[LoongArch] Lower [x]vshuf.d to [x]vshuf4i.d if possible.

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

@@ -1026,46 +1026,54 @@ static SDValue lowerVECTOR_SHUFFLE_VSHUF4I(const SDLoc &DL, ArrayRef<int> Mask,
                                            MVT VT, SDValue V1, SDValue V2,
                                            SelectionDAG &DAG) {
 
-  // When the size is less than 4, lower cost instructions may be used.
-  if (Mask.size() < 4)
-    return SDValue();
+  unsigned SubVecSize = 4;
+  if (VT == MVT::v2f64 || VT == MVT::v2i64 || VT == MVT::v4f64 ||
+      VT == MVT::v4i64) {
+    SubVecSize = 2;
+  }
 
   int SubMask[4] = {-1, -1, -1, -1};
-  for (unsigned i = 0; i < 4; ++i) {
-    for (unsigned j = i; j < Mask.size(); j += 4) {
-      int Idx = Mask[j];
+  for (unsigned i = 0; i < SubVecSize; ++i) {
+    for (unsigned j = i; j < Mask.size(); j += SubVecSize) {
+      int M = Mask[j];
 
       // Convert from vector index to 4-element subvector index
       // If an index refers to an element outside of the subvector then give up
-      if (Idx != -1) {
-        Idx -= 4 * (j / 4);
-        if (Idx < 0 || Idx >= 4)
+      if (M != -1) {
+        M -= 4 * (j / SubVecSize);
+        if (M < 0 || M >= 4)
           return SDValue();
       }
 
       // If the mask has an undef, replace it with the current index.
       // Note that it might still be undef if the current index is also undef
       if (SubMask[i] == -1)
-        SubMask[i] = Idx;
+        SubMask[i] = M;
       // Check that non-undef values are the same as in the mask. If they
       // aren't then give up
-      else if (Idx != -1 && Idx != SubMask[i])
+      else if (M != -1 && M != SubMask[i])
         return SDValue();
     }
   }
 
   // Calculate the immediate. Replace any remaining undefs with zero
   APInt Imm(64, 0);
-  for (int i = 3; i >= 0; --i) {
-    int Idx = SubMask[i];
+  for (int i = SubVecSize - 1; i >= 0; --i) {
+    int M = SubMask[i];
 
-    if (Idx == -1)
-      Idx = 0;
+    if (M == -1)
+      M = 0;
 
     Imm <<= 2;
-    Imm |= Idx & 0x3;
+    Imm |= M & 0x3;
   }
 
+  // Return vshuf4i.d and xvshuf4i.d
+  if (VT == MVT::v2f64 || VT == MVT::v2i64 || VT == MVT::v4f64 ||
+      VT == MVT::v4i64)
+    return DAG.getNode(LoongArchISD::VSHUF4I, DL, VT, V1, V2,
+                       DAG.getConstant(Imm, DL, MVT::i64));
+
   return DAG.getNode(LoongArchISD::VSHUF4I, DL, VT, V1,
                      DAG.getConstant(Imm, DL, MVT::i64));
 }
@@ -1389,6 +1397,9 @@ static SDValue lower128BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
     return Result;
   if ((Result = lowerVECTOR_SHUFFLE_VPICKOD(DL, Mask, VT, V1, V2, DAG)))
     return Result;
+  if ((VT.SimpleTy == MVT::v2i64 || VT.SimpleTy == MVT::v2f64) &&
+      (Result = lowerVECTOR_SHUFFLE_VSHUF4I(DL, Mask, VT, V1, V2, DAG)))
+    return Result;
   if ((Result = lowerVECTOR_SHUFFLEAsZeroOrAnyExtend(DL, Mask, VT, V1, V2, DAG,
                                                      Zeroable)))
     return Result;
@@ -1447,10 +1458,6 @@ static SDValue lowerVECTOR_SHUFFLE_XVREPLVEI(const SDLoc &DL,
 static SDValue lowerVECTOR_SHUFFLE_XVSHUF4I(const SDLoc &DL, ArrayRef<int> Mask,
                                             MVT VT, SDValue V1, SDValue V2,
                                             SelectionDAG &DAG) {
-  // When the size is less than or equal to 4, lower cost instructions may be
-  // used.
-  if (Mask.size() <= 4)
-    return SDValue();
   return lowerVECTOR_SHUFFLE_VSHUF4I(DL, Mask, VT, V1, V2, DAG);
 }
 
