diff --git a/clang-tools-extra/clang-include-fixer/find-all-symbols/STLPostfixHeaderMap.cpp b/clang-tools-extra/clang-include-fixer/find-all-symbols/STLPostfixHeaderMap.cpp
--- a/clang-tools-extra/clang-include-fixer/find-all-symbols/STLPostfixHeaderMap.cpp
+++ b/clang-tools-extra/clang-include-fixer/find-all-symbols/STLPostfixHeaderMap.cpp
@@ -57,6 +57,16 @@
       {"include/prfchwintrin.h$", "<x86intrin.h>"},
       {"include/rdseedintrin.h$", "<x86intrin.h>"},
       {"include/rtmintrin.h$", "<immintrin.h>"},
+      {"include/rv32bintrin-builtins.h$", "<rv32bintrin-builtins.h>"},
+      {"include/rv32bintrin-emulation.h$", "<rv32bintrin-emulation.h>"},
+      {"include/rv32bintrin.h$", "<rv32bintrin.h>"},
+      {"include/rv64bintrin-asm.h$", "<rv64bintrin-asm.h>"},
+      {"include/rv64bintrin-builtins.h$", "<rv64bintrin-builtins.h>"},
+      {"include/rv64bintrin-emulation.h$", "<rv64bintrin-emulation.h>"},
+      {"include/rv64bintrin.h$", "<rv64bintrin.h>"},
+      {"include/rvbintrin-asm.h$", "<rvbintrin-asm.h>"},
+      {"include/rvbintrin-emulation.h$", "<rvbintrin-emulation.h>"},
+      {"include/rvintrin.h$", "<rvintrin.h>"},
       {"include/shaintrin.h$", "<immintrin.h>"},
       {"include/smmintrin.h$", "<smmintrin.h>"},
       {"include/stdalign.h$", "<cstdalign>"},
diff --git a/clang-tools-extra/clangd/index/CanonicalIncludes.cpp b/clang-tools-extra/clangd/index/CanonicalIncludes.cpp
--- a/clang-tools-extra/clangd/index/CanonicalIncludes.cpp
+++ b/clang-tools-extra/clangd/index/CanonicalIncludes.cpp
@@ -150,6 +150,16 @@
       {"include/prfchwintrin.h", "<x86intrin.h>"},
       {"include/rdseedintrin.h", "<x86intrin.h>"},
       {"include/rtmintrin.h", "<immintrin.h>"},
+      {"include/rv32bintrin-builtins.h", "<rv32bintrin-builtins.h>"},
+      {"include/rv32bintrin-emulation.h", "<rv32bintrin-emulation.h>"},
+      {"include/rv32bintrin.h", "<rv32bintrin.h>"},
+      {"include/rv64bintrin-asm.h", "<rv64bintrin-asm.h>"},
+      {"include/rv64bintrin-builtins.h", "<rv64bintrin-builtins.h>"},
+      {"include/rv64bintrin-emulation.h", "<rv64bintrin-emulation.h>"},
+      {"include/rv64bintrin.h", "<rv64bintrin.h>"},
+      {"include/rvbintrin-asm.h", "<rvbintrin-asm.h>"},
+      {"include/rvbintrin-emulation.h", "<rvbintrin-emulation.h>"},
+      {"include/rvintrin.h", "<rvintrin.h>"},
       {"include/shaintrin.h", "<immintrin.h>"},
       {"include/smmintrin.h", "<smmintrin.h>"},
       {"include/stdalign.h", "<cstdalign>"},
diff --git a/clang/lib/Headers/CMakeLists.txt b/clang/lib/Headers/CMakeLists.txt
--- a/clang/lib/Headers/CMakeLists.txt
+++ b/clang/lib/Headers/CMakeLists.txt
@@ -86,6 +86,16 @@
   ptwriteintrin.h
   rdseedintrin.h
   rtmintrin.h
+  rv32bintrin-builtins.h
+  rv32bintrin-emulation.h
+  rv32bintrin.h
+  rv64bintrin-asm.h
+  rv64bintrin-builtins.h
+  rv64bintrin-emulation.h
+  rv64bintrin.h
+  rvbintrin-asm.h
+  rvbintrin-emulation.h
+  rvintrin.h
   sgxintrin.h
   s390intrin.h
   shaintrin.h
diff --git a/clang/lib/Headers/intrin.h b/clang/lib/Headers/intrin.h
--- a/clang/lib/Headers/intrin.h
+++ b/clang/lib/Headers/intrin.h
@@ -28,6 +28,10 @@
 #include <arm64intr.h>
 #endif
 
+#if defined(__riscv)
+#include <rvintrin.h>
+#endif
+
 /* For the definition of jmp_buf. */
 #if __STDC_HOSTED__
 #include <setjmp.h>
diff --git a/clang/lib/Headers/rv32bintrin-builtins.h b/clang/lib/Headers/rv32bintrin-builtins.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rv32bintrin-builtins.h
@@ -0,0 +1,65 @@
+/* ===-------- rv32bintrin-builtins.h --------------------------------------===
+ *
+ * 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 __RV32BINTRIN_BUILTINS_H
+#define __RV32BINTRIN_BUILTINS_H
+
+#ifndef __RVINTRIN_H
+#error                                                                         \
+    "Never use <rv32bintrin-builtins.h> directly; include <rvintrin.h> instead."
+#endif
+
+// These are available regardless of whether or not the Bitmanip extension is
+// enabled or not.
+
+// Our approach to these functions is to use inline asm when the Bitmanip
+// extension is enabled and use replacment C code when it's disabled. However,
+// where LLVM builtins are already avaiable we are making use of them.
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_clz(const uint_xlen_t rs1) {
+  // Calling these builtins with 0 results in undefined behaviour.
+  if (rs1 == 0) {
+    return XLEN;
+  }
+  return __builtin_clz(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_ctz(const uint_xlen_t rs1) {
+  if (rs1 == 0) {
+    return XLEN;
+  }
+  return __builtin_ctz(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_pcnt(const uint_xlen_t rs1) {
+  return __builtin_popcount(rs1);
+}
+
+// Genric aliases
+// e.g. _rv_* is an alias of _rv64_*
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_clz(const uint_xlen_t rs1) {
+  return _rv32_clz(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_ctz(const uint_xlen_t rs1) {
+  return _rv32_ctz(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_pcnt(const uint_xlen_t rs1) {
+  return _rv32_pcnt(rs1);
+}
+
+#endif //__RV32BINTRIN_BUILTINS_H
diff --git a/clang/lib/Headers/rv32bintrin-emulation.h b/clang/lib/Headers/rv32bintrin-emulation.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rv32bintrin-emulation.h
@@ -0,0 +1,106 @@
+/* ===-------- rv32bintrin-emulation.h -------------------------------------===
+ *
+ * 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 __RV32BINTRIN_EMULATION_H
+#define __RV32BINTRIN_EMULATION_H
+
+#ifndef __RVINTRIN_H
+#error                                                                         \
+    "Never use <rv32bintrin-emulation.h> directly; include <rvintrin.h> instead."
+#endif
+
+#include <stdint.h>
+
+#include "rvbintrin-emulation.h"
+
+// If the builtins are unavailable then these provide the same functionality.
