StringRef'ize EmitSourceFileHeader().
[oota-llvm.git] / include / llvm / IntrinsicsX86.td
index 6a35b7cbb40145859d3a0213ef0b19879d7e7ab9..f4abba98c083a05a7255e23718fa5d36ecd1d687 100644 (file)
@@ -1,16 +1,98 @@
 //===- IntrinsicsX86.td - Defines X86 intrinsics -----------*- tablegen -*-===//
-// 
+//
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
-// 
+//
 //===----------------------------------------------------------------------===//
 //
 // This file defines all of the X86-specific intrinsics.
 //
 //===----------------------------------------------------------------------===//
 
+//===----------------------------------------------------------------------===//
+// Interrupt traps
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_int : Intrinsic<[], [llvm_i8_ty]>;
+}
+
+//===----------------------------------------------------------------------===//
+// 3DNow!
+
+let TargetPrefix = "x86" in {
+  def int_x86_3dnow_pavgusb : GCCBuiltin<"__builtin_ia32_pavgusb">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pf2id : GCCBuiltin<"__builtin_ia32_pf2id">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
+  def int_x86_3dnow_pfacc : GCCBuiltin<"__builtin_ia32_pfacc">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfadd : GCCBuiltin<"__builtin_ia32_pfadd">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfcmpeq : GCCBuiltin<"__builtin_ia32_pfcmpeq">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfcmpge : GCCBuiltin<"__builtin_ia32_pfcmpge">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfcmpgt : GCCBuiltin<"__builtin_ia32_pfcmpgt">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfmax : GCCBuiltin<"__builtin_ia32_pfmax">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfmin : GCCBuiltin<"__builtin_ia32_pfmin">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfmul : GCCBuiltin<"__builtin_ia32_pfmul">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfrcp : GCCBuiltin<"__builtin_ia32_pfrcp">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
+  def int_x86_3dnow_pfrcpit1 : GCCBuiltin<"__builtin_ia32_pfrcpit1">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfrcpit2 : GCCBuiltin<"__builtin_ia32_pfrcpit2">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfrsqrt : GCCBuiltin<"__builtin_ia32_pfrsqrt">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
+  def int_x86_3dnow_pfrsqit1 : GCCBuiltin<"__builtin_ia32_pfrsqit1">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfsub : GCCBuiltin<"__builtin_ia32_pfsub">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pfsubr : GCCBuiltin<"__builtin_ia32_pfsubr">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnow_pi2fd : GCCBuiltin<"__builtin_ia32_pi2fd">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
+  def int_x86_3dnow_pmulhrw : GCCBuiltin<"__builtin_ia32_pmulhrw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+}
+
+//===----------------------------------------------------------------------===//
+// 3DNow! extensions
+
+let TargetPrefix = "x86" in {
+  def int_x86_3dnowa_pf2iw : GCCBuiltin<"__builtin_ia32_pf2iw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
+  def int_x86_3dnowa_pfnacc : GCCBuiltin<"__builtin_ia32_pfnacc">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnowa_pfpnacc : GCCBuiltin<"__builtin_ia32_pfpnacc">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_3dnowa_pi2fw : GCCBuiltin<"__builtin_ia32_pi2fw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
+  def int_x86_3dnowa_pswapd :
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
+}
 
 //===----------------------------------------------------------------------===//
 // SSE1
 // Arithmetic ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse_add_ss : GCCBuiltin<"__builtin_ia32_addss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
-                         llvm_v4f32_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
+                         llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_sub_ss : GCCBuiltin<"__builtin_ia32_subss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_mul_ss : GCCBuiltin<"__builtin_ia32_mulss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
-                         llvm_v4f32_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
+                         llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_div_ss : GCCBuiltin<"__builtin_ia32_divss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_sqrt_ss : GCCBuiltin<"__builtin_ia32_sqrtss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty],
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty],
                         [IntrNoMem]>;
   def int_x86_sse_sqrt_ps : GCCBuiltin<"__builtin_ia32_sqrtps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty],
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty],
                         [IntrNoMem]>;
   def int_x86_sse_rcp_ss : GCCBuiltin<"__builtin_ia32_rcpss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty],
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty],
                         [IntrNoMem]>;
   def int_x86_sse_rcp_ps : GCCBuiltin<"__builtin_ia32_rcpps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty],
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty],
                         [IntrNoMem]>;
   def int_x86_sse_rsqrt_ss : GCCBuiltin<"__builtin_ia32_rsqrtss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty],
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty],
                         [IntrNoMem]>;
   def int_x86_sse_rsqrt_ps : GCCBuiltin<"__builtin_ia32_rsqrtps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty],
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty],
                         [IntrNoMem]>;
   def int_x86_sse_min_ss : GCCBuiltin<"__builtin_ia32_minss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_min_ps : GCCBuiltin<"__builtin_ia32_minps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_max_ss : GCCBuiltin<"__builtin_ia32_maxss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_max_ps : GCCBuiltin<"__builtin_ia32_maxps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
 }
 
 // Comparison ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse_cmp_ss :
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+  def int_x86_sse_cmp_ss : GCCBuiltin<"__builtin_ia32_cmpss">,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>;
-  def int_x86_sse_cmp_ps :
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+  def int_x86_sse_cmp_ps : GCCBuiltin<"__builtin_ia32_cmpps">,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>;
   def int_x86_sse_comieq_ss : GCCBuiltin<"__builtin_ia32_comieq">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_comilt_ss : GCCBuiltin<"__builtin_ia32_comilt">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_comile_ss : GCCBuiltin<"__builtin_ia32_comile">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_comigt_ss : GCCBuiltin<"__builtin_ia32_comigt">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_comige_ss : GCCBuiltin<"__builtin_ia32_comige">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_comineq_ss : GCCBuiltin<"__builtin_ia32_comineq">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_ucomieq_ss : GCCBuiltin<"__builtin_ia32_ucomieq">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_ucomilt_ss : GCCBuiltin<"__builtin_ia32_ucomilt">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_ucomile_ss : GCCBuiltin<"__builtin_ia32_ucomile">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_ucomigt_ss : GCCBuiltin<"__builtin_ia32_ucomigt">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_ucomige_ss : GCCBuiltin<"__builtin_ia32_ucomige">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_ucomineq_ss : GCCBuiltin<"__builtin_ia32_ucomineq">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
 }
 
@@ -111,62 +193,53 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
 // Conversion ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse_cvtss2si : GCCBuiltin<"__builtin_ia32_cvtss2si">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_cvtss2si64 : GCCBuiltin<"__builtin_ia32_cvtss2si64">,
-              Intrinsic<[llvm_i64_tyllvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_cvttss2si : GCCBuiltin<"__builtin_ia32_cvttss2si">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_cvttss2si64 : GCCBuiltin<"__builtin_ia32_cvttss2si64">,
-              Intrinsic<[llvm_i64_tyllvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_cvtsi2ss : GCCBuiltin<"__builtin_ia32_cvtsi2ss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse_cvtsi642ss : GCCBuiltin<"__builtin_ia32_cvtsi642ss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_i64_ty], [IntrNoMem]>;
   def int_x86_sse_cvtps2pi : GCCBuiltin<"__builtin_ia32_cvtps2pi">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_cvttps2pi: GCCBuiltin<"__builtin_ia32_cvttps2pi">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_cvtpi2ps : GCCBuiltin<"__builtin_ia32_cvtpi2ps">,
-              Intrinsic<[llvm_v4f32_ty, llvm_v4f32_ty, 
-                         llvm_v2i32_ty], [IntrNoMem]>;
-}
-
-// SIMD load ops
-let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse_loadu_ps : GCCBuiltin<"__builtin_ia32_loadups">,
-              Intrinsic<[llvm_v4f32_ty, llvm_ptr_ty], [IntrReadMem]>;
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
 }
 
 // SIMD store ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse_storeu_ps : GCCBuiltin<"__builtin_ia32_storeups">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_v4f32_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptr_ty,
+                         llvm_v4f32_ty], []>;
 }
 
 // Cacheability support ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse_movnt_ps : GCCBuiltin<"__builtin_ia32_movntps">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_v4f32_ty], [IntrWriteMem]>;
   def int_x86_sse_sfence : GCCBuiltin<"__builtin_ia32_sfence">,
-              Intrinsic<[llvm_void_ty], [IntrWriteMem]>;
+              Intrinsic<[], [], []>;
 }
 
 // Control register.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse_stmxcsr :
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptr_ty], []>;
   def int_x86_sse_ldmxcsr :
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptr_ty], []>;
 }
 
 // Misc.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse_movmsk_ps : GCCBuiltin<"__builtin_ia32_movmskps">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
 }
 
 //===----------------------------------------------------------------------===//
@@ -175,336 +248,293 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
 // FP arithmetic ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_add_sd : GCCBuiltin<"__builtin_ia32_addsd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
-                         llvm_v2f64_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
+                         llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_sub_sd : GCCBuiltin<"__builtin_ia32_subsd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_mul_sd : GCCBuiltin<"__builtin_ia32_mulsd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
-                         llvm_v2f64_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
+                         llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_div_sd : GCCBuiltin<"__builtin_ia32_divsd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_sqrt_sd : GCCBuiltin<"__builtin_ia32_sqrtsd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty],
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty],
                         [IntrNoMem]>;
   def int_x86_sse2_sqrt_pd : GCCBuiltin<"__builtin_ia32_sqrtpd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty],
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty],
                         [IntrNoMem]>;
   def int_x86_sse2_min_sd : GCCBuiltin<"__builtin_ia32_minsd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_min_pd : GCCBuiltin<"__builtin_ia32_minpd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_max_sd : GCCBuiltin<"__builtin_ia32_maxsd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_max_pd : GCCBuiltin<"__builtin_ia32_maxpd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
 }
 
 // FP comparison ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse2_cmp_sd :
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+  def int_x86_sse2_cmp_sd : GCCBuiltin<"__builtin_ia32_cmpsd">,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>;
-  def int_x86_sse2_cmp_pd :
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+  def int_x86_sse2_cmp_pd : GCCBuiltin<"__builtin_ia32_cmppd">,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>;
   def int_x86_sse2_comieq_sd : GCCBuiltin<"__builtin_ia32_comisdeq">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_comilt_sd : GCCBuiltin<"__builtin_ia32_comisdlt">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_comile_sd : GCCBuiltin<"__builtin_ia32_comisdle">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_comigt_sd : GCCBuiltin<"__builtin_ia32_comisdgt">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_comige_sd : GCCBuiltin<"__builtin_ia32_comisdge">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_comineq_sd : GCCBuiltin<"__builtin_ia32_comisdneq">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_ucomieq_sd : GCCBuiltin<"__builtin_ia32_ucomisdeq">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_ucomilt_sd : GCCBuiltin<"__builtin_ia32_ucomisdlt">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_ucomile_sd : GCCBuiltin<"__builtin_ia32_ucomisdle">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_ucomigt_sd : GCCBuiltin<"__builtin_ia32_ucomisdgt">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_ucomige_sd : GCCBuiltin<"__builtin_ia32_ucomisdge">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_ucomineq_sd : GCCBuiltin<"__builtin_ia32_ucomisdneq">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
 }
 
 // Integer arithmetic ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_padds_b : GCCBuiltin<"__builtin_ia32_paddsb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_padds_w : GCCBuiltin<"__builtin_ia32_paddsw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_paddus_b : GCCBuiltin<"__builtin_ia32_paddusb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_paddus_w : GCCBuiltin<"__builtin_ia32_paddusw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_psubs_b : GCCBuiltin<"__builtin_ia32_psubsb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem]>;
   def int_x86_sse2_psubs_w : GCCBuiltin<"__builtin_ia32_psubsw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_psubus_b : GCCBuiltin<"__builtin_ia32_psubusb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem]>;
   def int_x86_sse2_psubus_w : GCCBuiltin<"__builtin_ia32_psubusw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_pmulhu_w : GCCBuiltin<"__builtin_ia32_pmulhuw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pmulh_w : GCCBuiltin<"__builtin_ia32_pmulhw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pmulu_dq : GCCBuiltin<"__builtin_ia32_pmuludq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pmadd_wd : GCCBuiltin<"__builtin_ia32_pmaddwd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pavg_b : GCCBuiltin<"__builtin_ia32_pavgb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pavg_w : GCCBuiltin<"__builtin_ia32_pavgw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pmaxu_b : GCCBuiltin<"__builtin_ia32_pmaxub128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pmaxs_w : GCCBuiltin<"__builtin_ia32_pmaxsw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pminu_b : GCCBuiltin<"__builtin_ia32_pminub128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pmins_w : GCCBuiltin<"__builtin_ia32_pminsw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_psad_bw : GCCBuiltin<"__builtin_ia32_psadbw128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem, Commutative]>;
 }
 
