[LoongArch] Add patterns for [x]vclo.{b/h/w/d} instructions

Author: zhaoqi5

llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td

@@ -1443,6 +1443,12 @@ defm : PatXrXr<sra, "XVSRA">;
 defm : PatShiftXrXr<sra, "XVSRA">;
 defm : PatShiftXrSplatUimm<sra, "XVSRAI">;
 
+// XVCLO_{B/H/W/D}
+def : Pat<(ctlz (vnot v32i8:$xj)), (XVCLO_B v32i8:$xj)>;
+def : Pat<(ctlz (vnot v16i16:$xj)), (XVCLO_H v16i16:$xj)>;
+def : Pat<(ctlz (vnot v8i32:$xj)), (XVCLO_W v8i32:$xj)>;
+def : Pat<(ctlz (vnot v4i64:$xj)), (XVCLO_D v4i64:$xj)>;
+
 // XVCLZ_{B/H/W/D}
 defm : PatXr<ctlz, "XVCLZ">;
 

llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td

@@ -1645,6 +1645,12 @@ defm : PatVrVr<sra, "VSRA">;
 defm : PatShiftVrVr<sra, "VSRA">;
 defm : PatShiftVrSplatUimm<sra, "VSRAI">;
 
+// VCLO_{B/H/W/D}
+def : Pat<(ctlz (vnot v16i8:$vj)), (VCLO_B v16i8:$vj)>;
+def : Pat<(ctlz (vnot v8i16:$vj)), (VCLO_H v8i16:$vj)>;
+def : Pat<(ctlz (vnot v4i32:$vj)), (VCLO_W v4i32:$vj)>;
+def : Pat<(ctlz (vnot v2i64:$vj)), (VCLO_D v2i64:$vj)>;
+
 // VCLZ_{B/H/W/D}
 defm : PatVr<ctlz, "VCLZ">;
 

llvm/test/CodeGen/LoongArch/lasx/ctpop-ctlz.ll

@@ -110,8 +110,7 @@ define void @not_ctlz_v32i8(ptr %src, ptr %dst) nounwind {
 ; CHECK-LABEL: not_ctlz_v32i8:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xvld $xr0, $a0, 0
-; CHECK-NEXT:    xvxori.b $xr0, $xr0, 255
-; CHECK-NEXT:    xvclz.b $xr0, $xr0
+; CHECK-NEXT:    xvclo.b $xr0, $xr0
 ; CHECK-NEXT:    xvst $xr0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load <32 x i8>, ptr %src
@@ -125,9 +124,7 @@ define void @not_ctlz_v16i16(ptr %src, ptr %dst) nounwind {
 ; CHECK-LABEL: not_ctlz_v16i16:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xvld $xr0, $a0, 0
-; CHECK-NEXT:    xvrepli.b $xr1, -1
-; CHECK-NEXT:    xvxor.v $xr0, $xr0, $xr1
-; CHECK-NEXT:    xvclz.h $xr0, $xr0
+; CHECK-NEXT:    xvclo.h $xr0, $xr0
 ; CHECK-NEXT:    xvst $xr0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load <16 x i16>, ptr %src
@@ -141,9 +138,7 @@ define void @not_ctlz_v8i32(ptr %src, ptr %dst) nounwind {
 ; CHECK-LABEL: not_ctlz_v8i32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xvld $xr0, $a0, 0
-; CHECK-NEXT:    xvrepli.b $xr1, -1
-; CHECK-NEXT:    xvxor.v $xr0, $xr0, $xr1
-; CHECK-NEXT:    xvclz.w $xr0, $xr0
+; CHECK-NEXT:    xvclo.w $xr0, $xr0
 ; CHECK-NEXT:    xvst $xr0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load <8 x i32>, ptr %src
@@ -157,9 +152,7 @@ define void @not_ctlz_v4i64(ptr %src, ptr %dst) nounwind {
 ; CHECK-LABEL: not_ctlz_v4i64:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xvld $xr0, $a0, 0
-; CHECK-NEXT:    xvrepli.b $xr1, -1
-; CHECK-NEXT:    xvxor.v $xr0, $xr0, $xr1
-; CHECK-NEXT:    xvclz.d $xr0, $xr0
+; CHECK-NEXT:    xvclo.d $xr0, $xr0
 ; CHECK-NEXT:    xvst $xr0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load <4 x i64>, ptr %src

llvm/test/CodeGen/LoongArch/lsx/ctpop-ctlz.ll

@@ -110,8 +110,7 @@ define void @not_ctlz_v16i8(ptr %src, ptr %dst) nounwind {
 ; CHECK-LABEL: not_ctlz_v16i8:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vld $vr0, $a0, 0
-; CHECK-NEXT:    vxori.b $vr0, $vr0, 255
-; CHECK-NEXT:    vclz.b $vr0, $vr0
+; CHECK-NEXT:    vclo.b $vr0, $vr0
 ; CHECK-NEXT:    vst $vr0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load <16 x i8>, ptr %src
@@ -125,9 +124,7 @@ define void @not_ctlz_v8i16(ptr %src, ptr %dst) nounwind {
 ; CHECK-LABEL: not_ctlz_v8i16:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vld $vr0, $a0, 0
-; CHECK-NEXT:    vrepli.b $vr1, -1
-; CHECK-NEXT:    vxor.v $vr0, $vr0, $vr1
-; CHECK-NEXT:    vclz.h $vr0, $vr0
+; CHECK-NEXT:    vclo.h $vr0, $vr0
 ; CHECK-NEXT:    vst $vr0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load <8 x i16>, ptr %src
@@ -141,9 +138,7 @@ define void @not_ctlz_v4i32(ptr %src, ptr %dst) nounwind {
 ; CHECK-LABEL: not_ctlz_v4i32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vld $vr0, $a0, 0
-; CHECK-NEXT:    vrepli.b $vr1, -1
-; CHECK-NEXT:    vxor.v $vr0, $vr0, $vr1
-; CHECK-NEXT:    vclz.w $vr0, $vr0
+; CHECK-NEXT:    vclo.w $vr0, $vr0
 ; CHECK-NEXT:    vst $vr0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load <4 x i32>, ptr %src
@@ -157,9 +152,7 @@ define void @not_ctlz_v2i64(ptr %src, ptr %dst) nounwind {
 ; CHECK-LABEL: not_ctlz_v2i64:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vld $vr0, $a0, 0
-; CHECK-NEXT:    vrepli.b $vr1, -1
-; CHECK-NEXT:    vxor.v $vr0, $vr0, $vr1
-; CHECK-NEXT:    vclz.d $vr0, $vr0
+; CHECK-NEXT:    vclo.d $vr0, $vr0
 ; CHECK-NEXT:    vst $vr0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load <2 x i64>, ptr %src