diff --git a/llvm/include/llvm/IR/IntrinsicsX86.td b/llvm/include/llvm/IR/IntrinsicsX86.td
--- a/llvm/include/llvm/IR/IntrinsicsX86.td
+++ b/llvm/include/llvm/IR/IntrinsicsX86.td
@@ -2787,13 +2787,13 @@
           [IntrNoMem, ImmArg<ArgIndex<3>>]>;
 
   def int_x86_avx512_mask_cvtsd2ss_round :
-        GCCBuiltin<"__builtin_ia32_cvtsd2ss_round_mask">,
+        GCCBuiltin<"__builtin_ia32_cvtsd2ss_mask_round">,
           Intrinsic<[llvm_v4f32_ty],
           [llvm_v4f32_ty, llvm_v2f64_ty, llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty],
           [IntrNoMem, ImmArg<ArgIndex<4>>]>;
 
   def int_x86_avx512_mask_cvtss2sd_round :
-        GCCBuiltin<"__builtin_ia32_cvtss2sd_round_mask">,
+        GCCBuiltin<"__builtin_ia32_cvtss2sd_mask_round">,
           Intrinsic<[llvm_v2f64_ty],
           [llvm_v2f64_ty, llvm_v4f32_ty, llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty],
           [IntrNoMem, ImmArg<ArgIndex<4>>]>;
@@ -3211,56 +3211,56 @@
           Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty,
                      llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<2>>]>;
 
-  def int_x86_avx512_mask_add_ss_round : GCCBuiltin<"__builtin_ia32_addss_round_mask">,
+  def int_x86_avx512_mask_add_ss_round : GCCBuiltin<"__builtin_ia32_addss_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                      llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_div_ss_round : GCCBuiltin<"__builtin_ia32_divss_round_mask">,
+  def int_x86_avx512_mask_div_ss_round : GCCBuiltin<"__builtin_ia32_divss_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                      llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_mul_ss_round : GCCBuiltin<"__builtin_ia32_mulss_round_mask">,
+  def int_x86_avx512_mask_mul_ss_round : GCCBuiltin<"__builtin_ia32_mulss_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                      llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_sub_ss_round : GCCBuiltin<"__builtin_ia32_subss_round_mask">,
+  def int_x86_avx512_mask_sub_ss_round : GCCBuiltin<"__builtin_ia32_subss_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                      llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_max_ss_round : GCCBuiltin<"__builtin_ia32_maxss_round_mask">,
+  def int_x86_avx512_mask_max_ss_round : GCCBuiltin<"__builtin_ia32_maxss_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                      llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_min_ss_round : GCCBuiltin<"__builtin_ia32_minss_round_mask">,
+  def int_x86_avx512_mask_min_ss_round : GCCBuiltin<"__builtin_ia32_minss_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                      llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_add_sd_round : GCCBuiltin<"__builtin_ia32_addsd_round_mask">,
+  def int_x86_avx512_mask_add_sd_round : GCCBuiltin<"__builtin_ia32_addsd_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                      llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_div_sd_round : GCCBuiltin<"__builtin_ia32_divsd_round_mask">,
+  def int_x86_avx512_mask_div_sd_round : GCCBuiltin<"__builtin_ia32_divsd_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                      llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_mul_sd_round : GCCBuiltin<"__builtin_ia32_mulsd_round_mask">,
+  def int_x86_avx512_mask_mul_sd_round : GCCBuiltin<"__builtin_ia32_mulsd_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                      llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_sub_sd_round : GCCBuiltin<"__builtin_ia32_subsd_round_mask">,
+  def int_x86_avx512_mask_sub_sd_round : GCCBuiltin<"__builtin_ia32_subsd_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                      llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_max_sd_round : GCCBuiltin<"__builtin_ia32_maxsd_round_mask">,
+  def int_x86_avx512_mask_max_sd_round : GCCBuiltin<"__builtin_ia32_maxsd_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                      llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_min_sd_round : GCCBuiltin<"__builtin_ia32_minsd_round_mask">,
+  def int_x86_avx512_mask_min_sd_round : GCCBuiltin<"__builtin_ia32_minsd_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                      llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<4>>]>;
 