 // Integer shift ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_psll_w : GCCBuiltin<"__builtin_ia32_psllw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_psll_d : GCCBuiltin<"__builtin_ia32_pslld128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psll_q : GCCBuiltin<"__builtin_ia32_psllq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v2i64_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty,
                          llvm_v2i64_ty], [IntrNoMem]>;
   def int_x86_sse2_psrl_w : GCCBuiltin<"__builtin_ia32_psrlw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_psrl_d : GCCBuiltin<"__builtin_ia32_psrld128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrl_q : GCCBuiltin<"__builtin_ia32_psrlq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v2i64_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty,
                          llvm_v2i64_ty], [IntrNoMem]>;
   def int_x86_sse2_psra_w : GCCBuiltin<"__builtin_ia32_psraw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_psra_d : GCCBuiltin<"__builtin_ia32_psrad128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem]>;
 
   def int_x86_sse2_pslli_w : GCCBuiltin<"__builtin_ia32_psllwi128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_pslli_d : GCCBuiltin<"__builtin_ia32_pslldi128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_pslli_q : GCCBuiltin<"__builtin_ia32_psllqi128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v2i64_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrli_w : GCCBuiltin<"__builtin_ia32_psrlwi128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrli_d : GCCBuiltin<"__builtin_ia32_psrldi128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrli_q : GCCBuiltin<"__builtin_ia32_psrlqi128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v2i64_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrai_w : GCCBuiltin<"__builtin_ia32_psrawi128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrai_d : GCCBuiltin<"__builtin_ia32_psradi128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_i32_ty], [IntrNoMem]>;
 
   def int_x86_sse2_psll_dq : GCCBuiltin<"__builtin_ia32_pslldqi128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v2i64_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrl_dq : GCCBuiltin<"__builtin_ia32_psrldqi128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v2i64_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psll_dq_bs : GCCBuiltin<"__builtin_ia32_pslldqi128_byteshift">,
-              Intrinsic<[llvm_v2i64_tyllvm_v2i64_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrl_dq_bs : GCCBuiltin<"__builtin_ia32_psrldqi128_byteshift">,
-              Intrinsic<[llvm_v2i64_tyllvm_v2i64_ty,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
 }
 
-// Integer comparison ops
-let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse2_pcmpeq_b : GCCBuiltin<"__builtin_ia32_pcmpeqb128">,
-              Intrinsic<[llvm_v16i8_ty, llvm_v16i8_ty,
-                         llvm_v16i8_ty], [IntrNoMem]>;
-  def int_x86_sse2_pcmpeq_w : GCCBuiltin<"__builtin_ia32_pcmpeqw128">,
-              Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
-                         llvm_v8i16_ty], [IntrNoMem]>;
-  def int_x86_sse2_pcmpeq_d : GCCBuiltin<"__builtin_ia32_pcmpeqd128">,
-              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty,
-                         llvm_v4i32_ty], [IntrNoMem]>;
-  def int_x86_sse2_pcmpgt_b : GCCBuiltin<"__builtin_ia32_pcmpgtb128">,
-              Intrinsic<[llvm_v16i8_ty, llvm_v16i8_ty,
-                         llvm_v16i8_ty], [IntrNoMem]>;
-  def int_x86_sse2_pcmpgt_w : GCCBuiltin<"__builtin_ia32_pcmpgtw128">,
-              Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
-                         llvm_v8i16_ty], [IntrNoMem]>;
-  def int_x86_sse2_pcmpgt_d : GCCBuiltin<"__builtin_ia32_pcmpgtd128">,
-              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty,
-                         llvm_v4i32_ty], [IntrNoMem]>;
-}
-
 // Conversion ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_cvtdq2pd : GCCBuiltin<"__builtin_ia32_cvtdq2pd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v4i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v2f64_ty], [llvm_v4i32_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtdq2ps : GCCBuiltin<"__builtin_ia32_cvtdq2ps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4i32_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtpd2dq : GCCBuiltin<"__builtin_ia32_cvtpd2dq">,
-              Intrinsic<[llvm_v4i32_tyllvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v4i32_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvttpd2dq : GCCBuiltin<"__builtin_ia32_cvttpd2dq">,
-              Intrinsic<[llvm_v4i32_tyllvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v4i32_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtpd2ps : GCCBuiltin<"__builtin_ia32_cvtpd2ps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v4f32_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtps2dq : GCCBuiltin<"__builtin_ia32_cvtps2dq">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse2_cvttps2dq : GCCBuiltin<"__builtin_ia32_cvttps2dq">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtps2pd : GCCBuiltin<"__builtin_ia32_cvtps2pd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v4f32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v2f64_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtsd2si : GCCBuiltin<"__builtin_ia32_cvtsd2si">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtsd2si64 : GCCBuiltin<"__builtin_ia32_cvtsd2si64">,
-              Intrinsic<[llvm_i64_tyllvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvttsd2si : GCCBuiltin<"__builtin_ia32_cvttsd2si">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvttsd2si64 : GCCBuiltin<"__builtin_ia32_cvttsd2si64">,
-              Intrinsic<[llvm_i64_tyllvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtsi2sd : GCCBuiltin<"__builtin_ia32_cvtsi2sd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtsi642sd : GCCBuiltin<"__builtin_ia32_cvtsi642sd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_i64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtsd2ss : GCCBuiltin<"__builtin_ia32_cvtsd2ss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_cvtss2sd : GCCBuiltin<"__builtin_ia32_cvtss2sd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse_cvtpd2pi : GCCBuiltin<"__builtin_ia32_cvtpd2pi">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse_cvttpd2pi: GCCBuiltin<"__builtin_ia32_cvttpd2pi">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse_cvtpi2pd : GCCBuiltin<"__builtin_ia32_cvtpi2pd">,
-              Intrinsic<[llvm_v2f64_ty, llvm_v2i32_ty], [IntrNoMem]>;
-}
-
-// SIMD load ops
-let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse2_loadu_pd : GCCBuiltin<"__builtin_ia32_loadupd">,
-              Intrinsic<[llvm_v2f64_ty, llvm_ptr_ty], [IntrReadMem]>;
-  def int_x86_sse2_loadu_dq : GCCBuiltin<"__builtin_ia32_loaddqu">,
-              Intrinsic<[llvm_v16i8_ty, llvm_ptr_ty], [IntrReadMem]>;
+              Intrinsic<[llvm_v2f64_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
 }
 
 // SIMD store ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_storeu_pd : GCCBuiltin<"__builtin_ia32_storeupd">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_v2f64_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptr_ty,
+                         llvm_v2f64_ty], []>;
   def int_x86_sse2_storeu_dq : GCCBuiltin<"__builtin_ia32_storedqu">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_v16i8_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptr_ty,
+                         llvm_v16i8_ty], []>;
   def int_x86_sse2_storel_dq : GCCBuiltin<"__builtin_ia32_storelv4si">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_v4i32_ty], [IntrWriteMem]>;
-}
-
-// Cacheability support ops
-let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse2_movnt_dq : GCCBuiltin<"__builtin_ia32_movntdq">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_v2i64_ty], [IntrWriteMem]>;
-  def int_x86_sse2_movnt_pd : GCCBuiltin<"__builtin_ia32_movntpd">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_v2f64_ty], [IntrWriteMem]>;
-  def int_x86_sse2_movnt_i : GCCBuiltin<"__builtin_ia32_movnti">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_i32_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptr_ty,
+                         llvm_v4i32_ty], []>;
 }
 
 // Misc.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_packsswb_128 : GCCBuiltin<"__builtin_ia32_packsswb128">,
-              Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_packssdw_128 : GCCBuiltin<"__builtin_ia32_packssdw128">,
-              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem]>;
   def int_x86_sse2_packuswb_128 : GCCBuiltin<"__builtin_ia32_packuswb128">,
-              Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_movmsk_pd : GCCBuiltin<"__builtin_ia32_movmskpd">,
-              Intrinsic<[llvm_i32_tyllvm_v2f64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse2_pmovmskb_128 : GCCBuiltin<"__builtin_ia32_pmovmskb128">,
-              Intrinsic<[llvm_i32_tyllvm_v16i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty], [IntrNoMem]>;
   def int_x86_sse2_maskmov_dqu : GCCBuiltin<"__builtin_ia32_maskmovdqu">,
-              Intrinsic<[llvm_void_ty, llvm_v16i8_ty,
-                         llvm_v16i8_ty, llvm_ptr_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_v16i8_ty,
+                         llvm_v16i8_ty, llvm_ptr_ty], []>;
   def int_x86_sse2_clflush : GCCBuiltin<"__builtin_ia32_clflush">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptr_ty], []>;
   def int_x86_sse2_lfence : GCCBuiltin<"__builtin_ia32_lfence">,
-              Intrinsic<[llvm_void_ty], [IntrWriteMem]>;
+              Intrinsic<[], [], []>;
   def int_x86_sse2_mfence : GCCBuiltin<"__builtin_ia32_mfence">,
-              Intrinsic<[llvm_void_ty], [IntrWriteMem]>;
+              Intrinsic<[], [], []>;
 }
 
 //===----------------------------------------------------------------------===//
@@ -513,43 +543,43 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
 // Addition / subtraction ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse3_addsub_ps : GCCBuiltin<"__builtin_ia32_addsubps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse3_addsub_pd : GCCBuiltin<"__builtin_ia32_addsubpd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
 }
 
 // Horizontal ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse3_hadd_ps : GCCBuiltin<"__builtin_ia32_haddps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse3_hadd_pd : GCCBuiltin<"__builtin_ia32_haddpd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
   def int_x86_sse3_hsub_ps : GCCBuiltin<"__builtin_ia32_hsubps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_v4f32_ty], [IntrNoMem]>;
   def int_x86_sse3_hsub_pd : GCCBuiltin<"__builtin_ia32_hsubpd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_v2f64_ty], [IntrNoMem]>;
 }
 
 // Specialized unaligned load.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse3_ldu_dq : GCCBuiltin<"__builtin_ia32_lddqu">,
-              Intrinsic<[llvm_v16i8_tyllvm_ptr_ty], [IntrReadMem]>;
+              Intrinsic<[llvm_v16i8_ty], [llvm_ptr_ty], [IntrReadMem]>;
 }
 
 // Thread synchronization ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse3_monitor : GCCBuiltin<"__builtin_ia32_monitor">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_i32_ty, llvm_i32_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptr_ty,
+                         llvm_i32_ty, llvm_i32_ty], []>;
   def int_x86_sse3_mwait : GCCBuiltin<"__builtin_ia32_mwait">,
-              Intrinsic<[llvm_void_ty, llvm_i32_ty,
-                         llvm_i32_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_i32_ty,
+                         llvm_i32_ty], []>;
 }
 
 //===----------------------------------------------------------------------===//
@@ -558,125 +588,118 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
 // Horizontal arithmetic ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_ssse3_phadd_w         : GCCBuiltin<"__builtin_ia32_phaddw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_phadd_w_128     : GCCBuiltin<"__builtin_ia32_phaddw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_phadd_d         : GCCBuiltin<"__builtin_ia32_phaddd">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v2i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_phadd_d_128     : GCCBuiltin<"__builtin_ia32_phaddd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_phadd_sw        : GCCBuiltin<"__builtin_ia32_phaddsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_phadd_sw_128    : GCCBuiltin<"__builtin_ia32_phaddsw128">,
-              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty,
-                         llvm_v4i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
+                         llvm_v8i16_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_phsub_w         : GCCBuiltin<"__builtin_ia32_phsubw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_phsub_w_128     : GCCBuiltin<"__builtin_ia32_phsubw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_phsub_d         : GCCBuiltin<"__builtin_ia32_phsubd">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v2i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_phsub_d_128     : GCCBuiltin<"__builtin_ia32_phsubd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_phsub_sw        : GCCBuiltin<"__builtin_ia32_phsubsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_phsub_sw_128    : GCCBuiltin<"__builtin_ia32_phsubsw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_pmadd_ub_sw     : GCCBuiltin<"__builtin_ia32_pmaddubsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_pmadd_ub_sw_128 : GCCBuiltin<"__builtin_ia32_pmaddubsw128">,
-              Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
-                         llvm_v8i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v8i16_ty], [llvm_v16i8_ty,
+                         llvm_v16i8_ty], [IntrNoMem]>;
 }
 