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS _rv_gorc(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint32_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 31;
+  if (shamt & 1) {
+    x |= ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
+  }
+  if (shamt & 2) {
+    x |= ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
+  }
+  if (shamt & 4) {
+    x |= ((x & 0x0F0F0F0F) << 4) | ((x & 0xF0F0F0F0) >> 4);
+  }
+  if (shamt & 8) {
+    x |= ((x & 0x00FF00FF) << 8) | ((x & 0xFF00FF00) >> 8);
+  }
+  if (shamt & 16) {
+    x |= ((x & 0x0000FFFF) << 16) | ((x & 0xFFFF0000) >> 16);
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS _rv_grev(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint32_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 31;
+  if (shamt & 1) {
+    x = ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
+  }
+  if (shamt & 2) {
+    x = ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
+  }
+  if (shamt & 4) {
+    x = ((x & 0x0F0F0F0F) << 4) | ((x & 0xF0F0F0F0) >> 4);
+  }
+  if (shamt & 8) {
+    x = ((x & 0x00FF00FF) << 8) | ((x & 0xFF00FF00) >> 8);
+  }
+  if (shamt & 16) {
+    x = ((x & 0x0000FFFF) << 16) | ((x & 0xFFFF0000) >> 16);
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS _rv_shfl(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint32_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 15;
+  if (shamt & 8) {
+    x = shuffle_stage(x, 0x00ff0000, 0x0000ff00, 8);
+  }
+  if (shamt & 4) {
+    x = shuffle_stage(x, 0x0f000f00, 0x00f000f0, 4);
+  }
+  if (shamt & 2) {
+    x = shuffle_stage(x, 0x30303030, 0x0c0c0c0c, 2);
+  }
+  if (shamt & 1) {
+    x = shuffle_stage(x, 0x44444444, 0x22222222, 1);
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unshfl(const uint32_t rs1, const uint32_t rs2) {
+  uint32_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 15;
+  if (shamt & 1) {
+    x = shuffle_stage(x, 0x44444444, 0x22222222, 1);
+  }
+  if (shamt & 2) {
+    x = shuffle_stage(x, 0x30303030, 0x0c0c0c0c, 2);
+  }
+  if (shamt & 4) {
+    x = shuffle_stage(x, 0x0f000f00, 0x00f000f0, 4);
+  }
+  if (shamt & 8) {
+    x = shuffle_stage(x, 0x00ff0000, 0x0000ff00, 8);
+  }
+  return x;
+}
+
+#endif //__RV32BINTRIN_EMULATION_H
diff --git a/clang/lib/Headers/rv32bintrin.h b/clang/lib/Headers/rv32bintrin.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rv32bintrin.h
@@ -0,0 +1,748 @@
+/* ===-------- rv32bintrin.h -----------------------------------------------===
+ *
+ * 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 __RV32BINTRIN_H
+#define __RV32BINTRIN_H
+
+#ifndef __RVINTRIN_H
+#error "Never use <rv32bintrin.h> directly; include <rvintrin.h> instead."
+#endif
+
+#include "rv32bintrin-builtins.h"
+
+#if defined(__riscv_bitmanip)
+
+#include "rvbintrin-asm.h"
+
+#else // Bitmanip extension is disabled.
+
+#include "rv32bintrin-emulation.h"
+
+#endif // defined(__riscv_bitmanip)
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_andn(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_andn(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_fsl(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return _rv_fsl(rs1, rs2, rs3);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_fsr(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return _rv_fsr(rs1, rs2, rs3);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_max(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_max(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_maxu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_maxu(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_min(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_min(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_minu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_minu(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orn(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_orn(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rol(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_rol(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_ror(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_ror(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_bdep(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_bdep(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_bext(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_bext(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_bfp(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_bfp(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_clmul(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_clmul(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_clmulh(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_clmulh(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_clmulr(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_clmulr(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS _rv32_cmix(
+    const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return _rv_cmix(rs1, rs2, rs3);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS _rv32_cmov(
+    const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return _rv_cmov(rs1, rs2, rs3);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_crc32_b(const uint_xlen_t rs1) {
+  return _rv_crc32_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_crc32_h(const uint_xlen_t rs1) {
+  return _rv_crc32_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_crc32_w(const uint_xlen_t rs1) {
+  return _rv_crc32_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_crc32c_b(const uint_xlen_t rs1) {
+  return _rv_crc32c_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_crc32c_h(const uint_xlen_t rs1) {
+  return _rv_crc32c_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_crc32c_w(const uint_xlen_t rs1) {
+  return _rv_crc32c_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_gorc(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_gorc(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_grev(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_grev(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_pack(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_pack(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_packu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_packu(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_packh(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_packh(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_sext_b(const uint_xlen_t rs1) {
+  return _rv_sext_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_sext_h(const uint_xlen_t rs1) {
+  return _rv_sext_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_sbclr(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sbclr(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_sbext(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sbext(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_sbinv(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sbinv(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_sbset(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sbset(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_shfl(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_shfl(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_slo(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_slo(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_sro(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sro(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unshfl(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_unshfl(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_xnor(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_xnor(rs1, rs2);
+}
+
+// Pseudo Instructions
+
+// REV (GREVI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev_p(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b00001);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev2_n(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b00010);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev_n(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b00011);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev4_b(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b00100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev2_b(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b00110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev_b(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b00111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev8_h(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b01000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev4_h(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b01100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev2_h(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b01110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev_h(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b01111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev16(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b10000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev8(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b11000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev4(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b11100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev2(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b11110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_rev(const uint_xlen_t rs1) {
+  return _rv32_grev(rs1, 0b11111);
+}
+
+// ORC (GORCI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc_p(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b00001);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc2_n(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b00010);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc_n(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b00011);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc4_b(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b00100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc2_b(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b00110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc_b(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b00111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc8_h(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b01000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc4_h(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b01100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc2_h(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b01110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc_h(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b01111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc16(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b10000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc8(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b11000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc4(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b11100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc2(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b11110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_orc(const uint_xlen_t rs1) {
+  return _rv32_gorc(rs1, 0b11111);
+}
+
+// ZIP (SHFLI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip_n(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b0001);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip2_b(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b0010);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip_b(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b0011);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip4_h(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b0100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip2_h(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b0110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip_h(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b0111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip8(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b1000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip4(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b1100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip2(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b1110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_zip(const uint_xlen_t rs1) {
+  return _rv32_shfl(rs1, 0b1111);
+}
+
+// UNZIP (UNSHFL) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip_n(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b0001);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip2_b(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b0010);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip_b(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b0011);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip4_h(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b0100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip2_h(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b0110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip_h(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b0111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip8(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b1000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip4(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b1100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip2(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b1110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv32_unzip(const uint_xlen_t rs1) {
+  return _rv32_unshfl(rs1, 0b1111);
+}
+
+// Genric aliases for 32 bit pseudo instructions
+// e.g. _rv_* is an alias of _rv32_*
