x c -E -dM %s -o - \
+// RUN:   | FileCheck --match-full-lines --check-prefix=MNO-LSX %s
+// MNO-LSX-NOT: #define __loongarch_asx
+// MNO-LSX-NOT: #define __loongarch_sx

>From 2812c79f08452059a598d1a7bd656d86a5fae953 Mon Sep 17 00:00:00 2001
From: licongtian <licongtian at loongson.cn>
Date: Wed, 25 Oct 2023 17:35:32 +0800
Subject: [PATCH 2/4] Add ABI implementation of passing vectors

---
 clang/lib/CodeGen/Targets/LoongArch.cpp | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/clang/lib/CodeGen/Targets/LoongArch.cpp b/clang/lib/CodeGen/Targets/LoongArch.cpp
index 7483bf6d6d1e8e2..26c68c3583b2a19 100644
--- a/clang/lib/CodeGen/Targets/LoongArch.cpp
+++ b/clang/lib/CodeGen/Targets/LoongArch.cpp
@@ -321,6 +321,13 @@ ABIArgInfo LoongArchABIInfo::classifyArgumentType(QualType Ty, bool IsFixed,
     return ABIArgInfo::getDirect();
   }
 
+  // Pass 128-bit/256-bit vector values via vector registers directly.
+  if (Ty->isVectorType() && (((getContext().getTypeSize(Ty) == 128) &&
+                              (getTarget().hasFeature("lsx"))) ||
+                             ((getContext().getTypeSize(Ty) == 256) &&
+                              getTarget().hasFeature("lasx"))))
+    return ABIArgInfo::getDirect();
+
   // Complex types for the *f or *d ABI must be passed directly rather than
   // using CoerceAndExpand.
   if (IsFixed && Ty->isComplexType() && FRLen && FARsLeft >= 2) {

>From 798968604ee06e888f16f2a7f7fe3c268a7df1de Mon Sep 17 00:00:00 2001
From: licongtian <licongtian at loongson.cn>
Date: Wed, 25 Oct 2023 17:41:03 +0800
Subject: [PATCH 3/4] Support the builtin functions for LSX

This patch does the following work:
- Define the builtin functions for LSX
- Add the header file lsxintrin.h
- Add the immediate number range checking for LSX builtins
---
 .../include/clang/Basic/BuiltinsLoongArch.def |   43 +-
 .../clang/Basic/BuiltinsLoongArchBase.def     |   53 +
 .../clang/Basic/BuiltinsLoongArchLSX.def      |  953 +++++
 clang/lib/Headers/CMakeLists.txt              |    1 +
 clang/lib/Headers/lsxintrin.h                 | 3726 +++++++++++++++++
 clang/lib/Sema/SemaChecking.cpp               |  229 +-
 6 files changed, 4965 insertions(+), 40 deletions(-)
 create mode 100644 clang/include/clang/Basic/BuiltinsLoongArchBase.def
 create mode 100644 clang/include/clang/Basic/BuiltinsLoongArchLSX.def
 create mode 100644 clang/lib/Headers/lsxintrin.h

diff --git a/clang/include/clang/Basic/BuiltinsLoongArch.def b/clang/include/clang/Basic/BuiltinsLoongArch.def
index 20510e18fe58c17..9ec19c31095afff 100644
--- a/clang/include/clang/Basic/BuiltinsLoongArch.def
+++ b/clang/include/clang/Basic/BuiltinsLoongArch.def
@@ -15,46 +15,11 @@
 #   define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) BUILTIN(ID, TYPE, ATTRS)
 #endif
 
-// TODO: Support more builtins.
-TARGET_BUILTIN(__builtin_loongarch_cacop_d, "vWiUWiWi", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_cacop_w, "viUii", "nc", "32bit")
-TARGET_BUILTIN(__builtin_loongarch_dbar, "vIUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_ibar, "vIUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_movfcsr2gr, "UiIUi", "nc", "f")
-TARGET_BUILTIN(__builtin_loongarch_movgr2fcsr, "vIUiUi", "nc", "f")
-TARGET_BUILTIN(__builtin_loongarch_break, "vIUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_syscall, "vIUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_cpucfg, "UiUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_asrtle_d, "vWiWi", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_asrtgt_d, "vWiWi", "nc", "64bit")
+// Definition of LoongArch basic builtins.
+#include "clang/Basic/BuiltinsLoongArchBase.def"
 
-TARGET_BUILTIN(__builtin_loongarch_crc_w_b_w, "iii", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_crc_w_h_w, "iii", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_crc_w_w_w, "iii", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_crc_w_d_w, "iWii", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_crcc_w_b_w, "iii", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_crcc_w_h_w, "iii", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_crcc_w_w_w, "iii", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_crcc_w_d_w, "iWii", "nc", "64bit")
-
-TARGET_BUILTIN(__builtin_loongarch_csrrd_w, "UiIUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_csrrd_d, "UWiIUi", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_csrwr_w, "UiUiIUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_csrwr_d, "UWiUWiIUi", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_csrxchg_w, "UiUiUiIUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_csrxchg_d, "UWiUWiUWiIUi", "nc", "64bit")
-
-TARGET_BUILTIN(__builtin_loongarch_iocsrrd_b, "UiUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_iocsrrd_h, "UiUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_iocsrrd_w, "UiUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_iocsrrd_d, "UWiUi", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_iocsrwr_b, "vUiUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_iocsrwr_h, "vUiUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_iocsrwr_w, "vUiUi", "nc", "")
-TARGET_BUILTIN(__builtin_loongarch_iocsrwr_d, "vUWiUi", "nc", "64bit")
-
-TARGET_BUILTIN(__builtin_loongarch_lddir_d, "WiWiIUWi", "nc", "64bit")
-TARGET_BUILTIN(__builtin_loongarch_ldpte_d, "vWiIUWi", "nc", "64bit")
+// Definition of LSX builtins.
+#include "clang/Basic/BuiltinsLoongArchLSX.def"
 
 #undef BUILTIN
 #undef TARGET_BUILTIN
diff --git a/clang/include/clang/Basic/BuiltinsLoongArchBase.def b/clang/include/clang/Basic/BuiltinsLoongArchBase.def
new file mode 100644
index 000000000000000..cbb239223aae3b2
--- /dev/null
+++ b/clang/include/clang/Basic/BuiltinsLoongArchBase.def
@@ -0,0 +1,53 @@
+//============------------ BuiltinsLoongArchBase.def -------------*- C++ -*-==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the LoongArch-specific basic builtin function database.
+// Users of this file must define the BUILTIN macro to make use of this
+// information.
+//
+//===----------------------------------------------------------------------===//
+
+TARGET_BUILTIN(__builtin_loongarch_cacop_d, "vWiUWiWi", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_cacop_w, "viUii", "nc", "32bit")
+TARGET_BUILTIN(__builtin_loongarch_dbar, "vIUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_ibar, "vIUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_movfcsr2gr, "UiIUi", "nc", "f")
+TARGET_BUILTIN(__builtin_loongarch_movgr2fcsr, "vIUiUi", "nc", "f")
+TARGET_BUILTIN(__builtin_loongarch_break, "vIUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_syscall, "vIUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_cpucfg, "UiUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_asrtle_d, "vWiWi", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_asrtgt_d, "vWiWi", "nc", "64bit")
+
+TARGET_BUILTIN(__builtin_loongarch_crc_w_b_w, "iii", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_crc_w_h_w, "iii", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_crc_w_w_w, "iii", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_crc_w_d_w, "iWii", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_crcc_w_b_w, "iii", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_crcc_w_h_w, "iii", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_crcc_w_w_w, "iii", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_crcc_w_d_w, "iWii", "nc", "64bit")
+
+TARGET_BUILTIN(__builtin_loongarch_csrrd_w, "UiIUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_csrrd_d, "UWiIUi", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_csrwr_w, "UiUiIUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_csrwr_d, "UWiUWiIUi", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_csrxchg_w, "UiUiUiIUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_csrxchg_d, "UWiUWiUWiIUi", "nc", "64bit")
+
+TARGET_BUILTIN(__builtin_loongarch_iocsrrd_b, "UiUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_iocsrrd_h, "UiUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_iocsrrd_w, "UiUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_iocsrrd_d, "UWiUi", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_iocsrwr_b, "vUiUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_iocsrwr_h, "vUiUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_iocsrwr_w, "vUiUi", "nc", "")
+TARGET_BUILTIN(__builtin_loongarch_iocsrwr_d, "vUWiUi", "nc", "64bit")
+
+TARGET_BUILTIN(__builtin_loongarch_lddir_d, "WiWiIUWi", "nc", "64bit")
+TARGET_BUILTIN(__builtin_loongarch_ldpte_d, "vWiIUWi", "nc", "64bit")
diff --git a/clang/include/clang/Basic/BuiltinsLoongArchLSX.def b/clang/include/clang/Basic/BuiltinsLoongArchLSX.def
new file mode 100644
index 000000000000000..8e6aec886c50cd9
--- /dev/null
+++ b/clang/include/clang/Basic/BuiltinsLoongArchLSX.def
@@ -0,0 +1,953 @@
+//=============------------- BuiltinsLoongArchLSX.def --------------- C++ -*-=//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the LoongArch-specific LSX builtin function database.
+// Users of this file must define the BUILTIN macro to make use of this
+// information.
+//
+//===----------------------------------------------------------------------===//
+
+TARGET_BUILTIN(__builtin_lsx_vadd_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vadd_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vadd_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vadd_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vadd_q, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsub_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsub_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsub_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsub_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsub_q, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vaddi_bu, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddi_hu, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddi_wu, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddi_du, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsubi_bu, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubi_hu, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubi_wu, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubi_du, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vneg_b, "V16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vneg_h, "V8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vneg_w, "V4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vneg_d, "V2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsadd_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsadd_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsadd_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsadd_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsadd_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsadd_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsadd_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsadd_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssub_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssub_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssub_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssub_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssub_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssub_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssub_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssub_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vhaddw_h_b, "V8SsV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhaddw_w_h, "V4SiV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhaddw_d_w, "V2SLLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhaddw_q_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vhaddw_hu_bu, "V8UsV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhaddw_wu_hu, "V4UiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhaddw_du_wu, "V2ULLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhaddw_qu_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vhsubw_h_b, "V8SsV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhsubw_w_h, "V4SiV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhsubw_d_w, "V2SLLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhsubw_q_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vhsubw_hu_bu, "V8UsV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhsubw_wu_hu, "V4UiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhsubw_du_wu, "V2ULLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vhsubw_qu_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vaddwev_h_b, "V8sV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_w_h, "V4SiV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_d_w, "V2LLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_q_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vaddwod_h_b, "V8sV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_w_h, "V4SiV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_d_w, "V2LLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_q_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsubwev_h_b, "V8sV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwev_w_h, "V4SiV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwev_d_w, "V2LLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwev_q_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsubwod_h_b, "V8sV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwod_w_h, "V4SiV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwod_d_w, "V2LLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwod_q_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vaddwev_h_bu, "V8sV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_w_hu, "V4SiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_d_wu, "V2LLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_q_du, "V2LLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vaddwod_h_bu, "V8sV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_w_hu, "V4SiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_d_wu, "V2LLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_q_du, "V2LLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsubwev_h_bu, "V8sV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwev_w_hu, "V4SiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwev_d_wu, "V2LLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwev_q_du, "V2LLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsubwod_h_bu, "V8sV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwod_w_hu, "V4SiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwod_d_wu, "V2LLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsubwod_q_du, "V2LLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vaddwev_h_bu_b, "V8sV16UcV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_w_hu_h, "V4SiV8UsV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_d_wu_w, "V2LLiV4UiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwev_q_du_d, "V2LLiV2ULLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vaddwod_h_bu_b, "V8sV16UcV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_w_hu_h, "V4SiV8UsV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_d_wu_w, "V2LLiV4UiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vaddwod_q_du_d, "V2LLiV2ULLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vavg_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavg_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavg_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavg_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vavg_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavg_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavg_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavg_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vavgr_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavgr_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavgr_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavgr_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vavgr_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavgr_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavgr_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vavgr_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vabsd_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vabsd_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vabsd_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vabsd_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vabsd_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vabsd_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vabsd_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vabsd_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vadda_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vadda_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vadda_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vadda_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmax_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmax_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmax_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmax_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmaxi_b, "V16ScV16ScIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaxi_h, "V8SsV8SsIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaxi_w, "V4SiV4SiIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaxi_d, "V2SLLiV2SLLiIi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmax_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmax_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmax_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmax_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmaxi_bu, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaxi_hu, "V8UsV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaxi_wu, "V4UiV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaxi_du, "V2ULLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmin_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmin_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmin_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmin_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmini_b, "V16ScV16ScIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmini_h, "V8SsV8SsIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmini_w, "V4SiV4SiIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmini_d, "V2SLLiV2SLLiIi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmin_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmin_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmin_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmin_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmini_bu, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmini_hu, "V8UsV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmini_wu, "V4UiV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmini_du, "V2ULLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmul_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmul_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmul_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmul_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmuh_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmuh_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmuh_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmuh_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmuh_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmuh_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmuh_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmuh_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmulwev_h_b, "V8sV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_w_h, "V4SiV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_d_w, "V2LLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_q_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmulwod_h_b, "V8sV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_w_h, "V4SiV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_d_w, "V2LLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_q_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmulwev_h_bu, "V8sV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_w_hu, "V4SiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_d_wu, "V2LLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_q_du, "V2LLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmulwod_h_bu, "V8sV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_w_hu, "V4SiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_d_wu, "V2LLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_q_du, "V2LLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmulwev_h_bu_b, "V8sV16UcV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_w_hu_h, "V4SiV8UsV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_d_wu_w, "V2LLiV4UiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwev_q_du_d, "V2LLiV2ULLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmulwod_h_bu_b, "V8sV16UcV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_w_hu_h, "V4SiV8UsV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_d_wu_w, "V2LLiV4UiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmulwod_q_du_d, "V2LLiV2ULLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmadd_b, "V16ScV16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmadd_h, "V8SsV8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmadd_w, "V4SiV4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmadd_d, "V2SLLiV2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmsub_b, "V16ScV16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmsub_h, "V8SsV8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmsub_w, "V4SiV4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmsub_d, "V2SLLiV2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_h_b, "V8sV8sV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_w_h, "V4SiV4SiV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_d_w, "V2LLiV2LLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_q_d, "V2LLiV2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_h_b, "V8sV8sV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_w_h, "V4SiV4SiV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_d_w, "V2LLiV2LLiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_q_d, "V2LLiV2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_h_bu, "V8UsV8UsV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_w_hu, "V4UiV4UiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_d_wu, "V2ULLiV2ULLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_q_du, "V2ULLiV2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_h_bu, "V8UsV8UsV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_w_hu, "V4UiV4UiV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_d_wu, "V2ULLiV2ULLiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_q_du, "V2ULLiV2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_h_bu_b, "V8sV8sV16UcV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_w_hu_h, "V4SiV4SiV8UsV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_d_wu_w, "V2LLiV2LLiV4UiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwev_q_du_d, "V2LLiV2LLiV2ULLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_h_bu_b, "V8sV8sV16UcV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_w_hu_h, "V4SiV4SiV8UsV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_d_wu_w, "V2LLiV2LLiV4UiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmaddwod_q_du_d, "V2LLiV2LLiV2ULLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vdiv_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vdiv_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vdiv_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vdiv_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vdiv_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vdiv_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vdiv_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vdiv_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmod_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmod_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmod_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmod_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+
+TARGET_BUILTIN(__builtin_lsx_vmod_bu, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmod_hu, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmod_wu, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmod_du, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsat_b, "V16ScV16ScIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsat_h, "V8SsV8SsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsat_w, "V4SiV4SiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsat_d, "V2SLLiV2SLLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsat_bu, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsat_hu, "V8UsV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsat_wu, "V4UiV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsat_du, "V2ULLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vexth_h_b, "V8sV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vexth_w_h, "V4SiV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vexth_d_w, "V2LLiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vexth_q_d, "V2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vexth_hu_bu, "V8UsV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vexth_wu_hu, "V4UiV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vexth_du_wu, "V2ULLiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vexth_qu_du, "V2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsigncov_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsigncov_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsigncov_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsigncov_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmskltz_b, "V16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmskltz_h, "V8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmskltz_w, "V4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmskltz_d, "V2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vmskgez_b, "V16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vmsknz_b, "V8sV8s", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vldi, "V2LLiIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrepli_b, "V16cIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrepli_h, "V8sIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrepli_w, "V4iIi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrepli_d, "V2LLiIi", "nc", "lsx")
+
+
+TARGET_BUILTIN(__builtin_lsx_vand_v, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vor_v, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vxor_v, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vnor_v, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vandn_v, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vorn_v, "V16ScV16ScV16Sc", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vandi_b, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vori_b, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vxori_b, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vnori_b, "V16UcV16UcIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsll_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsll_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsll_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsll_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vslli_b, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslli_h, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslli_w, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslli_d, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrl_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrl_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrl_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrl_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrli_b, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrli_h, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrli_w, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrli_d, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsra_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsra_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsra_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsra_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrai_b, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrai_h, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrai_w, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrai_d, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vrotr_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrotr_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrotr_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrotr_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vrotri_b, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrotri_h, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrotri_w, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vrotri_d, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsllwil_h_b, "V8sV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsllwil_w_h, "V4SiV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsllwil_d_w, "V2LLiV4SiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vextl_q_d, "V2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsllwil_hu_bu, "V8UsV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsllwil_wu_hu, "V4UiV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsllwil_du_wu, "V2ULLiV4UiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vextl_qu_du, "V2LLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrlr_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlr_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlr_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlr_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrlri_b, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlri_h, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlri_w, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlri_d, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrar_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrar_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrar_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrar_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrari_b, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrari_h, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrari_w, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrari_d, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrln_b_h, "V16ScV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrln_h_w, "V8sV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrln_w_d, "V4SiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsran_b_h, "V16ScV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsran_h_w, "V8sV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsran_w_d, "V4SiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrlni_b_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlni_h_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlni_w_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlni_d_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrani_b_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrani_h_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrani_w_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrani_d_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrlrn_b_h, "V16ScV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlrn_h_w, "V8sV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlrn_w_d, "V4SiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrarn_b_h, "V16ScV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrarn_h_w, "V8sV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrarn_w_d, "V4SiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrlrni_b_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlrni_h_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlrni_w_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrlrni_d_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsrarni_b_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrarni_h_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrarni_w_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsrarni_d_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrln_b_h, "V16ScV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrln_h_w, "V8sV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrln_w_d, "V4SiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssran_b_h, "V16ScV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssran_h_w, "V8sV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssran_w_d, "V4SiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrln_bu_h, "V16UcV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrln_hu_w, "V8UsV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrln_wu_d, "V4UiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssran_bu_h, "V16UcV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssran_hu_w, "V8UsV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssran_wu_d, "V4UiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrlni_b_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlni_h_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlni_w_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlni_d_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrani_b_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrani_h_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrani_w_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrani_d_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrlrni_bu_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrni_hu_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrni_wu_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrni_du_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrani_bu_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrani_hu_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrani_wu_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrani_du_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrlrn_b_h, "V16ScV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrn_h_w, "V8sV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrn_w_d, "V4SiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrarn_b_h, "V16ScV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarn_h_w, "V8sV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarn_w_d, "V4SiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrlrn_bu_h, "V16UcV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrn_hu_w, "V8UsV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrn_wu_d, "V4UiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrarn_bu_h, "V16UcV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarn_hu_w, "V8UsV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarn_wu_d, "V4UiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrlrni_b_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrni_h_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrni_w_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlrni_d_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrarni_b_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarni_h_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarni_w_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarni_d_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrlni_bu_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlni_hu_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlni_wu_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrlni_du_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vssrarni_bu_h, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarni_hu_w, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarni_wu_d, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vssrarni_du_q, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vclo_b, "V16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vclo_h, "V8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vclo_w, "V4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vclo_d, "V2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vclz_b, "V16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vclz_h, "V8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vclz_w, "V4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vclz_d, "V2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vpcnt_b, "V16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpcnt_h, "V8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpcnt_w, "V4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpcnt_d, "V2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbitclr_b, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitclr_h, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitclr_w, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitclr_d, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbitclri_b, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitclri_h, "V8UsV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitclri_w, "V4UiV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitclri_d, "V2ULLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbitset_b, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitset_h, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitset_w, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitset_d, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbitseti_b, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitseti_h, "V8UsV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitseti_w, "V4UiV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitseti_d, "V2ULLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbitrev_b, "V16UcV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitrev_h, "V8UsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitrev_w, "V4UiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitrev_d, "V2ULLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbitrevi_b, "V16UcV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitrevi_h, "V8UsV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitrevi_w, "V4UiV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbitrevi_d, "V2ULLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrstp_b, "V16ScV16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrstp_h, "V8SsV8SsV8SsV8Ss", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrstpi_b, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrstpi_h, "V8sV8sV8sIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfadd_s, "V4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfadd_d, "V2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfsub_s, "V4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfsub_d, "V2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfmul_s, "V4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfmul_d, "V2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfdiv_s, "V4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfdiv_d, "V2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfmadd_s, "V4fV4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfmadd_d, "V2dV2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfmsub_s, "V4fV4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfmsub_d, "V2dV2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfnmadd_s, "V4fV4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfnmadd_d, "V2dV2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfnmsub_s, "V4fV4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfnmsub_d, "V2dV2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfmax_s, "V4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfmax_d, "V2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfmin_s, "V4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfmin_d, "V2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfmaxa_s, "V4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfmaxa_d, "V2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfmina_s, "V4fV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfmina_d, "V2dV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vflogb_s, "V4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vflogb_d, "V2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfclass_s, "V4iV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfclass_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfsqrt_s, "V4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfsqrt_d, "V2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrecip_s, "V4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrecip_d, "V2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrsqrt_s, "V4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrsqrt_d, "V2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcvtl_s_h, "V4fV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcvtl_d_s, "V2dV4f", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcvth_s_h, "V4fV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcvth_d_s, "V2dV4f", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcvt_h_s, "V8sV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcvt_s_d, "V4fV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrintrne_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrintrne_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrintrz_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrintrz_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrintrp_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrintrp_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrintrm_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrintrm_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfrint_s, "V4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfrint_d, "V2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrne_w_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrne_l_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrz_w_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrz_l_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrp_w_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrp_l_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrm_w_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrm_l_d, "V2LLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftint_w_s, "V4SiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftint_l_d, "V2SLLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrz_wu_s, "V4UiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrz_lu_d, "V2ULLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftint_wu_s, "V4UiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftint_lu_d, "V2ULLiV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrne_w_d, "V4SiV2dV2d", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrz_w_d, "V4SiV2dV2d", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrp_w_d, "V4SiV2dV2d", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrm_w_d, "V4SiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftint_w_d, "V4SiV2dV2d", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrnel_l_s, "V2LLiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrneh_l_s, "V2LLiV4f", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrzl_l_s, "V2LLiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrzh_l_s, "V2LLiV4f", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrpl_l_s, "V2LLiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrph_l_s, "V2LLiV4f", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintrml_l_s, "V2LLiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftintrmh_l_s, "V2LLiV4f", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vftintl_l_s, "V2LLiV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vftinth_l_s, "V2LLiV4f", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vffint_s_w, "V4fV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vffint_d_l, "V2dV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vffint_s_wu, "V4fV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vffint_d_lu, "V2dV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vffintl_d_w, "V2dV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vffinth_d_w, "V2dV4Si", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vffint_s_l, "V4fV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vseq_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vseq_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vseq_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vseq_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vseqi_b, "V16ScV16ScISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vseqi_h, "V8SsV8SsISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vseqi_w, "V4SiV4SiISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vseqi_d, "V2SLLiV2SLLiISi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsle_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsle_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsle_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsle_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vslei_b, "V16ScV16ScISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslei_h, "V8SsV8SsISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslei_w, "V4SiV4SiISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslei_d, "V2SLLiV2SLLiISi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vsle_bu, "V16ScV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsle_hu, "V8SsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsle_wu, "V4SiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vsle_du, "V2SLLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vslei_bu, "V16ScV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslei_hu, "V8SsV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslei_wu, "V4SiV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslei_du, "V2SLLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vslt_b, "V16ScV16ScV16Sc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslt_h, "V8SsV8SsV8Ss", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslt_w, "V4SiV4SiV4Si", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslt_d, "V2SLLiV2SLLiV2SLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vslti_b, "V16ScV16ScISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslti_h, "V8SsV8SsISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslti_w, "V4SiV4SiISi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslti_d, "V2SLLiV2SLLiISi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vslt_bu, "V16ScV16UcV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslt_hu, "V8SsV8UsV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslt_wu, "V4SiV4UiV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslt_du, "V2SLLiV2ULLiV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vslti_bu, "V16ScV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslti_hu, "V8SsV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslti_wu, "V4SiV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vslti_du, "V2SLLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_caf_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_caf_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cun_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cun_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_ceq_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_ceq_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cueq_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cueq_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_clt_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_clt_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cult_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cult_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cle_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cle_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cule_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cule_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cne_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cne_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cor_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cor_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cune_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_cune_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_saf_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_saf_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sun_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sun_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_seq_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_seq_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sueq_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sueq_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_slt_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_slt_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sult_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sult_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sle_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sle_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sule_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sule_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sne_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sne_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sor_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sor_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sune_s, "V4SiV4fV4f", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vfcmp_sune_d, "V2SLLiV2dV2d", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbitsel_v, "V16UcV16UcV16UcV16Uc", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbitseli_b, "V16UcV16UcV16UcIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vreplgr2vr_b, "V16Sci", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplgr2vr_h, "V8Ssi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplgr2vr_w, "V4Sii", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplgr2vr_d, "V2SLLiLLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vinsgr2vr_b, "V16ScV16SciIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vinsgr2vr_h, "V8SsV8SsiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vinsgr2vr_w, "V4SiV4SiiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vinsgr2vr_d, "V2SLLiV2SLLiLLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vpickve2gr_b, "iV16ScIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickve2gr_h, "iV8SsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickve2gr_w, "iV4SiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickve2gr_d, "LLiV2SLLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vpickve2gr_bu, "iV16UcIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickve2gr_hu, "iV8UsIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickve2gr_wu, "iV4UiIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickve2gr_du, "LLiV2ULLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vreplve_b, "V16cV16cUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplve_h, "V8sV8sUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplve_w, "V4iV4iUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplve_d, "V2LLiV2LLiUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vreplvei_b, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplvei_h, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplvei_w, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vreplvei_d, "V2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vbsll_v, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vbsrl_v, "V16cV16cIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vpackev_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpackev_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpackev_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpackev_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vpackod_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpackod_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpackod_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpackod_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vpickev_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickev_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickev_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickev_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vpickod_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickod_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickod_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vpickod_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vilvl_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vilvl_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vilvl_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vilvl_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vilvh_b, "V16cV16cV16c", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vilvh_h, "V8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vilvh_w, "V4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vilvh_d, "V2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vshuf_b, "V16UcV16UcV16UcV16Uc", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vshuf_h, "V8sV8sV8sV8s", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vshuf_w, "V4iV4iV4iV4i", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vshuf_d, "V2LLiV2LLiV2LLiV2LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vshuf4i_b, "V16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vshuf4i_h, "V8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vshuf4i_w, "V4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vshuf4i_d, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vpermi_w, "V4iV4iV4iIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vextrins_b, "V16cV16cV16cIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vextrins_h, "V8sV8sV8sIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vextrins_w, "V4iV4iV4iIUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vextrins_d, "V2LLiV2LLiV2LLiIUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vld, "V16ScvC*Ii", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vst, "vV16Scv*Ii", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vldx, "V16ScvC*LLi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vstx, "vV16Scv*LLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vldrepl_b, "V16cvC*Ii", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vldrepl_h, "V8svC*Ii", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vldrepl_w, "V4ivC*Ii", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vldrepl_d, "V2LLivC*Ii", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_vstelm_b, "vV16Scv*IiUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vstelm_h, "vV8Ssv*IiUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vstelm_w, "vV4Siv*IiUi", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_vstelm_d, "vV2SLLiv*IiUi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_bz_v, "iV16Uc", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_bnz_v, "iV16Uc", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_bz_b, "iV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_bz_h, "iV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_bz_w, "iV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_bz_d, "iV2ULLi", "nc", "lsx")
+
+TARGET_BUILTIN(__builtin_lsx_bnz_b, "iV16Uc", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_bnz_h, "iV8Us", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_bnz_w, "iV4Ui", "nc", "lsx")
+TARGET_BUILTIN(__builtin_lsx_bnz_d, "iV2ULLi", "nc", "lsx")
diff --git a/clang/lib/Headers/CMakeLists.txt b/clang/lib/Headers/CMakeLists.txt
index 02a0c81644b6c6d..cee61a7e4fb0883 100644
--- a/clang/lib/Headers/CMakeLists.txt
+++ b/clang/lib/Headers/CMakeLists.txt
@@ -93,6 +93,7 @@ set(hlsl_files
 