 // Packed multiply high with round and scale
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_ssse3_pmul_hr_sw      : GCCBuiltin<"__builtin_ia32_pmulhrsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_ssse3_pmul_hr_sw_128  : GCCBuiltin<"__builtin_ia32_pmulhrsw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
 }
 
 // Shuffle ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_ssse3_pshuf_b         : GCCBuiltin<"__builtin_ia32_pshufb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_pshuf_b_128     : GCCBuiltin<"__builtin_ia32_pshufb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_sse_pshuf_w           : GCCBuiltin<"__builtin_ia32_pshufw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_i8_ty],
+                         [IntrNoMem]>;
 }
 
 // Sign ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_ssse3_psign_b         : GCCBuiltin<"__builtin_ia32_psignb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_psign_b_128     : GCCBuiltin<"__builtin_ia32_psignb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
                          llvm_v16i8_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_psign_w         : GCCBuiltin<"__builtin_ia32_psignw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_psign_w_128     : GCCBuiltin<"__builtin_ia32_psignw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_psign_d         : GCCBuiltin<"__builtin_ia32_psignd">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v2i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_psign_d_128     : GCCBuiltin<"__builtin_ia32_psignd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem]>;
 }
 
 // Absolute value ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_ssse3_pabs_b     : GCCBuiltin<"__builtin_ia32_pabsb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_pabs_b_128 : GCCBuiltin<"__builtin_ia32_pabsb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_pabs_w     : GCCBuiltin<"__builtin_ia32_pabsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_pabs_w_128 : GCCBuiltin<"__builtin_ia32_pabsw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty], [IntrNoMem]>;
 
   def int_x86_ssse3_pabs_d     : GCCBuiltin<"__builtin_ia32_pabsd">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_ssse3_pabs_d_128 : GCCBuiltin<"__builtin_ia32_pabsd128">,
-              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty], [IntrNoMem]>;
-}
-
-// Align ops
-let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_ssse3_palign_r        : GCCBuiltin<"__builtin_ia32_palignr">,
-              Intrinsic<[llvm_v1i64_ty, llvm_v1i64_ty,
-                         llvm_v1i64_ty, llvm_i16_ty], [IntrNoMem]>;
-  def int_x86_ssse3_palign_r_128    : GCCBuiltin<"__builtin_ia32_palignr128">,
-              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
-                         llvm_v2i64_ty, llvm_i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty], [IntrNoMem]>;
 }
 
 //===----------------------------------------------------------------------===//
@@ -685,188 +708,1663 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
 // FP rounding ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_round_ss        : GCCBuiltin<"__builtin_ia32_roundss">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty, llvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse41_round_ps        : GCCBuiltin<"__builtin_ia32_roundps">,
-              Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse41_round_sd        : GCCBuiltin<"__builtin_ia32_roundsd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty, llvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse41_round_pd        : GCCBuiltin<"__builtin_ia32_roundpd">,
-              Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
                          llvm_i32_ty], [IntrNoMem]>;
 }
 
 // Vector sign and zero extend
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_pmovsxbd        : GCCBuiltin<"__builtin_ia32_pmovsxbd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v16i8_ty],
+              Intrinsic<[llvm_v4i32_ty], [llvm_v16i8_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovsxbq        : GCCBuiltin<"__builtin_ia32_pmovsxbq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v16i8_ty],
+              Intrinsic<[llvm_v2i64_ty], [llvm_v16i8_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovsxbw        : GCCBuiltin<"__builtin_ia32_pmovsxbw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v16i8_ty],
+              Intrinsic<[llvm_v8i16_ty], [llvm_v16i8_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovsxdq        : GCCBuiltin<"__builtin_ia32_pmovsxdq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v4i32_ty],
+              Intrinsic<[llvm_v2i64_ty], [llvm_v4i32_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovsxwd        : GCCBuiltin<"__builtin_ia32_pmovsxwd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v8i16_ty],
+              Intrinsic<[llvm_v4i32_ty], [llvm_v8i16_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovsxwq        : GCCBuiltin<"__builtin_ia32_pmovsxwq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v8i16_ty],
+              Intrinsic<[llvm_v2i64_ty], [llvm_v8i16_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovzxbd        : GCCBuiltin<"__builtin_ia32_pmovzxbd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v16i8_ty],
+              Intrinsic<[llvm_v4i32_ty], [llvm_v16i8_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovzxbq        : GCCBuiltin<"__builtin_ia32_pmovzxbq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v16i8_ty],
+              Intrinsic<[llvm_v2i64_ty], [llvm_v16i8_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovzxbw        : GCCBuiltin<"__builtin_ia32_pmovzxbw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v16i8_ty],
+              Intrinsic<[llvm_v8i16_ty], [llvm_v16i8_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovzxdq        : GCCBuiltin<"__builtin_ia32_pmovzxdq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v4i32_ty],
+              Intrinsic<[llvm_v2i64_ty], [llvm_v4i32_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovzxwd        : GCCBuiltin<"__builtin_ia32_pmovzxwd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v8i16_ty],
+              Intrinsic<[llvm_v4i32_ty], [llvm_v8i16_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pmovzxwq        : GCCBuiltin<"__builtin_ia32_pmovzxwq128">,
-              Intrinsic<[llvm_v2i64_tyllvm_v8i16_ty],
+              Intrinsic<[llvm_v2i64_ty], [llvm_v8i16_ty],
                         [IntrNoMem]>;
 }
 
 // Vector min element
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_phminposuw     : GCCBuiltin<"__builtin_ia32_phminposuw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty],
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty],
                         [IntrNoMem]>;
 }
 
 // Vector compare, min, max
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse41_pcmpeqq         : GCCBuiltin<"__builtin_ia32_pcmpeqq">,
-              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty, llvm_v2i64_ty],
-                        [IntrNoMem, Commutative]>;
-  def int_x86_sse42_pcmpgtq         : GCCBuiltin<"__builtin_ia32_pcmpgtq">,
-              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty, llvm_v2i64_ty],
-                        [IntrNoMem]>;
   def int_x86_sse41_pmaxsb          : GCCBuiltin<"__builtin_ia32_pmaxsb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty, llvm_v16i8_ty],
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
                         [IntrNoMem, Commutative]>;
   def int_x86_sse41_pmaxsd          : GCCBuiltin<"__builtin_ia32_pmaxsd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty, llvm_v4i32_ty],
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
                         [IntrNoMem, Commutative]>;
   def int_x86_sse41_pmaxud          : GCCBuiltin<"__builtin_ia32_pmaxud128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty, llvm_v4i32_ty],
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
                         [IntrNoMem, Commutative]>;
   def int_x86_sse41_pmaxuw          : GCCBuiltin<"__builtin_ia32_pmaxuw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty, llvm_v8i16_ty],
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
                         [IntrNoMem, Commutative]>;
   def int_x86_sse41_pminsb          : GCCBuiltin<"__builtin_ia32_pminsb128">,
-              Intrinsic<[llvm_v16i8_tyllvm_v16i8_ty, llvm_v16i8_ty],
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
                         [IntrNoMem, Commutative]>;
   def int_x86_sse41_pminsd          : GCCBuiltin<"__builtin_ia32_pminsd128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty, llvm_v4i32_ty],
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
                         [IntrNoMem, Commutative]>;
   def int_x86_sse41_pminud          : GCCBuiltin<"__builtin_ia32_pminud128">,
-              Intrinsic<[llvm_v4i32_tyllvm_v4i32_ty, llvm_v4i32_ty],
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
                         [IntrNoMem, Commutative]>;
   def int_x86_sse41_pminuw          : GCCBuiltin<"__builtin_ia32_pminuw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty, llvm_v8i16_ty],
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
                         [IntrNoMem, Commutative]>;
 }
 
+// Advanced Encryption Standard (AES) Instructions
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_aesni_aesimc          : GCCBuiltin<"__builtin_ia32_aesimc128">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_aesni_aesenc          : GCCBuiltin<"__builtin_ia32_aesenc128">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_aesni_aesenclast : GCCBuiltin<"__builtin_ia32_aesenclast128">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_aesni_aesdec          : GCCBuiltin<"__builtin_ia32_aesdec128">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_aesni_aesdeclast : GCCBuiltin<"__builtin_ia32_aesdeclast128">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_aesni_aeskeygenassist :
+              GCCBuiltin<"__builtin_ia32_aeskeygenassist128">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_i8_ty],
+                        [IntrNoMem]>;
+}
+
 // Vector pack
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_packusdw        : GCCBuiltin<"__builtin_ia32_packusdw128">,
-              Intrinsic<[llvm_v8i16_tyllvm_v4i32_ty, llvm_v4i32_ty],
+              Intrinsic<[llvm_v8i16_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
                         [IntrNoMem]>;
 }
 
 // Vector multiply
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_pmuldq          : GCCBuiltin<"__builtin_ia32_pmuldq128">,
-              Intrinsic<[llvm_v2i64_ty, llvm_v4i32_ty, llvm_v4i32_ty],
-                        [IntrNoMem, Commutative]>;
-  def int_x86_sse41_pmulld          : GCCBuiltin<"__builtin_ia32_pmulld128">,
-              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty, llvm_v4i32_ty],
+              Intrinsic<[llvm_v2i64_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
                         [IntrNoMem, Commutative]>;
 }
 
 // Vector extract
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_pextrb         :
-              Intrinsic<[llvm_i32_tyllvm_v16i8_ty, llvm_i32_ty],
+              Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_i32_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pextrd         :
-              Intrinsic<[llvm_i32_tyllvm_v4i32_ty, llvm_i32_ty],
+              Intrinsic<[llvm_i32_ty], [llvm_v4i32_ty, llvm_i32_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_pextrq         :
-              Intrinsic<[llvm_i64_tyllvm_v2i64_ty, llvm_i32_ty],
+              Intrinsic<[llvm_i64_ty], [llvm_v2i64_ty, llvm_i32_ty],
                         [IntrNoMem]>;
   def int_x86_sse41_extractps      : GCCBuiltin<"__builtin_ia32_extractps128">,
-              Intrinsic<[llvm_i32_tyllvm_v4f32_ty, llvm_i32_ty],
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty],
                         [IntrNoMem]>;
 }
 
 // Vector insert
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_sse41_pinsrb         : GCCBuiltin<"__builtin_ia32_vec_set_v16qi">,
-          Intrinsic<[llvm_v16i8_ty, llvm_v16i8_ty, llvm_i32_ty, llvm_i32_ty],
-                    [IntrNoMem]>;
   def int_x86_sse41_insertps       : GCCBuiltin<"__builtin_ia32_insertps128">,
-          Intrinsic<[llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i32_ty],
+          Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,llvm_i32_ty],
                     [IntrNoMem]>;
 }
 
 // Vector blend
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_pblendvb         : GCCBuiltin<"__builtin_ia32_pblendvb128">,
-        Intrinsic<[llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty],
+        Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty,llvm_v16i8_ty],
                   [IntrNoMem]>;
   def int_x86_sse41_pblendw          : GCCBuiltin<"__builtin_ia32_pblendw128">,
-        Intrinsic<[llvm_v8i16_tyllvm_v8i16_ty, llvm_v8i16_ty, llvm_i32_ty],
+        Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty, llvm_i32_ty],
                   [IntrNoMem]>;
   def int_x86_sse41_blendpd          : GCCBuiltin<"__builtin_ia32_blendpd">,
-        Intrinsic<[llvm_v2f64_tyllvm_v2f64_ty, llvm_v2f64_ty, llvm_i32_ty],
+        Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_i32_ty],
                   [IntrNoMem]>;
   def int_x86_sse41_blendps          : GCCBuiltin<"__builtin_ia32_blendps">,
-        Intrinsic<[llvm_v4f32_tyllvm_v4f32_ty, llvm_v4f32_ty, llvm_i32_ty],
+        Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_i32_ty],
                   [IntrNoMem]>;
   def int_x86_sse41_blendvpd         : GCCBuiltin<"__builtin_ia32_blendvpd">,
-        Intrinsic<[llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+        Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,llvm_v2f64_ty],
                   [IntrNoMem]>;
   def int_x86_sse41_blendvps         : GCCBuiltin<"__builtin_ia32_blendvps">,
-        Intrinsic<[llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+        Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,llvm_v4f32_ty],
                   [IntrNoMem]>;
 }
 