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_p(const uint_xlen_t rs1) {
+  return _rv32_rev_p(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2_n(const uint_xlen_t rs1) {
+  return _rv32_rev2_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_n(const uint_xlen_t rs1) {
+  return _rv32_rev_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev4_b(const uint_xlen_t rs1) {
+  return _rv32_rev4_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2_b(const uint_xlen_t rs1) {
+  return _rv32_rev2_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_b(const uint_xlen_t rs1) {
+  return _rv32_rev_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev8_h(const uint_xlen_t rs1) {
+  return _rv32_rev8_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev4_h(const uint_xlen_t rs1) {
+  return _rv32_rev4_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2_h(const uint_xlen_t rs1) {
+  return _rv32_rev2_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_h(const uint_xlen_t rs1) {
+  return _rv32_rev_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev16(const uint_xlen_t rs1) {
+  return _rv32_rev16(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev8(const uint_xlen_t rs1) {
+  return _rv32_rev8(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev4(const uint_xlen_t rs1) {
+  return _rv32_rev4(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2(const uint_xlen_t rs1) {
+  return _rv32_rev2(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev(const uint_xlen_t rs1) {
+  return _rv32_rev(rs1);
+}
+
+// ORC (GORCI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_p(const uint_xlen_t rs1) {
+  return _rv32_orc_p(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2_n(const uint_xlen_t rs1) {
+  return _rv32_orc2_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_n(const uint_xlen_t rs1) {
+  return _rv32_orc_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc4_b(const uint_xlen_t rs1) {
+  return _rv32_orc4_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2_b(const uint_xlen_t rs1) {
+  return _rv32_orc2_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_b(const uint_xlen_t rs1) {
+  return _rv32_orc_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc8_h(const uint_xlen_t rs1) {
+  return _rv32_orc8_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc4_h(const uint_xlen_t rs1) {
+  return _rv32_orc4_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2_h(const uint_xlen_t rs1) {
+  return _rv32_orc2_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_h(const uint_xlen_t rs1) {
+  return _rv32_orc_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc16(const uint_xlen_t rs1) {
+  return _rv32_orc16(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc8(const uint_xlen_t rs1) {
+  return _rv32_orc8(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc4(const uint_xlen_t rs1) {
+  return _rv32_orc4(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2(const uint_xlen_t rs1) {
+  return _rv32_orc2(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc(const uint_xlen_t rs1) {
+  return _rv32_orc(rs1);
+}
+
+// ZIP (SHFLI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip_n(const uint_xlen_t rs1) {
+  return _rv32_zip_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip2_b(const uint_xlen_t rs1) {
+  return _rv32_zip2_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip_b(const uint_xlen_t rs1) {
+  return _rv32_zip_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip4_h(const uint_xlen_t rs1) {
+  return _rv32_zip4_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip2_h(const uint_xlen_t rs1) {
+  return _rv32_zip2_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip_h(const uint_xlen_t rs1) {
+  return _rv32_zip_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip8(const uint_xlen_t rs1) {
+  return _rv32_zip8(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip4(const uint_xlen_t rs1) {
+  return _rv32_zip4(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip2(const uint_xlen_t rs1) {
+  return _rv32_zip2(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip(const uint_xlen_t rs1) {
+  return _rv32_zip(rs1);
+}
+
+// UNZIP (UNSHFL) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip_n(const uint_xlen_t rs1) {
+  return _rv32_unzip_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip2_b(const uint_xlen_t rs1) {
+  return _rv32_unzip2_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip_b(const uint_xlen_t rs1) {
+  return _rv32_unzip_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip4_h(const uint_xlen_t rs1) {
+  return _rv32_unzip4_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip2_h(const uint_xlen_t rs1) {
+  return _rv32_unzip2_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip_h(const uint_xlen_t rs1) {
+  return _rv32_unzip_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip8(const uint_xlen_t rs1) {
+  return _rv32_unzip8(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip4(const uint_xlen_t rs1) {
+  return _rv32_unzip4(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip2(const uint_xlen_t rs1) {
+  return _rv32_unzip2(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip(const uint_xlen_t rs1) {
+  return _rv32_unzip(rs1);
+}
+
+#endif // __RV32BINTRIN_H
diff --git a/clang/lib/Headers/rv64bintrin-asm.h b/clang/lib/Headers/rv64bintrin-asm.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rv64bintrin-asm.h
@@ -0,0 +1,277 @@
+/* ===-------- rv64bintrin-asm.h -------------------------------------------===
+ *
+ * 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 __RV64BINTRIN_ASM_H
+#define __RV64BINTRIN_ASM_H
+
+#ifndef __RVINTRIN_H
+#error "Never use <rv64bintrin-asm.h> directly; include <rvintrin.h> instead."
+#endif
+
+#include <stdint.h>
+
+#include "rvbintrin-asm.h"
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_fslw(const uint32_t rs1,
+                                                       const uint32_t rs2,
+                                                       const uint32_t rs3) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("fsriw  %0, %1, %2, %3" : "=r"(rd) : "r"(rs1), "r"(rs2), "i"(rs3));
+  } else {
+    __asm__("fslw  %0, %1, %2, %3" : "=r"(rd) : "r"(rs1), "r"(rs2), "r"(rs3));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_fsrw(const uint32_t rs1,
+                                                       const uint32_t rs2,
+                                                       const uint32_t rs3) {
+  uint_xlen_t rd;
+  __asm__("fsrw  %0, %1, %2, %3" : "=r"(rd) : "r"(rs1), "r"(rs2), "r"(rs3));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_rolw(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("rolw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_rorw(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("roriw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("rorw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_addwu(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("addiwu  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("addwu  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_addu_w(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("addu.w  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bdepw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("bdepw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bfpw(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("bfpw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bextw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("bextw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bmatflip(const uint32_t rs1) {
+  uint_xlen_t rd;
+  __asm__("bmatflip  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bmator(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("bmator  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bmatxor(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("bmatxor  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_clmulhw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("clmulhw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_clmulrw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("clmulrw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_clmulw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("clmulw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_crc32_d(const uint32_t rs1) {
+  uint_xlen_t rd;
+  __asm__("crc32.d  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_crc32c_d(const uint32_t rs1) {
+  uint_xlen_t rd;
+  __asm__("crc32c.d  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_gorcw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("gorciw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("gorcw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_grevw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("greviw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("grevw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_packw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("packw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_packuw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("packuw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_sbclrw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sbclriw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("sbclrw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_sbextw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("sbextw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_sbinvw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sbinviw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("sbinvw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_sbsetw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sbsetiw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("sbsetw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_shflw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("shflw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_slow(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sloiw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("slow  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_srow(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sroiw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("srow  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_subu_w(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("subu.w  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_subwu(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("subwu  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_unshflw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint_xlen_t rd;
+  __asm__("unshflw  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+#endif //__RV64BINTRIN_ASM_H
diff --git a/clang/lib/Headers/rv64bintrin-builtins.h b/clang/lib/Headers/rv64bintrin-builtins.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rv64bintrin-builtins.h
@@ -0,0 +1,96 @@
+/* ===-------- rv64bintrin-builtins.h --------------------------------------===
+ *
+ * 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 __RV64BINTRIN_BUILTINS_H
+#define __RV64BINTRIN_BUILTINS_H
+
+#ifndef __RVINTRIN_H
+#error                                                                         \
+    "Never use <rv64bintrin-builtins.h> directly; include <rvintrin.h> instead."
+#endif
+
+#include <stdint.h>
+
+// These are available regardless of whether or not the Bitmanip extension is
+// enabled or not.
+
+// Our approach to these functions is to use inline asm when the Bitmanip
+// extension is enabled and use replacment C code when it's disabled. However,
+// where LLVM builtins are already avaiable we are making use of them.
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_clz(const uint_xlen_t rs1) {
+  if (rs1 == 0) {
+    return XLEN;
+  }
+  return __builtin_clzll(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_ctz(const uint_xlen_t rs1) {
+  if (rs1 == 0) {
+    return XLEN;
+  }
+  return __builtin_ctzll(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_pcnt(const uint_xlen_t rs1) {
+  return __builtin_popcountll(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_clzw(const uint32_t rs1) {
+  if (rs1 == 0) {
+    return XLEN;
+  }
+  return __builtin_clz(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_ctzw(const uint32_t rs1) {
+  if (rs1 == 0) {
+    return XLEN;
+  }
+  return __builtin_ctz(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_pcntw(const uint32_t rs1) {
+  return __builtin_popcount(rs1);
+}
+
+// Genric aliases
+// e.g. _rv_* is an alias of _rv64_*
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_clz(const uint_xlen_t rs1) {
+  return _rv64_clz(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_ctz(const uint_xlen_t rs1) {
+  return _rv64_ctz(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_pcnt(const uint_xlen_t rs1) {
+  return _rv64_pcnt(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_clzw(const uint32_t rs1) {
+  return _rv64_clzw(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_ctzw(const uint32_t rs1) {
+  return _rv64_ctzw(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_pcntw(const uint32_t rs1) {
+  return _rv64_pcntw(rs1);
+}
+
+#endif //__RV64BINTRIN_BUILTINS_H
diff --git a/clang/lib/Headers/rv64bintrin-emulation.h b/clang/lib/Headers/rv64bintrin-emulation.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rv64bintrin-emulation.h
@@ -0,0 +1,433 @@
+/* ===-------- rv64bintrin-emulation.h -------------------------------------===
+ *
+ * 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 __RV64INTRIN_EMULATION_H
+#define __RV64INTRIN_EMULATION_H
+
+#ifndef __RVINTRIN_H
+#error                                                                         \
+    "Never use <rv64bintrin-emulation.h> directly; include <rvintrin.h> instead."
