optimize extracting two elements when lasx supported

Author: zhaoqi5

llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td

@@ -1282,6 +1282,32 @@ multiclass PatCCXrXrF<CondCode CC, string Inst> {
             (!cast<LAInst>(Inst#"_D") LASX256:$xj, LASX256:$xk)>;
 }
 
+multiclass PairInsertExtractPatV8<ValueType vecty, ValueType elemty> {
+  foreach imm1 = 0...3 in {
+    foreach imm2 = 0...3 in {
+      defvar Imm = !or(!shl(imm2, 4), imm1);
+      def : Pat<(vector_insert (vector_insert vecty:$xd,
+                    (elemty (vector_extract vecty:$xj, imm1)), imm2),
+                    (elemty (vector_extract vecty:$xj, !add(imm1, 4))),
+                    !add(imm2, 4)),
+                (XVEXTRINS_W $xd, $xj, Imm)>;
+    }
+  }
+}
+
+multiclass PairInsertExtractPatV4<ValueType vecty, ValueType elemty> {
+  foreach imm1 = 0...1 in {
+    foreach imm2 = 0...1 in {
+      defvar Imm = !or(!shl(imm2, 4), imm1);
+      def : Pat<(vector_insert (vector_insert vecty:$xd,
+                    (elemty (vector_extract vecty:$xj, imm1)), imm2),
+                    (elemty (vector_extract vecty:$xj, !add(imm1, 2))),
+                    !add(imm2, 2)),
+                (XVEXTRINS_D $xd, $xj, Imm)>;
+    }
+  }
+}
+
 let Predicates = [HasExtLASX] in {
 
 // XVADD_{B/H/W/D}
@@ -1582,6 +1608,38 @@ defm : PatCCXrXrF<SETUNE, "XVFCMP_CUNE">;
 defm : PatCCXrXrF<SETO, "XVFCMP_COR">;
 defm : PatCCXrXrF<SETUO, "XVFCMP_CUN">;
 
+// Insert two elements extracted from vector into vector. (The positions
+// of the two elements must be same in the source or destination vector's
+// front and back 128bits.)
+// 2*XVPICKVE2GR_{W/D} + 2*XVINSGR2VR_{W/D} -> XVEXTRINS_{W/D}
+// XVPERMI_D + 2*XVPICKVE2GR_{B/H} + 2*PseudoXVINSGR2VR_{B/H} -> XVEXTRINS_{W/D}
+foreach imm1 = 0...15 in {
+  foreach imm2 = 0...15 in {
+    defvar Imm = !or(!shl(imm2, 4), imm1);
+    def : Pat<(vector_insert (vector_insert v32i8:$xd,
+                  (GRLenVT (vector_extract v32i8:$xj, imm1)), imm2),
+                  (GRLenVT (vector_extract v32i8:$xj, !add(imm1, 16))),
+                  !add(imm2, 16)),
+              (XVEXTRINS_B $xd, $xj, Imm)>;
+  }
+}
+
+foreach imm1 = 0...7 in {
+  foreach imm2 = 0...7 in {
+    defvar Imm = !or(!shl(imm2, 4), imm1);
+    def : Pat<(vector_insert (vector_insert v16i16:$xd,
+                  (GRLenVT (vector_extract v16i16:$xj, imm1)), imm2),
+                  (GRLenVT (vector_extract v16i16:$xj, !add(imm1, 8))),
+                  !add(imm2, 8)),
+              (XVEXTRINS_H $xd, $xj, Imm)>;
+  }
+}
+
+defm : PairInsertExtractPatV8<v8i32, GRLenVT>;
+defm : PairInsertExtractPatV8<v8f32, f32>;
+defm : PairInsertExtractPatV4<v4i64, GRLenVT>;
+defm : PairInsertExtractPatV4<v4f64, f64>;
+
 // PseudoXVINSGR2VR_{B/H}
 def : Pat<(vector_insert v32i8:$xd, GRLenVT:$rj, uimm5:$imm),
           (PseudoXVINSGR2VR_B v32i8:$xd, GRLenVT:$rj, uimm5:$imm)>;