 // Vector dot product
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_dppd            : GCCBuiltin<"__builtin_ia32_dppd">,
-          Intrinsic<[llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, llvm_i32_ty],
+          Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,llvm_i32_ty],
                     [IntrNoMem, Commutative]>;
   def int_x86_sse41_dpps            : GCCBuiltin<"__builtin_ia32_dpps">,
-          Intrinsic<[llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i32_ty],
+          Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,llvm_i32_ty],
                     [IntrNoMem, Commutative]>;
 }
 
 // Vector sum of absolute differences
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_mpsadbw         : GCCBuiltin<"__builtin_ia32_mpsadbw128">,
-          Intrinsic<[llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty, llvm_i32_ty],
+          Intrinsic<[llvm_v8i16_ty], [llvm_v16i8_ty, llvm_v16i8_ty,llvm_i32_ty],
                     [IntrNoMem, Commutative]>;
 }
 
 // Cacheability support ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse41_movntdqa        : GCCBuiltin<"__builtin_ia32_movntdqa">,
-          Intrinsic<[llvm_v2i64_ty, llvm_ptr_ty], [IntrReadMem]>;
+          Intrinsic<[llvm_v2i64_ty], [llvm_ptr_ty], [IntrReadMem]>;
+}
+
+// Test instruction with bitwise comparison.
+let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.".
+  def int_x86_sse41_ptestz          : GCCBuiltin<"__builtin_ia32_ptestz128">,
+          Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_v4f32_ty],
+                    [IntrNoMem]>;
+  def int_x86_sse41_ptestc          : GCCBuiltin<"__builtin_ia32_ptestc128">,
+          Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_v4f32_ty],
+                    [IntrNoMem]>;
+  def int_x86_sse41_ptestnzc        : GCCBuiltin<"__builtin_ia32_ptestnzc128">,
+          Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_v4f32_ty],
+                    [IntrNoMem]>;
+}
+
+//===----------------------------------------------------------------------===//
+// SSE4.2
+
+// Miscellaneous
+// CRC Instruction
+let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.".
+  def int_x86_sse42_crc32_32_8       : GCCBuiltin<"__builtin_ia32_crc32qi">,
+          Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i8_ty],
+                    [IntrNoMem]>;
+  def int_x86_sse42_crc32_32_16      : GCCBuiltin<"__builtin_ia32_crc32hi">,
+          Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i16_ty],
+                    [IntrNoMem]>;
+  def int_x86_sse42_crc32_32_32      : GCCBuiltin<"__builtin_ia32_crc32si">,
+          Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty],
+                    [IntrNoMem]>;
+  def int_x86_sse42_crc32_64_8       :
+          Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i8_ty],
+                    [IntrNoMem]>;
+  def int_x86_sse42_crc32_64_64      : GCCBuiltin<"__builtin_ia32_crc32di">,
+          Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty],
+                    [IntrNoMem]>;
+}
+
+// String/text processing ops.
+let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.".
+  def int_x86_sse42_pcmpistrm128  : GCCBuiltin<"__builtin_ia32_pcmpistrm128">,
+    Intrinsic<[llvm_v16i8_ty],
+        [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpistri128  : GCCBuiltin<"__builtin_ia32_pcmpistri128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpistria128 : GCCBuiltin<"__builtin_ia32_pcmpistria128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpistric128 : GCCBuiltin<"__builtin_ia32_pcmpistric128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpistrio128 : GCCBuiltin<"__builtin_ia32_pcmpistrio128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpistris128 : GCCBuiltin<"__builtin_ia32_pcmpistris128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpistriz128 : GCCBuiltin<"__builtin_ia32_pcmpistriz128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpestrm128  : GCCBuiltin<"__builtin_ia32_pcmpestrm128">,
+    Intrinsic<[llvm_v16i8_ty],
+        [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty,
+         llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpestri128  : GCCBuiltin<"__builtin_ia32_pcmpestri128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty,
+         llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpestria128 : GCCBuiltin<"__builtin_ia32_pcmpestria128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty,
+         llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpestric128 : GCCBuiltin<"__builtin_ia32_pcmpestric128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty,
+         llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpestrio128 : GCCBuiltin<"__builtin_ia32_pcmpestrio128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty,
+         llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpestris128 : GCCBuiltin<"__builtin_ia32_pcmpestris128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty,
+         llvm_i8_ty],
+        [IntrNoMem]>;
+  def int_x86_sse42_pcmpestriz128 : GCCBuiltin<"__builtin_ia32_pcmpestriz128">,
+    Intrinsic<[llvm_i32_ty],
+        [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty,
+         llvm_i8_ty],
+        [IntrNoMem]>;
+}
+
+//===----------------------------------------------------------------------===//
+// AVX
+
+// Arithmetic ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_addsub_pd_256 : GCCBuiltin<"__builtin_ia32_addsubpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_addsub_ps_256 : GCCBuiltin<"__builtin_ia32_addsubps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty], [IntrNoMem]>;
+  def int_x86_avx_max_pd_256 : GCCBuiltin<"__builtin_ia32_maxpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_max_ps_256 : GCCBuiltin<"__builtin_ia32_maxps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty], [IntrNoMem]>;
+  def int_x86_avx_min_pd_256 : GCCBuiltin<"__builtin_ia32_minpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_min_ps_256 : GCCBuiltin<"__builtin_ia32_minps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty], [IntrNoMem]>;
+
+  def int_x86_avx_sqrt_pd_256 : GCCBuiltin<"__builtin_ia32_sqrtpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_sqrt_ps_256 : GCCBuiltin<"__builtin_ia32_sqrtps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty], [IntrNoMem]>;
+
+  def int_x86_avx_rsqrt_ps_256 : GCCBuiltin<"__builtin_ia32_rsqrtps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty], [IntrNoMem]>;
+
+  def int_x86_avx_rcp_ps_256 : GCCBuiltin<"__builtin_ia32_rcpps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty], [IntrNoMem]>;
+
+  def int_x86_avx_round_pd_256 : GCCBuiltin<"__builtin_ia32_roundpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx_round_ps_256 : GCCBuiltin<"__builtin_ia32_roundps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_i32_ty], [IntrNoMem]>;
+}
+
+// Horizontal ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_hadd_pd_256 : GCCBuiltin<"__builtin_ia32_haddpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_hsub_ps_256 : GCCBuiltin<"__builtin_ia32_hsubps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty], [IntrNoMem]>;
+  def int_x86_avx_hsub_pd_256 : GCCBuiltin<"__builtin_ia32_hsubpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_hadd_ps_256 : GCCBuiltin<"__builtin_ia32_haddps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty], [IntrNoMem]>;
+}
+
+// Vector permutation
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_vpermilvar_pd : GCCBuiltin<"__builtin_ia32_vpermilvarpd">,
+        Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
+                  llvm_v2i64_ty], [IntrNoMem]>;
+  def int_x86_avx_vpermilvar_ps : GCCBuiltin<"__builtin_ia32_vpermilvarps">,
+        Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
+                  llvm_v4i32_ty], [IntrNoMem]>;
+
+  def int_x86_avx_vpermilvar_pd_256 :
+        GCCBuiltin<"__builtin_ia32_vpermilvarpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_v4i64_ty], [IntrNoMem]>;
+  def int_x86_avx_vpermilvar_ps_256 :
+        GCCBuiltin<"__builtin_ia32_vpermilvarps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_v8i32_ty], [IntrNoMem]>;
+
+  def int_x86_avx_vperm2f128_pd_256 :
+        GCCBuiltin<"__builtin_ia32_vperm2f128_pd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_vperm2f128_ps_256 :
+        GCCBuiltin<"__builtin_ia32_vperm2f128_ps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_vperm2f128_si_256 :
+        GCCBuiltin<"__builtin_ia32_vperm2f128_si256">,
+        Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                  llvm_v8i32_ty, llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_avx_vpermil_pd :
+        Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_vpermil_ps :
+        Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_avx_vpermil_pd_256 :
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_vpermil_ps_256 :
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+}
+
+// Vector blend
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_blend_pd_256 : GCCBuiltin<"__builtin_ia32_blendpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx_blend_ps_256 : GCCBuiltin<"__builtin_ia32_blendps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx_blendv_pd_256 : GCCBuiltin<"__builtin_ia32_blendvpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty, llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_blendv_ps_256 : GCCBuiltin<"__builtin_ia32_blendvps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty, llvm_v8f32_ty], [IntrNoMem]>;
+}
+
+// Vector dot product
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_dp_ps_256 : GCCBuiltin<"__builtin_ia32_dpps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty, llvm_i32_ty], [IntrNoMem]>;
+}
+
+// Vector compare
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_cmp_pd_256 : GCCBuiltin<"__builtin_ia32_cmppd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_cmp_ps_256 : GCCBuiltin<"__builtin_ia32_cmpps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>;
+}
+
+// Vector extract and insert
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_vextractf128_pd_256 :
+        GCCBuiltin<"__builtin_ia32_vextractf128_pd256">,
+        Intrinsic<[llvm_v2f64_ty], [llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_vextractf128_ps_256 :
+        GCCBuiltin<"__builtin_ia32_vextractf128_ps256">,
+        Intrinsic<[llvm_v4f32_ty], [llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_vextractf128_si_256 :
+        GCCBuiltin<"__builtin_ia32_vextractf128_si256">,
+        Intrinsic<[llvm_v4i32_ty], [llvm_v8i32_ty, llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_avx_vinsertf128_pd_256 :
+        GCCBuiltin<"__builtin_ia32_vinsertf128_pd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
+                  llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_vinsertf128_ps_256 :
+        GCCBuiltin<"__builtin_ia32_vinsertf128_ps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
+                  llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx_vinsertf128_si_256 :
+        GCCBuiltin<"__builtin_ia32_vinsertf128_si256">,
+        Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                  llvm_v4i32_ty, llvm_i8_ty], [IntrNoMem]>;
 }
 