@@ -1832,6 +1839,9 @@ static SDValue lower256BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
     return Result;
   if ((Result = lowerVECTOR_SHUFFLE_XVPICKOD(DL, NewMask, VT, V1, V2, DAG)))
     return Result;
+  if ((VT.SimpleTy == MVT::v4i64 || VT.SimpleTy == MVT::v4f64) &&
+      (Result = lowerVECTOR_SHUFFLE_XVSHUF4I(DL, NewMask, VT, V1, V2, DAG)))
+    return Result;
   if ((Result =
            lowerVECTOR_SHUFFLEAsShift(DL, NewMask, VT, V1, V2, DAG, Zeroable)))
     return Result;

llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td

@@ -1729,6 +1729,10 @@ def : Pat<(loongarch_vshuf4i v8i32:$xj, immZExt8:$ui8),
         (XVSHUF4I_W v8i32:$xj, immZExt8:$ui8)>;
 def : Pat<(loongarch_vshuf4i v8f32:$xj, immZExt8:$ui8),
         (XVSHUF4I_W v8f32:$xj, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i_d v4i64:$xj, v4i64:$xk, immZExt8:$ui8),
+          (XVSHUF4I_D v4i64:$xj, v4i64:$xk, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i_d v4f64:$xj, v4f64:$xk, immZExt8:$ui8),
+          (XVSHUF4I_D v4f64:$xj, v4f64:$xk, immZExt8:$ui8)>;
 
 // XVREPL128VEI_{B/H/W/D}
 def : Pat<(loongarch_vreplvei v32i8:$xj, immZExt4:$ui4),

llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td

@@ -23,6 +23,10 @@ def SDT_LoongArchV2R : SDTypeProfile<1, 2, [SDTCisVec<0>,
                                      SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>]>;
 def SDT_LoongArchV1RUimm: SDTypeProfile<1, 2, [SDTCisVec<0>,
                                         SDTCisSameAs<0,1>, SDTCisVT<2, i64>]>;
+def SDT_LoongArchVShuf4i_D
+    : SDTypeProfile<1, 3,
+                    [SDTCisVec<0>, SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>,
+                     SDTCisVT<3, i64>]>;
 def SDT_LoongArchVreplgr2vr : SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisVec<0>, SDTCisInt<1>]>;
 def SDT_LoongArchVFRECIPE : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisVec<0>, SDTCisSameAs<0, 1>]>;
 def SDT_LoongArchVFRSQRTE : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisVec<0>, SDTCisSameAs<0, 1>]>;
@@ -53,6 +57,8 @@ def loongarch_vilvl: SDNode<"LoongArchISD::VILVL", SDT_LoongArchV2R>;
 def loongarch_vilvh: SDNode<"LoongArchISD::VILVH", SDT_LoongArchV2R>;
 
 def loongarch_vshuf4i: SDNode<"LoongArchISD::VSHUF4I", SDT_LoongArchV1RUimm>;
+def loongarch_vshuf4i_d
+    : SDNode<"LoongArchISD::VSHUF4I", SDT_LoongArchVShuf4i_D>;
 def loongarch_vreplvei: SDNode<"LoongArchISD::VREPLVEI", SDT_LoongArchV1RUimm>;
 def loongarch_vreplgr2vr: SDNode<"LoongArchISD::VREPLGR2VR", SDT_LoongArchVreplgr2vr>;
 
@@ -1914,6 +1920,10 @@ def : Pat<(loongarch_vshuf4i v4i32:$vj, immZExt8:$ui8),
         (VSHUF4I_W v4i32:$vj, immZExt8:$ui8)>;
 def : Pat<(loongarch_vshuf4i v4f32:$vj, immZExt8:$ui8),
         (VSHUF4I_W v4f32:$vj, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i_d v2i64:$vj, v2i64:$vk, immZExt8:$ui8),
+          (VSHUF4I_D v2i64:$vj, v2i64:$vk, immZExt8:$ui8)>;
+def : Pat<(loongarch_vshuf4i_d v2f64:$vj, v2f64:$vk, immZExt8:$ui8),
+          (VSHUF4I_D v2f64:$vj, v2f64:$vk, immZExt8:$ui8)>;
 