-  def int_x86_avx512_mask_rndscale_ss : GCCBuiltin<"__builtin_ia32_rndscaless_round_mask">,
+  def int_x86_avx512_mask_rndscale_ss : GCCBuiltin<"__builtin_ia32_rndscaless_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty,
                                      llvm_i8_ty, llvm_i32_ty, llvm_i32_ty],
                                      [IntrNoMem, ImmArg<ArgIndex<4>>, ImmArg<ArgIndex<5>>]>;
-  def int_x86_avx512_mask_rndscale_sd : GCCBuiltin<"__builtin_ia32_rndscalesd_round_mask">,
+  def int_x86_avx512_mask_rndscale_sd : GCCBuiltin<"__builtin_ia32_rndscalesd_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty,
                                       llvm_i8_ty, llvm_i32_ty, llvm_i32_ty],
                                      [IntrNoMem, ImmArg<ArgIndex<4>>, ImmArg<ArgIndex<5>>]>;
-  def int_x86_avx512_mask_range_ss : GCCBuiltin<"__builtin_ia32_rangess128_round_mask">,
+  def int_x86_avx512_mask_range_ss : GCCBuiltin<"__builtin_ia32_rangess128_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty,
                                      llvm_i8_ty, llvm_i32_ty, llvm_i32_ty],
                                      [IntrNoMem, ImmArg<ArgIndex<4>>, ImmArg<ArgIndex<5>>]>;
-  def int_x86_avx512_mask_range_sd : GCCBuiltin<"__builtin_ia32_rangesd128_round_mask">,
+  def int_x86_avx512_mask_range_sd : GCCBuiltin<"__builtin_ia32_rangesd128_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty,
                                       llvm_i8_ty, llvm_i32_ty, llvm_i32_ty],
                                      [IntrNoMem, ImmArg<ArgIndex<4>>, ImmArg<ArgIndex<5>>]>;
@@ -3272,11 +3272,11 @@
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty,
                                       llvm_i8_ty, llvm_i32_ty, llvm_i32_ty],
                                      [IntrNoMem, ImmArg<ArgIndex<4>>, ImmArg<ArgIndex<5>>]>;
-  def int_x86_avx512_mask_scalef_sd : GCCBuiltin<"__builtin_ia32_scalefsd_round_mask">,
+  def int_x86_avx512_mask_scalef_sd : GCCBuiltin<"__builtin_ia32_scalefsd_mask_round">,
           Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                                       llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty],
                                      [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_scalef_ss : GCCBuiltin<"__builtin_ia32_scalefss_round_mask">,
+  def int_x86_avx512_mask_scalef_ss : GCCBuiltin<"__builtin_ia32_scalefss_mask_round">,
           Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                                       llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty],
                                      [IntrNoMem, ImmArg<ArgIndex<4>>]>;
@@ -3417,11 +3417,11 @@
                                      llvm_i16_ty, llvm_i32_ty],
                   [IntrNoMem, ImmArg<ArgIndex<3>>]>;
 
-  def int_x86_avx512_mask_getexp_ss : GCCBuiltin<"__builtin_ia32_getexpss128_round_mask">,
+  def int_x86_avx512_mask_getexp_ss : GCCBuiltin<"__builtin_ia32_getexpss128_mask_round">,
         Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty,
                                     llvm_i8_ty, llvm_i32_ty],
                   [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_mask_getexp_sd : GCCBuiltin<"__builtin_ia32_getexpsd128_round_mask">,
+  def int_x86_avx512_mask_getexp_sd : GCCBuiltin<"__builtin_ia32_getexpsd128_mask_round">,
         Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty,
                                     llvm_i8_ty, llvm_i32_ty],
                   [IntrNoMem, ImmArg<ArgIndex<4>>]>;
@@ -3463,13 +3463,13 @@
           [IntrNoMem, ImmArg<ArgIndex<1>>, ImmArg<ArgIndex<4>>]>;
 
   def int_x86_avx512_mask_getmant_ss :
-         GCCBuiltin<"__builtin_ia32_getmantss_round_mask">,
+         GCCBuiltin<"__builtin_ia32_getmantss_mask_round">,
           Intrinsic<[llvm_v4f32_ty],
           [llvm_v4f32_ty, llvm_v4f32_ty, llvm_i32_ty, llvm_v4f32_ty,
            llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<2>>, ImmArg<ArgIndex<5>>]>;
 
   def int_x86_avx512_mask_getmant_sd :
-         GCCBuiltin<"__builtin_ia32_getmantsd_round_mask">,
+         GCCBuiltin<"__builtin_ia32_getmantsd_mask_round">,
           Intrinsic<[llvm_v2f64_ty],
           [llvm_v2f64_ty, llvm_v2f64_ty, llvm_i32_ty, llvm_v2f64_ty,
            llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<2>>, ImmArg<ArgIndex<5>>]>;
@@ -3538,11 +3538,11 @@
             Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty,
                                         llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<ArgIndex<3>>]>;
 
