[AVX512] Intrinsics for vextract*x4
[oota-llvm.git] / include / llvm / IR / IntrinsicsX86.td
index 3012cdf7db4d893bfde9bd7432dcd9cb0dcc60ca..87c7c870033f72f87336bdbb3888278a00735a2d 100644 (file)
@@ -1791,6 +1791,23 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx2_vinserti128 : GCCBuiltin<"__builtin_ia32_insert128i256">,
               Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty,
                          llvm_v2i64_ty, llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_avx512_mask_vextractf32x4_512 :
+      GCCBuiltin<"__builtin_ia32_extractf32x4_mask">,
+                 Intrinsic<[llvm_v4f32_ty], [llvm_v16f32_ty, llvm_i8_ty,
+                           llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_vextracti32x4_512 :
+      GCCBuiltin<"__builtin_ia32_extracti32x4_mask">,
+                 Intrinsic<[llvm_v4i32_ty], [llvm_v16i32_ty, llvm_i8_ty,
+                           llvm_v4i32_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_vextractf64x4_512 :
+      GCCBuiltin<"__builtin_ia32_extractf64x4_mask">,
+                 Intrinsic<[llvm_v4f64_ty], [llvm_v8f64_ty, llvm_i8_ty,
+                           llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_vextracti64x4_512 :
+      GCCBuiltin<"__builtin_ia32_extracti64x4_mask">,
+                 Intrinsic<[llvm_v4i64_ty], [llvm_v8i64_ty, llvm_i8_ty,
+                           llvm_v4i64_ty, llvm_i8_ty], [IntrNoMem]>;
 }
 
 // Conditional load ops
@@ -3249,6 +3266,151 @@ let TargetPrefix = "x86" in {
   def int_x86_avx512_mask_pcmpeq_q_512 : GCCBuiltin<"__builtin_ia32_pcmpeqq512_mask">,
             Intrinsic<[llvm_i8_ty], [llvm_v8i64_ty, llvm_v8i64_ty, llvm_i8_ty],
                       [IntrNoMem]>;
+
+  def int_x86_avx512_mask_pcmpgt_b_512: GCCBuiltin<"__builtin_ia32_pcmpgtb512_mask">,
+        Intrinsic<[llvm_i64_ty], [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i64_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_w_512: GCCBuiltin<"__builtin_ia32_pcmpgtw512_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i16_ty, llvm_v32i16_ty, llvm_i32_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_d_512: GCCBuiltin<"__builtin_ia32_pcmpgtd512_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i32_ty, llvm_v16i32_ty, llvm_i16_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_q_512: GCCBuiltin<"__builtin_ia32_pcmpgtq512_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i64_ty, llvm_v8i64_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+
+  def int_x86_avx512_mask_cmp_b_512: GCCBuiltin<"__builtin_ia32_cmpb512_mask">,
+        Intrinsic<[llvm_i64_ty], [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i32_ty,
+                  llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_cmp_w_512: GCCBuiltin<"__builtin_ia32_cmpw512_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i16_ty, llvm_v32i16_ty, llvm_i32_ty,
+                  llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_cmp_d_512: GCCBuiltin<"__builtin_ia32_cmpd512_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i32_ty, llvm_v16i32_ty, llvm_i32_ty,
+                  llvm_i16_ty], [IntrNoMem ]>;
+  def int_x86_avx512_mask_cmp_q_512: GCCBuiltin<"__builtin_ia32_cmpq512_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i64_ty, llvm_v8i64_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_avx512_mask_ucmp_b_512: GCCBuiltin<"__builtin_ia32_ucmpb512_mask">,
+        Intrinsic<[llvm_i64_ty], [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i32_ty,
+                  llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_w_512: GCCBuiltin<"__builtin_ia32_ucmpw512_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i16_ty, llvm_v32i16_ty, llvm_i32_ty,
+                  llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_d_512: GCCBuiltin<"__builtin_ia32_ucmpd512_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i32_ty, llvm_v16i32_ty, llvm_i32_ty,
+                  llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_q_512: GCCBuiltin<"__builtin_ia32_ucmpq512_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i64_ty, llvm_v8i64_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+
+  // 256-bit
+  def int_x86_avx512_mask_pcmpeq_b_256 : GCCBuiltin<"__builtin_ia32_pcmpeqb256_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpeq_w_256 : GCCBuiltin<"__builtin_ia32_pcmpeqw256_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i16_ty, llvm_v16i16_ty, llvm_i16_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpeq_d_256 : GCCBuiltin<"__builtin_ia32_pcmpeqd256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i32_ty, llvm_v8i32_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpeq_q_256 : GCCBuiltin<"__builtin_ia32_pcmpeqq256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i64_ty, llvm_v4i64_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+