 // VREPLVEI_{B/H/W/D}
 def : Pat<(loongarch_vreplvei v16i8:$vj, immZExt4:$ui4),

llvm/test/CodeGen/LoongArch/lasx/ir-instruction/shuffle-as-xvshuf4i.ll

@@ -41,3 +41,23 @@ define <8 x float> @shufflevector_xvshuf4i_v8f32(<8 x float> %a, <8 x float> %b)
     %c = shufflevector <8 x float> %a, <8 x float> %b, <8 x i32> <i32 3, i32 2, i32 1, i32 0, i32 7, i32 6, i32 5, i32 4>
     ret <8 x float> %c
 }
+
+;; xvshuf4i.d
+define <4 x i64> @shufflevector_xvshuf4i_v4d64(<4 x i64> %a, <4 x i64> %b) {
+; CHECK-LABEL: shufflevector_xvshuf4i_v4d64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    xvshuf4i.d $xr0, $xr1, 9
+; CHECK-NEXT:    ret
+    %c = shufflevector <4 x i64> %a, <4 x i64> %b, <4 x i32> <i32 1, i32 2, i32 5, i32 6>
+    ret <4 x i64> %c
+}
+
+;; xvshuf4i.d
+define <4 x double> @shufflevector_xvshuf4i_v4f64(<4 x double> %a, <4 x double> %b) {
+; CHECK-LABEL: shufflevector_xvshuf4i_v4f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    xvshuf4i.d $xr0, $xr1, 9
+; CHECK-NEXT:    ret
+    %c = shufflevector <4 x double> %a, <4 x double> %b, <4 x i32> <i32 1, i32 2, i32 5, i32 6>
+    ret <4 x double> %c
+}

llvm/test/CodeGen/LoongArch/lsx/ir-instruction/shuffle-as-vshuf.ll

@@ -42,10 +42,7 @@ define <4 x i32> @shufflevector_v4i32(<4 x i32> %a, <4 x i32> %b) {
 define <2 x i64> @shufflevector_v2i64(<2 x i64> %a, <2 x i64> %b) {
 ; CHECK-LABEL: shufflevector_v2i64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI3_0)
-; CHECK-NEXT:    vld $vr2, $a0, %pc_lo12(.LCPI3_0)
-; CHECK-NEXT:    vshuf.d $vr2, $vr1, $vr0
-; CHECK-NEXT:    vori.b $vr0, $vr2, 0
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 12
 ; CHECK-NEXT:    ret
     %c = shufflevector <2 x i64> %a, <2 x i64> %b, <2 x i32> <i32 0, i32 3>
     ret <2 x i64> %c
@@ -68,10 +65,7 @@ define <4 x float> @shufflevector_v4f32(<4 x float> %a, <4 x float> %b) {
 define <2 x double> @shufflevector_v2f64(<2 x double> %a, <2 x double> %b) {
 ; CHECK-LABEL: shufflevector_v2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI5_0)
-; CHECK-NEXT:    vld $vr2, $a0, %pc_lo12(.LCPI5_0)
-; CHECK-NEXT:    vshuf.d $vr2, $vr1, $vr0
-; CHECK-NEXT:    vori.b $vr0, $vr2, 0
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 12
 ; CHECK-NEXT:    ret
     %c = shufflevector <2 x double> %a, <2 x double> %b, <2 x i32> <i32 0, i32 3>
     ret <2 x double> %c

llvm/test/CodeGen/LoongArch/lsx/ir-instruction/shuffle-as-vshuf4i.ll

@@ -1,7 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
 ; RUN: llc --mtriple=loongarch64 --mattr=+lsx %s -o - | FileCheck %s
 
-;; vilvh.b
+;; vshuf4i.b
 define <16 x i8> @shufflevector_vshuf4i_v16i8(<16 x i8> %a, <16 x i8> %b) {
 ; CHECK-LABEL: shufflevector_vshuf4i_v16i8:
 ; CHECK:       # %bb.0:
@@ -11,7 +11,7 @@ define <16 x i8> @shufflevector_vshuf4i_v16i8(<16 x i8> %a, <16 x i8> %b) {
     ret <16 x i8> %c
 }
 