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

@@ -4,14 +4,7 @@
 define <32 x i8> @insert_extract_v32i8(<32 x i8> %a) nounwind {
 ; CHECK-LABEL: insert_extract_v32i8:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vpickve2gr.b $a0, $vr0, 15
-; CHECK-NEXT:    xvpermi.d $xr1, $xr0, 14
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a0, 1
-; CHECK-NEXT:    vpickve2gr.b $a0, $vr1, 15
-; CHECK-NEXT:    xvori.b $xr1, $xr0, 0
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 1
-; CHECK-NEXT:    vinsgr2vr.b $vr1, $a0, 1
-; CHECK-NEXT:    xvpermi.q $xr0, $xr1, 2
+; CHECK-NEXT:    xvextrins.b $xr0, $xr0, 31
 ; CHECK-NEXT:    ret
 entry:
   %b_lo = extractelement <32 x i8> %a, i32 15
@@ -24,14 +17,7 @@ entry:
 define <16 x i16> @insert_extract_v16i16(<16 x i16> %a) nounwind {
 ; CHECK-LABEL: insert_extract_v16i16:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vpickve2gr.h $a0, $vr0, 7
-; CHECK-NEXT:    xvpermi.d $xr1, $xr0, 14
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a0, 1
-; CHECK-NEXT:    vpickve2gr.h $a0, $vr1, 7
-; CHECK-NEXT:    xvori.b $xr1, $xr0, 0
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 1
-; CHECK-NEXT:    vinsgr2vr.h $vr1, $a0, 1
-; CHECK-NEXT:    xvpermi.q $xr0, $xr1, 2
+; CHECK-NEXT:    xvextrins.h $xr0, $xr0, 23
 ; CHECK-NEXT:    ret
 entry:
   %b_lo = extractelement <16 x i16> %a, i32 7
@@ -44,10 +30,7 @@ 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, 3
-; CHECK-NEXT:    xvpickve2gr.w $a1, $xr0, 7
-; CHECK-NEXT:    xvinsgr2vr.w $xr0, $a0, 1
-; CHECK-NEXT:    xvinsgr2vr.w $xr0, $a1, 5
+; CHECK-NEXT:    xvextrins.w $xr0, $xr0, 19
 ; CHECK-NEXT:    ret
 entry:
   %b_lo = extractelement <8 x i32> %a, i32 3
@@ -60,10 +43,7 @@ entry:
 define <8 x float> @insert_extract_v8f32(<8 x float> %a) nounwind {
 ; CHECK-LABEL: insert_extract_v8f32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvpickve2gr.w $a0, $xr0, 3
-; CHECK-NEXT:    xvpickve2gr.w $a1, $xr0, 7
-; CHECK-NEXT:    xvinsgr2vr.w $xr0, $a0, 1
-; CHECK-NEXT:    xvinsgr2vr.w $xr0, $a1, 5
+; CHECK-NEXT:    xvextrins.w $xr0, $xr0, 19
 ; CHECK-NEXT:    ret
 entry:
   %b_lo = extractelement <8 x float> %a, i32 3
@@ -76,10 +56,7 @@ 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, 1
-; CHECK-NEXT:    xvpickve2gr.d $a1, $xr0, 3
-; CHECK-NEXT:    xvinsgr2vr.d $xr0, $a0, 0
-; CHECK-NEXT:    xvinsgr2vr.d $xr0, $a1, 2
+; CHECK-NEXT:    xvextrins.d $xr0, $xr0, 1
 ; CHECK-NEXT:    ret
 entry:
   %b_lo = extractelement <4 x i64> %a, i32 1
@@ -92,10 +69,7 @@ entry:
 define <4 x double> @insert_extract_v4f64(<4 x double> %a) nounwind {
 ; CHECK-LABEL: insert_extract_v4f64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
-; CHECK-NEXT:    xvpickve2gr.d $a1, $xr0, 3
-; CHECK-NEXT:    xvinsgr2vr.d $xr0, $a0, 0
-; CHECK-NEXT:    xvinsgr2vr.d $xr0, $a1, 2
+; CHECK-NEXT:    xvextrins.d $xr0, $xr0, 1
 ; CHECK-NEXT:    ret
 entry:
   %b_lo = extractelement <4 x double> %a, i32 1