[AArch64] Eliminate redundant setcc on vector comparison results

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -4746,6 +4746,19 @@ SDValue TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
     }
   }
 
+  // setcc X, 0, setlt --> X  (when X is all sign bits)
+  //
+  // When we know that X has 0 or -1 in each lane, this comparison will produce
+  // X. This is only true when boolean contents are represented via 0s and -1s.
+  if (OpVT.isVector() && VT == OpVT &&
+      // Check that the result of setcc is 0 and -1.
+      getBooleanContents(VT) == ZeroOrNegativeOneBooleanContent &&
+      // Match only for checks X < 0
+      Cond == ISD::SETLT && isNullOrNullSplat(N1) &&
+      // The identity holds iff we know all sign bits for all lanes.
+      DAG.ComputeNumSignBits(N0) == N0.getScalarValueSizeInBits())
+    return N0;
+
   // FIXME: Support vectors.
   if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
     const APInt &C1 = N1C->getAPIntValue();

llvm/test/CodeGen/AArch64/setcc-redundant-cmlt.ll

@@ -0,0 +1,118 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=aarch64 < %s | FileCheck %s
+
+define <4 x i32> @direct_setcc_lt0(<4 x i32> %a, <4 x i32> %b, <4 x i32> %x, <4 x i32> %y) {
+; CHECK-LABEL: direct_setcc_lt0:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmgt v0.4s, v1.4s, v0.4s
+; CHECK-NEXT:    bsl v0.16b, v2.16b, v3.16b
+; CHECK-NEXT:    ret
+  %cmp = icmp slt <4 x i32> %a, %b
+  %sext = sext <4 x i1> %cmp to <4 x i32>
+  %lt0 = icmp slt <4 x i32> %sext, zeroinitializer
+  %sel = select <4 x i1> %lt0, <4 x i32> %x, <4 x i32> %y
+  ret <4 x i32> %sel
+}
+
+define <4 x i32> @shuffle_setcc_lt0(<4 x i32> %a, <4 x i32> %b, <4 x i32> %x, <4 x i32> %y) {
+; CHECK-LABEL: shuffle_setcc_lt0:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmgt v0.4s, v1.4s, v0.4s
+; CHECK-NEXT:    dup v0.4s, v0.s[2]
+; CHECK-NEXT:    bsl v0.16b, v2.16b, v3.16b
+; CHECK-NEXT:    ret
+  %cmp = icmp slt <4 x i32> %a, %b
+  %sext = sext <4 x i1> %cmp to <4 x i32>
+  %dup = shufflevector <4 x i32> %sext, <4 x i32> poison, <4 x i32> <i32 2, i32 2, i32 2, i32 2>
+  %lt0 = icmp slt <4 x i32> %dup, zeroinitializer
+  %sel = select <4 x i1> %lt0, <4 x i32> %x, <4 x i32> %y
+  ret <4 x i32> %sel
+}
+
+define <4 x i32> @direct_setcc_0gt(<4 x i32> %a, <4 x i32> %b, <4 x i32> %x, <4 x i32> %y) {
+; CHECK-LABEL: direct_setcc_0gt:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmgt v0.4s, v1.4s, v0.4s
+; CHECK-NEXT:    bsl v0.16b, v2.16b, v3.16b
+; CHECK-NEXT:    ret
+  %cmp = icmp slt <4 x i32> %a, %b
+  %sext = sext <4 x i1> %cmp to <4 x i32>
+  %gt0 = icmp sgt <4 x i32> zeroinitializer, %sext
+  %sel = select <4 x i1> %gt0, <4 x i32> %x, <4 x i32> %y
+  ret <4 x i32> %sel
+}
+
+define <8 x i16> @direct_setcc_lt0_v8i16(<8 x i16> %a, <8 x i16> %b, <8 x i16> %x, <8 x i16> %y) {