-;; vilvh.h
+;; vshuf4i.h
 define <8 x i16> @shufflevector_vshuf4i_v8i4(<8 x i16> %a, <8 x i16> %b) {
 ; CHECK-LABEL: shufflevector_vshuf4i_v8i4:
 ; CHECK:       # %bb.0:
@@ -21,7 +21,7 @@ define <8 x i16> @shufflevector_vshuf4i_v8i4(<8 x i16> %a, <8 x i16> %b) {
     ret <8 x i16> %c
 }
 
-;; vilvh.w
+;; vshuf4i.w
 define <4 x i32> @shufflevector_vshuf4i_v4i32(<4 x i32> %a, <4 x i32> %b) {
 ; CHECK-LABEL: shufflevector_vshuf4i_v4i32:
 ; CHECK:       # %bb.0:
@@ -31,7 +31,7 @@ define <4 x i32> @shufflevector_vshuf4i_v4i32(<4 x i32> %a, <4 x i32> %b) {
     ret <4 x i32> %c
 }
 
-;; vilvh.w
+;; vshuf4i.w
 define <4 x float> @shufflevector_vshuf4i_v4f32(<4 x float> %a, <4 x float> %b) {
 ; CHECK-LABEL: shufflevector_vshuf4i_v4f32:
 ; CHECK:       # %bb.0:
@@ -40,3 +40,23 @@ define <4 x float> @shufflevector_vshuf4i_v4f32(<4 x float> %a, <4 x float> %b)
     %c = shufflevector <4 x float> %a, <4 x float> %b, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
     ret <4 x float> %c
 }
+
+;; vshuf4i.d
+define <2 x i64> @shufflevector_vshuf4i_v2d64(<2 x i64> %a, <2 x i64> %b) {
+; CHECK-LABEL: shufflevector_vshuf4i_v2d64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 9
+; CHECK-NEXT:    ret
+    %c = shufflevector <2 x i64> %a, <2 x i64> %b, <2 x i32> <i32 1, i32 2>
+    ret <2 x i64> %c
+}
+
+;; vshuf4i.d
+define <2 x double> @shufflevector_vshuf4i_v2f64(<2 x double> %a, <2 x double> %b) {
+; CHECK-LABEL: shufflevector_vshuf4i_v2f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 9
+; CHECK-NEXT:    ret
+    %c = shufflevector <2 x double> %a, <2 x double> %b, <2 x i32> <i32 1, i32 2>
+    ret <2 x double> %c
+}

llvm/test/CodeGen/LoongArch/lsx/vec-shuffle-byte-rotate.ll

@@ -103,9 +103,7 @@ define <4 x i32> @byte_rotate_v4i32_3(<4 x i32> %a) nounwind {
 define <2 x i64> @byte_rotate_v2i64_1(<2 x i64> %a, <2 x i64> %b) nounwind {
 ; CHECK-LABEL: byte_rotate_v2i64_1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vbsrl.v $vr1, $vr1, 8
-; CHECK-NEXT:    vbsll.v $vr0, $vr0, 8
-; CHECK-NEXT:    vor.v $vr0, $vr0, $vr1
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 3
 ; CHECK-NEXT:    ret
     %shuffle = shufflevector <2 x i64> %a, <2 x i64> %b, <2 x i32> <i32 3, i32 0>
     ret <2 x i64> %shuffle
@@ -114,9 +112,7 @@ define <2 x i64> @byte_rotate_v2i64_1(<2 x i64> %a, <2 x i64> %b) nounwind {
 define <2 x i64> @byte_rotate_v2i64_2(<2 x i64> %a, <2 x i64> %b) nounwind {
 ; CHECK-LABEL: byte_rotate_v2i64_2:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vbsrl.v $vr0, $vr0, 8
-; CHECK-NEXT:    vbsll.v $vr1, $vr1, 8
-; CHECK-NEXT:    vor.v $vr0, $vr1, $vr0
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 9
 ; CHECK-NEXT:    ret
     %shuffle = shufflevector <2 x i64> %a, <2 x i64> %b, <2 x i32> <i32 1, i32 2>
     ret <2 x i64> %shuffle
@@ -125,9 +121,7 @@ define <2 x i64> @byte_rotate_v2i64_2(<2 x i64> %a, <2 x i64> %b) nounwind {
 define <2 x i64> @byte_rotate_v2i64_3(<2 x i64> %a) nounwind {
 ; CHECK-LABEL: byte_rotate_v2i64_3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vbsrl.v $vr1, $vr0, 8
-; CHECK-NEXT:    vbsll.v $vr0, $vr0, 8
-; CHECK-NEXT:    vor.v $vr0, $vr0, $vr1
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr0, 1
 ; CHECK-NEXT:    ret
     %shuffle = shufflevector <2 x i64> %a, <2 x i64> poison, <2 x i32> <i32 1, i32 0>
     ret <2 x i64> %shuffle