+#endif
+
+#include <stdint.h>
+
+#include "rvbintrin-emulation.h"
+
+static __inline__ uint32_t shuffle32_stage(const uint32_t src,
+                                           const uint32_t maskL,
+                                           const uint32_t maskR,
+                                           const uint_xlen_t N) {
+  uint32_t x = src & ~(maskL | maskR);
+  x |= ((src << N) & maskL) | ((src >> N) & maskR);
+  return x;
+}
+
+static __inline__ uint64_t __DEFAULT_FN_ATTRS _rv_shfl(const uint64_t rs1,
+                                                       const uint64_t rs2) {
+  uint64_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 31;
+  if (shamt & 16) {
+    x = shuffle_stage(x, 0x0000ffff00000000LL, 0x00000000ffff0000LL, 16);
+  }
+  if (shamt & 8) {
+    x = shuffle_stage(x, 0x00ff000000ff0000LL, 0x0000ff000000ff00LL, 8);
+  }
+  if (shamt & 4) {
+    x = shuffle_stage(x, 0x0f000f000f000f00LL, 0x00f000f000f000f0LL, 4);
+  }
+  if (shamt & 2) {
+    x = shuffle_stage(x, 0x3030303030303030LL, 0x0c0c0c0c0c0c0c0cLL, 2);
+  }
+  if (shamt & 1) {
+    x = shuffle_stage(x, 0x4444444444444444LL, 0x2222222222222222LL, 1);
+  }
+
+  return x;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_shflw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint32_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 15;
+  if (shamt & 8) {
+    x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
+  }
+  if (shamt & 4) {
+    x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
+  }
+  if (shamt & 2) {
+    x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
+  }
+  if (shamt & 1) {
+    x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
+  }
+  return x;
+}
+
+static __inline__ uint64_t __DEFAULT_FN_ATTRS _rv_bmatflip(const uint64_t rs1) {
+  uint64_t x = rs1;
+  x = _rv_shfl(x, 31);
+  x = _rv_shfl(x, 31);
+  x = _rv_shfl(x, 31);
+  return x;
+}
+
+static __inline__ uint64_t __DEFAULT_FN_ATTRS _rv_bmator(const uint64_t rs1,
+                                                         const uint64_t rs2) {
+  // transpose of rs2
+  const uint64_t rs2t = _rv_bmatflip(rs2);
+  uint8_t u[8]; // rows of rs1
+  uint8_t v[8]; // cols of rs2
+  for (uint_xlen_t i = 0; i < 8; i++) {
+    u[i] = rs1 >> (i * 8);
+    v[i] = rs2t >> (i * 8);
+  }
+  uint64_t x = 0;
+  for (uint_xlen_t i = 0; i < 8; i++) {
+    if ((u[i / 8] & v[i % 8]) != 0) {
+      x |= 1LL << i;
+    }
+  }
+  return x;
+}
+
+static __inline__ uint64_t __DEFAULT_FN_ATTRS _rv_bmatxor(const uint64_t rs1,
+                                                          const uint64_t rs2) {
+  // transpose of rs2
+  const uint64_t rs2t = _rv_bmatflip(rs2);
+  uint8_t u[8]; // rows of rs1
+  uint8_t v[8]; // cols of rs2
+  for (uint_xlen_t i = 0; i < 8; i++) {
+    u[i] = rs1 >> (i * 8);
+    v[i] = rs2t >> (i * 8);
+  }
+  uint64_t x = 0;
+  for (uint_xlen_t i = 0; i < 64; i++) {
+    if (_rv_pcnt(u[i / 8] & v[i % 8]) & 1) {
+      x |= 1LL << i;
+    }
+  }
+  return x;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_fslw(const uint32_t rs1,
+                                                       const uint32_t rs2,
+                                                       const uint32_t rs3) {
+  uint32_t shamt = rs2 & (2 * XLEN - 1);
+  uint32_t A = rs1;
+  uint32_t B = rs3;
+  if (shamt >= XLEN) {
+    shamt -= XLEN;
+    A = rs3;
+    B = rs1;
+  }
+  return shamt ? (A << shamt) | (B >> (XLEN - shamt)) : A;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_fsrw(const uint32_t rs1,
+                                                       const uint32_t rs2,
+                                                       const uint32_t rs3) {
+  uint32_t shamt = rs2 & (2 * XLEN - 1);
+  uint32_t A = rs1;
+  uint32_t B = rs3;
+  if (shamt >= XLEN) {
+    shamt -= XLEN;
+    A = rs3;
+    B = rs1;
+  }
+  return shamt ? (A >> shamt) | (B << (XLEN - shamt)) : A;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_rolw(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  const uint32_t shamt = rs2 & (XLEN - 1);
+  return (rs1 << shamt) | (rs1 >> ((XLEN - shamt) & (XLEN - 1)));
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_rorw(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  const uint32_t shamt = rs2 & (XLEN - 1);
+  return (rs1 >> shamt) | (rs1 << ((XLEN - shamt) & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_addu_w(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 + (rs2 & 0xFFFFFFFF);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_addwu(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  return (rs1 + rs2) & 0xFFFFFFFF;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bdepw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint32_t r = 0;
+  for (uint_xlen_t i = 0, j = 0; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      if ((rs1 >> j) & 1) {
+        r |= 1 << i;
+      }
+      j++;
+    }
+  }
+  return r;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bextw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint32_t r = 0;
+  for (uint_xlen_t i = 0, j = 0; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      if ((rs1 >> i) & 1) {
+        r |= 1 << j;
+      }
+      j++;
+    }
+  }
+  return r;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_bfpw(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  uint_xlen_t len = (rs2 >> 24) & 15;
+  const uint32_t off = (rs2 >> 16) & (XLEN - 1);
+  len = len ? len : 16;
+  const uint32_t mask = _rv_rol(_rv_slo(0, len), off);
+  const uint32_t data = _rv_rol(rs2, off);
+  return (data & mask) | (rs1 & ~mask);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_clmulhw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint32_t x = 0;
+  for (uint_xlen_t i = 1; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      x ^= rs1 >> (XLEN - i);
+    }
+  }
+  return x;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_clmulrw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint32_t x = 0;
+  for (uint_xlen_t i = 0; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      x ^= rs1 >> (XLEN - i - 1);
+    }
+  }
+  return x;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_clmulw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  uint32_t x = 0;
+  for (uint_xlen_t i = 0; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      x ^= rs1 << i;
+    }
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32_d(const uint_xlen_t rs1) {
+  return crc32(rs1, 64);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32c_d(const uint_xlen_t rs1) {
+  return crc32c(rs1, 64);
+}
+
+static __inline__ uint64_t __DEFAULT_FN_ATTRS _rv_gorc(const uint64_t rs1,
+                                                       const uint64_t rs2) {
+  uint64_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 63;
+  if (shamt & 1)
+    x |= ((x & 0x5555555555555555LL) << 1) | ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
+  if (shamt & 2)
+    x |= ((x & 0x3333333333333333LL) << 2) | ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
+  if (shamt & 4)
+    x |= ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) | ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
+  if (shamt & 8)
+    x |= ((x & 0x00FF00FF00FF00FFLL) << 8) | ((x & 0xFF00FF00FF00FF00LL) >> 8);
+  if (shamt & 16)
+    x |=
+        ((x & 0x0000FFFF0000FFFFLL) << 16) | ((x & 0xFFFF0000FFFF0000LL) >> 16);
+  if (shamt & 32)
+    x |=
+        ((x & 0x00000000FFFFFFFFLL) << 32) | ((x & 0xFFFFFFFF00000000LL) >> 32);
+  return x;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_gorcw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint32_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 31;
+  if (shamt & 1) {
+    x |= ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
+  }
+  if (shamt & 2) {
+    x |= ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
+  }
+  if (shamt & 4) {
+    x |= ((x & 0x0F0F0F0F) << 4) | ((x & 0xF0F0F0F0) >> 4);
+  }
+  if (shamt & 8) {
+    x |= ((x & 0x00FF00FF) << 8) | ((x & 0xFF00FF00) >> 8);
+  }
+  if (shamt & 16) {
+    x |= ((x & 0x0000FFFF) << 16) | ((x & 0xFFFF0000) >> 16);
+  }
+  return x;
+}
+
+static __inline__ uint64_t __DEFAULT_FN_ATTRS _rv_grev(const uint64_t rs1,
+                                                       const uint64_t rs2) {
+  uint64_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 63;
+  if (shamt & 1) {
+    x = ((x & 0x5555555555555555LL) << 1) | ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
+  }
+  if (shamt & 2) {
+    x = ((x & 0x3333333333333333LL) << 2) | ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
+  }
+  if (shamt & 4) {
+    x = ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) | ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
+  }
+  if (shamt & 8) {
+    x = ((x & 0x00FF00FF00FF00FFLL) << 8) | ((x & 0xFF00FF00FF00FF00LL) >> 8);
+  }
+  if (shamt & 16) {
+    x = ((x & 0x0000FFFF0000FFFFLL) << 16) | ((x & 0xFFFF0000FFFF0000LL) >> 16);
+  }
+  if (shamt & 32) {
+    x = ((x & 0x00000000FFFFFFFFLL) << 32) | ((x & 0xFFFFFFFF00000000LL) >> 32);
+  }
+  return x;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_grevw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  uint32_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 31;
+  if (shamt & 1) {
+    x = ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
+  }
+  if (shamt & 2) {
+    x = ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
+  }
+  if (shamt & 4) {
+    x = ((x & 0x0F0F0F0F) << 4) | ((x & 0xF0F0F0F0) >> 4);
+  }
+  if (shamt & 8) {
+    x = ((x & 0x00FF00FF) << 8) | ((x & 0xFF00FF00) >> 8);
+  }
+  if (shamt & 16) {
+    x = ((x & 0x0000FFFF) << 16) | ((x & 0xFFFF0000) >> 16);
+  }
+  return x;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_packw(const uint32_t rs1,
+                                                        const uint32_t rs2) {
+  // XLEN / 4 to get half of a 32 bit value on riscv64
+  return (rs2 << XLEN / 4) | ((rs1 << XLEN / 4) >> XLEN / 4);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_packuw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  // XLEN / 4 to get half of a 32 bit value on riscv64
+  return (rs1 >> XLEN / 4) | ((rs2 >> XLEN / 4) << XLEN / 4)
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_sbclrw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return rs1 & ~(1 << (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_sbextw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return 1 & (rs1 >> (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_sbinvw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return rs1 ^ (1 << (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_sbsetw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return rs1 | (1 << (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_slow(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  return ~(~rs1 << (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_srow(const uint32_t rs1,
+                                                       const uint32_t rs2) {
+  return ~(~rs1 >> (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_subu_w(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 - (rs2 & 0xFFFFFFFF);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_subwu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return (rs1 - rs2) & 0xFFFFFFFF;
+}
+
+static __inline__ uint64_t __DEFAULT_FN_ATTRS _rv_unshfl(const uint64_t rs1,
+                                                         const uint64_t rs2) {
+  uint64_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 31;
+  if (shamt & 1) {
+    x = shuffle_stage(x, 0x4444444444444444LL, 0x2222222222222222LL, 1);
+  }
+  if (shamt & 2) {
+    x = shuffle_stage(x, 0x3030303030303030LL, 0x0c0c0c0c0c0c0c0cLL, 2);
+  }
+  if (shamt & 4) {
+    x = shuffle_stage(x, 0x0f000f000f000f00LL, 0x00f000f000f000f0LL, 4);
+  }
+  if (shamt & 8) {
+    x = shuffle_stage(x, 0x00ff000000ff0000LL, 0x0000ff000000ff00LL, 8);
+  }
+  if (shamt & 16) {
+    x = shuffle_stage(x, 0x0000ffff00000000LL, 0x00000000ffff0000LL, 16);
+  }
+  return x;
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv_unshflw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  uint32_t x = rs1;
+  const uint_xlen_t shamt = rs2 & 15;
+  if (shamt & 1) {
+    x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
+  }
+  if (shamt & 2) {
+    x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
+  }
+  if (shamt & 4) {
+    x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
+  }
+  if (shamt & 8) {
+    x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
+  }
+  return x;
+}
+
+#endif //__RV64INTRIN_EMULATION_H
diff --git a/clang/lib/Headers/rv64bintrin.h b/clang/lib/Headers/rv64bintrin.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rv64bintrin.h
@@ -0,0 +1,1127 @@
+/* ===-------- rv64bintrin.h -----------------------------------------------===
+ *
+ * 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 __RV64BINTRIN_H
+#define __RV64BINTRIN_H
+
+#ifndef __RVINTRIN_H
+#error "Never use <rv64bintrin.h> directly; include <rvintrin.h> instead."
+#endif
+
+#include <stdint.h>
+
+#include "rv64bintrin-builtins.h"
+
+#if defined(__riscv_bitmanip)
+
+#include "rv64bintrin-asm.h"