 set(loongarch_files
   larchintrin.h
+  lsxintrin.h
   )
 
 set(mips_msa_files
diff --git a/clang/lib/Headers/lsxintrin.h b/clang/lib/Headers/lsxintrin.h
new file mode 100644
index 000000000000000..a29bc7757ab5680
--- /dev/null
+++ b/clang/lib/Headers/lsxintrin.h
@@ -0,0 +1,3726 @@
+/*===------------- lsxintrin.h - LoongArch LSX intrinsics ------------------===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+ *
+ *===-----------------------------------------------------------------------===
+ */
+
+#ifndef _LOONGSON_SXINTRIN_H
+#define _LOONGSON_SXINTRIN_H 1
+
+#if defined(__loongarch_sx)
+typedef signed char v16i8 __attribute__((vector_size(16), aligned(16)));
+typedef signed char v16i8_b __attribute__((vector_size(16), aligned(1)));
+typedef unsigned char v16u8 __attribute__((vector_size(16), aligned(16)));
+typedef unsigned char v16u8_b __attribute__((vector_size(16), aligned(1)));
+typedef short v8i16 __attribute__((vector_size(16), aligned(16)));
+typedef short v8i16_h __attribute__((vector_size(16), aligned(2)));
+typedef unsigned short v8u16 __attribute__((vector_size(16), aligned(16)));
+typedef unsigned short v8u16_h __attribute__((vector_size(16), aligned(2)));
+typedef int v4i32 __attribute__((vector_size(16), aligned(16)));
+typedef int v4i32_w __attribute__((vector_size(16), aligned(4)));
+typedef unsigned int v4u32 __attribute__((vector_size(16), aligned(16)));
+typedef unsigned int v4u32_w __attribute__((vector_size(16), aligned(4)));
+typedef long long v2i64 __attribute__((vector_size(16), aligned(16)));
+typedef long long v2i64_d __attribute__((vector_size(16), aligned(8)));
+typedef unsigned long long v2u64 __attribute__((vector_size(16), aligned(16)));
+typedef unsigned long long v2u64_d __attribute__((vector_size(16), aligned(8)));
+typedef float v4f32 __attribute__((vector_size(16), aligned(16)));
+typedef float v4f32_w __attribute__((vector_size(16), aligned(4)));
+typedef double v2f64 __attribute__((vector_size(16), aligned(16)));
+typedef double v2f64_d __attribute__((vector_size(16), aligned(8)));
+
+typedef long long __m128i __attribute__((__vector_size__(16), __may_alias__));
+typedef float __m128 __attribute__((__vector_size__(16), __may_alias__));
+typedef double __m128d __attribute__((__vector_size__(16), __may_alias__));
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsll_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsll_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsll_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsll_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsll_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsll_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsll_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsll_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vslli_b(/*__m128i*/ _1, /*ui3*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslli_b((v16i8)(_1), (_2)))
+
+#define __lsx_vslli_h(/*__m128i*/ _1, /*ui4*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslli_h((v8i16)(_1), (_2)))
+
+#define __lsx_vslli_w(/*__m128i*/ _1, /*ui5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslli_w((v4i32)(_1), (_2)))
+
+#define __lsx_vslli_d(/*__m128i*/ _1, /*ui6*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslli_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsra_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsra_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsra_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsra_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsra_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsra_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsra_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsra_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vsrai_b(/*__m128i*/ _1, /*ui3*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsrai_b((v16i8)(_1), (_2)))
+
+#define __lsx_vsrai_h(/*__m128i*/ _1, /*ui4*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsrai_h((v8i16)(_1), (_2)))
+
+#define __lsx_vsrai_w(/*__m128i*/ _1, /*ui5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsrai_w((v4i32)(_1), (_2)))
+
+#define __lsx_vsrai_d(/*__m128i*/ _1, /*ui6*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsrai_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrar_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrar_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrar_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrar_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrar_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrar_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrar_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrar_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vsrari_b(/*__m128i*/ _1, /*ui3*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsrari_b((v16i8)(_1), (_2)))
+
+#define __lsx_vsrari_h(/*__m128i*/ _1, /*ui4*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsrari_h((v8i16)(_1), (_2)))
+
+#define __lsx_vsrari_w(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsrari_w((v4i32)(_1), (_2)))
+
+#define __lsx_vsrari_d(/*__m128i*/ _1, /*ui6*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsrari_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrl_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrl_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrl_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrl_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrl_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrl_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrl_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrl_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vsrli_b(/*__m128i*/ _1, /*ui3*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsrli_b((v16i8)(_1), (_2)))
+
+#define __lsx_vsrli_h(/*__m128i*/ _1, /*ui4*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsrli_h((v8i16)(_1), (_2)))
+
+#define __lsx_vsrli_w(/*__m128i*/ _1, /*ui5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsrli_w((v4i32)(_1), (_2)))
+
+#define __lsx_vsrli_d(/*__m128i*/ _1, /*ui6*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsrli_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrlr_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrlr_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrlr_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrlr_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrlr_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrlr_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrlr_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrlr_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vsrlri_b(/*__m128i*/ _1, /*ui3*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsrlri_b((v16i8)(_1), (_2)))
+
+#define __lsx_vsrlri_h(/*__m128i*/ _1, /*ui4*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsrlri_h((v8i16)(_1), (_2)))
+
+#define __lsx_vsrlri_w(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsrlri_w((v4i32)(_1), (_2)))
+
+#define __lsx_vsrlri_d(/*__m128i*/ _1, /*ui6*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsrlri_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitclr_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitclr_b((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitclr_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitclr_h((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitclr_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitclr_w((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitclr_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitclr_d((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vbitclri_b(/*__m128i*/ _1, /*ui3*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitclri_b((v16u8)(_1), (_2)))
+
+#define __lsx_vbitclri_h(/*__m128i*/ _1, /*ui4*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitclri_h((v8u16)(_1), (_2)))
+
+#define __lsx_vbitclri_w(/*__m128i*/ _1, /*ui5*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitclri_w((v4u32)(_1), (_2)))
+
+#define __lsx_vbitclri_d(/*__m128i*/ _1, /*ui6*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitclri_d((v2u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitset_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitset_b((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitset_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitset_h((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitset_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitset_w((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitset_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitset_d((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vbitseti_b(/*__m128i*/ _1, /*ui3*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitseti_b((v16u8)(_1), (_2)))
+
+#define __lsx_vbitseti_h(/*__m128i*/ _1, /*ui4*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitseti_h((v8u16)(_1), (_2)))
+
+#define __lsx_vbitseti_w(/*__m128i*/ _1, /*ui5*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitseti_w((v4u32)(_1), (_2)))
+
+#define __lsx_vbitseti_d(/*__m128i*/ _1, /*ui6*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitseti_d((v2u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitrev_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitrev_b((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitrev_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitrev_h((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitrev_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitrev_w((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitrev_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vbitrev_d((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vbitrevi_b(/*__m128i*/ _1, /*ui3*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitrevi_b((v16u8)(_1), (_2)))
+
+#define __lsx_vbitrevi_h(/*__m128i*/ _1, /*ui4*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitrevi_h((v8u16)(_1), (_2)))
+
+#define __lsx_vbitrevi_w(/*__m128i*/ _1, /*ui5*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitrevi_w((v4u32)(_1), (_2)))
+
+#define __lsx_vbitrevi_d(/*__m128i*/ _1, /*ui6*/ _2)                           \
+  ((__m128i)__builtin_lsx_vbitrevi_d((v2u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadd_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadd_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadd_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadd_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadd_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadd_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadd_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadd_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vaddi_bu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vaddi_bu((v16i8)(_1), (_2)))
+
+#define __lsx_vaddi_hu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vaddi_hu((v8i16)(_1), (_2)))
+
+#define __lsx_vaddi_wu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vaddi_wu((v4i32)(_1), (_2)))
+
+#define __lsx_vaddi_du(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vaddi_du((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsub_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsub_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsub_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsub_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsub_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsub_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsub_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsub_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vsubi_bu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsubi_bu((v16i8)(_1), (_2)))
+
+#define __lsx_vsubi_hu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsubi_hu((v8i16)(_1), (_2)))
+
+#define __lsx_vsubi_wu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsubi_wu((v4i32)(_1), (_2)))
+
+#define __lsx_vsubi_du(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vsubi_du((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmax_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmax_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmax_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmax_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmax_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmax_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmax_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmax_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vmaxi_b(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vmaxi_b((v16i8)(_1), (_2)))
+
+#define __lsx_vmaxi_h(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vmaxi_h((v8i16)(_1), (_2)))
+
+#define __lsx_vmaxi_w(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vmaxi_w((v4i32)(_1), (_2)))
+
+#define __lsx_vmaxi_d(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vmaxi_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmax_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmax_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmax_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmax_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmax_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmax_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmax_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmax_du((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vmaxi_bu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vmaxi_bu((v16u8)(_1), (_2)))
+
+#define __lsx_vmaxi_hu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vmaxi_hu((v8u16)(_1), (_2)))
+
+#define __lsx_vmaxi_wu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vmaxi_wu((v4u32)(_1), (_2)))
+
+#define __lsx_vmaxi_du(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vmaxi_du((v2u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmin_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmin_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmin_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmin_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmin_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmin_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmin_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmin_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vmini_b(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vmini_b((v16i8)(_1), (_2)))
+
+#define __lsx_vmini_h(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vmini_h((v8i16)(_1), (_2)))
+
+#define __lsx_vmini_w(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vmini_w((v4i32)(_1), (_2)))
+
+#define __lsx_vmini_d(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vmini_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmin_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmin_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmin_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmin_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmin_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmin_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmin_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmin_du((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vmini_bu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vmini_bu((v16u8)(_1), (_2)))
+
+#define __lsx_vmini_hu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vmini_hu((v8u16)(_1), (_2)))
+
+#define __lsx_vmini_wu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vmini_wu((v4u32)(_1), (_2)))
+
+#define __lsx_vmini_du(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vmini_du((v2u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vseq_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vseq_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vseq_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vseq_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vseq_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vseq_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vseq_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vseq_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vseqi_b(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vseqi_b((v16i8)(_1), (_2)))
+
+#define __lsx_vseqi_h(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vseqi_h((v8i16)(_1), (_2)))
+
+#define __lsx_vseqi_w(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vseqi_w((v4i32)(_1), (_2)))
+
+#define __lsx_vseqi_d(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vseqi_d((v2i64)(_1), (_2)))
+
+#define __lsx_vslti_b(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslti_b((v16i8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vslt_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vslt_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vslt_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vslt_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vslt_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vslt_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vslt_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vslt_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vslti_h(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslti_h((v8i16)(_1), (_2)))
+
+#define __lsx_vslti_w(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslti_w((v4i32)(_1), (_2)))
+
+#define __lsx_vslti_d(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslti_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vslt_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vslt_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vslt_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vslt_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vslt_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vslt_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vslt_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vslt_du((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vslti_bu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vslti_bu((v16u8)(_1), (_2)))
+
+#define __lsx_vslti_hu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vslti_hu((v8u16)(_1), (_2)))
+
+#define __lsx_vslti_wu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vslti_wu((v4u32)(_1), (_2)))
+
+#define __lsx_vslti_du(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vslti_du((v2u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsle_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsle_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsle_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsle_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsle_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsle_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsle_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsle_d((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vslei_b(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslei_b((v16i8)(_1), (_2)))
+
+#define __lsx_vslei_h(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslei_h((v8i16)(_1), (_2)))
+
+#define __lsx_vslei_w(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslei_w((v4i32)(_1), (_2)))
+
+#define __lsx_vslei_d(/*__m128i*/ _1, /*si5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vslei_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsle_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsle_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsle_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsle_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsle_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsle_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsle_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsle_du((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vslei_bu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vslei_bu((v16u8)(_1), (_2)))
+
+#define __lsx_vslei_hu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vslei_hu((v8u16)(_1), (_2)))
+
+#define __lsx_vslei_wu(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vslei_wu((v4u32)(_1), (_2)))
+
+#define __lsx_vslei_du(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vslei_du((v2u64)(_1), (_2)))
+
+#define __lsx_vsat_b(/*__m128i*/ _1, /*ui3*/ _2)                               \
+  ((__m128i)__builtin_lsx_vsat_b((v16i8)(_1), (_2)))
+
+#define __lsx_vsat_h(/*__m128i*/ _1, /*ui4*/ _2)                               \
+  ((__m128i)__builtin_lsx_vsat_h((v8i16)(_1), (_2)))
+
+#define __lsx_vsat_w(/*__m128i*/ _1, /*ui5*/ _2)                               \
+  ((__m128i)__builtin_lsx_vsat_w((v4i32)(_1), (_2)))
+
+#define __lsx_vsat_d(/*__m128i*/ _1, /*ui6*/ _2)                               \
+  ((__m128i)__builtin_lsx_vsat_d((v2i64)(_1), (_2)))
+
+#define __lsx_vsat_bu(/*__m128i*/ _1, /*ui3*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsat_bu((v16u8)(_1), (_2)))
+
+#define __lsx_vsat_hu(/*__m128i*/ _1, /*ui4*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsat_hu((v8u16)(_1), (_2)))
+
+#define __lsx_vsat_wu(/*__m128i*/ _1, /*ui5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsat_wu((v4u32)(_1), (_2)))
+
+#define __lsx_vsat_du(/*__m128i*/ _1, /*ui6*/ _2)                              \
+  ((__m128i)__builtin_lsx_vsat_du((v2u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadda_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadda_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadda_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadda_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadda_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadda_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadda_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadda_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsadd_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsadd_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsadd_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsadd_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsadd_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsadd_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsadd_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsadd_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsadd_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsadd_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsadd_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsadd_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsadd_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsadd_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsadd_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsadd_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavg_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavg_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavg_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavg_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavg_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavg_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavg_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavg_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavg_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavg_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavg_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavg_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavg_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavg_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavg_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavg_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavgr_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavgr_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavgr_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavgr_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavgr_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavgr_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavgr_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavgr_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavgr_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavgr_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavgr_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavgr_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavgr_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavgr_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vavgr_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vavgr_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssub_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssub_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssub_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssub_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssub_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssub_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssub_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssub_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssub_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssub_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssub_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssub_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssub_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssub_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssub_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssub_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vabsd_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vabsd_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vabsd_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vabsd_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vabsd_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vabsd_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vabsd_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vabsd_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vabsd_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vabsd_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vabsd_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vabsd_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vabsd_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vabsd_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vabsd_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vabsd_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmul_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmul_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmul_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmul_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmul_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmul_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmul_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmul_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmadd_b(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmadd_b((v16i8)_1, (v16i8)_2, (v16i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmadd_h(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmadd_h((v8i16)_1, (v8i16)_2, (v8i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmadd_w(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmadd_w((v4i32)_1, (v4i32)_2, (v4i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmadd_d(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmadd_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmsub_b(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmsub_b((v16i8)_1, (v16i8)_2, (v16i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmsub_h(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmsub_h((v8i16)_1, (v8i16)_2, (v8i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmsub_w(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmsub_w((v4i32)_1, (v4i32)_2, (v4i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmsub_d(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmsub_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vdiv_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vdiv_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vdiv_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vdiv_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vdiv_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vdiv_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vdiv_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vdiv_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vdiv_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vdiv_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vdiv_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vdiv_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vdiv_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vdiv_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vdiv_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vdiv_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhaddw_h_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhaddw_h_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhaddw_w_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhaddw_w_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhaddw_d_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhaddw_d_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhaddw_hu_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhaddw_hu_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhaddw_wu_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhaddw_wu_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhaddw_du_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhaddw_du_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhsubw_h_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhsubw_h_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhsubw_w_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhsubw_w_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhsubw_d_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhsubw_d_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhsubw_hu_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhsubw_hu_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhsubw_wu_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhsubw_wu_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhsubw_du_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhsubw_du_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmod_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmod_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmod_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmod_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmod_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmod_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmod_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmod_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmod_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmod_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmod_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmod_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmod_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmod_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmod_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmod_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vreplve_b(__m128i _1, int _2) {
+  return (__m128i)__builtin_lsx_vreplve_b((v16i8)_1, (int)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vreplve_h(__m128i _1, int _2) {
+  return (__m128i)__builtin_lsx_vreplve_h((v8i16)_1, (int)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vreplve_w(__m128i _1, int _2) {
+  return (__m128i)__builtin_lsx_vreplve_w((v4i32)_1, (int)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vreplve_d(__m128i _1, int _2) {
+  return (__m128i)__builtin_lsx_vreplve_d((v2i64)_1, (int)_2);
+}
+
+#define __lsx_vreplvei_b(/*__m128i*/ _1, /*ui4*/ _2)                           \
+  ((__m128i)__builtin_lsx_vreplvei_b((v16i8)(_1), (_2)))
+
+#define __lsx_vreplvei_h(/*__m128i*/ _1, /*ui3*/ _2)                           \
+  ((__m128i)__builtin_lsx_vreplvei_h((v8i16)(_1), (_2)))
+
+#define __lsx_vreplvei_w(/*__m128i*/ _1, /*ui2*/ _2)                           \
+  ((__m128i)__builtin_lsx_vreplvei_w((v4i32)(_1), (_2)))
+
+#define __lsx_vreplvei_d(/*__m128i*/ _1, /*ui1*/ _2)                           \
+  ((__m128i)__builtin_lsx_vreplvei_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpickev_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpickev_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpickev_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpickev_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpickev_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpickev_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpickev_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpickev_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpickod_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpickod_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpickod_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpickod_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpickod_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpickod_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpickod_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpickod_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vilvh_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vilvh_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vilvh_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vilvh_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vilvh_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vilvh_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vilvh_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vilvh_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vilvl_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vilvl_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vilvl_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vilvl_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vilvl_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vilvl_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vilvl_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vilvl_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpackev_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpackev_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpackev_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpackev_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpackev_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpackev_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpackev_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpackev_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpackod_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpackod_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpackod_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpackod_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpackod_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpackod_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpackod_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vpackod_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vshuf_h(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vshuf_h((v8i16)_1, (v8i16)_2, (v8i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vshuf_w(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vshuf_w((v4i32)_1, (v4i32)_2, (v4i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vshuf_d(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vshuf_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vand_v(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vand_v((v16u8)_1, (v16u8)_2);
+}
+
+#define __lsx_vandi_b(/*__m128i*/ _1, /*ui8*/ _2)                              \
+  ((__m128i)__builtin_lsx_vandi_b((v16u8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vor_v(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vor_v((v16u8)_1, (v16u8)_2);
+}
+
+#define __lsx_vori_b(/*__m128i*/ _1, /*ui8*/ _2)                               \
+  ((__m128i)__builtin_lsx_vori_b((v16u8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vnor_v(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vnor_v((v16u8)_1, (v16u8)_2);
+}
+
+#define __lsx_vnori_b(/*__m128i*/ _1, /*ui8*/ _2)                              \
+  ((__m128i)__builtin_lsx_vnori_b((v16u8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vxor_v(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vxor_v((v16u8)_1, (v16u8)_2);
+}
+
+#define __lsx_vxori_b(/*__m128i*/ _1, /*ui8*/ _2)                              \
+  ((__m128i)__builtin_lsx_vxori_b((v16u8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vbitsel_v(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vbitsel_v((v16u8)_1, (v16u8)_2, (v16u8)_3);
+}
+
+#define __lsx_vbitseli_b(/*__m128i*/ _1, /*__m128i*/ _2, /*ui8*/ _3)           \
+  ((__m128i)__builtin_lsx_vbitseli_b((v16u8)(_1), (v16u8)(_2), (_3)))
+
+#define __lsx_vshuf4i_b(/*__m128i*/ _1, /*ui8*/ _2)                            \
+  ((__m128i)__builtin_lsx_vshuf4i_b((v16i8)(_1), (_2)))
+
+#define __lsx_vshuf4i_h(/*__m128i*/ _1, /*ui8*/ _2)                            \
+  ((__m128i)__builtin_lsx_vshuf4i_h((v8i16)(_1), (_2)))
+
+#define __lsx_vshuf4i_w(/*__m128i*/ _1, /*ui8*/ _2)                            \
+  ((__m128i)__builtin_lsx_vshuf4i_w((v4i32)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vreplgr2vr_b(int _1) {
+  return (__m128i)__builtin_lsx_vreplgr2vr_b((int)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vreplgr2vr_h(int _1) {
+  return (__m128i)__builtin_lsx_vreplgr2vr_h((int)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vreplgr2vr_w(int _1) {
+  return (__m128i)__builtin_lsx_vreplgr2vr_w((int)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vreplgr2vr_d(long int _1) {
+  return (__m128i)__builtin_lsx_vreplgr2vr_d((long int)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpcnt_b(__m128i _1) {
+  return (__m128i)__builtin_lsx_vpcnt_b((v16i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpcnt_h(__m128i _1) {
+  return (__m128i)__builtin_lsx_vpcnt_h((v8i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpcnt_w(__m128i _1) {
+  return (__m128i)__builtin_lsx_vpcnt_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vpcnt_d(__m128i _1) {
+  return (__m128i)__builtin_lsx_vpcnt_d((v2i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vclo_b(__m128i _1) {
+  return (__m128i)__builtin_lsx_vclo_b((v16i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vclo_h(__m128i _1) {
+  return (__m128i)__builtin_lsx_vclo_h((v8i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vclo_w(__m128i _1) {
+  return (__m128i)__builtin_lsx_vclo_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vclo_d(__m128i _1) {
+  return (__m128i)__builtin_lsx_vclo_d((v2i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vclz_b(__m128i _1) {
+  return (__m128i)__builtin_lsx_vclz_b((v16i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vclz_h(__m128i _1) {
+  return (__m128i)__builtin_lsx_vclz_h((v8i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vclz_w(__m128i _1) {
+  return (__m128i)__builtin_lsx_vclz_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vclz_d(__m128i _1) {
+  return (__m128i)__builtin_lsx_vclz_d((v2i64)_1);
+}
+
+#define __lsx_vpickve2gr_b(/*__m128i*/ _1, /*ui4*/ _2)                         \
+  ((int)__builtin_lsx_vpickve2gr_b((v16i8)(_1), (_2)))
+
+#define __lsx_vpickve2gr_h(/*__m128i*/ _1, /*ui3*/ _2)                         \
+  ((int)__builtin_lsx_vpickve2gr_h((v8i16)(_1), (_2)))
+
+#define __lsx_vpickve2gr_w(/*__m128i*/ _1, /*ui2*/ _2)                         \
+  ((int)__builtin_lsx_vpickve2gr_w((v4i32)(_1), (_2)))
+
+#define __lsx_vpickve2gr_d(/*__m128i*/ _1, /*ui1*/ _2)                         \
+  ((long int)__builtin_lsx_vpickve2gr_d((v2i64)(_1), (_2)))
+
+#define __lsx_vpickve2gr_bu(/*__m128i*/ _1, /*ui4*/ _2)                        \
+  ((unsigned int)__builtin_lsx_vpickve2gr_bu((v16i8)(_1), (_2)))
+
+#define __lsx_vpickve2gr_hu(/*__m128i*/ _1, /*ui3*/ _2)                        \
+  ((unsigned int)__builtin_lsx_vpickve2gr_hu((v8i16)(_1), (_2)))
+
+#define __lsx_vpickve2gr_wu(/*__m128i*/ _1, /*ui2*/ _2)                        \
+  ((unsigned int)__builtin_lsx_vpickve2gr_wu((v4i32)(_1), (_2)))
+
+#define __lsx_vpickve2gr_du(/*__m128i*/ _1, /*ui1*/ _2)                        \
+  ((unsigned long int)__builtin_lsx_vpickve2gr_du((v2i64)(_1), (_2)))
+
+#define __lsx_vinsgr2vr_b(/*__m128i*/ _1, /*int*/ _2, /*ui4*/ _3)              \
+  ((__m128i)__builtin_lsx_vinsgr2vr_b((v16i8)(_1), (int)(_2), (_3)))
+
+#define __lsx_vinsgr2vr_h(/*__m128i*/ _1, /*int*/ _2, /*ui3*/ _3)              \
+  ((__m128i)__builtin_lsx_vinsgr2vr_h((v8i16)(_1), (int)(_2), (_3)))
+
+#define __lsx_vinsgr2vr_w(/*__m128i*/ _1, /*int*/ _2, /*ui2*/ _3)              \
+  ((__m128i)__builtin_lsx_vinsgr2vr_w((v4i32)(_1), (int)(_2), (_3)))
+
+#define __lsx_vinsgr2vr_d(/*__m128i*/ _1, /*long int*/ _2, /*ui1*/ _3)         \
+  ((__m128i)__builtin_lsx_vinsgr2vr_d((v2i64)(_1), (long int)(_2), (_3)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfadd_s(__m128 _1, __m128 _2) {
+  return (__m128)__builtin_lsx_vfadd_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfadd_d(__m128d _1, __m128d _2) {
+  return (__m128d)__builtin_lsx_vfadd_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfsub_s(__m128 _1, __m128 _2) {
+  return (__m128)__builtin_lsx_vfsub_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfsub_d(__m128d _1, __m128d _2) {
+  return (__m128d)__builtin_lsx_vfsub_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfmul_s(__m128 _1, __m128 _2) {
+  return (__m128)__builtin_lsx_vfmul_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfmul_d(__m128d _1, __m128d _2) {
+  return (__m128d)__builtin_lsx_vfmul_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfdiv_s(__m128 _1, __m128 _2) {
+  return (__m128)__builtin_lsx_vfdiv_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfdiv_d(__m128d _1, __m128d _2) {
+  return (__m128d)__builtin_lsx_vfdiv_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcvt_h_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcvt_h_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfcvt_s_d(__m128d _1, __m128d _2) {
+  return (__m128)__builtin_lsx_vfcvt_s_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfmin_s(__m128 _1, __m128 _2) {
+  return (__m128)__builtin_lsx_vfmin_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfmin_d(__m128d _1, __m128d _2) {
+  return (__m128d)__builtin_lsx_vfmin_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfmina_s(__m128 _1, __m128 _2) {
+  return (__m128)__builtin_lsx_vfmina_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfmina_d(__m128d _1, __m128d _2) {
+  return (__m128d)__builtin_lsx_vfmina_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfmax_s(__m128 _1, __m128 _2) {
+  return (__m128)__builtin_lsx_vfmax_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfmax_d(__m128d _1, __m128d _2) {
+  return (__m128d)__builtin_lsx_vfmax_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfmaxa_s(__m128 _1, __m128 _2) {
+  return (__m128)__builtin_lsx_vfmaxa_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfmaxa_d(__m128d _1, __m128d _2) {
+  return (__m128d)__builtin_lsx_vfmaxa_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfclass_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vfclass_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfclass_d(__m128d _1) {
+  return (__m128i)__builtin_lsx_vfclass_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfsqrt_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vfsqrt_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfsqrt_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vfsqrt_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfrecip_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vfrecip_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfrecip_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vfrecip_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfrint_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vfrint_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfrint_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vfrint_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfrsqrt_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vfrsqrt_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfrsqrt_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vfrsqrt_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vflogb_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vflogb_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vflogb_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vflogb_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfcvth_s_h(__m128i _1) {
+  return (__m128)__builtin_lsx_vfcvth_s_h((v8i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfcvth_d_s(__m128 _1) {
+  return (__m128d)__builtin_lsx_vfcvth_d_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfcvtl_s_h(__m128i _1) {
+  return (__m128)__builtin_lsx_vfcvtl_s_h((v8i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfcvtl_d_s(__m128 _1) {
+  return (__m128d)__builtin_lsx_vfcvtl_d_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftint_w_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftint_w_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftint_l_d(__m128d _1) {
+  return (__m128i)__builtin_lsx_vftint_l_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftint_wu_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftint_wu_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftint_lu_d(__m128d _1) {
+  return (__m128i)__builtin_lsx_vftint_lu_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrz_w_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrz_w_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrz_l_d(__m128d _1) {
+  return (__m128i)__builtin_lsx_vftintrz_l_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrz_wu_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrz_wu_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrz_lu_d(__m128d _1) {
+  return (__m128i)__builtin_lsx_vftintrz_lu_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vffint_s_w(__m128i _1) {
+  return (__m128)__builtin_lsx_vffint_s_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vffint_d_l(__m128i _1) {
+  return (__m128d)__builtin_lsx_vffint_d_l((v2i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vffint_s_wu(__m128i _1) {
+  return (__m128)__builtin_lsx_vffint_s_wu((v4u32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vffint_d_lu(__m128i _1) {
