[AArch64][GlobalISel] Select G_ICMP Zero Instruction

Author: chuongg3

llvm/lib/Target/AArch64/AArch64InstrInfo.td

@@ -5450,6 +5450,52 @@ defm : SelectSetCCSwapOperands<setle, "CMGE">;
 defm : SelectSetCCSwapOperands<setult, "CMHI">;
 defm : SelectSetCCSwapOperands<setule, "CMHS">;
 
+multiclass SelectSetCCZeroRHS<PatFrags InFrag, string INST> {
+  def : Pat<(v8i8 (InFrag (v8i8 V64:$Rn), immAllZerosV)),
+            (v8i8 (!cast<Instruction>(INST # v8i8rz) (v8i8 V64:$Rn)))>;
+  def : Pat<(v16i8 (InFrag (v16i8 V128:$Rn), immAllZerosV)),
+            (v16i8 (!cast<Instruction>(INST # v16i8rz) (v16i8 V128:$Rn)))>;
+  def : Pat<(v4i16 (InFrag (v4i16 V64:$Rn), immAllZerosV)),
+            (v4i16 (!cast<Instruction>(INST # v4i16rz) (v4i16 V64:$Rn)))>;
+  def : Pat<(v8i16 (InFrag (v8i16 V128:$Rn), immAllZerosV)),
+            (v8i16 (!cast<Instruction>(INST # v8i16rz) (v8i16 V128:$Rn)))>;
+  def : Pat<(v2i32 (InFrag (v2i32 V64:$Rn), immAllZerosV)),
+            (v2i32 (!cast<Instruction>(INST # v2i32rz) (v2i32 V64:$Rn)))>;
+  def : Pat<(v4i32 (InFrag (v4i32 V128:$Rn), immAllZerosV)),
+            (v4i32 (!cast<Instruction>(INST # v4i32rz) (v4i32 V128:$Rn)))>;
+  def : Pat<(v2i64 (InFrag (v2i64 V128:$Rn), immAllZerosV)),
+            (v2i64 (!cast<Instruction>(INST # v2i64rz) (v2i64 V128:$Rn)))>;
+}
+
+defm : SelectSetCCZeroRHS<seteq, "CMEQ">;
+defm : SelectSetCCZeroRHS<setgt, "CMGT">;
+defm : SelectSetCCZeroRHS<setge, "CMGE">;
+defm : SelectSetCCZeroRHS<setlt, "CMLT">;
+defm : SelectSetCCZeroRHS<setle, "CMLE">;
+
+multiclass SelectSetCCZeroLHS<PatFrags InFrag, string INST> {
+  def : Pat<(v8i8 (InFrag immAllZerosV, (v8i8 V64:$Rn))),
+            (v8i8 (!cast<Instruction>(INST # v8i8rz) (v8i8 V64:$Rn)))>;
+  def : Pat<(v16i8 (InFrag immAllZerosV, (v16i8 V128:$Rn))),
+            (v16i8 (!cast<Instruction>(INST # v16i8rz) (v16i8 V128:$Rn)))>;
+  def : Pat<(v4i16 (InFrag immAllZerosV, (v4i16 V64:$Rn))),
+            (v4i16 (!cast<Instruction>(INST # v4i16rz) (v4i16 V64:$Rn)))>;
+  def : Pat<(v8i16 (InFrag immAllZerosV, (v8i16 V128:$Rn))),
+            (v8i16 (!cast<Instruction>(INST # v8i16rz) (v8i16 V128:$Rn)))>;
+  def : Pat<(v2i32 (InFrag immAllZerosV, (v2i32 V64:$Rn))),
+            (v2i32 (!cast<Instruction>(INST # v2i32rz) (v2i32 V64:$Rn)))>;
+  def : Pat<(v4i32 (InFrag immAllZerosV, (v4i32 V128:$Rn))),
+            (v4i32 (!cast<Instruction>(INST # v4i32rz) (v4i32 V128:$Rn)))>;
+  def : Pat<(v2i64 (InFrag immAllZerosV, (v2i64 V128:$Rn))),
+            (v2i64 (!cast<Instruction>(INST # v2i64rz) (v2i64 V128:$Rn)))>;
+}
+
+defm : SelectSetCCZeroLHS<seteq, "CMEQ">;
+defm : SelectSetCCZeroLHS<setgt, "CMLT">;
+defm : SelectSetCCZeroLHS<setge, "CMLE">;
+defm : SelectSetCCZeroLHS<setlt, "CMGT">;
+defm : SelectSetCCZeroLHS<setle, "CMGE">;
+
 let Predicates = [HasNEON] in {
 def : InstAlias<"mov{\t$dst.16b, $src.16b|.16b\t$dst, $src}",
                 (ORRv16i8 V128:$dst, V128:$src, V128:$src), 1>;

llvm/test/CodeGen/AArch64/aarch64-addv.ll

@@ -94,20 +94,19 @@ define i32 @oversized_ADDV_256(ptr noalias nocapture readonly %arg1, ptr noalias
 ;
 ; GISEL-LABEL: oversized_ADDV_256:
 ; GISEL:       // %bb.0: // %entry
-; GISEL-NEXT:    ldr d1, [x0]
-; GISEL-NEXT:    ldr d2, [x1]
-; GISEL-NEXT:    movi v0.2d, #0000000000000000
+; GISEL-NEXT:    ldr d0, [x0]
+; GISEL-NEXT:    ldr d1, [x1]
+; GISEL-NEXT:    ushll v0.8h, v0.8b, #0
 ; GISEL-NEXT:    ushll v1.8h, v1.8b, #0
-; GISEL-NEXT:    ushll v2.8h, v2.8b, #0
-; GISEL-NEXT:    usubl v3.4s, v1.4h, v2.4h
-; GISEL-NEXT:    usubl2 v1.4s, v1.8h, v2.8h
-; GISEL-NEXT:    cmgt v2.4s, v0.4s, v3.4s
-; GISEL-NEXT:    cmgt v0.4s, v0.4s, v1.4s
-; GISEL-NEXT:    neg v4.4s, v3.4s
-; GISEL-NEXT:    neg v5.4s, v1.4s
-; GISEL-NEXT:    bsl v2.16b, v4.16b, v3.16b
-; GISEL-NEXT:    bsl v0.16b, v5.16b, v1.16b
-; GISEL-NEXT:    add v0.4s, v2.4s, v0.4s
+; GISEL-NEXT:    usubl v2.4s, v0.4h, v1.4h
+; GISEL-NEXT:    usubl2 v0.4s, v0.8h, v1.8h
+; GISEL-NEXT:    cmlt v1.4s, v2.4s, #0
+; GISEL-NEXT:    cmlt v3.4s, v0.4s, #0
+; GISEL-NEXT:    neg v4.4s, v2.4s
+; GISEL-NEXT:    neg v5.4s, v0.4s
+; GISEL-NEXT:    bsl v1.16b, v4.16b, v2.16b
+; GISEL-NEXT:    bit v0.16b, v5.16b, v3.16b
+; GISEL-NEXT:    add v0.4s, v1.4s, v0.4s
 ; GISEL-NEXT:    addv s0, v0.4s
 ; GISEL-NEXT:    fmov w0, s0
 ; GISEL-NEXT:    ret