+// Vector convert
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_cvtdq2_pd_256 : GCCBuiltin<"__builtin_ia32_cvtdq2pd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4i32_ty], [IntrNoMem]>;
+  def int_x86_avx_cvtdq2_ps_256 : GCCBuiltin<"__builtin_ia32_cvtdq2ps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_v8i32_ty], [IntrNoMem]>;
+  def int_x86_avx_cvt_pd2_ps_256 : GCCBuiltin<"__builtin_ia32_cvtpd2ps256">,
+        Intrinsic<[llvm_v4f32_ty], [llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_cvt_ps2dq_256 : GCCBuiltin<"__builtin_ia32_cvtps2dq256">,
+        Intrinsic<[llvm_v8i32_ty], [llvm_v8f32_ty], [IntrNoMem]>;
+  def int_x86_avx_cvt_ps2_pd_256 : GCCBuiltin<"__builtin_ia32_cvtps2pd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_v4f32_ty], [IntrNoMem]>;
+  def int_x86_avx_cvtt_pd2dq_256 : GCCBuiltin<"__builtin_ia32_cvttpd2dq256">,
+        Intrinsic<[llvm_v4i32_ty], [llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_cvt_pd2dq_256 : GCCBuiltin<"__builtin_ia32_cvtpd2dq256">,
+        Intrinsic<[llvm_v4i32_ty], [llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_cvtt_ps2dq_256 : GCCBuiltin<"__builtin_ia32_cvttps2dq256">,
+        Intrinsic<[llvm_v8i32_ty], [llvm_v8f32_ty], [IntrNoMem]>;
+}
+
+// Vector bit test
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_vtestz_pd : GCCBuiltin<"__builtin_ia32_vtestzpd">,
+        Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
+                  llvm_v2f64_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestc_pd : GCCBuiltin<"__builtin_ia32_vtestcpd">,
+        Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
+                  llvm_v2f64_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestnzc_pd : GCCBuiltin<"__builtin_ia32_vtestnzcpd">,
+        Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
+                  llvm_v2f64_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestz_ps : GCCBuiltin<"__builtin_ia32_vtestzps">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
+                  llvm_v4f32_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestc_ps : GCCBuiltin<"__builtin_ia32_vtestcps">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
+                  llvm_v4f32_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestnzc_ps : GCCBuiltin<"__builtin_ia32_vtestnzcps">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
+                  llvm_v4f32_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestz_pd_256 : GCCBuiltin<"__builtin_ia32_vtestzpd256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestc_pd_256 : GCCBuiltin<"__builtin_ia32_vtestcpd256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestnzc_pd_256 : GCCBuiltin<"__builtin_ia32_vtestnzcpd256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4f64_ty,
+                  llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestz_ps_256 : GCCBuiltin<"__builtin_ia32_vtestzps256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestc_ps_256 : GCCBuiltin<"__builtin_ia32_vtestcps256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty], [IntrNoMem]>;
+  def int_x86_avx_vtestnzc_ps_256 : GCCBuiltin<"__builtin_ia32_vtestnzcps256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v8f32_ty,
+                  llvm_v8f32_ty], [IntrNoMem]>;
+  def int_x86_avx_ptestz_256 : GCCBuiltin<"__builtin_ia32_ptestz256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4i64_ty,
+                  llvm_v4i64_ty], [IntrNoMem]>;
+  def int_x86_avx_ptestc_256 : GCCBuiltin<"__builtin_ia32_ptestc256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4i64_ty,
+                  llvm_v4i64_ty], [IntrNoMem]>;
+  def int_x86_avx_ptestnzc_256 : GCCBuiltin<"__builtin_ia32_ptestnzc256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4i64_ty,
+                  llvm_v4i64_ty], [IntrNoMem]>;
+}
+
+// Vector extract sign mask
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_movmsk_pd_256 : GCCBuiltin<"__builtin_ia32_movmskpd256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_avx_movmsk_ps_256 : GCCBuiltin<"__builtin_ia32_movmskps256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v8f32_ty], [IntrNoMem]>;
+}
+
+// Vector zero
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_vzeroall : GCCBuiltin<"__builtin_ia32_vzeroall">,
+        Intrinsic<[], [], []>;
+  def int_x86_avx_vzeroupper : GCCBuiltin<"__builtin_ia32_vzeroupper">,
+        Intrinsic<[], [], []>;
+}
+
+// Vector load with broadcast
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_vbroadcast_ss :
+        GCCBuiltin<"__builtin_ia32_vbroadcastss">,
+        Intrinsic<[llvm_v4f32_ty], [llvm_ptr_ty], [IntrReadMem]>;
+  def int_x86_avx_vbroadcast_sd_256 :
+        GCCBuiltin<"__builtin_ia32_vbroadcastsd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_ptr_ty], [IntrReadMem]>;
+  def int_x86_avx_vbroadcast_ss_256 :
+        GCCBuiltin<"__builtin_ia32_vbroadcastss256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_ptr_ty], [IntrReadMem]>;
+  def int_x86_avx_vbroadcastf128_pd_256 :
+        GCCBuiltin<"__builtin_ia32_vbroadcastf128_pd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_ptr_ty], [IntrReadMem]>;
+  def int_x86_avx_vbroadcastf128_ps_256 :
+        GCCBuiltin<"__builtin_ia32_vbroadcastf128_ps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_ptr_ty], [IntrReadMem]>;
+}
+
+// SIMD load ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_ldu_dq_256 : GCCBuiltin<"__builtin_ia32_lddqu256">,
+        Intrinsic<[llvm_v32i8_ty], [llvm_ptr_ty], [IntrReadMem]>;
+}
+
+// SIMD store ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_storeu_pd_256 : GCCBuiltin<"__builtin_ia32_storeupd256">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v4f64_ty], []>;
+  def int_x86_avx_storeu_ps_256 : GCCBuiltin<"__builtin_ia32_storeups256">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v8f32_ty], []>;
+  def int_x86_avx_storeu_dq_256 : GCCBuiltin<"__builtin_ia32_storedqu256">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v32i8_ty], []>;
+}
+
+// Cacheability support ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_movnt_dq_256 : GCCBuiltin<"__builtin_ia32_movntdq256">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v4i64_ty], []>;
+  def int_x86_avx_movnt_pd_256 : GCCBuiltin<"__builtin_ia32_movntpd256">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v4f64_ty], []>;
+  def int_x86_avx_movnt_ps_256 : GCCBuiltin<"__builtin_ia32_movntps256">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v8f32_ty], []>;
+}
+
+// Conditional load ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_maskload_pd : GCCBuiltin<"__builtin_ia32_maskloadpd">,
+        Intrinsic<[llvm_v2f64_ty], [llvm_ptr_ty, llvm_v2f64_ty], [IntrReadMem]>;
+  def int_x86_avx_maskload_ps : GCCBuiltin<"__builtin_ia32_maskloadps">,
+        Intrinsic<[llvm_v4f32_ty], [llvm_ptr_ty, llvm_v4f32_ty], [IntrReadMem]>;
+  def int_x86_avx_maskload_pd_256 : GCCBuiltin<"__builtin_ia32_maskloadpd256">,
+        Intrinsic<[llvm_v4f64_ty], [llvm_ptr_ty, llvm_v4f64_ty], [IntrReadMem]>;
+  def int_x86_avx_maskload_ps_256 : GCCBuiltin<"__builtin_ia32_maskloadps256">,
+        Intrinsic<[llvm_v8f32_ty], [llvm_ptr_ty, llvm_v8f32_ty], [IntrReadMem]>;
+}
+
+// Conditional store ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx_maskstore_pd : GCCBuiltin<"__builtin_ia32_maskstorepd">,
+        Intrinsic<[], [llvm_ptr_ty,
+                  llvm_v2f64_ty, llvm_v2f64_ty], []>;
+  def int_x86_avx_maskstore_ps : GCCBuiltin<"__builtin_ia32_maskstoreps">,
+        Intrinsic<[], [llvm_ptr_ty,
+                  llvm_v4f32_ty, llvm_v4f32_ty], []>;
+  def int_x86_avx_maskstore_pd_256 :
+        GCCBuiltin<"__builtin_ia32_maskstorepd256">,
+        Intrinsic<[], [llvm_ptr_ty,
+                  llvm_v4f64_ty, llvm_v4f64_ty], []>;
+  def int_x86_avx_maskstore_ps_256 :
+        GCCBuiltin<"__builtin_ia32_maskstoreps256">,
+        Intrinsic<[], [llvm_ptr_ty,
+                  llvm_v8f32_ty, llvm_v8f32_ty], []>;
+}
+
+//===----------------------------------------------------------------------===//
+// AVX2
+
+// Integer arithmetic ops.
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_padds_b : GCCBuiltin<"__builtin_ia32_paddsb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_padds_w : GCCBuiltin<"__builtin_ia32_paddsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_paddus_b : GCCBuiltin<"__builtin_ia32_paddusb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_paddus_w : GCCBuiltin<"__builtin_ia32_paddusw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_psubs_b : GCCBuiltin<"__builtin_ia32_psubsb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_psubs_w : GCCBuiltin<"__builtin_ia32_psubsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_psubus_b : GCCBuiltin<"__builtin_ia32_psubusb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_psubus_w : GCCBuiltin<"__builtin_ia32_psubusw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_pmulhu_w : GCCBuiltin<"__builtin_ia32_pmulhuw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmulh_w : GCCBuiltin<"__builtin_ia32_pmulhw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmulu_dq : GCCBuiltin<"__builtin_ia32_pmuludq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmul_dq : GCCBuiltin<"__builtin_ia32_pmuldq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmadd_wd : GCCBuiltin<"__builtin_ia32_pmaddwd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pavg_b : GCCBuiltin<"__builtin_ia32_pavgb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pavg_w : GCCBuiltin<"__builtin_ia32_pavgw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_psad_bw : GCCBuiltin<"__builtin_ia32_psadbw256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
+}
+
+// Vector min, max
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_pmaxu_b : GCCBuiltin<"__builtin_ia32_pmaxub256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmaxu_w : GCCBuiltin<"__builtin_ia32_pmaxuw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmaxu_d : GCCBuiltin<"__builtin_ia32_pmaxud256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmaxs_b : GCCBuiltin<"__builtin_ia32_pmaxsb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmaxs_w : GCCBuiltin<"__builtin_ia32_pmaxsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmaxs_d : GCCBuiltin<"__builtin_ia32_pmaxsd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pminu_b : GCCBuiltin<"__builtin_ia32_pminub256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pminu_w : GCCBuiltin<"__builtin_ia32_pminuw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pminu_d : GCCBuiltin<"__builtin_ia32_pminud256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmins_b : GCCBuiltin<"__builtin_ia32_pminsb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmins_w : GCCBuiltin<"__builtin_ia32_pminsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_pmins_d : GCCBuiltin<"__builtin_ia32_pminsd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem, Commutative]>;
+}
+
+// Integer shift ops.
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_psll_w : GCCBuiltin<"__builtin_ia32_psllw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_psll_d : GCCBuiltin<"__builtin_ia32_pslld256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v4i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psll_q : GCCBuiltin<"__builtin_ia32_psllq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_v2i64_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrl_w : GCCBuiltin<"__builtin_ia32_psrlw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrl_d : GCCBuiltin<"__builtin_ia32_psrld256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v4i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrl_q : GCCBuiltin<"__builtin_ia32_psrlq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_v2i64_ty], [IntrNoMem]>;
+  def int_x86_avx2_psra_w : GCCBuiltin<"__builtin_ia32_psraw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_psra_d : GCCBuiltin<"__builtin_ia32_psrad256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v4i32_ty], [IntrNoMem]>;
+
+  def int_x86_avx2_pslli_w : GCCBuiltin<"__builtin_ia32_psllwi256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_pslli_d : GCCBuiltin<"__builtin_ia32_pslldi256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_pslli_q : GCCBuiltin<"__builtin_ia32_psllqi256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrli_w : GCCBuiltin<"__builtin_ia32_psrlwi256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrli_d : GCCBuiltin<"__builtin_ia32_psrldi256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrli_q : GCCBuiltin<"__builtin_ia32_psrlqi256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrai_w : GCCBuiltin<"__builtin_ia32_psrawi256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrai_d : GCCBuiltin<"__builtin_ia32_psradi256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+
+  def int_x86_avx2_psll_dq : GCCBuiltin<"__builtin_ia32_pslldqi256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrl_dq : GCCBuiltin<"__builtin_ia32_psrldqi256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psll_dq_bs : GCCBuiltin<"__builtin_ia32_pslldqi256_byteshift">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_psrl_dq_bs : GCCBuiltin<"__builtin_ia32_psrldqi256_byteshift">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+}
+
+// Pack ops.
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_packsswb : GCCBuiltin<"__builtin_ia32_packsswb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_packssdw : GCCBuiltin<"__builtin_ia32_packssdw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_packuswb : GCCBuiltin<"__builtin_ia32_packuswb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_packusdw : GCCBuiltin<"__builtin_ia32_packusdw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem]>;
+}
+
+// Absolute value ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_pabs_b : GCCBuiltin<"__builtin_ia32_pabsb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_pabs_w : GCCBuiltin<"__builtin_ia32_pabsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_pabs_d : GCCBuiltin<"__builtin_ia32_pabsd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty], [IntrNoMem]>;
+}
+
+// Horizontal arithmetic ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_phadd_w : GCCBuiltin<"__builtin_ia32_phaddw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_phadd_d : GCCBuiltin<"__builtin_ia32_phaddd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_phadd_sw : GCCBuiltin<"__builtin_ia32_phaddsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_phsub_w : GCCBuiltin<"__builtin_ia32_phsubw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_phsub_d : GCCBuiltin<"__builtin_ia32_phsubd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_phsub_sw : GCCBuiltin<"__builtin_ia32_phsubsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_pmadd_ub_sw : GCCBuiltin<"__builtin_ia32_pmaddubsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem]>;
+}
+
+// Sign ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_psign_b : GCCBuiltin<"__builtin_ia32_psignb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_psign_w : GCCBuiltin<"__builtin_ia32_psignw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_psign_d : GCCBuiltin<"__builtin_ia32_psignd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty,
+                         llvm_v8i32_ty], [IntrNoMem]>;
+}
+
+// Packed multiply high with round and scale
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_pmul_hr_sw : GCCBuiltin<"__builtin_ia32_pmulhrsw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
+                         llvm_v16i16_ty], [IntrNoMem, Commutative]>;
+}
+
+// Vector sign and zero extend
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_pmovsxbd : GCCBuiltin<"__builtin_ia32_pmovsxbd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovsxbq : GCCBuiltin<"__builtin_ia32_pmovsxbq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovsxbw : GCCBuiltin<"__builtin_ia32_pmovsxbw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovsxdq : GCCBuiltin<"__builtin_ia32_pmovsxdq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovsxwd : GCCBuiltin<"__builtin_ia32_pmovsxwd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovsxwq : GCCBuiltin<"__builtin_ia32_pmovsxwq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovzxbd : GCCBuiltin<"__builtin_ia32_pmovzxbd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovzxbq : GCCBuiltin<"__builtin_ia32_pmovzxbq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovzxbw : GCCBuiltin<"__builtin_ia32_pmovzxbw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovzxdq : GCCBuiltin<"__builtin_ia32_pmovzxdq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovzxwd : GCCBuiltin<"__builtin_ia32_pmovzxwd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_pmovzxwq : GCCBuiltin<"__builtin_ia32_pmovzxwq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v8i16_ty],
+                        [IntrNoMem]>;
+}
+
+// Vector blend
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_pblendvb : GCCBuiltin<"__builtin_ia32_pblendvb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty, llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_pblendw : GCCBuiltin<"__builtin_ia32_pblendw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty, llvm_v16i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_pblendd_128 : GCCBuiltin<"__builtin_ia32_pblendd128">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_pblendd_256 : GCCBuiltin<"__builtin_ia32_pblendd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_v8i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+}
+
+// Vector load with broadcast
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_vbroadcast_ss_ps :
+              GCCBuiltin<"__builtin_ia32_vbroadcastss_ps">,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
+  def int_x86_avx2_vbroadcast_sd_pd_256 :
+              GCCBuiltin<"__builtin_ia32_vbroadcastsd_pd256">,
+              Intrinsic<[llvm_v4f64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
+  def int_x86_avx2_vbroadcast_ss_ps_256 :
+              GCCBuiltin<"__builtin_ia32_vbroadcastss_ps256">,
+              Intrinsic<[llvm_v8f32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
+  def int_x86_avx2_vbroadcasti128 :
+              GCCBuiltin<"__builtin_ia32_vbroadcastsi256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_ptr_ty], [IntrReadMem]>;
+  def int_x86_avx2_pbroadcastb_128 :
+              GCCBuiltin<"__builtin_ia32_pbroadcastb128">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_pbroadcastb_256 :
+              GCCBuiltin<"__builtin_ia32_pbroadcastb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_pbroadcastw_128 :
+              GCCBuiltin<"__builtin_ia32_pbroadcastw128">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_pbroadcastw_256 :
+              GCCBuiltin<"__builtin_ia32_pbroadcastw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_avx2_pbroadcastd_128 :
+              GCCBuiltin<"__builtin_ia32_pbroadcastd128">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_pbroadcastd_256 :
+              GCCBuiltin<"__builtin_ia32_pbroadcastd256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v4i32_ty], [IntrNoMem]>;
+  def int_x86_avx2_pbroadcastq_128 :
+              GCCBuiltin<"__builtin_ia32_pbroadcastq128">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty], [IntrNoMem]>;
+  def int_x86_avx2_pbroadcastq_256 :
+              GCCBuiltin<"__builtin_ia32_pbroadcastq256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v2i64_ty], [IntrNoMem]>;
+}
+
+// Vector permutation
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_permd : GCCBuiltin<"__builtin_ia32_permvarsi256">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_v8i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_permq : GCCBuiltin<"__builtin_ia32_permdi256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_permps : GCCBuiltin<"__builtin_ia32_permvarsf256">,
+              Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_v8f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_permpd : GCCBuiltin<"__builtin_ia32_permdf256">,
+              Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_vperm2i128 : GCCBuiltin<"__builtin_ia32_permti256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_v4i64_ty, llvm_i8_ty], [IntrNoMem]>;
+}
+
+// Vector extract and insert
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_vextracti128 : GCCBuiltin<"__builtin_ia32_extract128i256">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v4i64_ty,
+                         llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_vinserti128 : GCCBuiltin<"__builtin_ia32_insert128i256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
+                         llvm_v2i64_ty, llvm_i8_ty], [IntrNoMem]>;
+}
+
+// Conditional load ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_maskload_d : GCCBuiltin<"__builtin_ia32_maskloadd">,
+        Intrinsic<[llvm_v4i32_ty], [llvm_ptr_ty, llvm_v4i32_ty], [IntrReadMem]>;
+  def int_x86_avx2_maskload_q : GCCBuiltin<"__builtin_ia32_maskloadq">,
+        Intrinsic<[llvm_v2i64_ty], [llvm_ptr_ty, llvm_v2i64_ty], [IntrReadMem]>;
+  def int_x86_avx2_maskload_d_256 : GCCBuiltin<"__builtin_ia32_maskloadd256">,
+        Intrinsic<[llvm_v8i32_ty], [llvm_ptr_ty, llvm_v8i32_ty], [IntrReadMem]>;
+  def int_x86_avx2_maskload_q_256 : GCCBuiltin<"__builtin_ia32_maskloadq256">,
+        Intrinsic<[llvm_v4i64_ty], [llvm_ptr_ty, llvm_v4i64_ty], [IntrReadMem]>;
+}
+
+// Conditional store ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_maskstore_d : GCCBuiltin<"__builtin_ia32_maskstored">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v4i32_ty, llvm_v4i32_ty], []>;
+  def int_x86_avx2_maskstore_q : GCCBuiltin<"__builtin_ia32_maskstoreq">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v2i64_ty, llvm_v2i64_ty], []>;
+  def int_x86_avx2_maskstore_d_256 :
+        GCCBuiltin<"__builtin_ia32_maskstored256">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v8i32_ty, llvm_v8i32_ty], []>;
+  def int_x86_avx2_maskstore_q_256 :
+        GCCBuiltin<"__builtin_ia32_maskstoreq256">,
+        Intrinsic<[], [llvm_ptr_ty, llvm_v4i64_ty, llvm_v4i64_ty], []>;
+}
+
+// Variable bit shift ops