llvm/test/CodeGen/LoongArch/lsx/vec-shuffle-byte-shift.ll

@@ -464,7 +464,8 @@ define <4 x i32> @shuffle_4i32_vbsrl_v_12(<4 x i32> %a) nounwind {
 define <2 x i64> @shuffle_2i64_vbsrl_v_8(<2 x i64> %a) nounwind {
 ; CHECK-LABEL: shuffle_2i64_vbsrl_v_8:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vbsrl.v $vr0, $vr0, 8
+; CHECK-NEXT:    vrepli.b $vr1, 0
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 9
 ; CHECK-NEXT:    ret
   %shuffle = shufflevector <2 x i64> %a, <2 x i64> zeroinitializer, <2 x i32> <i32 1, i32 2>
   ret <2 x i64> %shuffle

llvm/test/CodeGen/LoongArch/lsx/widen-shuffle-mask.ll

@@ -28,10 +28,7 @@ define <16 x i8> @widen_shuffle_mask_v16i8_to_v4i32(<16 x i8> %a, <16 x i8> %b)
 define <16 x i8> @widen_shuffle_mask_v16i8_to_v2i64(<16 x i8> %a, <16 x i8> %b) {
 ; CHECK-LABEL: widen_shuffle_mask_v16i8_to_v2i64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI2_0)
-; CHECK-NEXT:    vld $vr2, $a0, %pc_lo12(.LCPI2_0)
-; CHECK-NEXT:    vshuf.d $vr2, $vr1, $vr0
-; CHECK-NEXT:    vori.b $vr0, $vr2, 0
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 12
 ; CHECK-NEXT:    ret
     %r = shufflevector <16 x i8> %a, <16 x i8> %b, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
     ret <16 x i8> %r
@@ -52,10 +49,7 @@ define <8 x i16> @widen_shuffle_mask_v8i16_to_v4i32(<8 x i16> %a, <8 x i16> %b)
 define <8 x i16> @widen_shuffle_mask_v8i16_to_v2i64(<8 x i16> %a, <8 x i16> %b) {
 ; CHECK-LABEL: widen_shuffle_mask_v8i16_to_v2i64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI4_0)
-; CHECK-NEXT:    vld $vr2, $a0, %pc_lo12(.LCPI4_0)
-; CHECK-NEXT:    vshuf.d $vr2, $vr1, $vr0
-; CHECK-NEXT:    vori.b $vr0, $vr2, 0
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 12
 ; CHECK-NEXT:    ret
     %r = shufflevector <8 x i16> %a, <8 x i16> %b, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 12, i32 13, i32 14, i32 15>
     ret <8 x i16> %r
@@ -64,10 +58,7 @@ define <8 x i16> @widen_shuffle_mask_v8i16_to_v2i64(<8 x i16> %a, <8 x i16> %b)
 define <4 x i32> @widen_shuffle_mask_v4i32_to_v2i64(<4 x i32> %a, <4 x i32> %b) {
 ; CHECK-LABEL: widen_shuffle_mask_v4i32_to_v2i64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pcalau12i $a0, %pc_hi20(.LCPI5_0)
-; CHECK-NEXT:    vld $vr2, $a0, %pc_lo12(.LCPI5_0)
-; CHECK-NEXT:    vshuf.d $vr2, $vr1, $vr0
-; CHECK-NEXT:    vori.b $vr0, $vr2, 0
+; CHECK-NEXT:    vshuf4i.d $vr0, $vr1, 12
 ; CHECK-NEXT:    ret
     %r = shufflevector <4 x i32> %a, <4 x i32> %b, <4 x i32> <i32 0, i32 1, i32 6, i32 7>
     ret <4 x i32> %r