+
+#else // Bitmanip extension is disabled.
+
+#include "rv64bintrin-emulation.h"
+
+#endif // defined(__riscv_bitmanip)
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_andn(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_andn(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_fsl(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return _rv_fsl(rs1, rs2, rs3);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_fsr(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return _rv_fsr(rs1, rs2, rs3);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_max(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_max(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_maxu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_maxu(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_min(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_min(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_minu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_minu(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orn(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_orn(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rol(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_rol(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_ror(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_ror(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_bdep(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_bdep(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_bext(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_bext(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_bfp(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_bfp(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_clmul(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_clmul(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_clmulh(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_clmulh(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_clmulr(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_clmulr(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS _rv64_cmix(
+    const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return _rv_cmix(rs1, rs2, rs3);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS _rv64_cmov(
+    const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return _rv_cmov(rs1, rs2, rs3);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_crc32_b(const uint_xlen_t rs1) {
+  return _rv_crc32_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_crc32_h(const uint_xlen_t rs1) {
+  return _rv_crc32_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_crc32_w(const uint_xlen_t rs1) {
+  return _rv_crc32_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_crc32c_b(const uint_xlen_t rs1) {
+  return _rv_crc32c_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_crc32c_h(const uint_xlen_t rs1) {
+  return _rv_crc32c_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_crc32c_w(const uint_xlen_t rs1) {
+  return _rv_crc32c_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_gorc(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_gorc(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_grev(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_grev(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_pack(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_pack(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_packu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_packu(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_packh(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_packh(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_sext_b(const uint_xlen_t rs1) {
+  return _rv_sext_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_sext_h(const uint_xlen_t rs1) {
+  return _rv_sext_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_sbclr(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sbclr(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_sbext(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sbext(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_sbinv(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sbinv(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_sbset(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sbset(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_shfl(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_shfl(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_slo(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_slo(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_sro(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_sro(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unshfl(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_unshfl(rs1, rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_xnor(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return _rv_xnor(rs1, rs2);
+}
+
+// 64 bit only
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_fslw(const uint32_t rs1,
+                                                         const uint32_t rs2,
+                                                         const uint32_t rs3) {
+  return _rv_fslw(rs1, rs2, rs3);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_fsrw(const uint32_t rs1,
+                                                         const uint32_t rs2,
+                                                         const uint32_t rs3) {
+  return _rv_fsrw(rs1, rs2, rs3);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_rolw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return _rv_rolw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_rorw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return _rv_rorw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_addu_w(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_addu_w(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_addwu(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_addwu(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_bdepw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_bdepw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_bfpw(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return _rv_bfpw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_bextw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_bextw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS
+_rv64_bmatflip(const uint32_t rs1) {
+  return _rv_bmatflip(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_bmator(const uint32_t rs1,
+                                                           const uint32_t rs2) {
+  return _rv_bmator(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS
+_rv64_bmatxor(const uint32_t rs1, const uint32_t rs2) {
+  return _rv_bmatxor(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS
+_rv64_clmulhw(const uint32_t rs1, const uint32_t rs2) {
+  return _rv_clmulhw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS
+_rv64_clmulrw(const uint32_t rs1, const uint32_t rs2) {
+  return _rv_clmulrw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_clmulw(const uint32_t rs1,
+                                                           const uint32_t rs2) {
+  return _rv_clmulw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_crc32_d(const uint32_t rs1) {
+  return _rv_crc32_d(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS
+_rv64_crc32c_d(const uint32_t rs1) {
+  return _rv_crc32c_d(rs1);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_gorcw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_gorcw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_grevw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_grevw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_packw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_packw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_packuw(const uint32_t rs1,
+                                                           const uint32_t rs2) {
+  return _rv_packuw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_sbclrw(const uint32_t rs1,
+                                                           const uint32_t rs2) {
+  return _rv_sbclrw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_sbextw(const uint32_t rs1,
+                                                           const uint32_t rs2) {
+  return _rv_sbextw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_sbinvw(const uint32_t rs1,
+                                                           const uint32_t rs2) {
+  return _rv_sbinvw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_sbsetw(const uint32_t rs1,
+                                                           const uint32_t rs2) {
+  return _rv_sbsetw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_shflw(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_shflw(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_slow(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return _rv_slow(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_srow(const uint32_t rs1,
+                                                         const uint32_t rs2) {
+  return _rv_srow(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_subu_w(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_subu_w(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS _rv64_subwu(const uint32_t rs1,
+                                                          const uint32_t rs2) {
+  return _rv_subwu(rs1, rs2);
+}
+
+static __inline__ uint32_t __DEFAULT_FN_ATTRS
+_rv64_unshflw(const uint32_t rs1, const uint32_t rs2) {
+  return _rv_unshflw(rs1, rs2);
+}
+
+// Pseudo Instructions
+
+// REV (GREVI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev_p(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b000001);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev2_n(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b000010);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev_n(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b000011);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev4_b(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b000100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev2_b(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b000110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev_b(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b000111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev8_h(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b001000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev4_h(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b001100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev2_h(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b001110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev_h(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b001111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev16_w(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b010000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev8_w(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b011000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev4_w(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b011100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev2_w(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b011110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev_w(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b011111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev32(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b100000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev16(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b110000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev8(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b111000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev4(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b111100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev2(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b111110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_rev(const uint_xlen_t rs1) {
+  return _rv64_grev(rs1, 0b111111);
+}
+