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_psllv_d : GCCBuiltin<"__builtin_ia32_psllv4si">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_psllv_d_256 : GCCBuiltin<"__builtin_ia32_psllv8si">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_v8i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_psllv_q : GCCBuiltin<"__builtin_ia32_psllv2di">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_psllv_q_256 : GCCBuiltin<"__builtin_ia32_psllv4di">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_v4i64_ty],
+                        [IntrNoMem]>;
+
+  def int_x86_avx2_psrlv_d : GCCBuiltin<"__builtin_ia32_psrlv4si">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_psrlv_d_256 : GCCBuiltin<"__builtin_ia32_psrlv8si">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_v8i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_psrlv_q : GCCBuiltin<"__builtin_ia32_psrlv2di">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_psrlv_q_256 : GCCBuiltin<"__builtin_ia32_psrlv4di">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_v4i64_ty],
+                        [IntrNoMem]>;
+
+  def int_x86_avx2_psrav_d : GCCBuiltin<"__builtin_ia32_psrav4si">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_avx2_psrav_d_256 : GCCBuiltin<"__builtin_ia32_psrav8si">,
+              Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_v8i32_ty],
+                        [IntrNoMem]>;
+}
+
+// Misc.
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx2_pmovmskb : GCCBuiltin<"__builtin_ia32_pmovmskb256">,
+              Intrinsic<[llvm_i32_ty], [llvm_v32i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_pshuf_b : GCCBuiltin<"__builtin_ia32_pshufb256">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                         llvm_v32i8_ty], [IntrNoMem]>;
+  def int_x86_avx2_mpsadbw : GCCBuiltin<"__builtin_ia32_mpsadbw256">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v32i8_ty, llvm_v32i8_ty,
+                         llvm_i32_ty], [IntrNoMem, Commutative]>;
+  def int_x86_avx2_movntdqa : GCCBuiltin<"__builtin_ia32_movntdqa256">,
+              Intrinsic<[llvm_v4i64_ty], [llvm_ptr_ty], [IntrReadMem]>;
+}
+
+//===----------------------------------------------------------------------===//
+// FMA4
+
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_fma4_vfmadd_ss : GCCBuiltin<"__builtin_ia32_vfmaddss">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmadd_sd : GCCBuiltin<"__builtin_ia32_vfmaddsd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmadd_ps : GCCBuiltin<"__builtin_ia32_vfmaddps">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmadd_pd : GCCBuiltin<"__builtin_ia32_vfmaddpd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmadd_ps_256 : GCCBuiltin<"__builtin_ia32_vfmaddps256">,
+              Intrinsic<[llvm_v8f32_ty],
+                        [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmadd_pd_256 : GCCBuiltin<"__builtin_ia32_vfmaddpd256">,
+              Intrinsic<[llvm_v4f64_ty],
+                        [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsub_ss : GCCBuiltin<"__builtin_ia32_vfmsubss">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsub_sd : GCCBuiltin<"__builtin_ia32_vfmsubsd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsub_ps : GCCBuiltin<"__builtin_ia32_vfmsubps">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsub_pd : GCCBuiltin<"__builtin_ia32_vfmsubpd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsub_ps_256 : GCCBuiltin<"__builtin_ia32_vfmsubps256">,
+              Intrinsic<[llvm_v8f32_ty],
+                        [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsub_pd_256 : GCCBuiltin<"__builtin_ia32_vfmsubpd256">,
+              Intrinsic<[llvm_v4f64_ty],
+                        [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmadd_ss : GCCBuiltin<"__builtin_ia32_vfnmaddss">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmadd_sd : GCCBuiltin<"__builtin_ia32_vfnmaddsd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmadd_ps : GCCBuiltin<"__builtin_ia32_vfnmaddps">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmadd_pd : GCCBuiltin<"__builtin_ia32_vfnmaddpd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmadd_ps_256 : GCCBuiltin<"__builtin_ia32_vfnmaddps256">,
+              Intrinsic<[llvm_v8f32_ty],
+                        [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmadd_pd_256 : GCCBuiltin<"__builtin_ia32_vfnmaddpd256">,
+              Intrinsic<[llvm_v4f64_ty],
+                        [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmsub_ss : GCCBuiltin<"__builtin_ia32_vfnmsubss">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmsub_sd : GCCBuiltin<"__builtin_ia32_vfnmsubsd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmsub_ps : GCCBuiltin<"__builtin_ia32_vfnmsubps">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmsub_pd : GCCBuiltin<"__builtin_ia32_vfnmsubpd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmsub_ps_256 : GCCBuiltin<"__builtin_ia32_vfnmsubps256">,
+              Intrinsic<[llvm_v8f32_ty],
+                        [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfnmsub_pd_256 : GCCBuiltin<"__builtin_ia32_vfnmsubpd256">,
+              Intrinsic<[llvm_v4f64_ty],
+                        [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmaddsub_ps : GCCBuiltin<"__builtin_ia32_vfmaddsubps">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmaddsub_pd : GCCBuiltin<"__builtin_ia32_vfmaddsubpd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmaddsub_ps_256 :
+               GCCBuiltin<"__builtin_ia32_vfmaddsubps256">,
+              Intrinsic<[llvm_v8f32_ty],
+                        [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmaddsub_pd_256 :
+              GCCBuiltin<"__builtin_ia32_vfmaddsubpd256">,
+              Intrinsic<[llvm_v4f64_ty],
+                        [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsubadd_ps : GCCBuiltin<"__builtin_ia32_vfmsubaddps">,
+              Intrinsic<[llvm_v4f32_ty],
+                        [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsubadd_pd : GCCBuiltin<"__builtin_ia32_vfmsubaddpd">,
+              Intrinsic<[llvm_v2f64_ty],
+                        [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsubadd_ps_256 :
+              GCCBuiltin<"__builtin_ia32_vfmsubaddps256">,
+              Intrinsic<[llvm_v8f32_ty],
+                        [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_fma4_vfmsubadd_pd_256 :
+              GCCBuiltin<"__builtin_ia32_vfmsubaddpd256">,
+              Intrinsic<[llvm_v4f64_ty],
+                        [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4f64_ty],
+                        [IntrNoMem]>;
+}
+
+//===----------------------------------------------------------------------===//
+// XOP
+
+  def int_x86_xop_vpermil2pd : GCCBuiltin<"__builtin_ia32_vpermil2pd">,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
+                                          llvm_v2f64_ty, llvm_i8_ty],
+                        [IntrNoMem]>;
+
+  def int_x86_xop_vpermil2pd_256 :
+              GCCBuiltin<"__builtin_ia32_vpermil2pd256">,
+              Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_v4f64_ty,
+                                          llvm_v4f64_ty, llvm_i8_ty],
+                        [IntrNoMem]>;
+
+  def int_x86_xop_vpermil2ps : GCCBuiltin<"__builtin_ia32_vpermil2ps">,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
+                                          llvm_v4f32_ty, llvm_i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpermil2ps_256 :
+              GCCBuiltin<"__builtin_ia32_vpermil2ps256">,
+              Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_v8f32_ty,
+                                          llvm_v8f32_ty, llvm_i8_ty],
+                        [IntrNoMem]>;
+
+  def int_x86_xop_vfrcz_pd :
+              GCCBuiltin<"__builtin_ia32_vfrczpd">,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty], [IntrNoMem]>;
+  def int_x86_xop_vfrcz_ps :
+              GCCBuiltin<"__builtin_ia32_vfrczps">,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
+  def int_x86_xop_vfrcz_sd :
+              GCCBuiltin<"__builtin_ia32_vfrczsd">,
+              Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vfrcz_ss :
+              GCCBuiltin<"__builtin_ia32_vfrczss">,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vfrcz_pd_256 :
+              GCCBuiltin<"__builtin_ia32_vfrczpd256">,
+              Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty], [IntrNoMem]>;
+  def int_x86_xop_vfrcz_ps_256 :
+              GCCBuiltin<"__builtin_ia32_vfrczps256">,
+              Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty], [IntrNoMem]>;
+  def int_x86_xop_vpcmov :
+              GCCBuiltin<"__builtin_ia32_vpcmov">,
+              Intrinsic<[llvm_v2i64_ty],
+                        [llvm_v2i64_ty, llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcmov_256 :
+              GCCBuiltin<"__builtin_ia32_vpcmov_256">,
+              Intrinsic<[llvm_v4i64_ty],
+                        [llvm_v4i64_ty, llvm_v4i64_ty, llvm_v4i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomeqb :
+              GCCBuiltin<"__builtin_ia32_vpcomeqb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomeqw :
+              GCCBuiltin<"__builtin_ia32_vpcomeqw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomeqd :
+              GCCBuiltin<"__builtin_ia32_vpcomeqd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomeqq :
+              GCCBuiltin<"__builtin_ia32_vpcomeqq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomequb :
+              GCCBuiltin<"__builtin_ia32_vpcomequb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomequd :
+              GCCBuiltin<"__builtin_ia32_vpcomequd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomequq :
+              GCCBuiltin<"__builtin_ia32_vpcomequq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomequw :
+              GCCBuiltin<"__builtin_ia32_vpcomequw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomfalseb :
+              GCCBuiltin<"__builtin_ia32_vpcomfalseb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomfalsed :
+              GCCBuiltin<"__builtin_ia32_vpcomfalsed">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomfalseq :
+              GCCBuiltin<"__builtin_ia32_vpcomfalseq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomfalseub :
+              GCCBuiltin<"__builtin_ia32_vpcomfalseub">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomfalseud :
+              GCCBuiltin<"__builtin_ia32_vpcomfalseud">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomfalseuq :
+              GCCBuiltin<"__builtin_ia32_vpcomfalseuq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomfalseuw :
+              GCCBuiltin<"__builtin_ia32_vpcomfalseuw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomfalsew :
+              GCCBuiltin<"__builtin_ia32_vpcomfalsew">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgeb :
+              GCCBuiltin<"__builtin_ia32_vpcomgeb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomged :
+              GCCBuiltin<"__builtin_ia32_vpcomged">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgeq :
+              GCCBuiltin<"__builtin_ia32_vpcomgeq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgeub :
+              GCCBuiltin<"__builtin_ia32_vpcomgeub">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgeud :
+              GCCBuiltin<"__builtin_ia32_vpcomgeud">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgeuq :
+              GCCBuiltin<"__builtin_ia32_vpcomgeuq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgeuw :
+              GCCBuiltin<"__builtin_ia32_vpcomgeuw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgew :
+              GCCBuiltin<"__builtin_ia32_vpcomgew">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgtb :
+              GCCBuiltin<"__builtin_ia32_vpcomgtb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgtd :
+              GCCBuiltin<"__builtin_ia32_vpcomgtd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgtq :
+              GCCBuiltin<"__builtin_ia32_vpcomgtq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgtub :
+              GCCBuiltin<"__builtin_ia32_vpcomgtub">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgtud :
+              GCCBuiltin<"__builtin_ia32_vpcomgtud">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgtuq :
+              GCCBuiltin<"__builtin_ia32_vpcomgtuq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgtuw :
+              GCCBuiltin<"__builtin_ia32_vpcomgtuw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomgtw :
+              GCCBuiltin<"__builtin_ia32_vpcomgtw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomleb :
+              GCCBuiltin<"__builtin_ia32_vpcomleb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomled :
+              GCCBuiltin<"__builtin_ia32_vpcomled">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomleq :
+              GCCBuiltin<"__builtin_ia32_vpcomleq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomleub :
+              GCCBuiltin<"__builtin_ia32_vpcomleub">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomleud :
+              GCCBuiltin<"__builtin_ia32_vpcomleud">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomleuq :
+              GCCBuiltin<"__builtin_ia32_vpcomleuq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomleuw :
+              GCCBuiltin<"__builtin_ia32_vpcomleuw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomlew :
+              GCCBuiltin<"__builtin_ia32_vpcomlew">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomltb :
+              GCCBuiltin<"__builtin_ia32_vpcomltb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomltd :
+              GCCBuiltin<"__builtin_ia32_vpcomltd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomltq :
+              GCCBuiltin<"__builtin_ia32_vpcomltq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomltub :
+              GCCBuiltin<"__builtin_ia32_vpcomltub">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomltud :
+              GCCBuiltin<"__builtin_ia32_vpcomltud">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomltuq :
+              GCCBuiltin<"__builtin_ia32_vpcomltuq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomltuw :
+              GCCBuiltin<"__builtin_ia32_vpcomltuw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomltw :
+              GCCBuiltin<"__builtin_ia32_vpcomltw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomneb :
+              GCCBuiltin<"__builtin_ia32_vpcomneb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomned :
+              GCCBuiltin<"__builtin_ia32_vpcomned">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomneq :
+              GCCBuiltin<"__builtin_ia32_vpcomneq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomneub :
+              GCCBuiltin<"__builtin_ia32_vpcomneub">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomneud :
+              GCCBuiltin<"__builtin_ia32_vpcomneud">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomneuq :
+              GCCBuiltin<"__builtin_ia32_vpcomneuq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomneuw :
+              GCCBuiltin<"__builtin_ia32_vpcomneuw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomnew :
+              GCCBuiltin<"__builtin_ia32_vpcomnew">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomtrueb :
+              GCCBuiltin<"__builtin_ia32_vpcomtrueb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomtrued :
+              GCCBuiltin<"__builtin_ia32_vpcomtrued">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomtrueq :
+              GCCBuiltin<"__builtin_ia32_vpcomtrueq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomtrueub :
+              GCCBuiltin<"__builtin_ia32_vpcomtrueub">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomtrueud :
+              GCCBuiltin<"__builtin_ia32_vpcomtrueud">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomtrueuq :
+              GCCBuiltin<"__builtin_ia32_vpcomtrueuq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomtrueuw :
+              GCCBuiltin<"__builtin_ia32_vpcomtrueuw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpcomtruew :
+              GCCBuiltin<"__builtin_ia32_vpcomtruew">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vphaddbd :
+              GCCBuiltin<"__builtin_ia32_vphaddbd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_xop_vphaddbq :
+              GCCBuiltin<"__builtin_ia32_vphaddbq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_xop_vphaddbw :
+              GCCBuiltin<"__builtin_ia32_vphaddbw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_xop_vphadddq :
+              GCCBuiltin<"__builtin_ia32_vphadddq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v4i32_ty], [IntrNoMem]>;
+  def int_x86_xop_vphaddubd :
+              GCCBuiltin<"__builtin_ia32_vphaddubd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_xop_vphaddubq :
+              GCCBuiltin<"__builtin_ia32_vphaddubq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_xop_vphaddubw :
+              GCCBuiltin<"__builtin_ia32_vphaddubw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_xop_vphaddudq :
+              GCCBuiltin<"__builtin_ia32_vphaddudq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v4i32_ty], [IntrNoMem]>;
+  def int_x86_xop_vphadduwd :
+              GCCBuiltin<"__builtin_ia32_vphadduwd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_xop_vphadduwq :
+              GCCBuiltin<"__builtin_ia32_vphadduwq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_xop_vphaddwd :
+              GCCBuiltin<"__builtin_ia32_vphaddwd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_xop_vphaddwq :
+              GCCBuiltin<"__builtin_ia32_vphaddwq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_xop_vphsubbw :
+              GCCBuiltin<"__builtin_ia32_vphsubbw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v16i8_ty], [IntrNoMem]>;
+  def int_x86_xop_vphsubdq :
+              GCCBuiltin<"__builtin_ia32_vphsubdq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v4i32_ty], [IntrNoMem]>;
+  def int_x86_xop_vphsubwd :
+              GCCBuiltin<"__builtin_ia32_vphsubwd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_xop_vpmacsdd :
+              GCCBuiltin<"__builtin_ia32_vpmacsdd">,
+              Intrinsic<[llvm_v4i32_ty],
+                        [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacsdqh :
+              GCCBuiltin<"__builtin_ia32_vpmacsdqh">,
+              Intrinsic<[llvm_v2i64_ty],
+                        [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacsdql :
+              GCCBuiltin<"__builtin_ia32_vpmacsdql">,
+              Intrinsic<[llvm_v2i64_ty],
+                        [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacssdd :
+              GCCBuiltin<"__builtin_ia32_vpmacssdd">,
+              Intrinsic<[llvm_v4i32_ty],
+                        [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacssdqh :
+              GCCBuiltin<"__builtin_ia32_vpmacssdqh">,
+              Intrinsic<[llvm_v2i64_ty],
+                        [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacssdql :
+              GCCBuiltin<"__builtin_ia32_vpmacssdql">,
+              Intrinsic<[llvm_v2i64_ty],
+                        [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacsswd :
+              GCCBuiltin<"__builtin_ia32_vpmacsswd">,
+              Intrinsic<[llvm_v4i32_ty],
+                        [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacssww :
+              GCCBuiltin<"__builtin_ia32_vpmacssww">,
+              Intrinsic<[llvm_v8i16_ty],
+                        [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacswd :
+              GCCBuiltin<"__builtin_ia32_vpmacswd">,
+              Intrinsic<[llvm_v4i32_ty],
+                        [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmacsww :
+              GCCBuiltin<"__builtin_ia32_vpmacsww">,
+              Intrinsic<[llvm_v8i16_ty],
+                        [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmadcsswd :
+              GCCBuiltin<"__builtin_ia32_vpmadcsswd">,
+              Intrinsic<[llvm_v4i32_ty],
+                        [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpmadcswd :
+              GCCBuiltin<"__builtin_ia32_vpmadcswd">,
+              Intrinsic<[llvm_v4i32_ty],
+                        [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpperm :
+              GCCBuiltin<"__builtin_ia32_vpperm">,
+              Intrinsic<[llvm_v16i8_ty],
+                        [llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vprotb :
+              GCCBuiltin<"__builtin_ia32_vprotb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vprotd :
+              GCCBuiltin<"__builtin_ia32_vprotd">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vprotq :
+              GCCBuiltin<"__builtin_ia32_vprotq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vprotw :
+              GCCBuiltin<"__builtin_ia32_vprotw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpshab :
+              GCCBuiltin<"__builtin_ia32_vpshab">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpshad :
+              GCCBuiltin<"__builtin_ia32_vpshad">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpshaq :
+              GCCBuiltin<"__builtin_ia32_vpshaq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpshaw :
+              GCCBuiltin<"__builtin_ia32_vpshaw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpshlb :
+              GCCBuiltin<"__builtin_ia32_vpshlb">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpshld :
+              GCCBuiltin<"__builtin_ia32_vpshld">,
+              Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpshlq :
+              GCCBuiltin<"__builtin_ia32_vpshlq">,
+              Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty],
+                        [IntrNoMem]>;
+  def int_x86_xop_vpshlw :
+              GCCBuiltin<"__builtin_ia32_vpshlw">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty],
+                        [IntrNoMem]>;
 