+  return (__m128d)__builtin_lsx_vffint_d_lu((v2u64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vandn_v(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vandn_v((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vneg_b(__m128i _1) {
+  return (__m128i)__builtin_lsx_vneg_b((v16i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vneg_h(__m128i _1) {
+  return (__m128i)__builtin_lsx_vneg_h((v8i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vneg_w(__m128i _1) {
+  return (__m128i)__builtin_lsx_vneg_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vneg_d(__m128i _1) {
+  return (__m128i)__builtin_lsx_vneg_d((v2i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmuh_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmuh_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmuh_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmuh_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmuh_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmuh_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmuh_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmuh_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmuh_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmuh_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmuh_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmuh_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmuh_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmuh_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmuh_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmuh_du((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vsllwil_h_b(/*__m128i*/ _1, /*ui3*/ _2)                          \
+  ((__m128i)__builtin_lsx_vsllwil_h_b((v16i8)(_1), (_2)))
+
+#define __lsx_vsllwil_w_h(/*__m128i*/ _1, /*ui4*/ _2)                          \
+  ((__m128i)__builtin_lsx_vsllwil_w_h((v8i16)(_1), (_2)))
+
+#define __lsx_vsllwil_d_w(/*__m128i*/ _1, /*ui5*/ _2)                          \
+  ((__m128i)__builtin_lsx_vsllwil_d_w((v4i32)(_1), (_2)))
+
+#define __lsx_vsllwil_hu_bu(/*__m128i*/ _1, /*ui3*/ _2)                        \
+  ((__m128i)__builtin_lsx_vsllwil_hu_bu((v16u8)(_1), (_2)))
+
+#define __lsx_vsllwil_wu_hu(/*__m128i*/ _1, /*ui4*/ _2)                        \
+  ((__m128i)__builtin_lsx_vsllwil_wu_hu((v8u16)(_1), (_2)))
+
+#define __lsx_vsllwil_du_wu(/*__m128i*/ _1, /*ui5*/ _2)                        \
+  ((__m128i)__builtin_lsx_vsllwil_du_wu((v4u32)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsran_b_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsran_b_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsran_h_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsran_h_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsran_w_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsran_w_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssran_b_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssran_b_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssran_h_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssran_h_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssran_w_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssran_w_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssran_bu_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssran_bu_h((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssran_hu_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssran_hu_w((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssran_wu_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssran_wu_d((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrarn_b_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrarn_b_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrarn_h_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrarn_h_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrarn_w_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrarn_w_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrarn_b_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrarn_b_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrarn_h_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrarn_h_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrarn_w_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrarn_w_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrarn_bu_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrarn_bu_h((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrarn_hu_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrarn_hu_w((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrarn_wu_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrarn_wu_d((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrln_b_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrln_b_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrln_h_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrln_h_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrln_w_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrln_w_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrln_bu_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrln_bu_h((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrln_hu_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrln_hu_w((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrln_wu_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrln_wu_d((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrlrn_b_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrlrn_b_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrlrn_h_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrlrn_h_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsrlrn_w_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsrlrn_w_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrlrn_bu_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrlrn_bu_h((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrlrn_hu_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrlrn_hu_w((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrlrn_wu_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrlrn_wu_d((v2u64)_1, (v2u64)_2);
+}
+
+#define __lsx_vfrstpi_b(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)            \
+  ((__m128i)__builtin_lsx_vfrstpi_b((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vfrstpi_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)            \
+  ((__m128i)__builtin_lsx_vfrstpi_h((v8i16)(_1), (v8i16)(_2), (_3)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfrstp_b(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vfrstp_b((v16i8)_1, (v16i8)_2, (v16i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfrstp_h(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vfrstp_h((v8i16)_1, (v8i16)_2, (v8i16)_3);
+}
+
+#define __lsx_vshuf4i_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui8*/ _3)            \
+  ((__m128i)__builtin_lsx_vshuf4i_d((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vbsrl_v(/*__m128i*/ _1, /*ui5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vbsrl_v((v16i8)(_1), (_2)))
+
+#define __lsx_vbsll_v(/*__m128i*/ _1, /*ui5*/ _2)                              \
+  ((__m128i)__builtin_lsx_vbsll_v((v16i8)(_1), (_2)))
+
+#define __lsx_vextrins_b(/*__m128i*/ _1, /*__m128i*/ _2, /*ui8*/ _3)           \
+  ((__m128i)__builtin_lsx_vextrins_b((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vextrins_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui8*/ _3)           \
+  ((__m128i)__builtin_lsx_vextrins_h((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vextrins_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui8*/ _3)           \
+  ((__m128i)__builtin_lsx_vextrins_w((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vextrins_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui8*/ _3)           \
+  ((__m128i)__builtin_lsx_vextrins_d((v2i64)(_1), (v2i64)(_2), (_3)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmskltz_b(__m128i _1) {
+  return (__m128i)__builtin_lsx_vmskltz_b((v16i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmskltz_h(__m128i _1) {
+  return (__m128i)__builtin_lsx_vmskltz_h((v8i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmskltz_w(__m128i _1) {
+  return (__m128i)__builtin_lsx_vmskltz_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmskltz_d(__m128i _1) {
+  return (__m128i)__builtin_lsx_vmskltz_d((v2i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsigncov_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsigncov_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsigncov_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsigncov_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsigncov_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsigncov_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsigncov_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsigncov_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfmadd_s(__m128 _1, __m128 _2, __m128 _3) {
+  return (__m128)__builtin_lsx_vfmadd_s((v4f32)_1, (v4f32)_2, (v4f32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfmadd_d(__m128d _1, __m128d _2, __m128d _3) {
+  return (__m128d)__builtin_lsx_vfmadd_d((v2f64)_1, (v2f64)_2, (v2f64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfmsub_s(__m128 _1, __m128 _2, __m128 _3) {
+  return (__m128)__builtin_lsx_vfmsub_s((v4f32)_1, (v4f32)_2, (v4f32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfmsub_d(__m128d _1, __m128d _2, __m128d _3) {
+  return (__m128d)__builtin_lsx_vfmsub_d((v2f64)_1, (v2f64)_2, (v2f64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfnmadd_s(__m128 _1, __m128 _2, __m128 _3) {
+  return (__m128)__builtin_lsx_vfnmadd_s((v4f32)_1, (v4f32)_2, (v4f32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfnmadd_d(__m128d _1, __m128d _2, __m128d _3) {
+  return (__m128d)__builtin_lsx_vfnmadd_d((v2f64)_1, (v2f64)_2, (v2f64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfnmsub_s(__m128 _1, __m128 _2, __m128 _3) {
+  return (__m128)__builtin_lsx_vfnmsub_s((v4f32)_1, (v4f32)_2, (v4f32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfnmsub_d(__m128d _1, __m128d _2, __m128d _3) {
+  return (__m128d)__builtin_lsx_vfnmsub_d((v2f64)_1, (v2f64)_2, (v2f64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrne_w_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrne_w_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrne_l_d(__m128d _1) {
+  return (__m128i)__builtin_lsx_vftintrne_l_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrp_w_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrp_w_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrp_l_d(__m128d _1) {
+  return (__m128i)__builtin_lsx_vftintrp_l_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrm_w_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrm_w_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrm_l_d(__m128d _1) {
+  return (__m128i)__builtin_lsx_vftintrm_l_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftint_w_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vftint_w_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vffint_s_l(__m128i _1, __m128i _2) {
+  return (__m128)__builtin_lsx_vffint_s_l((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrz_w_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vftintrz_w_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrp_w_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vftintrp_w_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrm_w_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vftintrm_w_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrne_w_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vftintrne_w_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintl_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintl_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftinth_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftinth_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vffinth_d_w(__m128i _1) {
+  return (__m128d)__builtin_lsx_vffinth_d_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vffintl_d_w(__m128i _1) {
+  return (__m128d)__builtin_lsx_vffintl_d_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrzl_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrzl_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrzh_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrzh_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrpl_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrpl_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrph_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrph_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrml_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrml_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrmh_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrmh_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrnel_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrnel_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vftintrneh_l_s(__m128 _1) {
+  return (__m128i)__builtin_lsx_vftintrneh_l_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfrintrne_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vfrintrne_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfrintrne_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vfrintrne_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfrintrz_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vfrintrz_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfrintrz_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vfrintrz_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfrintrp_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vfrintrp_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfrintrp_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vfrintrp_d((v2f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128
+    __lsx_vfrintrm_s(__m128 _1) {
+  return (__m128)__builtin_lsx_vfrintrm_s((v4f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128d
+    __lsx_vfrintrm_d(__m128d _1) {
+  return (__m128d)__builtin_lsx_vfrintrm_d((v2f64)_1);
+}
+
+#define __lsx_vstelm_b(/*__m128i*/ _1, /*void **/ _2, /*si8*/ _3, /*idx*/ _4)  \
+  ((void)__builtin_lsx_vstelm_b((v16i8)(_1), (void *)(_2), (_3), (_4)))
+
+#define __lsx_vstelm_h(/*__m128i*/ _1, /*void **/ _2, /*si8*/ _3, /*idx*/ _4)  \
+  ((void)__builtin_lsx_vstelm_h((v8i16)(_1), (void *)(_2), (_3), (_4)))
+
+#define __lsx_vstelm_w(/*__m128i*/ _1, /*void **/ _2, /*si8*/ _3, /*idx*/ _4)  \
+  ((void)__builtin_lsx_vstelm_w((v4i32)(_1), (void *)(_2), (_3), (_4)))
+
+#define __lsx_vstelm_d(/*__m128i*/ _1, /*void **/ _2, /*si8*/ _3, /*idx*/ _4)  \
+  ((void)__builtin_lsx_vstelm_d((v2i64)(_1), (void *)(_2), (_3), (_4)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_d_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_d_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_w_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_w_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_h_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_h_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_d_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_d_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_w_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_w_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_h_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_h_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_d_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_d_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_w_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_w_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_h_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_h_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_d_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_d_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_w_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_w_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_h_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_h_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_d_wu_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_d_wu_w((v4u32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_w_hu_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_w_hu_h((v8u16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_h_bu_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_h_bu_b((v16u8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_d_wu_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_d_wu_w((v4u32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_w_hu_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_w_hu_h((v8u16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_h_bu_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_h_bu_b((v16u8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwev_d_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwev_d_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwev_w_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwev_w_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwev_h_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwev_h_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwod_d_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwod_d_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwod_w_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwod_w_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwod_h_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwod_h_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwev_d_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwev_d_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwev_w_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwev_w_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwev_h_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwev_h_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwod_d_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwod_d_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwod_w_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwod_w_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwod_h_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwod_h_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_q_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_q_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_q_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_q_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_q_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_q_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_q_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_q_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwev_q_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwev_q_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwod_q_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwod_q_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwev_q_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwev_q_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsubwod_q_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsubwod_q_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwev_q_du_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwev_q_du_d((v2u64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vaddwod_q_du_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vaddwod_q_du_d((v2u64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_d_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_d_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_w_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_w_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_h_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_h_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_d_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_d_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_w_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_w_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_h_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_h_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_d_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_d_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_w_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_w_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_h_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_h_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_d_wu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_d_wu((v4u32)_1, (v4u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_w_hu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_w_hu((v8u16)_1, (v8u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_h_bu(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_h_bu((v16u8)_1, (v16u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_d_wu_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_d_wu_w((v4u32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_w_hu_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_w_hu_h((v8u16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_h_bu_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_h_bu_b((v16u8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_d_wu_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_d_wu_w((v4u32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_w_hu_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_w_hu_h((v8u16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_h_bu_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_h_bu_b((v16u8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_q_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_q_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_q_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_q_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_q_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_q_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_q_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_q_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwev_q_du_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwev_q_du_d((v2u64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmulwod_q_du_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vmulwod_q_du_d((v2u64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhaddw_q_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhaddw_q_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhaddw_qu_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhaddw_qu_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhsubw_q_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhsubw_q_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vhsubw_qu_du(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vhsubw_qu_du((v2u64)_1, (v2u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_d_w(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_d_w((v2i64)_1, (v4i32)_2, (v4i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_w_h(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_w_h((v4i32)_1, (v8i16)_2, (v8i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_h_b(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_h_b((v8i16)_1, (v16i8)_2, (v16i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_d_wu(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_d_wu((v2u64)_1, (v4u32)_2, (v4u32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_w_hu(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_w_hu((v4u32)_1, (v8u16)_2, (v8u16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_h_bu(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_h_bu((v8u16)_1, (v16u8)_2, (v16u8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_d_w(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_d_w((v2i64)_1, (v4i32)_2, (v4i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_w_h(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_w_h((v4i32)_1, (v8i16)_2, (v8i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_h_b(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_h_b((v8i16)_1, (v16i8)_2, (v16i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_d_wu(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_d_wu((v2u64)_1, (v4u32)_2, (v4u32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_w_hu(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_w_hu((v4u32)_1, (v8u16)_2, (v8u16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_h_bu(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_h_bu((v8u16)_1, (v16u8)_2, (v16u8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_d_wu_w(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_d_wu_w((v2i64)_1, (v4u32)_2,
+                                                (v4i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_w_hu_h(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_w_hu_h((v4i32)_1, (v8u16)_2,
+                                                (v8i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_h_bu_b(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_h_bu_b((v8i16)_1, (v16u8)_2,
+                                                (v16i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_d_wu_w(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_d_wu_w((v2i64)_1, (v4u32)_2,
+                                                (v4i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_w_hu_h(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_w_hu_h((v4i32)_1, (v8u16)_2,
+                                                (v8i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_h_bu_b(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_h_bu_b((v8i16)_1, (v16u8)_2,
+                                                (v16i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_q_d(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_q_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_q_d(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_q_d((v2i64)_1, (v2i64)_2, (v2i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_q_du(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_q_du((v2u64)_1, (v2u64)_2, (v2u64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_q_du(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_q_du((v2u64)_1, (v2u64)_2, (v2u64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwev_q_du_d(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwev_q_du_d((v2i64)_1, (v2u64)_2,
+                                                (v2i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmaddwod_q_du_d(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vmaddwod_q_du_d((v2i64)_1, (v2u64)_2,
+                                                (v2i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vrotr_b(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vrotr_b((v16i8)_1, (v16i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vrotr_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vrotr_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vrotr_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vrotr_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vrotr_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vrotr_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vadd_q(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vadd_q((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vsub_q(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vsub_q((v2i64)_1, (v2i64)_2);
+}
+
+#define __lsx_vldrepl_b(/*void **/ _1, /*si12*/ _2)                            \
+  ((__m128i)__builtin_lsx_vldrepl_b((void const *)(_1), (_2)))
+
+#define __lsx_vldrepl_h(/*void **/ _1, /*si11*/ _2)                            \
+  ((__m128i)__builtin_lsx_vldrepl_h((void const *)(_1), (_2)))
+
+#define __lsx_vldrepl_w(/*void **/ _1, /*si10*/ _2)                            \
+  ((__m128i)__builtin_lsx_vldrepl_w((void const *)(_1), (_2)))
+
+#define __lsx_vldrepl_d(/*void **/ _1, /*si9*/ _2)                             \
+  ((__m128i)__builtin_lsx_vldrepl_d((void const *)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmskgez_b(__m128i _1) {
+  return (__m128i)__builtin_lsx_vmskgez_b((v16i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vmsknz_b(__m128i _1) {
+  return (__m128i)__builtin_lsx_vmsknz_b((v16i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vexth_h_b(__m128i _1) {
+  return (__m128i)__builtin_lsx_vexth_h_b((v16i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vexth_w_h(__m128i _1) {
+  return (__m128i)__builtin_lsx_vexth_w_h((v8i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vexth_d_w(__m128i _1) {
+  return (__m128i)__builtin_lsx_vexth_d_w((v4i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vexth_q_d(__m128i _1) {
+  return (__m128i)__builtin_lsx_vexth_q_d((v2i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vexth_hu_bu(__m128i _1) {
+  return (__m128i)__builtin_lsx_vexth_hu_bu((v16u8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vexth_wu_hu(__m128i _1) {
+  return (__m128i)__builtin_lsx_vexth_wu_hu((v8u16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vexth_du_wu(__m128i _1) {
+  return (__m128i)__builtin_lsx_vexth_du_wu((v4u32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vexth_qu_du(__m128i _1) {
+  return (__m128i)__builtin_lsx_vexth_qu_du((v2u64)_1);
+}
+
+#define __lsx_vrotri_b(/*__m128i*/ _1, /*ui3*/ _2)                             \
+  ((__m128i)__builtin_lsx_vrotri_b((v16i8)(_1), (_2)))
+
+#define __lsx_vrotri_h(/*__m128i*/ _1, /*ui4*/ _2)                             \
+  ((__m128i)__builtin_lsx_vrotri_h((v8i16)(_1), (_2)))
+
+#define __lsx_vrotri_w(/*__m128i*/ _1, /*ui5*/ _2)                             \
+  ((__m128i)__builtin_lsx_vrotri_w((v4i32)(_1), (_2)))
+
+#define __lsx_vrotri_d(/*__m128i*/ _1, /*ui6*/ _2)                             \
+  ((__m128i)__builtin_lsx_vrotri_d((v2i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vextl_q_d(__m128i _1) {
+  return (__m128i)__builtin_lsx_vextl_q_d((v2i64)_1);
+}
+
+#define __lsx_vsrlni_b_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)           \
+  ((__m128i)__builtin_lsx_vsrlni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vsrlni_h_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)           \
+  ((__m128i)__builtin_lsx_vsrlni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vsrlni_w_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)           \
+  ((__m128i)__builtin_lsx_vsrlni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vsrlni_d_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)           \
+  ((__m128i)__builtin_lsx_vsrlni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vsrlrni_b_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)          \
+  ((__m128i)__builtin_lsx_vsrlrni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vsrlrni_h_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)          \
+  ((__m128i)__builtin_lsx_vsrlrni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vsrlrni_w_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)          \
+  ((__m128i)__builtin_lsx_vsrlrni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vsrlrni_d_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)          \
+  ((__m128i)__builtin_lsx_vsrlrni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vssrlni_b_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)          \
+  ((__m128i)__builtin_lsx_vssrlni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vssrlni_h_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)          \
+  ((__m128i)__builtin_lsx_vssrlni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vssrlni_w_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)          \
+  ((__m128i)__builtin_lsx_vssrlni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vssrlni_d_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)          \
+  ((__m128i)__builtin_lsx_vssrlni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vssrlni_bu_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrlni_bu_h((v16u8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vssrlni_hu_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrlni_hu_w((v8u16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vssrlni_wu_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrlni_wu_d((v4u32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vssrlni_du_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrlni_du_q((v2u64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vssrlrni_b_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrlrni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vssrlrni_h_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrlrni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vssrlrni_w_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrlrni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vssrlrni_d_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrlrni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vssrlrni_bu_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)        \
+  ((__m128i)__builtin_lsx_vssrlrni_bu_h((v16u8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vssrlrni_hu_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)        \
+  ((__m128i)__builtin_lsx_vssrlrni_hu_w((v8u16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vssrlrni_wu_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)        \
+  ((__m128i)__builtin_lsx_vssrlrni_wu_d((v4u32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vssrlrni_du_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)        \
+  ((__m128i)__builtin_lsx_vssrlrni_du_q((v2u64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vsrani_b_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)           \
+  ((__m128i)__builtin_lsx_vsrani_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vsrani_h_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)           \
+  ((__m128i)__builtin_lsx_vsrani_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vsrani_w_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)           \
+  ((__m128i)__builtin_lsx_vsrani_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vsrani_d_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)           \
+  ((__m128i)__builtin_lsx_vsrani_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vsrarni_b_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)          \
+  ((__m128i)__builtin_lsx_vsrarni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vsrarni_h_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)          \
+  ((__m128i)__builtin_lsx_vsrarni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vsrarni_w_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)          \
+  ((__m128i)__builtin_lsx_vsrarni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vsrarni_d_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)          \
+  ((__m128i)__builtin_lsx_vsrarni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vssrani_b_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)          \
+  ((__m128i)__builtin_lsx_vssrani_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vssrani_h_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)          \
+  ((__m128i)__builtin_lsx_vssrani_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vssrani_w_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)          \
+  ((__m128i)__builtin_lsx_vssrani_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vssrani_d_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)          \
+  ((__m128i)__builtin_lsx_vssrani_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vssrani_bu_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrani_bu_h((v16u8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vssrani_hu_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrani_hu_w((v8u16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vssrani_wu_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrani_wu_d((v4u32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vssrani_du_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrani_du_q((v2u64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vssrarni_b_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrarni_b_h((v16i8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vssrarni_h_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrarni_h_w((v8i16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vssrarni_w_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrarni_w_d((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vssrarni_d_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)         \
+  ((__m128i)__builtin_lsx_vssrarni_d_q((v2i64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vssrarni_bu_h(/*__m128i*/ _1, /*__m128i*/ _2, /*ui4*/ _3)        \
+  ((__m128i)__builtin_lsx_vssrarni_bu_h((v16u8)(_1), (v16i8)(_2), (_3)))
+
+#define __lsx_vssrarni_hu_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui5*/ _3)        \
+  ((__m128i)__builtin_lsx_vssrarni_hu_w((v8u16)(_1), (v8i16)(_2), (_3)))
+
+#define __lsx_vssrarni_wu_d(/*__m128i*/ _1, /*__m128i*/ _2, /*ui6*/ _3)        \
+  ((__m128i)__builtin_lsx_vssrarni_wu_d((v4u32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vssrarni_du_q(/*__m128i*/ _1, /*__m128i*/ _2, /*ui7*/ _3)        \
+  ((__m128i)__builtin_lsx_vssrarni_du_q((v2u64)(_1), (v2i64)(_2), (_3)))
+
+#define __lsx_vpermi_w(/*__m128i*/ _1, /*__m128i*/ _2, /*ui8*/ _3)             \
+  ((__m128i)__builtin_lsx_vpermi_w((v4i32)(_1), (v4i32)(_2), (_3)))
+
+#define __lsx_vld(/*void **/ _1, /*si12*/ _2)                                  \
+  ((__m128i)__builtin_lsx_vld((void const *)(_1), (_2)))
+
+#define __lsx_vst(/*__m128i*/ _1, /*void **/ _2, /*si12*/ _3)                  \
+  ((void)__builtin_lsx_vst((v16i8)(_1), (void *)(_2), (_3)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrlrn_b_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrlrn_b_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrlrn_h_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrlrn_h_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrlrn_w_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrlrn_w_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrln_b_h(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrln_b_h((v8i16)_1, (v8i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrln_h_w(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrln_h_w((v4i32)_1, (v4i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vssrln_w_d(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vssrln_w_d((v2i64)_1, (v2i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vorn_v(__m128i _1, __m128i _2) {
+  return (__m128i)__builtin_lsx_vorn_v((v16i8)_1, (v16i8)_2);
+}
+
+#define __lsx_vldi(/*i13*/ _1) ((__m128i)__builtin_lsx_vldi((_1)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vshuf_b(__m128i _1, __m128i _2, __m128i _3) {
+  return (__m128i)__builtin_lsx_vshuf_b((v16i8)_1, (v16i8)_2, (v16i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vldx(void const *_1, long int _2) {
+  return (__m128i)__builtin_lsx_vldx((void const *)_1, (long int)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) void
+    __lsx_vstx(__m128i _1, void *_2, long int _3) {
+  return (void)__builtin_lsx_vstx((v16i8)_1, (void *)_2, (long int)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vextl_qu_du(__m128i _1) {
+  return (__m128i)__builtin_lsx_vextl_qu_du((v2u64)_1);
+}
+
+#define __lsx_bnz_b(/*__m128i*/ _1) ((int)__builtin_lsx_bnz_b((v16u8)(_1)))
+
+#define __lsx_bnz_d(/*__m128i*/ _1) ((int)__builtin_lsx_bnz_d((v2u64)(_1)))
+
+#define __lsx_bnz_h(/*__m128i*/ _1) ((int)__builtin_lsx_bnz_h((v8u16)(_1)))
+
+#define __lsx_bnz_v(/*__m128i*/ _1) ((int)__builtin_lsx_bnz_v((v16u8)(_1)))
+
+#define __lsx_bnz_w(/*__m128i*/ _1) ((int)__builtin_lsx_bnz_w((v4u32)(_1)))
+
+#define __lsx_bz_b(/*__m128i*/ _1) ((int)__builtin_lsx_bz_b((v16u8)(_1)))
+
+#define __lsx_bz_d(/*__m128i*/ _1) ((int)__builtin_lsx_bz_d((v2u64)(_1)))
+
+#define __lsx_bz_h(/*__m128i*/ _1) ((int)__builtin_lsx_bz_h((v8u16)(_1)))
+
+#define __lsx_bz_v(/*__m128i*/ _1) ((int)__builtin_lsx_bz_v((v16u8)(_1)))
+
+#define __lsx_bz_w(/*__m128i*/ _1) ((int)__builtin_lsx_bz_w((v4u32)(_1)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_caf_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_caf_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_caf_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_caf_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_ceq_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_ceq_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_ceq_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_ceq_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cle_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cle_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cle_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cle_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_clt_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_clt_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_clt_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_clt_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cne_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cne_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cne_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cne_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cor_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cor_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cor_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cor_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cueq_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cueq_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cueq_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cueq_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cule_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cule_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cule_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cule_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cult_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cult_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cult_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cult_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cun_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cun_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cune_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cune_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cune_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cune_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_cun_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_cun_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_saf_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_saf_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_saf_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_saf_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_seq_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_seq_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_seq_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_seq_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sle_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sle_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sle_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sle_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_slt_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_slt_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_slt_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_slt_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sne_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sne_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sne_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sne_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sor_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sor_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sor_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sor_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sueq_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sueq_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sueq_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sueq_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sule_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sule_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sule_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sule_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sult_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sult_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sult_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sult_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sun_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sun_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sune_d(__m128d _1, __m128d _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sune_d((v2f64)_1, (v2f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sune_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sune_s((v4f32)_1, (v4f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m128i
+    __lsx_vfcmp_sun_s(__m128 _1, __m128 _2) {
+  return (__m128i)__builtin_lsx_vfcmp_sun_s((v4f32)_1, (v4f32)_2);
+}
+
+#define __lsx_vrepli_b(/*si10*/ _1) ((__m128i)__builtin_lsx_vrepli_b((_1)))
+
+#define __lsx_vrepli_d(/*si10*/ _1) ((__m128i)__builtin_lsx_vrepli_d((_1)))
+
+#define __lsx_vrepli_h(/*si10*/ _1) ((__m128i)__builtin_lsx_vrepli_h((_1)))
+
+#define __lsx_vrepli_w(/*si10*/ _1) ((__m128i)__builtin_lsx_vrepli_w((_1)))
+
+#endif /* defined(__loongarch_sx) */
+#endif /* _LOONGSON_SXINTRIN_H */
diff --git a/clang/lib/Sema/SemaChecking.cpp b/clang/lib/Sema/SemaChecking.cpp
index 7972919d1420116..e131061eb88548b 100644
--- a/clang/lib/Sema/SemaChecking.cpp
+++ b/clang/lib/Sema/SemaChecking.cpp
@@ -3928,6 +3928,7 @@ bool Sema::CheckLoongArchBuiltinFunctionCall(const TargetInfo &TI,
   switch (BuiltinID) {
   default:
     break;
+  // Basic intrinsics.
   case LoongArch::BI__builtin_loongarch_cacop_d:
   case LoongArch::BI__builtin_loongarch_cacop_w: {
     SemaBuiltinConstantArgRange(TheCall, 0, 0, llvm::maxUIntN(5));
@@ -3956,8 +3957,234 @@ bool Sema::CheckLoongArchBuiltinFunctionCall(const TargetInfo &TI,
   case LoongArch::BI__builtin_loongarch_movfcsr2gr:
   case LoongArch::BI__builtin_loongarch_movgr2fcsr:
     return SemaBuiltinConstantArgRange(TheCall, 0, 0, llvm::maxUIntN(2));
-  }
 
