From: Adam Nemet <anemet@apple.com>
Date: Thu, 2 Oct 2014 23:18:30 +0000 (+0000)
Subject: [AVX512] Pull pattern for subvector insert into the instruction definition
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=6955c9d1acb0c5550abb566e993ec74aee50e994;p=llvm

[AVX512] Pull pattern for subvector insert into the instruction definition

No functional change intended.

Very similar to the change I made for subvector extract in r218480.

test/CodeGen/X86/avx512-insert-extract.ll covers this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@218928 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index f3c87ccc8eb..10055d065dc 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -309,7 +309,10 @@ multiclass vinsert_for_size<int Opcode,
                (ins VR512:$src1, From.RC:$src2, i8imm:$src3),
                "vinsert" # From.EltTypeName # "x4\t{$src3, $src2, $src1, $dst|"
                                                    "$dst, $src1, $src2, $src3}",
-               []>, EVEX_4V, EVEX_V512;
+               [(set To.RC:$dst, (vinsert_insert:$src3 (To.VT VR512:$src1),
+                                                       (From.VT From.RC:$src2),
+                                                       (iPTR imm)))]>,
+             EVEX_4V, EVEX_V512;
 
     let mayLoad = 1 in
     def rm : AVX512AIi8<Opcode, MRMSrcMem, (outs VR512:$dst),
@@ -319,13 +322,6 @@ multiclass vinsert_for_size<int Opcode,
                []>, EVEX_4V, EVEX_V512, EVEX_CD8<From.EltSize, CD8VT4>;
   }
 
-  // Codegen pattern, e.g. v4i32 -> v16i32 for vinserti32x4
-  def : Pat<(vinsert_insert:$ins
-               (To.VT VR512:$src1), (From.VT From.RC:$src2), (iPTR imm)),
-            (To.VT (!cast<Instruction>(NAME # From.EltSize # "x4rr")
-                          VR512:$src1, From.RC:$src2,
-                          (INSERT_get_vinsert_imm VR512:$ins)))>;
-
   // Codegen pattern with the alternative types, e.g. v2i64 -> v8i64 for
   // vinserti32x4
   def : Pat<(vinsert_insert:$ins