+// ORC (GORCI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc_p(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b000001);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc2_n(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b000010);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc_n(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b000011);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc4_b(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b000100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc2_b(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b000110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc_b(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b000111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc8_h(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b001000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc4_h(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b001100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc2_h(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b001110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc_h(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b001111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc16_w(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b010000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc8_w(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b011000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc4_w(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b011100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc2_w(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b011110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc_w(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b011111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc32(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b100000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc16(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b110000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc8(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b111000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc4(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b111100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc2(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b111110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_orc(const uint_xlen_t rs1) {
+  return _rv64_gorc(rs1, 0b111111);
+}
+
+// ZIP (SHFLI) pseudo instructions
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip_n(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b00001);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip2_b(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b00010);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip_b(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b00011);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip4_h(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b00100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip2_h(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b00110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip_h(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b00111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip8_w(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b01000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip4_w(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b01100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip2_w(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b01110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip_w(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b01111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip16(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b10000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip8(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b11000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip4(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b11100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip2(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b11110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_zip(const uint_xlen_t rs1) {
+  return _rv64_shfl(rs1, 0b11111);
+}
+
+// UNZIP (UNSHFL) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip_n(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b00001);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip2_b(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b00010);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip_b(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b00011);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip4_h(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b00100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip2_h(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b00110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip_h(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b00111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip8_w(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b01000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip4_w(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b01100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip2_w(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b01110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip_w(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b01111);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip16(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b11000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip8(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b11000);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip4(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b11100);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip2(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b11110);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv64_unzip(const uint_xlen_t rs1) {
+  return _rv64_unshfl(rs1, 0b11111);
+}
+
+// Genric aliases for 64 bit pseudo instructions
+// e.g. _rv_* is an alias of _rv64_*
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_p(const uint_xlen_t rs1) {
+  return _rv64_rev_p(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2_n(const uint_xlen_t rs1) {
+  return _rv64_rev2_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_n(const uint_xlen_t rs1) {
+  return _rv64_rev_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev4_b(const uint_xlen_t rs1) {
+  return _rv64_rev4_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2_b(const uint_xlen_t rs1) {
+  return _rv64_rev2_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_b(const uint_xlen_t rs1) {
+  return _rv64_rev_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev8_h(const uint_xlen_t rs1) {
+  return _rv64_rev8_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev4_h(const uint_xlen_t rs1) {
+  return _rv64_rev4_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2_h(const uint_xlen_t rs1) {
+  return _rv64_rev2_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_h(const uint_xlen_t rs1) {
+  return _rv64_rev_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev16_w(const uint_xlen_t rs1) {
+  return _rv64_rev16_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev8_w(const uint_xlen_t rs1) {
+  return _rv64_rev8_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev4_w(const uint_xlen_t rs1) {
+  return _rv64_rev4_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2_w(const uint_xlen_t rs1) {
+  return _rv64_rev2_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev_w(const uint_xlen_t rs1) {
+  return _rv64_rev_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev32(const uint_xlen_t rs1) {
+  return _rv64_rev32(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev16(const uint_xlen_t rs1) {
+  return _rv64_rev16(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev8(const uint_xlen_t rs1) {
+  return _rv64_rev8(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev4(const uint_xlen_t rs1) {
+  return _rv64_rev4(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev2(const uint_xlen_t rs1) {
+  return _rv64_rev2(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rev(const uint_xlen_t rs1) {
+  return _rv64_rev(rs1);
+}
+
+// ORC (GORCI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_p(const uint_xlen_t rs1) {
+  return _rv64_orc_p(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2_n(const uint_xlen_t rs1) {
+  return _rv64_orc2_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_n(const uint_xlen_t rs1) {
+  return _rv64_orc_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc4_b(const uint_xlen_t rs1) {
+  return _rv64_orc4_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2_b(const uint_xlen_t rs1) {
+  return _rv64_orc2_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_b(const uint_xlen_t rs1) {
+  return _rv64_orc_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc8_h(const uint_xlen_t rs1) {
+  return _rv64_orc8_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc4_h(const uint_xlen_t rs1) {
+  return _rv64_orc4_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2_h(const uint_xlen_t rs1) {
+  return _rv64_orc2_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_h(const uint_xlen_t rs1) {
+  return _rv64_orc_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc16_w(const uint_xlen_t rs1) {
+  return _rv64_orc16_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc8_w(const uint_xlen_t rs1) {
+  return _rv64_orc8_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc4_w(const uint_xlen_t rs1) {
+  return _rv64_orc4_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2_w(const uint_xlen_t rs1) {
+  return _rv64_orc2_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc_w(const uint_xlen_t rs1) {
+  return _rv64_orc_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc32(const uint_xlen_t rs1) {
+  return _rv64_orc32(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc16(const uint_xlen_t rs1) {
+  return _rv64_orc16(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc8(const uint_xlen_t rs1) {
+  return _rv64_orc8(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc4(const uint_xlen_t rs1) {
+  return _rv64_orc4(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc2(const uint_xlen_t rs1) {
+  return _rv64_orc2(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orc(const uint_xlen_t rs1) {
+  return _rv64_orc(rs1);
+}
+
+// ZIP (SHFLI) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip_n(const uint_xlen_t rs1) {
+  return _rv64_zip_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip2_b(const uint_xlen_t rs1) {
+  return _rv64_zip2_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip_b(const uint_xlen_t rs1) {
+  return _rv64_zip_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip4_h(const uint_xlen_t rs1) {
+  return _rv64_zip4_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip2_h(const uint_xlen_t rs1) {
+  return _rv64_zip2_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip_h(const uint_xlen_t rs1) {
+  return _rv64_zip_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip8_w(const uint_xlen_t rs1) {
+  return _rv64_zip8_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip4_w(const uint_xlen_t rs1) {
+  return _rv64_zip4_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip2_w(const uint_xlen_t rs1) {
+  return _rv64_zip2_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip_w(const uint_xlen_t rs1) {
+  return _rv64_zip_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip16(const uint_xlen_t rs1) {
+  return _rv64_zip16(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip8(const uint_xlen_t rs1) {
+  return _rv64_zip8(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip4(const uint_xlen_t rs1) {
+  return _rv64_zip4(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip2(const uint_xlen_t rs1) {
+  return _rv64_zip2(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_zip(const uint_xlen_t rs1) {
+  return _rv64_zip(rs1);
+}
+
+// UNZIP (UNSHFL) pseudo instructions
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip_n(const uint_xlen_t rs1) {
+  return _rv64_unzip_n(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip2_b(const uint_xlen_t rs1) {
+  return _rv64_unzip2_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip_b(const uint_xlen_t rs1) {
+  return _rv64_unzip_b(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip4_h(const uint_xlen_t rs1) {
+  return _rv64_unzip4_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip2_h(const uint_xlen_t rs1) {
+  return _rv64_unzip2_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip_h(const uint_xlen_t rs1) {
+  return _rv64_unzip_h(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip8_w(const uint_xlen_t rs1) {
+  return _rv64_unzip8_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip4_w(const uint_xlen_t rs1) {
+  return _rv64_unzip4_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip2_w(const uint_xlen_t rs1) {
+  return _rv64_unzip2_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip_w(const uint_xlen_t rs1) {
+  return _rv64_unzip_w(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip16(const uint_xlen_t rs1) {
+  return _rv64_unzip16(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip8(const uint_xlen_t rs1) {
+  return _rv64_unzip8(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip4(const uint_xlen_t rs1) {
+  return _rv64_unzip4(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip2(const uint_xlen_t rs1) {
+  return _rv64_unzip2(rs1);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unzip(const uint_xlen_t rs1) {
+  return _rv64_unzip(rs1);
+}
+
+#endif // __RV64BINTRIN_H
diff --git a/clang/lib/Headers/rvbintrin-asm.h b/clang/lib/Headers/rvbintrin-asm.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rvbintrin-asm.h
@@ -0,0 +1,343 @@
+/* ===-------- rvbintrin-asm.h ---------------------------------------------===
+ *
+ * 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 __RVBINTRIN_ASM_H
+#define __RVBINTRIN_ASM_H
+
+#ifndef __RVINTRIN_H
+#error "Never use <rvbintrin-asm.h> directly; include <rvintrin.h> instead."