+  // LSX intrinsics.
+  case LoongArch::BI__builtin_lsx_vbitclri_b:
+  case LoongArch::BI__builtin_lsx_vbitrevi_b:
+  case LoongArch::BI__builtin_lsx_vbitseti_b:
+  case LoongArch::BI__builtin_lsx_vsat_b:
+  case LoongArch::BI__builtin_lsx_vsat_bu:
+  case LoongArch::BI__builtin_lsx_vslli_b:
+  case LoongArch::BI__builtin_lsx_vsrai_b:
+  case LoongArch::BI__builtin_lsx_vsrari_b:
+  case LoongArch::BI__builtin_lsx_vsrli_b:
+  case LoongArch::BI__builtin_lsx_vsllwil_h_b:
+  case LoongArch::BI__builtin_lsx_vsllwil_hu_bu:
+  case LoongArch::BI__builtin_lsx_vrotri_b:
+  case LoongArch::BI__builtin_lsx_vsrlri_b:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 7);
+  case LoongArch::BI__builtin_lsx_vbitclri_h:
+  case LoongArch::BI__builtin_lsx_vbitrevi_h:
+  case LoongArch::BI__builtin_lsx_vbitseti_h:
+  case LoongArch::BI__builtin_lsx_vsat_h:
+  case LoongArch::BI__builtin_lsx_vsat_hu:
+  case LoongArch::BI__builtin_lsx_vslli_h:
+  case LoongArch::BI__builtin_lsx_vsrai_h:
+  case LoongArch::BI__builtin_lsx_vsrari_h:
+  case LoongArch::BI__builtin_lsx_vsrli_h:
+  case LoongArch::BI__builtin_lsx_vsllwil_w_h:
+  case LoongArch::BI__builtin_lsx_vsllwil_wu_hu:
+  case LoongArch::BI__builtin_lsx_vrotri_h:
+  case LoongArch::BI__builtin_lsx_vsrlri_h:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 15);
+  case LoongArch::BI__builtin_lsx_vssrarni_b_h:
+  case LoongArch::BI__builtin_lsx_vssrarni_bu_h:
+  case LoongArch::BI__builtin_lsx_vssrani_b_h:
+  case LoongArch::BI__builtin_lsx_vssrani_bu_h:
+  case LoongArch::BI__builtin_lsx_vsrarni_b_h:
+  case LoongArch::BI__builtin_lsx_vsrlni_b_h:
+  case LoongArch::BI__builtin_lsx_vsrlrni_b_h:
+  case LoongArch::BI__builtin_lsx_vssrlni_b_h:
+  case LoongArch::BI__builtin_lsx_vssrlni_bu_h:
+  case LoongArch::BI__builtin_lsx_vssrlrni_b_h:
+  case LoongArch::BI__builtin_lsx_vssrlrni_bu_h:
+  case LoongArch::BI__builtin_lsx_vsrani_b_h:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 15);
+  case LoongArch::BI__builtin_lsx_vslei_bu:
+  case LoongArch::BI__builtin_lsx_vslei_hu:
+  case LoongArch::BI__builtin_lsx_vslei_wu:
+  case LoongArch::BI__builtin_lsx_vslei_du:
+  case LoongArch::BI__builtin_lsx_vslti_bu:
+  case LoongArch::BI__builtin_lsx_vslti_hu:
+  case LoongArch::BI__builtin_lsx_vslti_wu:
+  case LoongArch::BI__builtin_lsx_vslti_du:
+  case LoongArch::BI__builtin_lsx_vmaxi_bu:
+  case LoongArch::BI__builtin_lsx_vmaxi_hu:
+  case LoongArch::BI__builtin_lsx_vmaxi_wu:
+  case LoongArch::BI__builtin_lsx_vmaxi_du:
+  case LoongArch::BI__builtin_lsx_vmini_bu:
+  case LoongArch::BI__builtin_lsx_vmini_hu:
+  case LoongArch::BI__builtin_lsx_vmini_wu:
+  case LoongArch::BI__builtin_lsx_vmini_du:
+  case LoongArch::BI__builtin_lsx_vaddi_bu:
+  case LoongArch::BI__builtin_lsx_vaddi_hu:
+  case LoongArch::BI__builtin_lsx_vaddi_wu:
+  case LoongArch::BI__builtin_lsx_vaddi_du:
+  case LoongArch::BI__builtin_lsx_vbitclri_w:
+  case LoongArch::BI__builtin_lsx_vbitrevi_w:
+  case LoongArch::BI__builtin_lsx_vbitseti_w:
+  case LoongArch::BI__builtin_lsx_vsat_w:
+  case LoongArch::BI__builtin_lsx_vsat_wu:
+  case LoongArch::BI__builtin_lsx_vslli_w:
+  case LoongArch::BI__builtin_lsx_vsrai_w:
+  case LoongArch::BI__builtin_lsx_vsrari_w:
+  case LoongArch::BI__builtin_lsx_vsrli_w:
+  case LoongArch::BI__builtin_lsx_vsllwil_d_w:
+  case LoongArch::BI__builtin_lsx_vsllwil_du_wu:
+  case LoongArch::BI__builtin_lsx_vsrlri_w:
+  case LoongArch::BI__builtin_lsx_vrotri_w:
+  case LoongArch::BI__builtin_lsx_vsubi_bu:
+  case LoongArch::BI__builtin_lsx_vsubi_hu:
+  case LoongArch::BI__builtin_lsx_vbsrl_v:
+  case LoongArch::BI__builtin_lsx_vbsll_v:
+  case LoongArch::BI__builtin_lsx_vsubi_wu:
+  case LoongArch::BI__builtin_lsx_vsubi_du:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 31);
+  case LoongArch::BI__builtin_lsx_vssrarni_h_w:
+  case LoongArch::BI__builtin_lsx_vssrarni_hu_w:
+  case LoongArch::BI__builtin_lsx_vssrani_h_w:
+  case LoongArch::BI__builtin_lsx_vssrani_hu_w:
+  case LoongArch::BI__builtin_lsx_vsrarni_h_w:
+  case LoongArch::BI__builtin_lsx_vsrani_h_w:
+  case LoongArch::BI__builtin_lsx_vfrstpi_b:
+  case LoongArch::BI__builtin_lsx_vfrstpi_h:
+  case LoongArch::BI__builtin_lsx_vsrlni_h_w:
+  case LoongArch::BI__builtin_lsx_vsrlrni_h_w:
+  case LoongArch::BI__builtin_lsx_vssrlni_h_w:
+  case LoongArch::BI__builtin_lsx_vssrlni_hu_w:
+  case LoongArch::BI__builtin_lsx_vssrlrni_h_w:
+  case LoongArch::BI__builtin_lsx_vssrlrni_hu_w:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 31);
+  case LoongArch::BI__builtin_lsx_vbitclri_d:
+  case LoongArch::BI__builtin_lsx_vbitrevi_d:
+  case LoongArch::BI__builtin_lsx_vbitseti_d:
+  case LoongArch::BI__builtin_lsx_vsat_d:
+  case LoongArch::BI__builtin_lsx_vsat_du:
+  case LoongArch::BI__builtin_lsx_vslli_d:
+  case LoongArch::BI__builtin_lsx_vsrai_d:
+  case LoongArch::BI__builtin_lsx_vsrli_d:
+  case LoongArch::BI__builtin_lsx_vsrari_d:
+  case LoongArch::BI__builtin_lsx_vrotri_d:
+  case LoongArch::BI__builtin_lsx_vsrlri_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 63);
+  case LoongArch::BI__builtin_lsx_vssrarni_w_d:
+  case LoongArch::BI__builtin_lsx_vssrarni_wu_d:
+  case LoongArch::BI__builtin_lsx_vssrani_w_d:
+  case LoongArch::BI__builtin_lsx_vssrani_wu_d:
+  case LoongArch::BI__builtin_lsx_vsrarni_w_d:
+  case LoongArch::BI__builtin_lsx_vsrlni_w_d:
+  case LoongArch::BI__builtin_lsx_vsrlrni_w_d:
+  case LoongArch::BI__builtin_lsx_vssrlni_w_d:
+  case LoongArch::BI__builtin_lsx_vssrlni_wu_d:
+  case LoongArch::BI__builtin_lsx_vssrlrni_w_d:
+  case LoongArch::BI__builtin_lsx_vssrlrni_wu_d:
+  case LoongArch::BI__builtin_lsx_vsrani_w_d:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 63);
+  case LoongArch::BI__builtin_lsx_vssrarni_d_q:
+  case LoongArch::BI__builtin_lsx_vssrarni_du_q:
+  case LoongArch::BI__builtin_lsx_vssrani_d_q:
+  case LoongArch::BI__builtin_lsx_vssrani_du_q:
+  case LoongArch::BI__builtin_lsx_vsrarni_d_q:
+  case LoongArch::BI__builtin_lsx_vssrlni_d_q:
+  case LoongArch::BI__builtin_lsx_vssrlni_du_q:
+  case LoongArch::BI__builtin_lsx_vssrlrni_d_q:
+  case LoongArch::BI__builtin_lsx_vssrlrni_du_q:
+  case LoongArch::BI__builtin_lsx_vsrani_d_q:
+  case LoongArch::BI__builtin_lsx_vsrlrni_d_q:
+  case LoongArch::BI__builtin_lsx_vsrlni_d_q:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 127);
+  case LoongArch::BI__builtin_lsx_vseqi_b:
+  case LoongArch::BI__builtin_lsx_vseqi_h:
+  case LoongArch::BI__builtin_lsx_vseqi_w:
+  case LoongArch::BI__builtin_lsx_vseqi_d:
+  case LoongArch::BI__builtin_lsx_vslti_b:
+  case LoongArch::BI__builtin_lsx_vslti_h:
+  case LoongArch::BI__builtin_lsx_vslti_w:
+  case LoongArch::BI__builtin_lsx_vslti_d:
+  case LoongArch::BI__builtin_lsx_vslei_b:
+  case LoongArch::BI__builtin_lsx_vslei_h:
+  case LoongArch::BI__builtin_lsx_vslei_w:
+  case LoongArch::BI__builtin_lsx_vslei_d:
+  case LoongArch::BI__builtin_lsx_vmaxi_b:
+  case LoongArch::BI__builtin_lsx_vmaxi_h:
+  case LoongArch::BI__builtin_lsx_vmaxi_w:
+  case LoongArch::BI__builtin_lsx_vmaxi_d:
+  case LoongArch::BI__builtin_lsx_vmini_b:
+  case LoongArch::BI__builtin_lsx_vmini_h:
+  case LoongArch::BI__builtin_lsx_vmini_w:
+  case LoongArch::BI__builtin_lsx_vmini_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -16, 15);
+  case LoongArch::BI__builtin_lsx_vandi_b:
+  case LoongArch::BI__builtin_lsx_vnori_b:
+  case LoongArch::BI__builtin_lsx_vori_b:
+  case LoongArch::BI__builtin_lsx_vshuf4i_b:
+  case LoongArch::BI__builtin_lsx_vshuf4i_h:
+  case LoongArch::BI__builtin_lsx_vshuf4i_w:
+  case LoongArch::BI__builtin_lsx_vxori_b:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 255);
+  case LoongArch::BI__builtin_lsx_vbitseli_b:
+  case LoongArch::BI__builtin_lsx_vshuf4i_d:
+  case LoongArch::BI__builtin_lsx_vextrins_b:
+  case LoongArch::BI__builtin_lsx_vextrins_h:
+  case LoongArch::BI__builtin_lsx_vextrins_w:
+  case LoongArch::BI__builtin_lsx_vextrins_d:
+  case LoongArch::BI__builtin_lsx_vpermi_w:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 255);
+  case LoongArch::BI__builtin_lsx_vpickve2gr_b:
+  case LoongArch::BI__builtin_lsx_vpickve2gr_bu:
+  case LoongArch::BI__builtin_lsx_vreplvei_b:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 15);
+  case LoongArch::BI__builtin_lsx_vinsgr2vr_b:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 15);
+  case LoongArch::BI__builtin_lsx_vpickve2gr_h:
+  case LoongArch::BI__builtin_lsx_vpickve2gr_hu:
+  case LoongArch::BI__builtin_lsx_vreplvei_h:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 7);
+  case LoongArch::BI__builtin_lsx_vinsgr2vr_h:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 7);
+  case LoongArch::BI__builtin_lsx_vpickve2gr_w:
+  case LoongArch::BI__builtin_lsx_vpickve2gr_wu:
+  case LoongArch::BI__builtin_lsx_vreplvei_w:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 3);
+  case LoongArch::BI__builtin_lsx_vinsgr2vr_w:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 3);
+  case LoongArch::BI__builtin_lsx_vpickve2gr_d:
+  case LoongArch::BI__builtin_lsx_vpickve2gr_du:
+  case LoongArch::BI__builtin_lsx_vreplvei_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 1);
+  case LoongArch::BI__builtin_lsx_vinsgr2vr_d:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 1);
+  case LoongArch::BI__builtin_lsx_vstelm_b:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -128, 127) ||
+           SemaBuiltinConstantArgRange(TheCall, 3, 0, 15);
+  case LoongArch::BI__builtin_lsx_vstelm_h:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -256, 254) ||
+           SemaBuiltinConstantArgRange(TheCall, 3, 0, 7);
+  case LoongArch::BI__builtin_lsx_vstelm_w:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -512, 508) ||
+           SemaBuiltinConstantArgRange(TheCall, 3, 0, 3);
+  case LoongArch::BI__builtin_lsx_vstelm_d:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -1024, 1016) ||
+           SemaBuiltinConstantArgRange(TheCall, 3, 0, 1);
+  case LoongArch::BI__builtin_lsx_vldrepl_b:
+  case LoongArch::BI__builtin_lsx_vld:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -2048, 2047);
+  case LoongArch::BI__builtin_lsx_vldrepl_h:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -2048, 2046);
+  case LoongArch::BI__builtin_lsx_vldrepl_w:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -2048, 2044);
+  case LoongArch::BI__builtin_lsx_vldrepl_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -2048, 2040);
+  case LoongArch::BI__builtin_lsx_vst:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -2048, 2047);
+  case LoongArch::BI__builtin_lsx_vldi:
+    return SemaBuiltinConstantArgRange(TheCall, 0, -4096, 4095);
+  case LoongArch::BI__builtin_lsx_vrepli_b:
+  case LoongArch::BI__builtin_lsx_vrepli_h:
+  case LoongArch::BI__builtin_lsx_vrepli_w:
+  case LoongArch::BI__builtin_lsx_vrepli_d:
+    return SemaBuiltinConstantArgRange(TheCall, 0, -512, 511);
+  }
   return false;
 }
 