+; CHECK-LABEL: direct_setcc_lt0_v8i16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmgt v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    bsl v0.16b, v2.16b, v3.16b
+; CHECK-NEXT:    ret
+  %cmp = icmp slt <8 x i16> %a, %b
+  %sext = sext <8 x i1> %cmp to <8 x i16>
+  %lt0 = icmp slt <8 x i16> %sext, zeroinitializer
+  %sel = select <8 x i1> %lt0, <8 x i16> %x, <8 x i16> %y
+  ret <8 x i16> %sel
+}
+
+define <4 x i32> @non_splat_shuffle(<4 x i32> %a, <4 x i32> %b, <4 x i32> %x, <4 x i32> %y) {
+; CHECK-LABEL: non_splat_shuffle:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmgt v0.4s, v1.4s, v0.4s
+; CHECK-NEXT:    rev64 v0.4s, v0.4s
+; CHECK-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    bsl v0.16b, v2.16b, v3.16b
+; CHECK-NEXT:    ret
+  %cmp = icmp slt <4 x i32> %a, %b
+  %sext = sext <4 x i1> %cmp to <4 x i32>
+  %shuf = shufflevector <4 x i32> %sext, <4 x i32> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+  %lt0 = icmp slt <4 x i32> %shuf, zeroinitializer
+  %sel = select <4 x i1> %lt0, <4 x i32> %x, <4 x i32> %y
+  ret <4 x i32> %sel
+}
+
+define <16 x i8> @bitcast_narrow(<4 x i32> %a, <4 x i32> %b, <16 x i8> %x, <16 x i8> %y) {
+; CHECK-LABEL: bitcast_narrow:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmgt v0.4s, v1.4s, v0.4s
+; CHECK-NEXT:    bsl v0.16b, v2.16b, v3.16b
+; CHECK-NEXT:    ret
+  %cmp = icmp slt <4 x i32> %a, %b
+  %sext = sext <4 x i1> %cmp to <4 x i32>
+  %bc = bitcast <4 x i32> %sext to <16 x i8>
+  %lt0 = icmp slt <16 x i8> %bc, zeroinitializer
+  %sel = select <16 x i1> %lt0, <16 x i8> %x, <16 x i8> %y
+  ret <16 x i8> %sel
+}
+
+define <8 x i16> @chain_shuffle_bitcast(<4 x i32> %a, <4 x i32> %b, <8 x i16> %x, <8 x i16> %y) {
+; CHECK-LABEL: chain_shuffle_bitcast:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmgt v0.4s, v1.4s, v0.4s
+; CHECK-NEXT:    dup v0.4s, v0.s[2]
+; CHECK-NEXT:    bsl v0.16b, v2.16b, v3.16b
+; CHECK-NEXT:    ret
+  %cmp = icmp slt <4 x i32> %a, %b
+  %sext = sext <4 x i1> %cmp to <4 x i32>
+  %shuf = shufflevector <4 x i32> %sext, <4 x i32> poison, <4 x i32> <i32 2, i32 2, i32 2, i32 2>
+  %bc = bitcast <4 x i32> %shuf to <8 x i16>
+  %lt0 = icmp slt <8 x i16> %bc, zeroinitializer
+  %sel = select <8 x i1> %lt0, <8 x i16> %x, <8 x i16> %y
+  ret <8 x i16> %sel
+}
+
+; NEGATIVE TEST: Widening bitcast should NOT be optimized
+define <4 x i32> @bitcast_widen_negative(<16 x i8> %a, <16 x i8> %b, <4 x i32> %x, <4 x i32> %y) {
+; CHECK-LABEL: bitcast_widen_negative:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmgt v0.16b, v1.16b, v0.16b
+; CHECK-NEXT:    cmlt v0.4s, v0.4s, #0
+; CHECK-NEXT:    bsl v0.16b, v2.16b, v3.16b
+; CHECK-NEXT:    ret
+  %cmp = icmp slt <16 x i8> %a, %b
+  %sext = sext <16 x i1> %cmp to <16 x i8>
+  %bc = bitcast <16 x i8> %sext to <4 x i32>
+  %lt0 = icmp slt <4 x i32> %bc, zeroinitializer
+  %sel = select <4 x i1> %lt0, <4 x i32> %x, <4 x i32> %y
+  ret <4 x i32> %sel
+}