 //===----------------------------------------------------------------------===//
 // MMX
@@ -874,191 +2372,326 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
 // Empty MMX state op.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_mmx_emms  : GCCBuiltin<"__builtin_ia32_emms">,
-              Intrinsic<[llvm_void_ty], [IntrWriteMem]>;
+              Intrinsic<[], [], []>;
   def int_x86_mmx_femms : GCCBuiltin<"__builtin_ia32_femms">,
-              Intrinsic<[llvm_void_ty], [IntrWriteMem]>;
+              Intrinsic<[], [], []>;
 }
 
 // Integer arithmetic ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   // Addition
+  def int_x86_mmx_padd_b : GCCBuiltin<"__builtin_ia32_paddb">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_padd_w : GCCBuiltin<"__builtin_ia32_paddw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_padd_d : GCCBuiltin<"__builtin_ia32_paddd">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_padd_q : GCCBuiltin<"__builtin_ia32_paddq">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+
   def int_x86_mmx_padds_b : GCCBuiltin<"__builtin_ia32_paddsb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_padds_w : GCCBuiltin<"__builtin_ia32_paddsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
 
   def int_x86_mmx_paddus_b : GCCBuiltin<"__builtin_ia32_paddusb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_paddus_w : GCCBuiltin<"__builtin_ia32_paddusw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
 