>From f23acda54a39b30ed0f7e30be0a14e5c295b3621 Mon Sep 17 00:00:00 2001
From: licongtian <licongtian at loongson.cn>
Date: Wed, 25 Oct 2023 17:44:06 +0800
Subject: [PATCH 4/4] Support the builtin functions for LASX

This patch does the following work:
- Define the builtin functions for LASX
- Add the header files lasxintrin.h
---
 .../include/clang/Basic/BuiltinsLoongArch.def |    3 +
 .../clang/Basic/BuiltinsLoongArchLASX.def     |  982 +++++
 clang/lib/Headers/CMakeLists.txt              |    1 +
 clang/lib/Headers/lasxintrin.h                | 3860 +++++++++++++++++
 clang/lib/Sema/SemaChecking.cpp               |  227 +
 5 files changed, 5073 insertions(+)
 create mode 100644 clang/include/clang/Basic/BuiltinsLoongArchLASX.def
 create mode 100644 clang/lib/Headers/lasxintrin.h

diff --git a/clang/include/clang/Basic/BuiltinsLoongArch.def b/clang/include/clang/Basic/BuiltinsLoongArch.def
index 9ec19c31095afff..95359a3fdc711d4 100644
--- a/clang/include/clang/Basic/BuiltinsLoongArch.def
+++ b/clang/include/clang/Basic/BuiltinsLoongArch.def
@@ -21,5 +21,8 @@
 // Definition of LSX builtins.
 #include "clang/Basic/BuiltinsLoongArchLSX.def"
 
+// Definition of LASX builtins.
+#include "clang/Basic/BuiltinsLoongArchLASX.def"
+
 #undef BUILTIN
 #undef TARGET_BUILTIN
diff --git a/clang/include/clang/Basic/BuiltinsLoongArchLASX.def b/clang/include/clang/Basic/BuiltinsLoongArchLASX.def
new file mode 100644
index 000000000000000..3de200f665b680a
--- /dev/null
+++ b/clang/include/clang/Basic/BuiltinsLoongArchLASX.def
@@ -0,0 +1,982 @@
+//=BuiltinsLoongArchLASX.def - LoongArch Builtin function database -- C++ -*-=//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the LoongArch-specific LASX builtin function database.
+// Users of this file must define the BUILTIN macro to make use of this
+// information.
+//
+//===----------------------------------------------------------------------===//
+
+TARGET_BUILTIN(__builtin_lasx_xvadd_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvadd_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvadd_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvadd_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvadd_q, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsub_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsub_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsub_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsub_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsub_q, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvaddi_bu, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddi_hu, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddi_wu, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddi_du, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsubi_bu, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubi_hu, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubi_wu, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubi_du, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvneg_b, "V32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvneg_h, "V16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvneg_w, "V8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvneg_d, "V4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsadd_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsadd_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsadd_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsadd_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsadd_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsadd_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsadd_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsadd_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssub_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssub_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssub_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssub_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssub_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssub_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssub_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssub_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvhaddw_h_b, "V16SsV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhaddw_w_h, "V8SiV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhaddw_d_w, "V4SLLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhaddw_q_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvhaddw_hu_bu, "V16UsV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhaddw_wu_hu, "V8UiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhaddw_du_wu, "V4ULLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhaddw_qu_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvhsubw_h_b, "V16SsV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhsubw_w_h, "V8SiV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhsubw_d_w, "V4SLLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhsubw_q_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvhsubw_hu_bu, "V16UsV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhsubw_wu_hu, "V8UiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhsubw_du_wu, "V4ULLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvhsubw_qu_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_h_b, "V16sV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_w_h, "V8SiV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_d_w, "V4LLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_q_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_h_b, "V16sV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_w_h, "V8SiV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_d_w, "V4LLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_q_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsubwev_h_b, "V16sV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwev_w_h, "V8SiV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwev_d_w, "V4LLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwev_q_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsubwod_h_b, "V16sV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwod_w_h, "V8SiV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwod_d_w, "V4LLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwod_q_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_h_bu, "V16sV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_w_hu, "V8SiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_d_wu, "V4LLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_q_du, "V4LLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_h_bu, "V16sV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_w_hu, "V8SiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_d_wu, "V4LLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_q_du, "V4LLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsubwev_h_bu, "V16sV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwev_w_hu, "V8SiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwev_d_wu, "V4LLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwev_q_du, "V4LLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsubwod_h_bu, "V16sV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwod_w_hu, "V8SiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwod_d_wu, "V4LLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsubwod_q_du, "V4LLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_h_bu_b, "V16sV32UcV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_w_hu_h, "V8SiV16UsV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_d_wu_w, "V4LLiV8UiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwev_q_du_d, "V4LLiV4ULLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_h_bu_b, "V16sV32UcV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_w_hu_h, "V8SiV16UsV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_d_wu_w, "V4LLiV8UiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvaddwod_q_du_d, "V4LLiV4ULLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvavg_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavg_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavg_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavg_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvavg_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavg_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavg_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavg_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvavgr_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavgr_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavgr_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavgr_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvavgr_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavgr_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavgr_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvavgr_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvabsd_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvabsd_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvabsd_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvabsd_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvabsd_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvabsd_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvabsd_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvabsd_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvadda_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvadda_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvadda_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvadda_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmax_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmax_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmax_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmax_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmaxi_b, "V32ScV32ScIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaxi_h, "V16SsV16SsIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaxi_w, "V8SiV8SiIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaxi_d, "V4SLLiV4SLLiIi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmax_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmax_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmax_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmax_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmaxi_bu, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaxi_hu, "V16UsV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaxi_wu, "V8UiV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaxi_du, "V4ULLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmin_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmin_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmin_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmin_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmini_b, "V32ScV32ScIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmini_h, "V16SsV16SsIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmini_w, "V8SiV8SiIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmini_d, "V4SLLiV4SLLiIi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmin_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmin_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmin_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmin_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmini_bu, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmini_hu, "V16UsV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmini_wu, "V8UiV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmini_du, "V4ULLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmul_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmul_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmul_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmul_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmuh_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmuh_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmuh_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmuh_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmuh_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmuh_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmuh_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmuh_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_h_b, "V16sV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_w_h, "V8SiV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_d_w, "V4LLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_q_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_h_b, "V16sV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_w_h, "V8SiV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_d_w, "V4LLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_q_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_h_bu, "V16sV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_w_hu, "V8SiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_d_wu, "V4LLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_q_du, "V4LLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_h_bu, "V16sV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_w_hu, "V8SiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_d_wu, "V4LLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_q_du, "V4LLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_h_bu_b, "V16sV32UcV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_w_hu_h, "V8SiV16UsV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_d_wu_w, "V4LLiV8UiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwev_q_du_d, "V4LLiV4ULLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_h_bu_b, "V16sV32UcV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_w_hu_h, "V8SiV16UsV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_d_wu_w, "V4LLiV8UiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmulwod_q_du_d, "V4LLiV4ULLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmadd_b, "V32ScV32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmadd_h, "V16SsV16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmadd_w, "V8SiV8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmadd_d, "V4SLLiV4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmsub_b, "V32ScV32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmsub_h, "V16SsV16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmsub_w, "V8SiV8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmsub_d, "V4SLLiV4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_h_b, "V16sV16sV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_w_h, "V8SiV8SiV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_d_w, "V4LLiV4LLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_q_d, "V4LLiV4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_h_b, "V16sV16sV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_w_h, "V8SiV8SiV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_d_w, "V4LLiV4LLiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_q_d, "V4LLiV4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_h_bu, "V16UsV16UsV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_w_hu, "V8UiV8UiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_d_wu, "V4ULLiV4ULLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_q_du, "V4ULLiV4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_h_bu, "V16UsV16UsV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_w_hu, "V8UiV8UiV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_d_wu, "V4ULLiV4ULLiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_q_du, "V4ULLiV4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_h_bu_b, "V16sV16sV32UcV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_w_hu_h, "V8SiV8SiV16UsV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_d_wu_w, "V4LLiV4LLiV8UiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwev_q_du_d, "V4LLiV4LLiV4ULLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_h_bu_b, "V16sV16sV32UcV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_w_hu_h, "V8SiV8SiV16UsV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_d_wu_w, "V4LLiV4LLiV8UiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmaddwod_q_du_d, "V4LLiV4LLiV4ULLiV4LLi", "nc", "lasx")
+
+
+TARGET_BUILTIN(__builtin_lasx_xvdiv_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvdiv_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvdiv_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvdiv_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvdiv_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvdiv_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvdiv_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvdiv_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmod_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmod_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmod_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmod_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmod_bu, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmod_hu, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmod_wu, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmod_du, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsat_b, "V32ScV32ScIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsat_h, "V16SsV16SsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsat_w, "V8SiV8SiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsat_d, "V4SLLiV4SLLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsat_bu, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsat_hu, "V16UsV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsat_wu, "V8UiV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsat_du, "V4ULLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvexth_h_b, "V16sV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvexth_w_h, "V8SiV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvexth_d_w, "V4LLiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvexth_q_d, "V4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvexth_hu_bu, "V16UsV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvexth_wu_hu, "V8UiV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvexth_du_wu, "V4ULLiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvexth_qu_du, "V4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_vext2xv_h_b, "V16sV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_vext2xv_w_b, "V8SiV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_vext2xv_d_b, "V4LLiV32c", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_vext2xv_w_h, "V8SiV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_vext2xv_d_h, "V4LLiV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_vext2xv_d_w, "V4LLiV8Si", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_vext2xv_hu_bu, "V16sV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_vext2xv_wu_bu, "V8SiV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_vext2xv_du_bu, "V4LLiV32c", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_vext2xv_wu_hu, "V8SiV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_vext2xv_du_hu, "V4LLiV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_vext2xv_du_wu, "V4LLiV8Si", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsigncov_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsigncov_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsigncov_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsigncov_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmskltz_b, "V32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmskltz_h, "V16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmskltz_w, "V8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmskltz_d, "V4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvmskgez_b, "V32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvmsknz_b, "V16sV16s", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvldi, "V4LLiIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrepli_b, "V32cIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrepli_h, "V16sIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrepli_w, "V8iIi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrepli_d, "V4LLiIi", "nc", "lasx")
+
+
+TARGET_BUILTIN(__builtin_lasx_xvand_v, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvor_v, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvxor_v, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvnor_v, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvandn_v, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvorn_v, "V32ScV32ScV32Sc", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvandi_b, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvori_b, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvxori_b, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvnori_b, "V32UcV32UcIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsll_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsll_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsll_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsll_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvslli_b, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslli_h, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslli_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslli_d, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrl_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrl_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrl_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrl_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrli_b, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrli_h, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrli_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrli_d, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsra_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsra_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsra_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsra_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrai_b, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrai_h, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrai_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrai_d, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvrotr_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrotr_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrotr_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrotr_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvrotri_b, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrotri_h, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrotri_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrotri_d, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsllwil_h_b, "V16sV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsllwil_w_h, "V8SiV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsllwil_d_w, "V4LLiV8SiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvextl_q_d, "V4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsllwil_hu_bu, "V16UsV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsllwil_wu_hu, "V8UiV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsllwil_du_wu, "V4ULLiV8UiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvextl_qu_du, "V4LLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrlr_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlr_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlr_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlr_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrlri_b, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlri_h, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlri_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlri_d, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrar_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrar_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrar_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrar_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrari_b, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrari_h, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrari_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrari_d, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrln_b_h, "V32ScV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrln_h_w, "V16sV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrln_w_d, "V8SiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsran_b_h, "V32ScV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsran_h_w, "V16sV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsran_w_d, "V8SiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrlni_b_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlni_h_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlni_w_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlni_d_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrani_b_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrani_h_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrani_w_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrani_d_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrlrn_b_h, "V32ScV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlrn_h_w, "V16sV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlrn_w_d, "V8SiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrarn_b_h, "V32ScV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrarn_h_w, "V16sV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrarn_w_d, "V8SiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrlrni_b_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlrni_h_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlrni_w_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrlrni_d_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsrarni_b_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrarni_h_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrarni_w_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsrarni_d_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrln_b_h, "V32ScV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrln_h_w, "V16sV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrln_w_d, "V8SiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssran_b_h, "V32ScV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssran_h_w, "V16sV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssran_w_d, "V8SiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrln_bu_h, "V32UcV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrln_hu_w, "V16UsV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrln_wu_d, "V8UiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssran_bu_h, "V32UcV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssran_hu_w, "V16UsV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssran_wu_d, "V8UiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrlni_b_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlni_h_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlni_w_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlni_d_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrani_b_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrani_h_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrani_w_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrani_d_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrlrni_bu_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrni_hu_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrni_wu_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrni_du_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrani_bu_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrani_hu_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrani_wu_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrani_du_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrlrn_b_h, "V32ScV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrn_h_w, "V16sV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrn_w_d, "V8SiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrarn_b_h, "V32ScV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarn_h_w, "V16sV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarn_w_d, "V8SiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrlrn_bu_h, "V32UcV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrn_hu_w, "V16UsV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrn_wu_d, "V8UiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrarn_bu_h, "V32UcV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarn_hu_w, "V16UsV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarn_wu_d, "V8UiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrlrni_b_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrni_h_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrni_w_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlrni_d_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrarni_b_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarni_h_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarni_w_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarni_d_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrlni_bu_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlni_hu_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlni_wu_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrlni_du_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvssrarni_bu_h, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarni_hu_w, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarni_wu_d, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvssrarni_du_q, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvclo_b, "V32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvclo_h, "V16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvclo_w, "V8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvclo_d, "V4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvclz_b, "V32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvclz_h, "V16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvclz_w, "V8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvclz_d, "V4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpcnt_b, "V32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpcnt_h, "V16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpcnt_w, "V8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpcnt_d, "V4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbitclr_b, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitclr_h, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitclr_w, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitclr_d, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbitclri_b, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitclri_h, "V16UsV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitclri_w, "V8UiV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitclri_d, "V4ULLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbitset_b, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitset_h, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitset_w, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitset_d, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbitseti_b, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitseti_h, "V16UsV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitseti_w, "V8UiV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitseti_d, "V4ULLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbitrev_b, "V32UcV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitrev_h, "V16UsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitrev_w, "V8UiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitrev_d, "V4ULLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbitrevi_b, "V32UcV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitrevi_h, "V16UsV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitrevi_w, "V8UiV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbitrevi_d, "V4ULLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrstp_b, "V32ScV32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrstp_h, "V16SsV16SsV16SsV16Ss", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrstpi_b, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrstpi_h, "V16sV16sV16sIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfadd_s, "V8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfadd_d, "V4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfsub_s, "V8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfsub_d, "V4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfmul_s, "V8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfmul_d, "V4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfdiv_s, "V8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfdiv_d, "V4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfmadd_s, "V8fV8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfmadd_d, "V4dV4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfmsub_s, "V8fV8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfmsub_d, "V4dV4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfnmadd_s, "V8fV8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfnmadd_d, "V4dV4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfnmsub_s, "V8fV8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfnmsub_d, "V4dV4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfmax_s, "V8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfmax_d, "V4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfmin_s, "V8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfmin_d, "V4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfmaxa_s, "V8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfmaxa_d, "V4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfmina_s, "V8fV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfmina_d, "V4dV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvflogb_s, "V8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvflogb_d, "V4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfclass_s, "V8iV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfclass_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfsqrt_s, "V8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfsqrt_d, "V4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrecip_s, "V8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrecip_d, "V4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrsqrt_s, "V8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrsqrt_d, "V4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcvtl_s_h, "V8fV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcvth_s_h, "V8fV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcvtl_d_s, "V4dV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcvth_d_s, "V4dV8f", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcvt_h_s, "V16sV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcvt_s_d, "V8fV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrintrne_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrintrne_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrintrz_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrintrz_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrintrp_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrintrp_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrintrm_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrintrm_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfrint_s, "V8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfrint_d, "V4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrne_w_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrne_l_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrz_w_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrz_l_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrp_w_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrp_l_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrm_w_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrm_l_d, "V4LLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftint_w_s, "V8SiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftint_l_d, "V4SLLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrz_wu_s, "V8UiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrz_lu_d, "V4ULLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftint_wu_s, "V8UiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftint_lu_d, "V4ULLiV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrne_w_d, "V8SiV4dV4d", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrz_w_d, "V8SiV4dV4d", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrp_w_d, "V8SiV4dV4d", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrm_w_d, "V8SiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftint_w_d, "V8SiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrnel_l_s, "V4LLiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrneh_l_s, "V4LLiV8f", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrzl_l_s, "V4LLiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrzh_l_s, "V4LLiV8f", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrpl_l_s, "V4LLiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrph_l_s, "V4LLiV8f", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintrml_l_s, "V4LLiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftintrmh_l_s, "V4LLiV8f", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvftintl_l_s, "V4LLiV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvftinth_l_s, "V4LLiV8f", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvffint_s_w, "V8fV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvffint_d_l, "V4dV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvffint_s_wu, "V8fV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvffint_d_lu, "V4dV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvffintl_d_w, "V4dV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvffinth_d_w, "V4dV8Si", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvffint_s_l, "V8fV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvseq_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvseq_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvseq_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvseq_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvseqi_b, "V32ScV32ScISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvseqi_h, "V16SsV16SsISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvseqi_w, "V8SiV8SiISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvseqi_d, "V4SLLiV4SLLiISi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsle_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsle_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsle_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsle_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvslei_b, "V32ScV32ScISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslei_h, "V16SsV16SsISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslei_w, "V8SiV8SiISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslei_d, "V4SLLiV4SLLiISi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvsle_bu, "V32ScV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsle_hu, "V16SsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsle_wu, "V8SiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvsle_du, "V4SLLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvslei_bu, "V32ScV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslei_hu, "V16SsV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslei_wu, "V8SiV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslei_du, "V4SLLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvslt_b, "V32ScV32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslt_h, "V16SsV16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslt_w, "V8SiV8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslt_d, "V4SLLiV4SLLiV4SLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvslti_b, "V32ScV32ScISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslti_h, "V16SsV16SsISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslti_w, "V8SiV8SiISi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslti_d, "V4SLLiV4SLLiISi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvslt_bu, "V32ScV32UcV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslt_hu, "V16SsV16UsV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslt_wu, "V8SiV8UiV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslt_du, "V4SLLiV4ULLiV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvslti_bu, "V32ScV32UcIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslti_hu, "V16SsV16UsIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslti_wu, "V8SiV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvslti_du, "V4SLLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_caf_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_caf_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cun_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cun_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_ceq_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_ceq_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cueq_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cueq_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_clt_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_clt_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cult_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cult_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cle_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cle_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cule_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cule_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cne_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cne_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cor_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cor_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cune_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_cune_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_saf_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_saf_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sun_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sun_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_seq_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_seq_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sueq_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sueq_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_slt_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_slt_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sult_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sult_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sle_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sle_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sule_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sule_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sne_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sne_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sor_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sor_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sune_s, "V8SiV8fV8f", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvfcmp_sune_d, "V4SLLiV4dV4d", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbitsel_v, "V32UcV32UcV32UcV32Uc", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbitseli_b, "V32UcV32UcV32UcIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvreplgr2vr_b, "V32Sci", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplgr2vr_h, "V16Ssi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplgr2vr_w, "V8Sii", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplgr2vr_d, "V4SLLiLLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvinsgr2vr_w, "V8SiV8SiiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvinsgr2vr_d, "V4SLLiV4SLLiLLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpickve2gr_w, "iV8SiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickve2gr_d, "LLiV4SLLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpickve2gr_wu, "iV8UiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickve2gr_du, "LLiV4ULLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvreplve_b, "V32cV32cUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplve_h, "V16sV16sUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplve_w, "V8iV8iUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplve_d, "V4LLiV4LLiUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvrepl128vei_b, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrepl128vei_h, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrepl128vei_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvrepl128vei_d, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvreplve0_b, "V32ScV32Sc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplve0_h, "V16SsV16Ss", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplve0_w, "V8SiV8Si", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplve0_d, "V4SLLiV4SLLi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvreplve0_q, "V32ScV32Sc", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvinsve0_w, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvinsve0_d, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpickve_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickve_d, "V4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpickve_w_f, "V8fV8fIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickve_d_f, "V4dV4dIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvbsll_v, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvbsrl_v, "V32cV32cIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpackev_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpackev_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpackev_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpackev_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpackod_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpackod_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpackod_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpackod_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpickev_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickev_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickev_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickev_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpickod_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickod_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickod_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpickod_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvilvl_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvilvl_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvilvl_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvilvl_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvilvh_b, "V32cV32cV32c", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvilvh_h, "V16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvilvh_w, "V8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvilvh_d, "V4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvshuf_b, "V32UcV32UcV32UcV32Uc", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvshuf_h, "V16sV16sV16sV16s", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvshuf_w, "V8iV8iV8iV8i", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvshuf_d, "V4LLiV4LLiV4LLiV4LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvperm_w, "V8iV8iV8i", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvshuf4i_b, "V32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvshuf4i_h, "V16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvshuf4i_w, "V8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvshuf4i_d, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvpermi_w, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpermi_d, "V4LLiV4LLiIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvpermi_q, "V32cV32cV32cIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvextrins_b, "V32cV32cV32cIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvextrins_h, "V16sV16sV16sIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvextrins_w, "V8iV8iV8iIUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvextrins_d, "V4LLiV4LLiV4LLiIUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvld, "V32ScvC*Ii", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvst, "vV32Scv*Ii", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvldx, "V32ScvC*LLi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvstx, "vV32Scv*LLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvldrepl_b, "V32cvC*Ii", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvldrepl_h, "V16svC*Ii", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvldrepl_w, "V8ivC*Ii", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvldrepl_d, "V4LLivC*Ii", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xvstelm_b, "vV32Scv*IiUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvstelm_h, "vV16Ssv*IiUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvstelm_w, "vV8Siv*IiUi", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xvstelm_d, "vV4SLLiv*IiUi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xbz_v, "iV32Uc", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xbnz_v, "iV32Uc", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xbz_b, "iV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xbz_h, "iV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xbz_w, "iV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xbz_d, "iV4ULLi", "nc", "lasx")
+
+TARGET_BUILTIN(__builtin_lasx_xbnz_b, "iV32Uc", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xbnz_h, "iV16Us", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xbnz_w, "iV8Ui", "nc", "lasx")
+TARGET_BUILTIN(__builtin_lasx_xbnz_d, "iV4ULLi", "nc", "lasx")
diff --git a/clang/lib/Headers/CMakeLists.txt b/clang/lib/Headers/CMakeLists.txt
index cee61a7e4fb0883..8b1e2bc4afa4dcd 100644
--- a/clang/lib/Headers/CMakeLists.txt
+++ b/clang/lib/Headers/CMakeLists.txt
@@ -93,6 +93,7 @@ set(hlsl_files
 