+#endif
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_andn(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("andn  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_fsl(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  uint_xlen_t rd;
+  __asm__("fsl  %0, %1, %2, %3" : "=r"(rd) : "r"(rs1), "r"(rs2), "r"(rs3));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_fsr(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("fsri  %0, %1, %2, %3" : "=r"(rd) : "r"(rs1), "r"(rs2), "i"(rs3));
+  } else {
+    __asm__("fsr  %0, %1, %2, %3" : "=r"(rd) : "r"(rs1), "r"(rs2), "r"(rs3));
+  }
+  return rd;
+}
+
+static __inline__ int_xlen_t _rv_max(const int_xlen_t rs1,
+                                     const int_xlen_t rs2) {
+  int_xlen_t rd;
+  __asm__("max  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_maxu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("maxu  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ int_xlen_t _rv_min(const int_xlen_t rs1,
+                                     const int_xlen_t rs2) {
+  int_xlen_t rd;
+  __asm__("min  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_minu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("minu  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orn(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("orn  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rol(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("rol  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_ror(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("rori  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("ror  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_bdep(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("bdep  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_bext(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("bext  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_bfp(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("bfp  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_clmul(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("clmul  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_clmulh(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("clmulh  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_clmulr(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("clmulr  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_cmix(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  uint_xlen_t rd;
+  __asm__("cmix  %0, %1, %2, %3" : "=r"(rd) : "r"(rs1), "r"(rs2), "r"(rs3));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_cmov(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  uint_xlen_t rd;
+  __asm__("cmov  %0, %1, %2, %3" : "=r"(rd) : "r"(rs1), "r"(rs2), "r"(rs3));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32_b(const uint_xlen_t rs1) {
+  uint_xlen_t rd;
+  __asm__("crc32.b  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32_h(const uint_xlen_t rs1) {
+  uint_xlen_t rd;
+  __asm__("crc32.h  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32_w(const uint_xlen_t rs1) {
+  uint_xlen_t rd;
+  __asm__("crc32.w  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32c_b(const uint_xlen_t rs1) {
+  uint_xlen_t rd;
+  __asm__("crc32c.b  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32c_h(const uint_xlen_t rs1) {
+  uint_xlen_t rd;
+  __asm__("crc32c.h  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32c_w(const uint_xlen_t rs1) {
+  uint_xlen_t rd;
+  __asm__("crc32c.w  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_gorc(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("gorci  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("gorc  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_grev(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("grevi  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("grev  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_pack(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("pack  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_packu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("packu  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_packh(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  __asm__("packh  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sext_b(const uint_xlen_t rs1) {
+  uint_xlen_t rd;
+  __asm__("sext.b  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sext_h(const uint_xlen_t rs1) {
+  uint_xlen_t rd;
+  __asm__("sext.h  %0, %1" : "=r"(rd) : "r"(rs1));
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sbclr(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sbclri  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("sbclr  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sbext(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sbexti  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("sbext  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sbinv(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sbinvi  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("sbinv  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sbset(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sbseti  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("sbset  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_shfl(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("shfli  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("shfl  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_slo(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sloi  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("slo  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sro(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("sroi  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("sro  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_unshfl(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t rd;
+  if (__builtin_constant_p(rs2)) {
+    __asm__("unshfli  %0, %1, %2" : "=r"(rd) : "r"(rs1), "i"(rs2));
+  } else {
+    __asm__("unshfl  %0, %1, %2" : "=r"(rd) : "r"(rs1), "r"(rs2));
+  }
+  return rd;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_xnor(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 ^ ~rs2;
+}
+
+#endif // __RVBINTRIN_ASM_H
diff --git a/clang/lib/Headers/rvbintrin-emulation.h b/clang/lib/Headers/rvbintrin-emulation.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rvbintrin-emulation.h
@@ -0,0 +1,285 @@
+/* ===-------- rvbintrin-emulation.h ---------------------------------------===
+ *
+ * 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 __RVBINTRIN_EMULATION_H
+#define __RVBINTRIN_EMULATION_H
+
+#ifndef __RVINTRIN_H
+#error                                                                         \
+    "Never use <rvbintrin-emulation.h> directly; include <rvintrin.h> instead."
+#endif
+
+#include <stdint.h>
+
+// If the builtins are unavailable then these provide the same functionality.