+  def int_x86_avx512_mask_pcmpgt_b_256: GCCBuiltin<"__builtin_ia32_pcmpgtb256_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_w_256: GCCBuiltin<"__builtin_ia32_pcmpgtw256_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i16_ty, llvm_v16i16_ty, llvm_i16_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_d_256: GCCBuiltin<"__builtin_ia32_pcmpgtd256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i32_ty, llvm_v8i32_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_q_256: GCCBuiltin<"__builtin_ia32_pcmpgtq256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i64_ty, llvm_v4i64_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+
+  def int_x86_avx512_mask_cmp_b_256: GCCBuiltin<"__builtin_ia32_cmpb256_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty,
+                  llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_cmp_w_256: GCCBuiltin<"__builtin_ia32_cmpw256_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i16_ty, llvm_v16i16_ty, llvm_i32_ty,
+                  llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_cmp_d_256: GCCBuiltin<"__builtin_ia32_cmpd256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i32_ty, llvm_v8i32_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_cmp_q_256: GCCBuiltin<"__builtin_ia32_cmpq256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i64_ty, llvm_v4i64_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_avx512_mask_ucmp_b_256: GCCBuiltin<"__builtin_ia32_ucmpb256_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty,
+                  llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_w_256: GCCBuiltin<"__builtin_ia32_ucmpw256_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i16_ty, llvm_v16i16_ty, llvm_i32_ty,
+                  llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_d_256: GCCBuiltin<"__builtin_ia32_ucmpd256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i32_ty, llvm_v8i32_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_q_256: GCCBuiltin<"__builtin_ia32_ucmpq256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i64_ty, llvm_v4i64_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+
+  // 128-bit
+  def int_x86_avx512_mask_pcmpeq_b_128 : GCCBuiltin<"__builtin_ia32_pcmpeqb128_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i16_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpeq_w_128 : GCCBuiltin<"__builtin_ia32_pcmpeqw128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i16_ty, llvm_v8i16_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpeq_d_128 : GCCBuiltin<"__builtin_ia32_pcmpeqd128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i32_ty, llvm_v4i32_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpeq_q_128 : GCCBuiltin<"__builtin_ia32_pcmpeqq128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v2i64_ty, llvm_v2i64_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+
+  def int_x86_avx512_mask_pcmpgt_b_128: GCCBuiltin<"__builtin_ia32_pcmpgtb128_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i16_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_w_128: GCCBuiltin<"__builtin_ia32_pcmpgtw128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i16_ty, llvm_v8i16_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_d_128: GCCBuiltin<"__builtin_ia32_pcmpgtd128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i32_ty, llvm_v4i32_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_q_128: GCCBuiltin<"__builtin_ia32_pcmpgtq128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v2i64_ty, llvm_v2i64_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+
+  def int_x86_avx512_mask_cmp_b_128: GCCBuiltin<"__builtin_ia32_cmpb128_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i32_ty,
+                  llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_cmp_w_128: GCCBuiltin<"__builtin_ia32_cmpw128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i16_ty, llvm_v8i16_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_cmp_d_128: GCCBuiltin<"__builtin_ia32_cmpd128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i32_ty, llvm_v4i32_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_cmp_q_128: GCCBuiltin<"__builtin_ia32_cmpq128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v2i64_ty, llvm_v2i64_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_avx512_mask_ucmp_b_128: GCCBuiltin<"__builtin_ia32_ucmpb128_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i32_ty,
+                  llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_w_128: GCCBuiltin<"__builtin_ia32_ucmpw128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i16_ty, llvm_v8i16_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_d_128: GCCBuiltin<"__builtin_ia32_ucmpd128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i32_ty, llvm_v4i32_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_ucmp_q_128: GCCBuiltin<"__builtin_ia32_ucmpq128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v2i64_ty, llvm_v2i64_ty, llvm_i32_ty,
+                  llvm_i8_ty], [IntrNoMem]>;
 }
 
 // Misc.