 set(loongarch_files
   larchintrin.h
+  lasxintrin.h
   lsxintrin.h
   )
 
diff --git a/clang/lib/Headers/lasxintrin.h b/clang/lib/Headers/lasxintrin.h
new file mode 100644
index 000000000000000..6b4d5012a24b589
--- /dev/null
+++ b/clang/lib/Headers/lasxintrin.h
@@ -0,0 +1,3860 @@
+/*===------------ lasxintrin.h - LoongArch LASX intrinsics -----------------===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+ *
+ *===-----------------------------------------------------------------------===
+ */
+
+#ifndef _LOONGSON_ASXINTRIN_H
+#define _LOONGSON_ASXINTRIN_H 1
+
+#if defined(__loongarch_asx)
+
+typedef signed char v32i8 __attribute__((vector_size(32), aligned(32)));
+typedef signed char v32i8_b __attribute__((vector_size(32), aligned(1)));
+typedef unsigned char v32u8 __attribute__((vector_size(32), aligned(32)));
+typedef unsigned char v32u8_b __attribute__((vector_size(32), aligned(1)));
+typedef short v16i16 __attribute__((vector_size(32), aligned(32)));
+typedef short v16i16_h __attribute__((vector_size(32), aligned(2)));
+typedef unsigned short v16u16 __attribute__((vector_size(32), aligned(32)));
+typedef unsigned short v16u16_h __attribute__((vector_size(32), aligned(2)));
+typedef int v8i32 __attribute__((vector_size(32), aligned(32)));
+typedef int v8i32_w __attribute__((vector_size(32), aligned(4)));
+typedef unsigned int v8u32 __attribute__((vector_size(32), aligned(32)));
+typedef unsigned int v8u32_w __attribute__((vector_size(32), aligned(4)));
+typedef long long v4i64 __attribute__((vector_size(32), aligned(32)));
+typedef long long v4i64_d __attribute__((vector_size(32), aligned(8)));
+typedef unsigned long long v4u64 __attribute__((vector_size(32), aligned(32)));
+typedef unsigned long long v4u64_d __attribute__((vector_size(32), aligned(8)));
+typedef float v8f32 __attribute__((vector_size(32), aligned(32)));
+typedef float v8f32_w __attribute__((vector_size(32), aligned(4)));
+typedef double v4f64 __attribute__((vector_size(32), aligned(32)));
+typedef double v4f64_d __attribute__((vector_size(32), aligned(8)));
+
+typedef double v4f64 __attribute__((vector_size(32), aligned(32)));
+typedef double v4f64_d __attribute__((vector_size(32), aligned(8)));
+
+typedef float __m256 __attribute__((__vector_size__(32), __may_alias__));
+typedef long long __m256i __attribute__((__vector_size__(32), __may_alias__));
+typedef double __m256d __attribute__((__vector_size__(32), __may_alias__));
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsll_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsll_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsll_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsll_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsll_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsll_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsll_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsll_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvslli_b(/*__m256i*/ _1, /*ui3*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslli_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvslli_h(/*__m256i*/ _1, /*ui4*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslli_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvslli_w(/*__m256i*/ _1, /*ui5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslli_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvslli_d(/*__m256i*/ _1, /*ui6*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslli_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsra_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsra_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsra_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsra_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsra_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsra_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsra_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsra_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvsrai_b(/*__m256i*/ _1, /*ui3*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsrai_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvsrai_h(/*__m256i*/ _1, /*ui4*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsrai_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvsrai_w(/*__m256i*/ _1, /*ui5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsrai_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvsrai_d(/*__m256i*/ _1, /*ui6*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsrai_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrar_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrar_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrar_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrar_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrar_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrar_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrar_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrar_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvsrari_b(/*__m256i*/ _1, /*ui3*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsrari_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvsrari_h(/*__m256i*/ _1, /*ui4*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsrari_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvsrari_w(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsrari_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvsrari_d(/*__m256i*/ _1, /*ui6*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsrari_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrl_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrl_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrl_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrl_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrl_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrl_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrl_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrl_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvsrli_b(/*__m256i*/ _1, /*ui3*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsrli_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvsrli_h(/*__m256i*/ _1, /*ui4*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsrli_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvsrli_w(/*__m256i*/ _1, /*ui5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsrli_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvsrli_d(/*__m256i*/ _1, /*ui6*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsrli_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrlr_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrlr_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrlr_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrlr_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrlr_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrlr_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrlr_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrlr_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvsrlri_b(/*__m256i*/ _1, /*ui3*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsrlri_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvsrlri_h(/*__m256i*/ _1, /*ui4*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsrlri_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvsrlri_w(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsrlri_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvsrlri_d(/*__m256i*/ _1, /*ui6*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsrlri_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitclr_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitclr_b((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitclr_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitclr_h((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitclr_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitclr_w((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitclr_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitclr_d((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvbitclri_b(/*__m256i*/ _1, /*ui3*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitclri_b((v32u8)(_1), (_2)))
+
+#define __lasx_xvbitclri_h(/*__m256i*/ _1, /*ui4*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitclri_h((v16u16)(_1), (_2)))
+
+#define __lasx_xvbitclri_w(/*__m256i*/ _1, /*ui5*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitclri_w((v8u32)(_1), (_2)))
+
+#define __lasx_xvbitclri_d(/*__m256i*/ _1, /*ui6*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitclri_d((v4u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitset_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitset_b((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitset_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitset_h((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitset_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitset_w((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitset_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitset_d((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvbitseti_b(/*__m256i*/ _1, /*ui3*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitseti_b((v32u8)(_1), (_2)))
+
+#define __lasx_xvbitseti_h(/*__m256i*/ _1, /*ui4*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitseti_h((v16u16)(_1), (_2)))
+
+#define __lasx_xvbitseti_w(/*__m256i*/ _1, /*ui5*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitseti_w((v8u32)(_1), (_2)))
+
+#define __lasx_xvbitseti_d(/*__m256i*/ _1, /*ui6*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitseti_d((v4u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitrev_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitrev_b((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitrev_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitrev_h((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitrev_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitrev_w((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitrev_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvbitrev_d((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvbitrevi_b(/*__m256i*/ _1, /*ui3*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitrevi_b((v32u8)(_1), (_2)))
+
+#define __lasx_xvbitrevi_h(/*__m256i*/ _1, /*ui4*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitrevi_h((v16u16)(_1), (_2)))
+
+#define __lasx_xvbitrevi_w(/*__m256i*/ _1, /*ui5*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitrevi_w((v8u32)(_1), (_2)))
+
+#define __lasx_xvbitrevi_d(/*__m256i*/ _1, /*ui6*/ _2)                         \
+  ((__m256i)__builtin_lasx_xvbitrevi_d((v4u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadd_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadd_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadd_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadd_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadd_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadd_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadd_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadd_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvaddi_bu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvaddi_bu((v32i8)(_1), (_2)))
+
+#define __lasx_xvaddi_hu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvaddi_hu((v16i16)(_1), (_2)))
+
+#define __lasx_xvaddi_wu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvaddi_wu((v8i32)(_1), (_2)))
+
+#define __lasx_xvaddi_du(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvaddi_du((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsub_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsub_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsub_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsub_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsub_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsub_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsub_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsub_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvsubi_bu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsubi_bu((v32i8)(_1), (_2)))
+
+#define __lasx_xvsubi_hu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsubi_hu((v16i16)(_1), (_2)))
+
+#define __lasx_xvsubi_wu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsubi_wu((v8i32)(_1), (_2)))
+
+#define __lasx_xvsubi_du(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvsubi_du((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmax_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmax_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmax_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmax_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmax_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmax_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmax_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmax_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvmaxi_b(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvmaxi_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvmaxi_h(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvmaxi_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvmaxi_w(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvmaxi_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvmaxi_d(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvmaxi_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmax_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmax_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmax_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmax_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmax_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmax_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmax_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmax_du((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvmaxi_bu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvmaxi_bu((v32u8)(_1), (_2)))
+
+#define __lasx_xvmaxi_hu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvmaxi_hu((v16u16)(_1), (_2)))
+
+#define __lasx_xvmaxi_wu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvmaxi_wu((v8u32)(_1), (_2)))
+
+#define __lasx_xvmaxi_du(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvmaxi_du((v4u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmin_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmin_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmin_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmin_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmin_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmin_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmin_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmin_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvmini_b(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvmini_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvmini_h(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvmini_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvmini_w(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvmini_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvmini_d(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvmini_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmin_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmin_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmin_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmin_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmin_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmin_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmin_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmin_du((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvmini_bu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvmini_bu((v32u8)(_1), (_2)))
+
+#define __lasx_xvmini_hu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvmini_hu((v16u16)(_1), (_2)))
+
+#define __lasx_xvmini_wu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvmini_wu((v8u32)(_1), (_2)))
+
+#define __lasx_xvmini_du(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvmini_du((v4u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvseq_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvseq_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvseq_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvseq_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvseq_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvseq_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvseq_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvseq_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvseqi_b(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvseqi_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvseqi_h(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvseqi_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvseqi_w(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvseqi_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvseqi_d(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvseqi_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvslt_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvslt_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvslt_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvslt_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvslt_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvslt_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvslt_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvslt_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvslti_b(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslti_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvslti_h(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslti_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvslti_w(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslti_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvslti_d(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslti_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvslt_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvslt_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvslt_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvslt_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvslt_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvslt_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvslt_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvslt_du((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvslti_bu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvslti_bu((v32u8)(_1), (_2)))
+
+#define __lasx_xvslti_hu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvslti_hu((v16u16)(_1), (_2)))
+
+#define __lasx_xvslti_wu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvslti_wu((v8u32)(_1), (_2)))
+
+#define __lasx_xvslti_du(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvslti_du((v4u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsle_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsle_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsle_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsle_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsle_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsle_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsle_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsle_d((v4i64)_1, (v4i64)_2);
+}
+
+#define __lasx_xvslei_b(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslei_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvslei_h(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslei_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvslei_w(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslei_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvslei_d(/*__m256i*/ _1, /*si5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvslei_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsle_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsle_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsle_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsle_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsle_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsle_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsle_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsle_du((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvslei_bu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvslei_bu((v32u8)(_1), (_2)))
+
+#define __lasx_xvslei_hu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvslei_hu((v16u16)(_1), (_2)))
+
+#define __lasx_xvslei_wu(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvslei_wu((v8u32)(_1), (_2)))
+
+#define __lasx_xvslei_du(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvslei_du((v4u64)(_1), (_2)))
+
+#define __lasx_xvsat_b(/*__m256i*/ _1, /*ui3*/ _2)                             \
+  ((__m256i)__builtin_lasx_xvsat_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvsat_h(/*__m256i*/ _1, /*ui4*/ _2)                             \
+  ((__m256i)__builtin_lasx_xvsat_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvsat_w(/*__m256i*/ _1, /*ui5*/ _2)                             \
+  ((__m256i)__builtin_lasx_xvsat_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvsat_d(/*__m256i*/ _1, /*ui6*/ _2)                             \
+  ((__m256i)__builtin_lasx_xvsat_d((v4i64)(_1), (_2)))
+
+#define __lasx_xvsat_bu(/*__m256i*/ _1, /*ui3*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsat_bu((v32u8)(_1), (_2)))
+
+#define __lasx_xvsat_hu(/*__m256i*/ _1, /*ui4*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsat_hu((v16u16)(_1), (_2)))
+
+#define __lasx_xvsat_wu(/*__m256i*/ _1, /*ui5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsat_wu((v8u32)(_1), (_2)))
+
+#define __lasx_xvsat_du(/*__m256i*/ _1, /*ui6*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvsat_du((v4u64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadda_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadda_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadda_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadda_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadda_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadda_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadda_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadda_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsadd_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsadd_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsadd_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsadd_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsadd_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsadd_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsadd_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsadd_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsadd_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsadd_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsadd_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsadd_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsadd_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsadd_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsadd_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsadd_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavg_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavg_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavg_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavg_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavg_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavg_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavg_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavg_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavg_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavg_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavg_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavg_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavg_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavg_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavg_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavg_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavgr_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavgr_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavgr_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavgr_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavgr_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavgr_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavgr_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavgr_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavgr_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavgr_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavgr_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavgr_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavgr_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavgr_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvavgr_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvavgr_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssub_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssub_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssub_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssub_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssub_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssub_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssub_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssub_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssub_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssub_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssub_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssub_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssub_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssub_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssub_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssub_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvabsd_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvabsd_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvabsd_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvabsd_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvabsd_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvabsd_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvabsd_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvabsd_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvabsd_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvabsd_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvabsd_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvabsd_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvabsd_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvabsd_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvabsd_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvabsd_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmul_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmul_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmul_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmul_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmul_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmul_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmul_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmul_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmadd_b(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmadd_b((v32i8)_1, (v32i8)_2, (v32i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmadd_h(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmadd_h((v16i16)_1, (v16i16)_2, (v16i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmadd_w(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmadd_w((v8i32)_1, (v8i32)_2, (v8i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmadd_d(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmadd_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmsub_b(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmsub_b((v32i8)_1, (v32i8)_2, (v32i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmsub_h(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmsub_h((v16i16)_1, (v16i16)_2, (v16i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmsub_w(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmsub_w((v8i32)_1, (v8i32)_2, (v8i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmsub_d(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmsub_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvdiv_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvdiv_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvdiv_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvdiv_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvdiv_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvdiv_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvdiv_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvdiv_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvdiv_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvdiv_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvdiv_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvdiv_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvdiv_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvdiv_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvdiv_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvdiv_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhaddw_h_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhaddw_h_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhaddw_w_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhaddw_w_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhaddw_d_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhaddw_d_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhaddw_hu_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhaddw_hu_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhaddw_wu_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhaddw_wu_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhaddw_du_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhaddw_du_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhsubw_h_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhsubw_h_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhsubw_w_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhsubw_w_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhsubw_d_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhsubw_d_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhsubw_hu_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhsubw_hu_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhsubw_wu_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhsubw_wu_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhsubw_du_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhsubw_du_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmod_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmod_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmod_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmod_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmod_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmod_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmod_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmod_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmod_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmod_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmod_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmod_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmod_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmod_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmod_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmod_du((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvrepl128vei_b(/*__m256i*/ _1, /*ui4*/ _2)                      \
+  ((__m256i)__builtin_lasx_xvrepl128vei_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvrepl128vei_h(/*__m256i*/ _1, /*ui3*/ _2)                      \
+  ((__m256i)__builtin_lasx_xvrepl128vei_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvrepl128vei_w(/*__m256i*/ _1, /*ui2*/ _2)                      \
+  ((__m256i)__builtin_lasx_xvrepl128vei_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvrepl128vei_d(/*__m256i*/ _1, /*ui1*/ _2)                      \
+  ((__m256i)__builtin_lasx_xvrepl128vei_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpickev_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpickev_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpickev_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpickev_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpickev_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpickev_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpickev_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpickev_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpickod_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpickod_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpickod_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpickod_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpickod_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpickod_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpickod_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpickod_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvilvh_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvilvh_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvilvh_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvilvh_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvilvh_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvilvh_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvilvh_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvilvh_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvilvl_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvilvl_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvilvl_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvilvl_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvilvl_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvilvl_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvilvl_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvilvl_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpackev_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpackev_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpackev_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpackev_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpackev_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpackev_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpackev_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpackev_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpackod_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpackod_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpackod_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpackod_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpackod_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpackod_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpackod_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvpackod_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvshuf_b(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvshuf_b((v32i8)_1, (v32i8)_2, (v32i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvshuf_h(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvshuf_h((v16i16)_1, (v16i16)_2, (v16i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvshuf_w(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvshuf_w((v8i32)_1, (v8i32)_2, (v8i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvshuf_d(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvshuf_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvand_v(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvand_v((v32u8)_1, (v32u8)_2);
+}
+
+#define __lasx_xvandi_b(/*__m256i*/ _1, /*ui8*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvandi_b((v32u8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvor_v(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvor_v((v32u8)_1, (v32u8)_2);