   // Subtraction
+  def int_x86_mmx_psub_b : GCCBuiltin<"__builtin_ia32_psubb">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_psub_w : GCCBuiltin<"__builtin_ia32_psubw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_psub_d : GCCBuiltin<"__builtin_ia32_psubd">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_psub_q : GCCBuiltin<"__builtin_ia32_psubq">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+
   def int_x86_mmx_psubs_b : GCCBuiltin<"__builtin_ia32_psubsb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_psubs_w : GCCBuiltin<"__builtin_ia32_psubsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
 
   def int_x86_mmx_psubus_b : GCCBuiltin<"__builtin_ia32_psubusb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_psubus_w : GCCBuiltin<"__builtin_ia32_psubusw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
 
   // Multiplication
   def int_x86_mmx_pmulh_w : GCCBuiltin<"__builtin_ia32_pmulhw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
+  def int_x86_mmx_pmull_w : GCCBuiltin<"__builtin_ia32_pmullw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_pmulhu_w : GCCBuiltin<"__builtin_ia32_pmulhuw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_pmulu_dq : GCCBuiltin<"__builtin_ia32_pmuludq">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v2i32_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_pmadd_wd : GCCBuiltin<"__builtin_ia32_pmaddwd">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
+
+  // Bitwise operations
+  def int_x86_mmx_pand : GCCBuiltin<"__builtin_ia32_pand">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_pandn : GCCBuiltin<"__builtin_ia32_pandn">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_por : GCCBuiltin<"__builtin_ia32_por">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_pxor : GCCBuiltin<"__builtin_ia32_pxor">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
 
   // Averages
   def int_x86_mmx_pavg_b : GCCBuiltin<"__builtin_ia32_pavgb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_pavg_w : GCCBuiltin<"__builtin_ia32_pavgw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
 
   // Maximum
   def int_x86_mmx_pmaxu_b : GCCBuiltin<"__builtin_ia32_pmaxub">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_pmaxs_w : GCCBuiltin<"__builtin_ia32_pmaxsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
 
   // Minimum
   def int_x86_mmx_pminu_b : GCCBuiltin<"__builtin_ia32_pminub">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_pmins_w : GCCBuiltin<"__builtin_ia32_pminsw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
 
   // Packed sum of absolute differences
   def int_x86_mmx_psad_bw : GCCBuiltin<"__builtin_ia32_psadbw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem, Commutative]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
 }
 
 // Integer shift ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   // Shift left logical
   def int_x86_mmx_psll_w : GCCBuiltin<"__builtin_ia32_psllw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v1i64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_psll_d : GCCBuiltin<"__builtin_ia32_pslld">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v1i64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_psll_q : GCCBuiltin<"__builtin_ia32_psllq">,
-              Intrinsic<[llvm_v1i64_ty, llvm_v1i64_ty,
-                         llvm_v1i64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
 
   def int_x86_mmx_psrl_w : GCCBuiltin<"__builtin_ia32_psrlw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v1i64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_psrl_d : GCCBuiltin<"__builtin_ia32_psrld">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v1i64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_psrl_q : GCCBuiltin<"__builtin_ia32_psrlq">,
-              Intrinsic<[llvm_v1i64_ty,   llvm_v1i64_ty,
-                         llvm_v1i64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
 
   def int_x86_mmx_psra_w : GCCBuiltin<"__builtin_ia32_psraw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v1i64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_psra_d : GCCBuiltin<"__builtin_ia32_psrad">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v1i64_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
 
   def int_x86_mmx_pslli_w : GCCBuiltin<"__builtin_ia32_psllwi">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_mmx_pslli_d : GCCBuiltin<"__builtin_ia32_pslldi">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_mmx_pslli_q : GCCBuiltin<"__builtin_ia32_psllqi">,
-              Intrinsic<[llvm_v1i64_ty, llvm_v1i64_ty,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
                          llvm_i32_ty], [IntrNoMem]>;
 
   def int_x86_mmx_psrli_w : GCCBuiltin<"__builtin_ia32_psrlwi">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_mmx_psrli_d : GCCBuiltin<"__builtin_ia32_psrldi">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_mmx_psrli_q : GCCBuiltin<"__builtin_ia32_psrlqi">,
-              Intrinsic<[llvm_v1i64_ty,   llvm_v1i64_ty,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
                          llvm_i32_ty], [IntrNoMem]>;
 
   def int_x86_mmx_psrai_w : GCCBuiltin<"__builtin_ia32_psrawi">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
                          llvm_i32_ty], [IntrNoMem]>;
   def int_x86_mmx_psrai_d : GCCBuiltin<"__builtin_ia32_psradi">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
                          llvm_i32_ty], [IntrNoMem]>;
 }
 
 // Pack ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_mmx_packsswb : GCCBuiltin<"__builtin_ia32_packsswb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_packssdw : GCCBuiltin<"__builtin_ia32_packssdw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v2i32_ty,
-                         llvm_v2i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_packuswb : GCCBuiltin<"__builtin_ia32_packuswb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
+}
+
+// Unpacking ops.
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_mmx_punpckhbw : GCCBuiltin<"__builtin_ia32_punpckhbw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_punpckhwd : GCCBuiltin<"__builtin_ia32_punpckhwd">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_punpckhdq : GCCBuiltin<"__builtin_ia32_punpckhdq">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_punpcklbw : GCCBuiltin<"__builtin_ia32_punpcklbw">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_punpcklwd : GCCBuiltin<"__builtin_ia32_punpcklwd">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
+  def int_x86_mmx_punpckldq : GCCBuiltin<"__builtin_ia32_punpckldq">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty],
+                        [IntrNoMem]>;
 }
 
 // Integer comparison ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_mmx_pcmpeq_b : GCCBuiltin<"__builtin_ia32_pcmpeqb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_pcmpeq_w : GCCBuiltin<"__builtin_ia32_pcmpeqw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
   def int_x86_mmx_pcmpeq_d : GCCBuiltin<"__builtin_ia32_pcmpeqd">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v2i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem, Commutative]>;
 
   def int_x86_mmx_pcmpgt_b : GCCBuiltin<"__builtin_ia32_pcmpgtb">,
-              Intrinsic<[llvm_v8i8_ty, llvm_v8i8_ty,
-                         llvm_v8i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_pcmpgt_w : GCCBuiltin<"__builtin_ia32_pcmpgtw">,
-              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
-                         llvm_v4i16_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
   def int_x86_mmx_pcmpgt_d : GCCBuiltin<"__builtin_ia32_pcmpgtd">,
-              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
-                         llvm_v2i32_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                         llvm_x86mmx_ty], [IntrNoMem]>;
 }
 
 // Misc.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_mmx_maskmovq : GCCBuiltin<"__builtin_ia32_maskmovq">,
-              Intrinsic<[llvm_void_ty, llvm_v8i8_ty, llvm_v8i8_ty, llvm_ptr_ty],
-                        [IntrWriteMem]>;
+              Intrinsic<[], [llvm_x86mmx_ty, llvm_x86mmx_ty, llvm_ptr_ty], []>;
 
   def int_x86_mmx_pmovmskb : GCCBuiltin<"__builtin_ia32_pmovmskb">,
-              Intrinsic<[llvm_i32_ty, llvm_v8i8_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i32_ty], [llvm_x86mmx_ty], [IntrNoMem]>;
 
   def int_x86_mmx_movnt_dq : GCCBuiltin<"__builtin_ia32_movntq">,
-              Intrinsic<[llvm_void_ty, llvm_ptr_ty,
-                         llvm_v1i64_ty], [IntrWriteMem]>;
+              Intrinsic<[], [llvm_ptrx86mmx_ty, llvm_x86mmx_ty], []>;
+
+  def int_x86_mmx_palignr_b : GCCBuiltin<"__builtin_ia32_palignr">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                        llvm_x86mmx_ty, llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_mmx_pextr_w : GCCBuiltin<"__builtin_ia32_vec_ext_v4hi">,
+              Intrinsic<[llvm_i32_ty], [llvm_x86mmx_ty, llvm_i32_ty],
+                        [IntrNoMem]>;
+
+  def int_x86_mmx_pinsr_w : GCCBuiltin<"__builtin_ia32_vec_set_v4hi">,
+              Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty,
+                        llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+}
+
+//===----------------------------------------------------------------------===//
+// BMI
+
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_bmi_bextr_32 : GCCBuiltin<"__builtin_ia32_bextr_u32">,
+              Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_bmi_bextr_64 : GCCBuiltin<"__builtin_ia32_bextr_u64">,
+              Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_bmi_bzhi_32 : GCCBuiltin<"__builtin_ia32_bzhi_si">,
+              Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_bmi_bzhi_64 : GCCBuiltin<"__builtin_ia32_bzhi_di">,
+              Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_bmi_pdep_32 : GCCBuiltin<"__builtin_ia32_pdep_si">,
+              Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_bmi_pdep_64 : GCCBuiltin<"__builtin_ia32_pdep_di">,
+              Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_bmi_pext_32 : GCCBuiltin<"__builtin_ia32_pext_si">,
+              Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_bmi_pext_64 : GCCBuiltin<"__builtin_ia32_pext_di">,
+              Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>;
+}
+
+//===----------------------------------------------------------------------===//
+// FS/GS Base
+
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_rdfsbase_32 : GCCBuiltin<"__builtin_ia32_rdfsbase32">,
+              Intrinsic<[llvm_i32_ty], []>;
+  def int_x86_rdgsbase_32 : GCCBuiltin<"__builtin_ia32_rdgsbase32">,
+              Intrinsic<[llvm_i32_ty], []>;
+  def int_x86_rdfsbase_64 : GCCBuiltin<"__builtin_ia32_rdfsbase64">,
+              Intrinsic<[llvm_i64_ty], []>;
+  def int_x86_rdgsbase_64 : GCCBuiltin<"__builtin_ia32_rdgsbase64">,
+              Intrinsic<[llvm_i64_ty], []>;
+  def int_x86_wrfsbase_32 : GCCBuiltin<"__builtin_ia32_wrfsbase32">,
+              Intrinsic<[], [llvm_i32_ty]>;
+  def int_x86_wrgsbase_32 : GCCBuiltin<"__builtin_ia32_wrgsbase32">,
+              Intrinsic<[], [llvm_i32_ty]>;
+  def int_x86_wrfsbase_64 : GCCBuiltin<"__builtin_ia32_wrfsbase64">,
+              Intrinsic<[], [llvm_i64_ty]>;
+  def int_x86_wrgsbase_64 : GCCBuiltin<"__builtin_ia32_wrgsbase64">,
+              Intrinsic<[], [llvm_i64_ty]>;
+}
+
+//===----------------------------------------------------------------------===//
+// Half float conversion
+
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_vcvtph2ps_128 : GCCBuiltin<"__builtin_ia32_vcvtph2ps">,
+              Intrinsic<[llvm_v4f32_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_vcvtph2ps_256 : GCCBuiltin<"__builtin_ia32_vcvtph2ps256">,
+              Intrinsic<[llvm_v8f32_ty], [llvm_v8i16_ty], [IntrNoMem]>;
+  def int_x86_vcvtps2ph_128 : GCCBuiltin<"__builtin_ia32_vcvtps2ph">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v4f32_ty, llvm_i32_ty],
+                        [IntrNoMem]>;
+  def int_x86_vcvtps2ph_256 : GCCBuiltin<"__builtin_ia32_vcvtps2ph256">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v8f32_ty, llvm_i32_ty],
+                        [IntrNoMem]>;
 }