Optimize inserting extracted element for v4i64/v8i32

Author: tangaac

llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td

@@ -1640,6 +1640,22 @@ defm : PairInsertExtractPatV8<v8f32, f32>;
 defm : PairInsertExtractPatV4<v4i64, GRLenVT>;
 defm : PairInsertExtractPatV4<v4f64, f64>;
 
+foreach imm1 = 0...7 in {
+  foreach imm2 = 0...7 in {
+    def : Pat<(vector_insert v8i32:$xd,
+                  (GRLenVT(vector_extract v8i32:$xj, imm1)), imm2),
+              (XVINSVE0_W v8i32:$xd, (XVPICKVE_W v8i32:$xj, imm1), imm2)>;
+  }
+}
+
+foreach imm1 = 0...3 in {
+  foreach imm2 = 0...3 in {
+    def : Pat<(vector_insert v4i64:$xd,
+                  (GRLenVT(vector_extract v4i64:$xj, imm1)), imm2),
+              (XVINSVE0_D v4i64:$xd, (XVPICKVE_D v4i64:$xj, imm1), imm2)>;
+  }
+}
+
 // PseudoXVINSGR2VR_{B/H}
 def : Pat<(vector_insert v32i8:$xd, GRLenVT:$rj, uimm5:$imm),
           (PseudoXVINSGR2VR_B v32i8:$xd, GRLenVT:$rj, uimm5:$imm)>;

llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insert-extract-element.ll

@@ -30,8 +30,8 @@ entry:
 define <8 x i32> @insert_extract_v8i32(<8 x i32> %a) nounwind {
 ; CHECK-LABEL: insert_extract_v8i32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvpickve2gr.w $a0, $xr0, 7
-; CHECK-NEXT:    xvinsgr2vr.w $xr0, $a0, 1
+; CHECK-NEXT:    xvpickve.w $xr1, $xr0, 7
+; CHECK-NEXT:    xvinsve0.w $xr0, $xr1, 1
 ; CHECK-NEXT:    ret
 entry:
   %b = extractelement <8 x i32> %a, i32 7
@@ -54,8 +54,8 @@ entry:
 define <4 x i64> @insert_extract_v4i64(<4 x i64> %a) nounwind {
 ; CHECK-LABEL: insert_extract_v4i64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 3
-; CHECK-NEXT:    xvinsgr2vr.d $xr0, $a0, 1
+; CHECK-NEXT:    xvpickve.d $xr1, $xr0, 3
+; CHECK-NEXT:    xvinsve0.d $xr0, $xr1, 1
 ; CHECK-NEXT:    ret
 entry:
   %b = extractelement <4 x i64> %a, i32 3