+}
+
+#define __lasx_xvori_b(/*__m256i*/ _1, /*ui8*/ _2)                             \
+  ((__m256i)__builtin_lasx_xvori_b((v32u8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvnor_v(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvnor_v((v32u8)_1, (v32u8)_2);
+}
+
+#define __lasx_xvnori_b(/*__m256i*/ _1, /*ui8*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvnori_b((v32u8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvxor_v(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvxor_v((v32u8)_1, (v32u8)_2);
+}
+
+#define __lasx_xvxori_b(/*__m256i*/ _1, /*ui8*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvxori_b((v32u8)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvbitsel_v(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvbitsel_v((v32u8)_1, (v32u8)_2, (v32u8)_3);
+}
+
+#define __lasx_xvbitseli_b(/*__m256i*/ _1, /*__m256i*/ _2, /*ui8*/ _3)         \
+  ((__m256i)__builtin_lasx_xvbitseli_b((v32u8)(_1), (v32u8)(_2), (_3)))
+
+#define __lasx_xvshuf4i_b(/*__m256i*/ _1, /*ui8*/ _2)                          \
+  ((__m256i)__builtin_lasx_xvshuf4i_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvshuf4i_h(/*__m256i*/ _1, /*ui8*/ _2)                          \
+  ((__m256i)__builtin_lasx_xvshuf4i_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvshuf4i_w(/*__m256i*/ _1, /*ui8*/ _2)                          \
+  ((__m256i)__builtin_lasx_xvshuf4i_w((v8i32)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplgr2vr_b(int _1) {
+  return (__m256i)__builtin_lasx_xvreplgr2vr_b((int)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplgr2vr_h(int _1) {
+  return (__m256i)__builtin_lasx_xvreplgr2vr_h((int)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplgr2vr_w(int _1) {
+  return (__m256i)__builtin_lasx_xvreplgr2vr_w((int)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplgr2vr_d(long int _1) {
+  return (__m256i)__builtin_lasx_xvreplgr2vr_d((long int)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpcnt_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvpcnt_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpcnt_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvpcnt_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpcnt_w(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvpcnt_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvpcnt_d(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvpcnt_d((v4i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvclo_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvclo_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvclo_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvclo_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvclo_w(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvclo_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvclo_d(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvclo_d((v4i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvclz_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvclz_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvclz_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvclz_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvclz_w(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvclz_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvclz_d(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvclz_d((v4i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfadd_s(__m256 _1, __m256 _2) {
+  return (__m256)__builtin_lasx_xvfadd_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfadd_d(__m256d _1, __m256d _2) {
+  return (__m256d)__builtin_lasx_xvfadd_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfsub_s(__m256 _1, __m256 _2) {
+  return (__m256)__builtin_lasx_xvfsub_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfsub_d(__m256d _1, __m256d _2) {
+  return (__m256d)__builtin_lasx_xvfsub_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfmul_s(__m256 _1, __m256 _2) {
+  return (__m256)__builtin_lasx_xvfmul_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfmul_d(__m256d _1, __m256d _2) {
+  return (__m256d)__builtin_lasx_xvfmul_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfdiv_s(__m256 _1, __m256 _2) {
+  return (__m256)__builtin_lasx_xvfdiv_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfdiv_d(__m256d _1, __m256d _2) {
+  return (__m256d)__builtin_lasx_xvfdiv_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcvt_h_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcvt_h_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfcvt_s_d(__m256d _1, __m256d _2) {
+  return (__m256)__builtin_lasx_xvfcvt_s_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfmin_s(__m256 _1, __m256 _2) {
+  return (__m256)__builtin_lasx_xvfmin_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfmin_d(__m256d _1, __m256d _2) {
+  return (__m256d)__builtin_lasx_xvfmin_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfmina_s(__m256 _1, __m256 _2) {
+  return (__m256)__builtin_lasx_xvfmina_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfmina_d(__m256d _1, __m256d _2) {
+  return (__m256d)__builtin_lasx_xvfmina_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfmax_s(__m256 _1, __m256 _2) {
+  return (__m256)__builtin_lasx_xvfmax_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfmax_d(__m256d _1, __m256d _2) {
+  return (__m256d)__builtin_lasx_xvfmax_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfmaxa_s(__m256 _1, __m256 _2) {
+  return (__m256)__builtin_lasx_xvfmaxa_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfmaxa_d(__m256d _1, __m256d _2) {
+  return (__m256d)__builtin_lasx_xvfmaxa_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfclass_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvfclass_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfclass_d(__m256d _1) {
+  return (__m256i)__builtin_lasx_xvfclass_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfsqrt_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvfsqrt_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfsqrt_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvfsqrt_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfrecip_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvfrecip_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfrecip_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvfrecip_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfrint_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvfrint_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfrint_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvfrint_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfrsqrt_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvfrsqrt_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfrsqrt_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvfrsqrt_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvflogb_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvflogb_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvflogb_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvflogb_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfcvth_s_h(__m256i _1) {
+  return (__m256)__builtin_lasx_xvfcvth_s_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfcvth_d_s(__m256 _1) {
+  return (__m256d)__builtin_lasx_xvfcvth_d_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfcvtl_s_h(__m256i _1) {
+  return (__m256)__builtin_lasx_xvfcvtl_s_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfcvtl_d_s(__m256 _1) {
+  return (__m256d)__builtin_lasx_xvfcvtl_d_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftint_w_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftint_w_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftint_l_d(__m256d _1) {
+  return (__m256i)__builtin_lasx_xvftint_l_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftint_wu_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftint_wu_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftint_lu_d(__m256d _1) {
+  return (__m256i)__builtin_lasx_xvftint_lu_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrz_w_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrz_w_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrz_l_d(__m256d _1) {
+  return (__m256i)__builtin_lasx_xvftintrz_l_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrz_wu_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrz_wu_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrz_lu_d(__m256d _1) {
+  return (__m256i)__builtin_lasx_xvftintrz_lu_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvffint_s_w(__m256i _1) {
+  return (__m256)__builtin_lasx_xvffint_s_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvffint_d_l(__m256i _1) {
+  return (__m256d)__builtin_lasx_xvffint_d_l((v4i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvffint_s_wu(__m256i _1) {
+  return (__m256)__builtin_lasx_xvffint_s_wu((v8u32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvffint_d_lu(__m256i _1) {
+  return (__m256d)__builtin_lasx_xvffint_d_lu((v4u64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve_b(__m256i _1, int _2) {
+  return (__m256i)__builtin_lasx_xvreplve_b((v32i8)_1, (int)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve_h(__m256i _1, int _2) {
+  return (__m256i)__builtin_lasx_xvreplve_h((v16i16)_1, (int)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve_w(__m256i _1, int _2) {
+  return (__m256i)__builtin_lasx_xvreplve_w((v8i32)_1, (int)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve_d(__m256i _1, int _2) {
+  return (__m256i)__builtin_lasx_xvreplve_d((v4i64)_1, (int)_2);
+}
+
+#define __lasx_xvpermi_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui8*/ _3)           \
+  ((__m256i)__builtin_lasx_xvpermi_w((v8i32)(_1), (v8i32)(_2), (_3)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvandn_v(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvandn_v((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvneg_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvneg_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvneg_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvneg_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvneg_w(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvneg_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvneg_d(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvneg_d((v4i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmuh_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmuh_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmuh_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmuh_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmuh_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmuh_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmuh_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmuh_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmuh_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmuh_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmuh_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmuh_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmuh_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmuh_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmuh_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmuh_du((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvsllwil_h_b(/*__m256i*/ _1, /*ui3*/ _2)                        \
+  ((__m256i)__builtin_lasx_xvsllwil_h_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvsllwil_w_h(/*__m256i*/ _1, /*ui4*/ _2)                        \
+  ((__m256i)__builtin_lasx_xvsllwil_w_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvsllwil_d_w(/*__m256i*/ _1, /*ui5*/ _2)                        \
+  ((__m256i)__builtin_lasx_xvsllwil_d_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvsllwil_hu_bu(/*__m256i*/ _1, /*ui3*/ _2)                      \
+  ((__m256i)__builtin_lasx_xvsllwil_hu_bu((v32u8)(_1), (_2)))
+
+#define __lasx_xvsllwil_wu_hu(/*__m256i*/ _1, /*ui4*/ _2)                      \
+  ((__m256i)__builtin_lasx_xvsllwil_wu_hu((v16u16)(_1), (_2)))
+
+#define __lasx_xvsllwil_du_wu(/*__m256i*/ _1, /*ui5*/ _2)                      \
+  ((__m256i)__builtin_lasx_xvsllwil_du_wu((v8u32)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsran_b_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsran_b_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsran_h_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsran_h_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsran_w_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsran_w_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssran_b_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssran_b_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssran_h_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssran_h_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssran_w_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssran_w_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssran_bu_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssran_bu_h((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssran_hu_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssran_hu_w((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssran_wu_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssran_wu_d((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrarn_b_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrarn_b_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrarn_h_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrarn_h_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrarn_w_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrarn_w_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrarn_b_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrarn_b_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrarn_h_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrarn_h_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrarn_w_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrarn_w_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrarn_bu_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrarn_bu_h((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrarn_hu_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrarn_hu_w((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrarn_wu_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrarn_wu_d((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrln_b_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrln_b_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrln_h_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrln_h_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrln_w_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrln_w_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrln_bu_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrln_bu_h((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrln_hu_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrln_hu_w((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrln_wu_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrln_wu_d((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrlrn_b_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrlrn_b_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrlrn_h_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrlrn_h_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsrlrn_w_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsrlrn_w_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrlrn_bu_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrlrn_bu_h((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrlrn_hu_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrlrn_hu_w((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrlrn_wu_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrlrn_wu_d((v4u64)_1, (v4u64)_2);
+}
+
+#define __lasx_xvfrstpi_b(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)          \
+  ((__m256i)__builtin_lasx_xvfrstpi_b((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvfrstpi_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)          \
+  ((__m256i)__builtin_lasx_xvfrstpi_h((v16i16)(_1), (v16i16)(_2), (_3)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfrstp_b(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvfrstp_b((v32i8)_1, (v32i8)_2, (v32i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfrstp_h(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvfrstp_h((v16i16)_1, (v16i16)_2, (v16i16)_3);
+}
+
+#define __lasx_xvshuf4i_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui8*/ _3)          \
+  ((__m256i)__builtin_lasx_xvshuf4i_d((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvbsrl_v(/*__m256i*/ _1, /*ui5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvbsrl_v((v32i8)(_1), (_2)))
+
+#define __lasx_xvbsll_v(/*__m256i*/ _1, /*ui5*/ _2)                            \
+  ((__m256i)__builtin_lasx_xvbsll_v((v32i8)(_1), (_2)))
+
+#define __lasx_xvextrins_b(/*__m256i*/ _1, /*__m256i*/ _2, /*ui8*/ _3)         \
+  ((__m256i)__builtin_lasx_xvextrins_b((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvextrins_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui8*/ _3)         \
+  ((__m256i)__builtin_lasx_xvextrins_h((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvextrins_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui8*/ _3)         \
+  ((__m256i)__builtin_lasx_xvextrins_w((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvextrins_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui8*/ _3)         \
+  ((__m256i)__builtin_lasx_xvextrins_d((v4i64)(_1), (v4i64)(_2), (_3)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmskltz_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvmskltz_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmskltz_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvmskltz_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmskltz_w(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvmskltz_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmskltz_d(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvmskltz_d((v4i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsigncov_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsigncov_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsigncov_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsigncov_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsigncov_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsigncov_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsigncov_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsigncov_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfmadd_s(__m256 _1, __m256 _2, __m256 _3) {
+  return (__m256)__builtin_lasx_xvfmadd_s((v8f32)_1, (v8f32)_2, (v8f32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfmadd_d(__m256d _1, __m256d _2, __m256d _3) {
+  return (__m256d)__builtin_lasx_xvfmadd_d((v4f64)_1, (v4f64)_2, (v4f64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfmsub_s(__m256 _1, __m256 _2, __m256 _3) {
+  return (__m256)__builtin_lasx_xvfmsub_s((v8f32)_1, (v8f32)_2, (v8f32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfmsub_d(__m256d _1, __m256d _2, __m256d _3) {
+  return (__m256d)__builtin_lasx_xvfmsub_d((v4f64)_1, (v4f64)_2, (v4f64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfnmadd_s(__m256 _1, __m256 _2, __m256 _3) {
+  return (__m256)__builtin_lasx_xvfnmadd_s((v8f32)_1, (v8f32)_2, (v8f32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfnmadd_d(__m256d _1, __m256d _2, __m256d _3) {
+  return (__m256d)__builtin_lasx_xvfnmadd_d((v4f64)_1, (v4f64)_2, (v4f64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfnmsub_s(__m256 _1, __m256 _2, __m256 _3) {
+  return (__m256)__builtin_lasx_xvfnmsub_s((v8f32)_1, (v8f32)_2, (v8f32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfnmsub_d(__m256d _1, __m256d _2, __m256d _3) {
+  return (__m256d)__builtin_lasx_xvfnmsub_d((v4f64)_1, (v4f64)_2, (v4f64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrne_w_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrne_w_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrne_l_d(__m256d _1) {
+  return (__m256i)__builtin_lasx_xvftintrne_l_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrp_w_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrp_w_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrp_l_d(__m256d _1) {
+  return (__m256i)__builtin_lasx_xvftintrp_l_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrm_w_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrm_w_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrm_l_d(__m256d _1) {
+  return (__m256i)__builtin_lasx_xvftintrm_l_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftint_w_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvftint_w_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvffint_s_l(__m256i _1, __m256i _2) {
+  return (__m256)__builtin_lasx_xvffint_s_l((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrz_w_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvftintrz_w_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrp_w_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvftintrp_w_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrm_w_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvftintrm_w_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrne_w_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvftintrne_w_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftinth_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftinth_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintl_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintl_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvffinth_d_w(__m256i _1) {
+  return (__m256d)__builtin_lasx_xvffinth_d_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvffintl_d_w(__m256i _1) {
+  return (__m256d)__builtin_lasx_xvffintl_d_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrzh_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrzh_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrzl_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrzl_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrph_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrph_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrpl_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrpl_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrmh_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrmh_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrml_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrml_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrneh_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrneh_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvftintrnel_l_s(__m256 _1) {
+  return (__m256i)__builtin_lasx_xvftintrnel_l_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfrintrne_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvfrintrne_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfrintrne_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvfrintrne_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfrintrz_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvfrintrz_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfrintrz_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvfrintrz_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfrintrp_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvfrintrp_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfrintrp_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvfrintrp_d((v4f64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256
+    __lasx_xvfrintrm_s(__m256 _1) {
+  return (__m256)__builtin_lasx_xvfrintrm_s((v8f32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256d
+    __lasx_xvfrintrm_d(__m256d _1) {
+  return (__m256d)__builtin_lasx_xvfrintrm_d((v4f64)_1);
+}
+
+#define __lasx_xvld(/*void **/ _1, /*si12*/ _2)                                \
+  ((__m256i)__builtin_lasx_xvld((void const *)(_1), (_2)))
+
+#define __lasx_xvst(/*__m256i*/ _1, /*void **/ _2, /*si12*/ _3)                \
+  ((void)__builtin_lasx_xvst((v32i8)(_1), (void *)(_2), (_3)))
+
+#define __lasx_xvstelm_b(/*__m256i*/ _1, /*void **/ _2, /*si8*/ _3,            \
+                         /*idx*/ _4)                                           \
+  ((void)__builtin_lasx_xvstelm_b((v32i8)(_1), (void *)(_2), (_3), (_4)))
+
+#define __lasx_xvstelm_h(/*__m256i*/ _1, /*void **/ _2, /*si8*/ _3,            \
+                         /*idx*/ _4)                                           \
+  ((void)__builtin_lasx_xvstelm_h((v16i16)(_1), (void *)(_2), (_3), (_4)))
+
+#define __lasx_xvstelm_w(/*__m256i*/ _1, /*void **/ _2, /*si8*/ _3,            \
+                         /*idx*/ _4)                                           \
+  ((void)__builtin_lasx_xvstelm_w((v8i32)(_1), (void *)(_2), (_3), (_4)))
+
+#define __lasx_xvstelm_d(/*__m256i*/ _1, /*void **/ _2, /*si8*/ _3,            \
+                         /*idx*/ _4)                                           \
+  ((void)__builtin_lasx_xvstelm_d((v4i64)(_1), (void *)(_2), (_3), (_4)))
+
+#define __lasx_xvinsve0_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui3*/ _3)          \
+  ((__m256i)__builtin_lasx_xvinsve0_w((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvinsve0_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui2*/ _3)          \
+  ((__m256i)__builtin_lasx_xvinsve0_d((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvpickve_w(/*__m256i*/ _1, /*ui3*/ _2)                          \
+  ((__m256i)__builtin_lasx_xvpickve_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvpickve_d(/*__m256i*/ _1, /*ui2*/ _2)                          \
+  ((__m256i)__builtin_lasx_xvpickve_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrlrn_b_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrlrn_b_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrlrn_h_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrlrn_h_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrlrn_w_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrlrn_w_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrln_b_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrln_b_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrln_h_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrln_h_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvssrln_w_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvssrln_w_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvorn_v(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvorn_v((v32i8)_1, (v32i8)_2);
+}
+
+#define __lasx_xvldi(/*i13*/ _1) ((__m256i)__builtin_lasx_xvldi((_1)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvldx(void const *_1, long int _2) {
+  return (__m256i)__builtin_lasx_xvldx((void const *)_1, (long int)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) void
+    __lasx_xvstx(__m256i _1, void *_2, long int _3) {
+  return (void)__builtin_lasx_xvstx((v32i8)_1, (void *)_2, (long int)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvextl_qu_du(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvextl_qu_du((v4u64)_1);
+}
+
+#define __lasx_xvinsgr2vr_w(/*__m256i*/ _1, /*int*/ _2, /*ui3*/ _3)            \
+  ((__m256i)__builtin_lasx_xvinsgr2vr_w((v8i32)(_1), (int)(_2), (_3)))
+
+#define __lasx_xvinsgr2vr_d(/*__m256i*/ _1, /*long int*/ _2, /*ui2*/ _3)       \
+  ((__m256i)__builtin_lasx_xvinsgr2vr_d((v4i64)(_1), (long int)(_2), (_3)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve0_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvreplve0_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve0_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvreplve0_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve0_w(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvreplve0_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve0_d(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvreplve0_d((v4i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvreplve0_q(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvreplve0_q((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_h_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_h_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_w_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_w_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_d_w(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_d_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_w_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_w_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_d_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_d_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_d_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_d_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_hu_bu(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_hu_bu((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_wu_hu(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_wu_hu((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_du_wu(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_du_wu((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_wu_bu(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_wu_bu((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_du_hu(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_du_hu((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_vext2xv_du_bu(__m256i _1) {
+  return (__m256i)__builtin_lasx_vext2xv_du_bu((v32i8)_1);
+}
+
+#define __lasx_xvpermi_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui8*/ _3)           \
+  ((__m256i)__builtin_lasx_xvpermi_q((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvpermi_d(/*__m256i*/ _1, /*ui8*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvpermi_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvperm_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvperm_w((v8i32)_1, (v8i32)_2);
+}
+
+#define __lasx_xvldrepl_b(/*void **/ _1, /*si12*/ _2)                          \
+  ((__m256i)__builtin_lasx_xvldrepl_b((void const *)(_1), (_2)))
+
+#define __lasx_xvldrepl_h(/*void **/ _1, /*si11*/ _2)                          \
+  ((__m256i)__builtin_lasx_xvldrepl_h((void const *)(_1), (_2)))
+
+#define __lasx_xvldrepl_w(/*void **/ _1, /*si10*/ _2)                          \
+  ((__m256i)__builtin_lasx_xvldrepl_w((void const *)(_1), (_2)))
+
+#define __lasx_xvldrepl_d(/*void **/ _1, /*si9*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvldrepl_d((void const *)(_1), (_2)))
+
+#define __lasx_xvpickve2gr_w(/*__m256i*/ _1, /*ui3*/ _2)                       \
+  ((int)__builtin_lasx_xvpickve2gr_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvpickve2gr_wu(/*__m256i*/ _1, /*ui3*/ _2)                      \
+  ((unsigned int)__builtin_lasx_xvpickve2gr_wu((v8i32)(_1), (_2)))
+
+#define __lasx_xvpickve2gr_d(/*__m256i*/ _1, /*ui2*/ _2)                       \
+  ((long int)__builtin_lasx_xvpickve2gr_d((v4i64)(_1), (_2)))
+
+#define __lasx_xvpickve2gr_du(/*__m256i*/ _1, /*ui2*/ _2)                      \
+  ((unsigned long int)__builtin_lasx_xvpickve2gr_du((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_q_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_q_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_d_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_d_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_w_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_w_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_h_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_h_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_q_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_q_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_d_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_d_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_w_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_w_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_h_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_h_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwev_q_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwev_q_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwev_d_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwev_d_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwev_w_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwev_w_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwev_h_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwev_h_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwev_q_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwev_q_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwev_d_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwev_d_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwev_w_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwev_w_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwev_h_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwev_h_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_q_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_q_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_d_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_d_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_w_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_w_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_h_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_h_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_q_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_q_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_d_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_d_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_w_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_w_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_h_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_h_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_q_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_q_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_d_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_d_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_w_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_w_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_h_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_h_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_q_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_q_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_d_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_d_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_w_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_w_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_h_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_h_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwod_q_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwod_q_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwod_d_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwod_d_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwod_w_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwod_w_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwod_h_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwod_h_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwod_q_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwod_q_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwod_d_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwod_d_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwod_w_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwod_w_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsubwod_h_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsubwod_h_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_q_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_q_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_d_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_d_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_w_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_w_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_h_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_h_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_q_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_q_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_d_wu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_d_wu((v8u32)_1, (v8u32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_w_hu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_w_hu((v16u16)_1, (v16u16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_h_bu(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_h_bu((v32u8)_1, (v32u8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_d_wu_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_d_wu_w((v8u32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_w_hu_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_w_hu_h((v16u16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_h_bu_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_h_bu_b((v32u8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_d_wu_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_d_wu_w((v8u32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_w_hu_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_w_hu_h((v16u16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_h_bu_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_h_bu_b((v32u8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_d_wu_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_d_wu_w((v8u32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_w_hu_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_w_hu_h((v16u16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_h_bu_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_h_bu_b((v32u8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_d_wu_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_d_wu_w((v8u32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_w_hu_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_w_hu_h((v16u16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_h_bu_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_h_bu_b((v32u8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhaddw_q_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhaddw_q_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhaddw_qu_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhaddw_qu_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhsubw_q_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhsubw_q_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvhsubw_qu_du(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvhsubw_qu_du((v4u64)_1, (v4u64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_q_d(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_q_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_d_w(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_d_w((v4i64)_1, (v8i32)_2, (v8i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_w_h(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_w_h((v8i32)_1, (v16i16)_2,