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_slo(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return ~(~rs1 << (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_andn(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 & ~rs2;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_fsl(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  uint_xlen_t shamt = rs2 & (2 * XLEN - 1);
+  uint_xlen_t A = rs1;
+  uint_xlen_t B = rs3;
+  if (shamt >= XLEN) {
+    shamt -= XLEN;
+    A = rs3;
+    B = rs1;
+  }
+  return shamt ? (A << shamt) | (B >> (XLEN - shamt)) : A;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_fsr(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  uint_xlen_t shamt = rs2 & (2 * XLEN - 1);
+  uint_xlen_t A = rs1;
+  uint_xlen_t B = rs3;
+  if (shamt >= XLEN) {
+    shamt -= XLEN;
+    A = rs3;
+    B = rs1;
+  }
+  return shamt ? (A >> shamt) | (B << (XLEN - shamt)) : A;
+}
+
+static __inline__ int_xlen_t _rv_max(const int_xlen_t rs1,
+                                     const int_xlen_t rs2) {
+  return rs1 > rs2 ? rs1 : rs2;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_maxu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 > rs2 ? rs1 : rs2;
+}
+
+static __inline__ int_xlen_t _rv_min(const int_xlen_t rs1,
+                                     const int_xlen_t rs2) {
+  return rs1 < rs2 ? rs1 : rs2;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_minu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 < rs2 ? rs1 : rs2;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_orn(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 | ~rs2;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_rol(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  const uint_xlen_t shamt = rs2 & (XLEN - 1);
+  return (rs1 << shamt) | (rs1 >> ((XLEN - shamt) & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_ror(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  const uint_xlen_t shamt = rs2 & (XLEN - 1);
+  return (rs1 >> shamt) | (rs1 << ((XLEN - shamt) & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_bdep(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t r = 0;
+  for (uint_xlen_t i = 0, j = 0; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      if ((rs1 >> j) & 1) {
+        r |= 1 << i;
+      }
+      j++;
+    }
+  }
+  return r;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_bext(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t r = 0;
+  for (uint_xlen_t i = 0, j = 0; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      if ((rs1 >> i) & 1) {
+        r |= 1 << j;
+      }
+      j++;
+    }
+  }
+  return r;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_bfp(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t len = (rs2 >> 24) & 15;
+  const uint_xlen_t off = (rs2 >> 16) & (XLEN - 1);
+  len = len ? len : 16;
+  const uint_xlen_t mask = _rv_rol(_rv_slo(0, len), off);
+  const uint_xlen_t data = _rv_rol(rs2, off);
+  return (data & mask) | (rs1 & ~mask);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_clmul(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t x = 0;
+  for (uint_xlen_t i = 0; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      x ^= rs1 << i;
+    }
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_clmulh(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t x = 0;
+  for (uint_xlen_t i = 1; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      x ^= rs1 >> (XLEN - i);
+    }
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_clmulr(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  uint_xlen_t x = 0;
+  for (uint_xlen_t i = 0; i < XLEN; i++) {
+    if ((rs2 >> i) & 1) {
+      x ^= rs1 >> (XLEN - i - 1);
+    }
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_cmix(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return (rs1 & rs2) | (rs3 & ~rs2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_cmov(const uint_xlen_t rs1, const uint_xlen_t rs2, const uint_xlen_t rs3) {
+  return rs2 ? rs1 : rs3;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+crc32(uint_xlen_t x, const uint_xlen_t nbits) {
+  for (uint_xlen_t i = 0; i < nbits; i++) {
+    x = (x >> 1) ^ (0xEDB88320 & ~((x & 1) - 1));
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+crc32c(uint_xlen_t x, const uint_xlen_t nbits) {
+  for (uint_xlen_t i = 0; i < nbits; i++) {
+    x = (x >> 1) ^ (0x82F63B78 & ~((x & 1) - 1));
+  }
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32_b(const uint_xlen_t rs1) {
+  return crc32(rs1, 8);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32_h(const uint_xlen_t rs1) {
+  return crc32(rs1, 16);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32_w(const uint_xlen_t rs1) {
+  return crc32(rs1, 32);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32c_b(const uint_xlen_t rs1) {
+  return crc32c(rs1, 8);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32c_h(const uint_xlen_t rs1) {
+  return crc32c(rs1, 16);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_crc32c_w(const uint_xlen_t rs1) {
+  return crc32c(rs1, 32);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_pack(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return (rs2 << XLEN / 2) | ((rs1 << XLEN / 2) >> XLEN / 2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_packu(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return (rs1 >> XLEN / 2) | ((rs2 >> XLEN / 2) << XLEN / 2);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_packh(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return (rs1 & 255) | ((rs2 & 255) << 8);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sext_b(const uint_xlen_t rs1) {
+  return int_xlen_t(x << (XLEN - 8)) >> (XLEN - 8);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sext_h(const uint_xlen_t rs1) {
+  return int_xlen_t(x << (XLEN - 16)) >> (XLEN - 16);
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sbclr(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 & ~(1 << (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sbext(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return 1 & (rs1 >> (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sbinv(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 ^ (1 << (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sbset(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 | (1 << (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+shuffle_stage(const uint_xlen_t src, const uint_xlen_t maskL,
+              const uint_xlen_t maskR, const uint_xlen_t N) {
+  uint_xlen_t x = src & ~(maskL | maskR);
+  x |= ((src << N) & maskL) | ((src >> N) & maskR);
+  return x;
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_sro(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return ~(~rs1 >> (rs2 & (XLEN - 1)));
+}
+
+static __inline__ uint_xlen_t __DEFAULT_FN_ATTRS
+_rv_xnor(const uint_xlen_t rs1, const uint_xlen_t rs2) {
+  return rs1 ^ ~rs2;
+}
+
+#endif //__RVBINTRIN_EMULATION_H
diff --git a/clang/lib/Headers/rvintrin.h b/clang/lib/Headers/rvintrin.h
new file mode 100644
--- /dev/null
+++ b/clang/lib/Headers/rvintrin.h
@@ -0,0 +1,39 @@
+/* ===-------- rvintrin.h --------------------------------------------------===
+ *
+ * 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 __RVINTRIN_H
+#define __RVINTRIN_H
+
+// Long is 32 bit on riscv32 and 64 bit on riscv64
+#define int_xlen_t long
+#define uint_xlen_t unsigned int_xlen_t
+#define XLEN __riscv_xlen
+
+_Static_assert(__riscv_xlen == sizeof(uint_xlen_t) * 8,
+               "uint_xlen_t is not __riscv_xlen bits long");
+
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __artificial__, __nodebug__))
+
+#if defined(__riscv32__)
+
+#include "rv32bintrin.h"
+
+#elif defined(__riscv64__)
+
+#include "rv64bintrin.h"
+
+#endif // defined(__riscv64__)
+
+#undef __DEFAULT_FN_ATTRS
+
+#undef XLEN
+#undef uint_xlen_t
+#undef int_xlen_t
+#endif // __RVINTRIN_H
diff --git a/clang/test/Headers/rvintrin.c b/clang/test/Headers/rvintrin.c
new file mode 100644
--- /dev/null
+++ b/clang/test/Headers/rvintrin.c
@@ -0,0 +1,37 @@
+// RUN: %clang -fsyntax-only               -ffreestanding --target=riscv32 -std=c89 -xc %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -std=c99 -xc %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -std=c11 -xc %s
+
+// RUN: %clang -fsyntax-only               -ffreestanding --target=riscv64 -std=c89 -xc %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -std=c99 -xc %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -std=c11 -xc %s
+
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -std=c++98 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -std=c++11 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -std=c++14 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -std=c++17 -xc++ %s
+
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -std=c++98 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -std=c++11 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -std=c++14 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -std=c++17 -xc++ %s
+
+// RUN: %clang -fsyntax-only               -ffreestanding --target=riscv32 -Xclang -target-feature -Xclang +b -std=c89 -xc %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -Xclang -target-feature -Xclang +b -std=c99 -xc %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -Xclang -target-feature -Xclang +b -std=c11 -xc %s
+
+// RUN: %clang -fsyntax-only               -ffreestanding --target=riscv64 -Xclang -target-feature -Xclang +b -std=c89 -xc %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -Xclang -target-feature -Xclang +b -std=c99 -xc %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -Xclang -target-feature -Xclang +b -std=c11 -xc %s
+
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -Xclang -target-feature -Xclang +b -std=c++98 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -Xclang -target-feature -Xclang +b -std=c++11 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -Xclang -target-feature -Xclang +b -std=c++14 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv32 -Xclang -target-feature -Xclang +b -std=c++17 -xc++ %s
+
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -Xclang -target-feature -Xclang +b -std=c++98 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -Xclang -target-feature -Xclang +b -std=c++11 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -Xclang -target-feature -Xclang +b -std=c++14 -xc++ %s
+// RUN: %clang -fsyntax-only -Wall -Werror -ffreestanding --target=riscv64 -Xclang -target-feature -Xclang +b -std=c++17 -xc++ %s
+
+#include <rvintrin.h>
diff --git a/llvm/utils/gn/secondary/clang/lib/Headers/BUILD.gn b/llvm/utils/gn/secondary/clang/lib/Headers/BUILD.gn
--- a/llvm/utils/gn/secondary/clang/lib/Headers/BUILD.gn
+++ b/llvm/utils/gn/secondary/clang/lib/Headers/BUILD.gn
@@ -132,6 +132,16 @@
     "ptwriteintrin.h",
     "rdseedintrin.h",
     "rtmintrin.h",
+    "rv32bintrin-builtins.h",
+    "rv32bintrin-emulation.h",
+    "rv32bintrin.h",
+    "rv64bintrin-asm.h",
+    "rv64bintrin-builtins.h",
+    "rv64bintrin-emulation.h",
+    "rv64bintrin.h",
+    "rvbintrin-asm.h",
+    "rvbintrin-emulation.h",
+    "rvintrin.h",
     "s390intrin.h",
     "sgxintrin.h",
     "shaintrin.h",