-  def int_x86_avx512_rcp28_ss : GCCBuiltin<"__builtin_ia32_rcp28ss_round_mask">,
+  def int_x86_avx512_rcp28_ss : GCCBuiltin<"__builtin_ia32_rcp28ss_mask_round">,
             Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                                         llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_rcp28_sd : GCCBuiltin<"__builtin_ia32_rcp28sd_round_mask">,
+  def int_x86_avx512_rcp28_sd : GCCBuiltin<"__builtin_ia32_rcp28sd_mask_round">,
             Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                                         llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem, ImmArg<ArgIndex<4>>]>;
@@ -3554,11 +3554,11 @@
             Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty,
                                         llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem, ImmArg<ArgIndex<3>>]>;
-  def int_x86_avx512_rsqrt28_ss : GCCBuiltin<"__builtin_ia32_rsqrt28ss_round_mask">,
+  def int_x86_avx512_rsqrt28_ss : GCCBuiltin<"__builtin_ia32_rsqrt28ss_mask_round">,
             Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                                         llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem, ImmArg<ArgIndex<4>>]>;
-  def int_x86_avx512_rsqrt28_sd : GCCBuiltin<"__builtin_ia32_rsqrt28sd_round_mask">,
+  def int_x86_avx512_rsqrt28_sd : GCCBuiltin<"__builtin_ia32_rsqrt28sd_mask_round">,
             Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                                         llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem, ImmArg<ArgIndex<4>>]>;
@@ -5188,37 +5188,37 @@
                   [ IntrNoMem, ImmArg<ArgIndex<2>> ]>;
 
   def int_x86_avx512fp16_mask_add_sh_round
-      : GCCBuiltin<"__builtin_ia32_addsh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_addsh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_sub_sh_round
-      : GCCBuiltin<"__builtin_ia32_subsh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_subsh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_mul_sh_round
-      : GCCBuiltin<"__builtin_ia32_mulsh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_mulsh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_div_sh_round
-      : GCCBuiltin<"__builtin_ia32_divsh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_divsh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_min_sh_round
-      : GCCBuiltin<"__builtin_ia32_minsh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_minsh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_max_sh_round
-      : GCCBuiltin<"__builtin_ia32_maxsh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_maxsh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
@@ -5288,25 +5288,25 @@
                   [ llvm_v8f16_ty, llvm_v8f64_ty, llvm_i8_ty, llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<3>> ]>;
   def int_x86_avx512fp16_mask_vcvtsh2ss_round
-      : GCCBuiltin<"__builtin_ia32_vcvtsh2ss_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_vcvtsh2ss_mask_round">,
         Intrinsic<[ llvm_v4f32_ty ],
                   [ llvm_v4f32_ty, llvm_v8f16_ty, llvm_v4f32_ty, llvm_i8_ty,
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_vcvtss2sh_round
-      : GCCBuiltin<"__builtin_ia32_vcvtss2sh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_vcvtss2sh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v4f32_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_vcvtsd2sh_round
-      : GCCBuiltin<"__builtin_ia32_vcvtsd2sh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_vcvtsd2sh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v2f64_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_vcvtsh2sd_round
-      : GCCBuiltin<"__builtin_ia32_vcvtsh2sd_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_vcvtsh2sd_mask_round">,
         Intrinsic<[ llvm_v2f64_ty ],
                   [ llvm_v2f64_ty, llvm_v8f16_ty, llvm_v2f64_ty, llvm_i8_ty,
                     llvm_i32_ty ],
@@ -5648,7 +5648,7 @@
                   [ llvm_v32f16_ty, llvm_v32f16_ty, llvm_i32_ty, llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<3>> ]>;
   def int_x86_avx512fp16_mask_getexp_sh
-      : GCCBuiltin<"__builtin_ia32_getexpsh128_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_getexpsh128_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],
@@ -5670,7 +5670,7 @@
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<1>>, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_getmant_sh
-      : GCCBuiltin<"__builtin_ia32_getmantsh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_getmantsh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_i32_ty, llvm_v8f16_ty,
                     llvm_i8_ty, llvm_i32_ty ],
@@ -5692,7 +5692,7 @@
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<1>>, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_rndscale_sh
-      : GCCBuiltin<"__builtin_ia32_rndscalesh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_rndscalesh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty, llvm_i32_ty ],
@@ -5714,7 +5714,7 @@
                     llvm_i32_ty ],
                   [ IntrNoMem, ImmArg<ArgIndex<4>> ]>;
   def int_x86_avx512fp16_mask_scalef_sh
-      : GCCBuiltin<"__builtin_ia32_scalefsh_round_mask">,
+      : GCCBuiltin<"__builtin_ia32_scalefsh_mask_round">,
         Intrinsic<[ llvm_v8f16_ty ],
                   [ llvm_v8f16_ty, llvm_v8f16_ty, llvm_v8f16_ty, llvm_i8_ty,
                     llvm_i32_ty ],