+                                               (v16i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_h_b(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_h_b((v16i16)_1, (v32i8)_2,
+                                               (v32i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_q_du(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_q_du((v4u64)_1, (v4u64)_2,
+                                                (v4u64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_d_wu(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_d_wu((v4u64)_1, (v8u32)_2,
+                                                (v8u32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_w_hu(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_w_hu((v8u32)_1, (v16u16)_2,
+                                                (v16u16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_h_bu(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_h_bu((v16u16)_1, (v32u8)_2,
+                                                (v32u8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_q_d(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_q_d((v4i64)_1, (v4i64)_2, (v4i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_d_w(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_d_w((v4i64)_1, (v8i32)_2, (v8i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_w_h(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_w_h((v8i32)_1, (v16i16)_2,
+                                               (v16i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_h_b(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_h_b((v16i16)_1, (v32i8)_2,
+                                               (v32i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_q_du(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_q_du((v4u64)_1, (v4u64)_2,
+                                                (v4u64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_d_wu(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_d_wu((v4u64)_1, (v8u32)_2,
+                                                (v8u32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_w_hu(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_w_hu((v8u32)_1, (v16u16)_2,
+                                                (v16u16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_h_bu(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_h_bu((v16u16)_1, (v32u8)_2,
+                                                (v32u8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_q_du_d(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_q_du_d((v4i64)_1, (v4u64)_2,
+                                                  (v4i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_d_wu_w(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_d_wu_w((v4i64)_1, (v8u32)_2,
+                                                  (v8i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_w_hu_h(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_w_hu_h((v8i32)_1, (v16u16)_2,
+                                                  (v16i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwev_h_bu_b(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwev_h_bu_b((v16i16)_1, (v32u8)_2,
+                                                  (v32i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_q_du_d(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_q_du_d((v4i64)_1, (v4u64)_2,
+                                                  (v4i64)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_d_wu_w(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_d_wu_w((v4i64)_1, (v8u32)_2,
+                                                  (v8i32)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_w_hu_h(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_w_hu_h((v8i32)_1, (v16u16)_2,
+                                                  (v16i16)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmaddwod_h_bu_b(__m256i _1, __m256i _2, __m256i _3) {
+  return (__m256i)__builtin_lasx_xvmaddwod_h_bu_b((v16i16)_1, (v32u8)_2,
+                                                  (v32i8)_3);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvrotr_b(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvrotr_b((v32i8)_1, (v32i8)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvrotr_h(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvrotr_h((v16i16)_1, (v16i16)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvrotr_w(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvrotr_w((v8i32)_1, (v8i32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvrotr_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvrotr_d((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvadd_q(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvadd_q((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvsub_q(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvsub_q((v4i64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwev_q_du_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwev_q_du_d((v4u64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvaddwod_q_du_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvaddwod_q_du_d((v4u64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwev_q_du_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwev_q_du_d((v4u64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmulwod_q_du_d(__m256i _1, __m256i _2) {
+  return (__m256i)__builtin_lasx_xvmulwod_q_du_d((v4u64)_1, (v4i64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmskgez_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvmskgez_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvmsknz_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvmsknz_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvexth_h_b(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvexth_h_b((v32i8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvexth_w_h(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvexth_w_h((v16i16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvexth_d_w(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvexth_d_w((v8i32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvexth_q_d(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvexth_q_d((v4i64)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvexth_hu_bu(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvexth_hu_bu((v32u8)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvexth_wu_hu(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvexth_wu_hu((v16u16)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvexth_du_wu(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvexth_du_wu((v8u32)_1);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvexth_qu_du(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvexth_qu_du((v4u64)_1);
+}
+
+#define __lasx_xvrotri_b(/*__m256i*/ _1, /*ui3*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvrotri_b((v32i8)(_1), (_2)))
+
+#define __lasx_xvrotri_h(/*__m256i*/ _1, /*ui4*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvrotri_h((v16i16)(_1), (_2)))
+
+#define __lasx_xvrotri_w(/*__m256i*/ _1, /*ui5*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvrotri_w((v8i32)(_1), (_2)))
+
+#define __lasx_xvrotri_d(/*__m256i*/ _1, /*ui6*/ _2)                           \
+  ((__m256i)__builtin_lasx_xvrotri_d((v4i64)(_1), (_2)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvextl_q_d(__m256i _1) {
+  return (__m256i)__builtin_lasx_xvextl_q_d((v4i64)_1);
+}
+
+#define __lasx_xvsrlni_b_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)         \
+  ((__m256i)__builtin_lasx_xvsrlni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvsrlni_h_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)         \
+  ((__m256i)__builtin_lasx_xvsrlni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvsrlni_w_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)         \
+  ((__m256i)__builtin_lasx_xvsrlni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvsrlni_d_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)         \
+  ((__m256i)__builtin_lasx_xvsrlni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvsrlrni_b_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)        \
+  ((__m256i)__builtin_lasx_xvsrlrni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvsrlrni_h_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)        \
+  ((__m256i)__builtin_lasx_xvsrlrni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvsrlrni_w_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)        \
+  ((__m256i)__builtin_lasx_xvsrlrni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvsrlrni_d_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)        \
+  ((__m256i)__builtin_lasx_xvsrlrni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvssrlni_b_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)        \
+  ((__m256i)__builtin_lasx_xvssrlni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvssrlni_h_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)        \
+  ((__m256i)__builtin_lasx_xvssrlni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvssrlni_w_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)        \
+  ((__m256i)__builtin_lasx_xvssrlni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvssrlni_d_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)        \
+  ((__m256i)__builtin_lasx_xvssrlni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvssrlni_bu_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrlni_bu_h((v32u8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvssrlni_hu_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrlni_hu_w((v16u16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvssrlni_wu_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrlni_wu_d((v8u32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvssrlni_du_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrlni_du_q((v4u64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvssrlrni_b_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrlrni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvssrlrni_h_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrlrni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvssrlrni_w_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrlrni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvssrlrni_d_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrlrni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvssrlrni_bu_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)      \
+  ((__m256i)__builtin_lasx_xvssrlrni_bu_h((v32u8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvssrlrni_hu_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)      \
+  ((__m256i)__builtin_lasx_xvssrlrni_hu_w((v16u16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvssrlrni_wu_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)      \
+  ((__m256i)__builtin_lasx_xvssrlrni_wu_d((v8u32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvssrlrni_du_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)      \
+  ((__m256i)__builtin_lasx_xvssrlrni_du_q((v4u64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvsrani_b_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)         \
+  ((__m256i)__builtin_lasx_xvsrani_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvsrani_h_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)         \
+  ((__m256i)__builtin_lasx_xvsrani_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvsrani_w_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)         \
+  ((__m256i)__builtin_lasx_xvsrani_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvsrani_d_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)         \
+  ((__m256i)__builtin_lasx_xvsrani_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvsrarni_b_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)        \
+  ((__m256i)__builtin_lasx_xvsrarni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvsrarni_h_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)        \
+  ((__m256i)__builtin_lasx_xvsrarni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvsrarni_w_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)        \
+  ((__m256i)__builtin_lasx_xvsrarni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvsrarni_d_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)        \
+  ((__m256i)__builtin_lasx_xvsrarni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvssrani_b_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)        \
+  ((__m256i)__builtin_lasx_xvssrani_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvssrani_h_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)        \
+  ((__m256i)__builtin_lasx_xvssrani_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvssrani_w_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)        \
+  ((__m256i)__builtin_lasx_xvssrani_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvssrani_d_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)        \
+  ((__m256i)__builtin_lasx_xvssrani_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvssrani_bu_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrani_bu_h((v32u8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvssrani_hu_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrani_hu_w((v16u16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvssrani_wu_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrani_wu_d((v8u32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvssrani_du_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrani_du_q((v4u64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvssrarni_b_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrarni_b_h((v32i8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvssrarni_h_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrarni_h_w((v16i16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvssrarni_w_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrarni_w_d((v8i32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvssrarni_d_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)       \
+  ((__m256i)__builtin_lasx_xvssrarni_d_q((v4i64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xvssrarni_bu_h(/*__m256i*/ _1, /*__m256i*/ _2, /*ui4*/ _3)      \
+  ((__m256i)__builtin_lasx_xvssrarni_bu_h((v32u8)(_1), (v32i8)(_2), (_3)))
+
+#define __lasx_xvssrarni_hu_w(/*__m256i*/ _1, /*__m256i*/ _2, /*ui5*/ _3)      \
+  ((__m256i)__builtin_lasx_xvssrarni_hu_w((v16u16)(_1), (v16i16)(_2), (_3)))
+
+#define __lasx_xvssrarni_wu_d(/*__m256i*/ _1, /*__m256i*/ _2, /*ui6*/ _3)      \
+  ((__m256i)__builtin_lasx_xvssrarni_wu_d((v8u32)(_1), (v8i32)(_2), (_3)))
+
+#define __lasx_xvssrarni_du_q(/*__m256i*/ _1, /*__m256i*/ _2, /*ui7*/ _3)      \
+  ((__m256i)__builtin_lasx_xvssrarni_du_q((v4u64)(_1), (v4i64)(_2), (_3)))
+
+#define __lasx_xbnz_b(/*__m256i*/ _1) ((int)__builtin_lasx_xbnz_b((v32u8)(_1)))
+
+#define __lasx_xbnz_d(/*__m256i*/ _1) ((int)__builtin_lasx_xbnz_d((v4u64)(_1)))
+
+#define __lasx_xbnz_h(/*__m256i*/ _1) ((int)__builtin_lasx_xbnz_h((v16u16)(_1)))
+
+#define __lasx_xbnz_v(/*__m256i*/ _1) ((int)__builtin_lasx_xbnz_v((v32u8)(_1)))
+
+#define __lasx_xbnz_w(/*__m256i*/ _1) ((int)__builtin_lasx_xbnz_w((v8u32)(_1)))
+
+#define __lasx_xbz_b(/*__m256i*/ _1) ((int)__builtin_lasx_xbz_b((v32u8)(_1)))
+
+#define __lasx_xbz_d(/*__m256i*/ _1) ((int)__builtin_lasx_xbz_d((v4u64)(_1)))
+
+#define __lasx_xbz_h(/*__m256i*/ _1) ((int)__builtin_lasx_xbz_h((v16u16)(_1)))
+
+#define __lasx_xbz_v(/*__m256i*/ _1) ((int)__builtin_lasx_xbz_v((v32u8)(_1)))
+
+#define __lasx_xbz_w(/*__m256i*/ _1) ((int)__builtin_lasx_xbz_w((v8u32)(_1)))
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_caf_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_caf_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_caf_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_caf_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_ceq_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_ceq_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_ceq_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_ceq_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cle_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cle_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cle_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cle_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_clt_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_clt_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_clt_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_clt_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cne_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cne_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cne_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cne_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cor_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cor_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cor_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cor_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cueq_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cueq_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cueq_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cueq_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cule_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cule_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cule_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cule_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cult_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cult_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cult_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cult_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cun_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cun_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cune_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cune_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cune_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cune_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_cun_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_cun_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_saf_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_saf_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_saf_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_saf_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_seq_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_seq_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_seq_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_seq_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sle_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sle_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sle_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sle_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_slt_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_slt_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_slt_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_slt_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sne_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sne_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sne_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sne_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sor_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sor_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sor_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sor_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sueq_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sueq_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sueq_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sueq_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sule_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sule_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sule_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sule_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sult_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sult_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sult_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sult_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sun_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sun_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sune_d(__m256d _1, __m256d _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sune_d((v4f64)_1, (v4f64)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sune_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sune_s((v8f32)_1, (v8f32)_2);
+}
+
+extern __inline
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__)) __m256i
+    __lasx_xvfcmp_sun_s(__m256 _1, __m256 _2) {
+  return (__m256i)__builtin_lasx_xvfcmp_sun_s((v8f32)_1, (v8f32)_2);
+}
+
+#define __lasx_xvpickve_d_f(/*__m256d*/ _1, /*ui2*/ _2)                        \
+  ((__m256d)__builtin_lasx_xvpickve_d_f((v4f64)(_1), (_2)))
+
+#define __lasx_xvpickve_w_f(/*__m256*/ _1, /*ui3*/ _2)                         \
+  ((__m256)__builtin_lasx_xvpickve_w_f((v8f32)(_1), (_2)))
+
+#define __lasx_xvrepli_b(/*si10*/ _1) ((__m256i)__builtin_lasx_xvrepli_b((_1)))
+
+#define __lasx_xvrepli_d(/*si10*/ _1) ((__m256i)__builtin_lasx_xvrepli_d((_1)))
+
+#define __lasx_xvrepli_h(/*si10*/ _1) ((__m256i)__builtin_lasx_xvrepli_h((_1)))
+
+#define __lasx_xvrepli_w(/*si10*/ _1) ((__m256i)__builtin_lasx_xvrepli_w((_1)))
+
+#endif /* defined(__loongarch_asx).  */
+#endif /* _LOONGSON_ASXINTRIN_H.  */
diff --git a/clang/lib/Sema/SemaChecking.cpp b/clang/lib/Sema/SemaChecking.cpp
index e131061eb88548b..44f698abdb9fe21 100644
--- a/clang/lib/Sema/SemaChecking.cpp
+++ b/clang/lib/Sema/SemaChecking.cpp
@@ -4184,6 +4184,233 @@ bool Sema::CheckLoongArchBuiltinFunctionCall(const TargetInfo &TI,
   case LoongArch::BI__builtin_lsx_vrepli_w:
   case LoongArch::BI__builtin_lsx_vrepli_d:
     return SemaBuiltinConstantArgRange(TheCall, 0, -512, 511);
+
+  // LASX intrinsics.
+  case LoongArch::BI__builtin_lasx_xvbitclri_b:
+  case LoongArch::BI__builtin_lasx_xvbitrevi_b:
+  case LoongArch::BI__builtin_lasx_xvbitseti_b:
+  case LoongArch::BI__builtin_lasx_xvsat_b:
+  case LoongArch::BI__builtin_lasx_xvsat_bu:
+  case LoongArch::BI__builtin_lasx_xvslli_b:
+  case LoongArch::BI__builtin_lasx_xvsrai_b:
+  case LoongArch::BI__builtin_lasx_xvsrari_b:
+  case LoongArch::BI__builtin_lasx_xvsrli_b:
+  case LoongArch::BI__builtin_lasx_xvsllwil_h_b:
+  case LoongArch::BI__builtin_lasx_xvsllwil_hu_bu:
+  case LoongArch::BI__builtin_lasx_xvrotri_b:
+  case LoongArch::BI__builtin_lasx_xvsrlri_b:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 7);
+  case LoongArch::BI__builtin_lasx_xvbitclri_h:
+  case LoongArch::BI__builtin_lasx_xvbitrevi_h:
+  case LoongArch::BI__builtin_lasx_xvbitseti_h:
+  case LoongArch::BI__builtin_lasx_xvsat_h:
+  case LoongArch::BI__builtin_lasx_xvsat_hu:
+  case LoongArch::BI__builtin_lasx_xvslli_h:
+  case LoongArch::BI__builtin_lasx_xvsrai_h:
+  case LoongArch::BI__builtin_lasx_xvsrari_h:
+  case LoongArch::BI__builtin_lasx_xvsrli_h:
+  case LoongArch::BI__builtin_lasx_xvsllwil_w_h:
+  case LoongArch::BI__builtin_lasx_xvsllwil_wu_hu:
+  case LoongArch::BI__builtin_lasx_xvrotri_h:
+  case LoongArch::BI__builtin_lasx_xvsrlri_h:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 15);
+  case LoongArch::BI__builtin_lasx_xvssrarni_b_h:
+  case LoongArch::BI__builtin_lasx_xvssrarni_bu_h:
+  case LoongArch::BI__builtin_lasx_xvssrani_b_h:
+  case LoongArch::BI__builtin_lasx_xvssrani_bu_h:
+  case LoongArch::BI__builtin_lasx_xvsrarni_b_h:
+  case LoongArch::BI__builtin_lasx_xvsrlni_b_h:
+  case LoongArch::BI__builtin_lasx_xvsrlrni_b_h:
+  case LoongArch::BI__builtin_lasx_xvssrlni_b_h:
+  case LoongArch::BI__builtin_lasx_xvssrlni_bu_h:
+  case LoongArch::BI__builtin_lasx_xvssrlrni_b_h:
+  case LoongArch::BI__builtin_lasx_xvssrlrni_bu_h:
+  case LoongArch::BI__builtin_lasx_xvsrani_b_h:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 15);
+  case LoongArch::BI__builtin_lasx_xvslei_bu:
+  case LoongArch::BI__builtin_lasx_xvslei_hu:
+  case LoongArch::BI__builtin_lasx_xvslei_wu:
+  case LoongArch::BI__builtin_lasx_xvslei_du:
+  case LoongArch::BI__builtin_lasx_xvslti_bu:
+  case LoongArch::BI__builtin_lasx_xvslti_hu:
+  case LoongArch::BI__builtin_lasx_xvslti_wu:
+  case LoongArch::BI__builtin_lasx_xvslti_du:
+  case LoongArch::BI__builtin_lasx_xvmaxi_bu:
+  case LoongArch::BI__builtin_lasx_xvmaxi_hu:
+  case LoongArch::BI__builtin_lasx_xvmaxi_wu:
+  case LoongArch::BI__builtin_lasx_xvmaxi_du:
+  case LoongArch::BI__builtin_lasx_xvmini_bu:
+  case LoongArch::BI__builtin_lasx_xvmini_hu:
+  case LoongArch::BI__builtin_lasx_xvmini_wu:
+  case LoongArch::BI__builtin_lasx_xvmini_du:
+  case LoongArch::BI__builtin_lasx_xvaddi_bu:
+  case LoongArch::BI__builtin_lasx_xvaddi_hu:
+  case LoongArch::BI__builtin_lasx_xvaddi_wu:
+  case LoongArch::BI__builtin_lasx_xvaddi_du:
+  case LoongArch::BI__builtin_lasx_xvbitclri_w:
+  case LoongArch::BI__builtin_lasx_xvbitrevi_w:
+  case LoongArch::BI__builtin_lasx_xvbitseti_w:
+  case LoongArch::BI__builtin_lasx_xvsat_w:
+  case LoongArch::BI__builtin_lasx_xvsat_wu:
+  case LoongArch::BI__builtin_lasx_xvslli_w:
+  case LoongArch::BI__builtin_lasx_xvsrai_w:
+  case LoongArch::BI__builtin_lasx_xvsrari_w:
+  case LoongArch::BI__builtin_lasx_xvsrli_w:
+  case LoongArch::BI__builtin_lasx_xvsllwil_d_w:
+  case LoongArch::BI__builtin_lasx_xvsllwil_du_wu:
+  case LoongArch::BI__builtin_lasx_xvsrlri_w:
+  case LoongArch::BI__builtin_lasx_xvrotri_w:
+  case LoongArch::BI__builtin_lasx_xvsubi_bu:
+  case LoongArch::BI__builtin_lasx_xvsubi_hu:
+  case LoongArch::BI__builtin_lasx_xvsubi_wu:
+  case LoongArch::BI__builtin_lasx_xvsubi_du:
+  case LoongArch::BI__builtin_lasx_xvbsrl_v:
+  case LoongArch::BI__builtin_lasx_xvbsll_v:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 31);
+  case LoongArch::BI__builtin_lasx_xvssrarni_h_w:
+  case LoongArch::BI__builtin_lasx_xvssrarni_hu_w:
+  case LoongArch::BI__builtin_lasx_xvssrani_h_w:
+  case LoongArch::BI__builtin_lasx_xvssrani_hu_w:
+  case LoongArch::BI__builtin_lasx_xvsrarni_h_w:
+  case LoongArch::BI__builtin_lasx_xvsrani_h_w:
+  case LoongArch::BI__builtin_lasx_xvfrstpi_b:
+  case LoongArch::BI__builtin_lasx_xvfrstpi_h:
+  case LoongArch::BI__builtin_lasx_xvsrlni_h_w:
+  case LoongArch::BI__builtin_lasx_xvsrlrni_h_w:
+  case LoongArch::BI__builtin_lasx_xvssrlni_h_w:
+  case LoongArch::BI__builtin_lasx_xvssrlni_hu_w:
+  case LoongArch::BI__builtin_lasx_xvssrlrni_h_w:
+  case LoongArch::BI__builtin_lasx_xvssrlrni_hu_w:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 31);
+  case LoongArch::BI__builtin_lasx_xvbitclri_d:
+  case LoongArch::BI__builtin_lasx_xvbitrevi_d:
+  case LoongArch::BI__builtin_lasx_xvbitseti_d:
+  case LoongArch::BI__builtin_lasx_xvsat_d:
+  case LoongArch::BI__builtin_lasx_xvsat_du:
+  case LoongArch::BI__builtin_lasx_xvslli_d:
+  case LoongArch::BI__builtin_lasx_xvsrai_d:
+  case LoongArch::BI__builtin_lasx_xvsrli_d:
+  case LoongArch::BI__builtin_lasx_xvsrari_d:
+  case LoongArch::BI__builtin_lasx_xvrotri_d:
+  case LoongArch::BI__builtin_lasx_xvsrlri_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 63);
+  case LoongArch::BI__builtin_lasx_xvssrarni_w_d:
+  case LoongArch::BI__builtin_lasx_xvssrarni_wu_d:
+  case LoongArch::BI__builtin_lasx_xvssrani_w_d:
+  case LoongArch::BI__builtin_lasx_xvssrani_wu_d:
+  case LoongArch::BI__builtin_lasx_xvsrarni_w_d:
+  case LoongArch::BI__builtin_lasx_xvsrlni_w_d:
+  case LoongArch::BI__builtin_lasx_xvsrlrni_w_d:
+  case LoongArch::BI__builtin_lasx_xvssrlni_w_d:
+  case LoongArch::BI__builtin_lasx_xvssrlni_wu_d:
+  case LoongArch::BI__builtin_lasx_xvssrlrni_w_d:
+  case LoongArch::BI__builtin_lasx_xvssrlrni_wu_d:
+  case LoongArch::BI__builtin_lasx_xvsrani_w_d:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 63);
+  case LoongArch::BI__builtin_lasx_xvssrarni_d_q:
+  case LoongArch::BI__builtin_lasx_xvssrarni_du_q:
+  case LoongArch::BI__builtin_lasx_xvssrani_d_q:
+  case LoongArch::BI__builtin_lasx_xvssrani_du_q:
+  case LoongArch::BI__builtin_lasx_xvsrarni_d_q:
+  case LoongArch::BI__builtin_lasx_xvssrlni_d_q:
+  case LoongArch::BI__builtin_lasx_xvssrlni_du_q:
+  case LoongArch::BI__builtin_lasx_xvssrlrni_d_q:
+  case LoongArch::BI__builtin_lasx_xvssrlrni_du_q:
+  case LoongArch::BI__builtin_lasx_xvsrani_d_q:
+  case LoongArch::BI__builtin_lasx_xvsrlni_d_q:
+  case LoongArch::BI__builtin_lasx_xvsrlrni_d_q:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 127);
+  case LoongArch::BI__builtin_lasx_xvseqi_b:
+  case LoongArch::BI__builtin_lasx_xvseqi_h:
+  case LoongArch::BI__builtin_lasx_xvseqi_w:
+  case LoongArch::BI__builtin_lasx_xvseqi_d:
+  case LoongArch::BI__builtin_lasx_xvslti_b:
+  case LoongArch::BI__builtin_lasx_xvslti_h:
+  case LoongArch::BI__builtin_lasx_xvslti_w:
+  case LoongArch::BI__builtin_lasx_xvslti_d:
+  case LoongArch::BI__builtin_lasx_xvslei_b:
+  case LoongArch::BI__builtin_lasx_xvslei_h:
+  case LoongArch::BI__builtin_lasx_xvslei_w:
+  case LoongArch::BI__builtin_lasx_xvslei_d:
+  case LoongArch::BI__builtin_lasx_xvmaxi_b:
+  case LoongArch::BI__builtin_lasx_xvmaxi_h:
+  case LoongArch::BI__builtin_lasx_xvmaxi_w:
+  case LoongArch::BI__builtin_lasx_xvmaxi_d:
+  case LoongArch::BI__builtin_lasx_xvmini_b:
+  case LoongArch::BI__builtin_lasx_xvmini_h:
+  case LoongArch::BI__builtin_lasx_xvmini_w:
+  case LoongArch::BI__builtin_lasx_xvmini_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -16, 15);
+  case LoongArch::BI__builtin_lasx_xvandi_b:
+  case LoongArch::BI__builtin_lasx_xvnori_b:
+  case LoongArch::BI__builtin_lasx_xvori_b:
+  case LoongArch::BI__builtin_lasx_xvshuf4i_b:
+  case LoongArch::BI__builtin_lasx_xvshuf4i_h:
+  case LoongArch::BI__builtin_lasx_xvshuf4i_w:
+  case LoongArch::BI__builtin_lasx_xvxori_b:
+  case LoongArch::BI__builtin_lasx_xvpermi_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 255);
+  case LoongArch::BI__builtin_lasx_xvbitseli_b:
+  case LoongArch::BI__builtin_lasx_xvshuf4i_d:
+  case LoongArch::BI__builtin_lasx_xvextrins_b:
+  case LoongArch::BI__builtin_lasx_xvextrins_h:
+  case LoongArch::BI__builtin_lasx_xvextrins_w:
+  case LoongArch::BI__builtin_lasx_xvextrins_d:
+  case LoongArch::BI__builtin_lasx_xvpermi_q:
+  case LoongArch::BI__builtin_lasx_xvpermi_w:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 255);
+  case LoongArch::BI__builtin_lasx_xvrepl128vei_b:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 15);
+  case LoongArch::BI__builtin_lasx_xvrepl128vei_h:
+  case LoongArch::BI__builtin_lasx_xvpickve2gr_w:
+  case LoongArch::BI__builtin_lasx_xvpickve2gr_wu:
+  case LoongArch::BI__builtin_lasx_xvpickve_w_f:
+  case LoongArch::BI__builtin_lasx_xvpickve_w:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 7);
+  case LoongArch::BI__builtin_lasx_xvinsgr2vr_w:
+  case LoongArch::BI__builtin_lasx_xvinsve0_w:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 7);
+  case LoongArch::BI__builtin_lasx_xvrepl128vei_w:
+  case LoongArch::BI__builtin_lasx_xvpickve2gr_d:
+  case LoongArch::BI__builtin_lasx_xvpickve2gr_du:
+  case LoongArch::BI__builtin_lasx_xvpickve_d_f:
+  case LoongArch::BI__builtin_lasx_xvpickve_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 3);
+  case LoongArch::BI__builtin_lasx_xvinsve0_d:
+  case LoongArch::BI__builtin_lasx_xvinsgr2vr_d:
+    return SemaBuiltinConstantArgRange(TheCall, 2, 0, 3);
+  case LoongArch::BI__builtin_lasx_xvstelm_b:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -128, 127) ||
+           SemaBuiltinConstantArgRange(TheCall, 3, 0, 31);
+  case LoongArch::BI__builtin_lasx_xvstelm_h:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -256, 254) ||
+           SemaBuiltinConstantArgRange(TheCall, 3, 0, 15);
+  case LoongArch::BI__builtin_lasx_xvstelm_w:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -512, 508) ||
+           SemaBuiltinConstantArgRange(TheCall, 3, 0, 7);
+  case LoongArch::BI__builtin_lasx_xvstelm_d:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -1024, 1016) ||
+           SemaBuiltinConstantArgRange(TheCall, 3, 0, 3);
+  case LoongArch::BI__builtin_lasx_xvrepl128vei_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, 0, 1);
+  case LoongArch::BI__builtin_lasx_xvldrepl_b:
+  case LoongArch::BI__builtin_lasx_xvld:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -2048, 2047);
+  case LoongArch::BI__builtin_lasx_xvldrepl_h:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -2048, 2046);
+  case LoongArch::BI__builtin_lasx_xvldrepl_w:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -2048, 2044);
+  case LoongArch::BI__builtin_lasx_xvldrepl_d:
+    return SemaBuiltinConstantArgRange(TheCall, 1, -2048, 2040);
+  case LoongArch::BI__builtin_lasx_xvst:
+    return SemaBuiltinConstantArgRange(TheCall, 2, -2048, 2047);
+  case LoongArch::BI__builtin_lasx_xvldi:
+    return SemaBuiltinConstantArgRange(TheCall, 0, -4096, 4095);
+  case LoongArch::BI__builtin_lasx_xvrepli_b:
+  case LoongArch::BI__builtin_lasx_xvrepli_h:
+  case LoongArch::BI__builtin_lasx_xvrepli_w:
+  case LoongArch::BI__builtin_lasx_xvrepli_d:
+    return SemaBuiltinConstantArgRange(TheCall, 0, -512, 511);
   }
   return false;
 }



More information about the cfe-commits mailing list