AVX-512: Added all SKX forms of GATHER instructions.

Added intrinsics. Added encoding and tests. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@240905 91177308-0d34-0410-b5e6-96231b3b80d8
2025-11-23 16:19:52 +00:00 · 2015-06-28 10:53:29 +00:00
parent 1e7ecc8e57
commit b23b2fbd3a
11 changed files with 935 additions and 120 deletions
--- a/include/llvm/IR/IntrinsicsX86.td
+++ b/include/llvm/IR/IntrinsicsX86.td
@@ -4264,6 +4264,102 @@ let TargetPrefix = "x86" in {
                     llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty],
                    [IntrReadArgMem]>;

+  def int_x86_avx512_gather3div2_df : 
+        GCCBuiltin<"__builtin_ia32_gather3div2df">,
+          Intrinsic<[llvm_v2f64_ty],
+          [llvm_v2f64_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3div2_di : 
+        GCCBuiltin<"__builtin_ia32_gather3div2di">,
+          Intrinsic<[llvm_v4i32_ty],
+          [llvm_v2i64_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3div4_df : 
+        GCCBuiltin<"__builtin_ia32_gather3div4df">,
+          Intrinsic<[llvm_v4f64_ty],
+          [llvm_v4f64_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3div4_di : 
+        GCCBuiltin<"__builtin_ia32_gather3div4di">,
+          Intrinsic<[llvm_v8i32_ty],
+          [llvm_v4i64_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3div4_sf : 
+        GCCBuiltin<"__builtin_ia32_gather3div4sf">,
+          Intrinsic<[llvm_v4f32_ty],
+          [llvm_v4f32_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3div4_si : 
+        GCCBuiltin<"__builtin_ia32_gather3div4si">,
+          Intrinsic<[llvm_v4i32_ty],
+          [llvm_v4i32_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3div8_sf : 
+        GCCBuiltin<"__builtin_ia32_gather3div8sf">,
+          Intrinsic<[llvm_v4f32_ty],
+          [llvm_v4f32_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3div8_si : 
+        GCCBuiltin<"__builtin_ia32_gather3div8si">,
+          Intrinsic<[llvm_v4i32_ty],
+          [llvm_v4i32_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3siv2_df : 
+        GCCBuiltin<"__builtin_ia32_gather3siv2df">,
+          Intrinsic<[llvm_v2f64_ty],
+          [llvm_v2f64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3siv2_di : 
+        GCCBuiltin<"__builtin_ia32_gather3siv2di">,
+          Intrinsic<[llvm_v4i32_ty],
+          [llvm_v2i64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3siv4_df : 
+        GCCBuiltin<"__builtin_ia32_gather3siv4df">,
+          Intrinsic<[llvm_v4f64_ty],
+          [llvm_v4f64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3siv4_di : 
+        GCCBuiltin<"__builtin_ia32_gather3siv4di">,
+          Intrinsic<[llvm_v8i32_ty],
+          [llvm_v4i64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3siv4_sf : 
+        GCCBuiltin<"__builtin_ia32_gather3siv4sf">,
+          Intrinsic<[llvm_v4f32_ty],
+          [llvm_v4f32_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3siv4_si : 
+        GCCBuiltin<"__builtin_ia32_gather3siv4si">,
+          Intrinsic<[llvm_v4i32_ty],
+          [llvm_v4i32_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3siv8_sf : 
+        GCCBuiltin<"__builtin_ia32_gather3siv8sf">,
+          Intrinsic<[llvm_v8f32_ty],
+          [llvm_v8f32_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
+  def int_x86_avx512_gather3siv8_si : 
+        GCCBuiltin<"__builtin_ia32_gather3siv8si">,
+          Intrinsic<[llvm_v8i32_ty],
+          [llvm_v8i32_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty],
+          [IntrReadArgMem]>;
+
 // scatter
  def int_x86_avx512_scatter_dpd_512  : GCCBuiltin<"__builtin_ia32_scattersiv8df">,
          Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty,