[WIP][DAG] combineVSelectWithAllOnesOrZeros - fold "select Cond, 0, x -> and not(Cond), x"

Extend #145298 to remove the x86 combineVSelectWithLastZeros special case

WIP - still a couple of x86 regressions to address wrt unnecessary duplicated comparisons

Fixes #144513

Author: RKSimon

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -13159,6 +13159,15 @@ static SDValue combineVSelectWithAllOnesOrZeros(SDValue Cond, SDValue TVal,
     return DAG.getBitcast(VT, And);
   }
 
+  // select Cond, 0, x -> and not(Cond), x
+  if (IsTAllZero &&
+      (isBitwiseNot(peekThroughBitcasts(Cond)) || TLI.hasAndNot(Cond))) {
+    SDValue X = DAG.getBitcast(CondVT, FVal);
+    SDValue And =
+        DAG.getNode(ISD::AND, DL, CondVT, DAG.getNOT(DL, Cond, CondVT), X);
+    return DAG.getBitcast(VT, And);
+  }
+
   return SDValue();
 }
 

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -47262,57 +47262,6 @@ static SDValue combineToExtendBoolVectorInReg(
                      DAG.getConstant(EltSizeInBits - 1, DL, VT));
 }
 
-/// If a vector select has an left operand that is 0, try to simplify the
-/// select to a bitwise logic operation.
-/// TODO: Move to DAGCombiner.combineVSelectWithAllOnesOrZeros, possibly using
-/// TargetLowering::hasAndNot()?
-static SDValue combineVSelectWithLastZeros(SDNode *N, SelectionDAG &DAG,
-                                           const SDLoc &DL,
-                                           TargetLowering::DAGCombinerInfo &DCI,
-                                           const X86Subtarget &Subtarget) {
-  SDValue Cond = N->getOperand(0);
-  SDValue LHS = N->getOperand(1);
-  SDValue RHS = N->getOperand(2);
-  EVT VT = LHS.getValueType();
-  EVT CondVT = Cond.getValueType();
-  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
-
-  if (N->getOpcode() != ISD::VSELECT)
-    return SDValue();
-
-  assert(CondVT.isVector() && "Vector select expects a vector selector!");
-
-  // To use the condition operand as a bitwise mask, it must have elements that
-  // are the same size as the select elements. Ie, the condition operand must
-  // have already been promoted from the IR select condition type <N x i1>.
-  // Don't check if the types themselves are equal because that excludes
-  // vector floating-point selects.
-  if (CondVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
-    return SDValue();
-
-  // Cond value must be 'sign splat' to be converted to a logical op.
-  if (DAG.ComputeNumSignBits(Cond) != CondVT.getScalarSizeInBits())
-    return SDValue();
-
-  if (!TLI.isTypeLegal(CondVT))
-    return SDValue();
-
-  // vselect Cond, 000..., X -> andn Cond, X
-  if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
-    SDValue CastRHS = DAG.getBitcast(CondVT, RHS);
-    SDValue AndN;
-    // The canonical form differs for i1 vectors - x86andnp is not used
-    if (CondVT.getScalarType() == MVT::i1)
-      AndN = DAG.getNode(ISD::AND, DL, CondVT, DAG.getNOT(DL, Cond, CondVT),
-                         CastRHS);
-    else
-      AndN = DAG.getNode(X86ISD::ANDNP, DL, CondVT, Cond, CastRHS);
-    return DAG.getBitcast(VT, AndN);
-  }
-
-  return SDValue();
-}
-
 /// If both arms of a vector select are concatenated vectors, split the select,
 /// and concatenate the result to eliminate a wide (256-bit) vector instruction:
 ///   vselect Cond, (concat T0, T1), (concat F0, F1) -->
@@ -48059,9 +48008,6 @@ static SDValue combineSelect(SDNode *N, SelectionDAG &DAG,
   if (!TLI.isTypeLegal(VT) || isSoftF16(VT, Subtarget))
     return SDValue();
 
-  if (SDValue V = combineVSelectWithLastZeros(N, DAG, DL, DCI, Subtarget))
-    return V;
-
   if (SDValue V = combineVSelectToBLENDV(N, DAG, DL, DCI, Subtarget))
     return V;
 

llvm/test/CodeGen/AArch64/sve-fixed-length-shuffles.ll

@@ -30,7 +30,9 @@ define void @crash_when_lowering_extract_shuffle(ptr %dst, i1 %cond) vscale_rang
 ; CHECK-NEXT:  // %bb.1: // %vector.body
 ; CHECK-NEXT:    movi v0.2d, #0000000000000000
 ; CHECK-NEXT:    movi v1.2d, #0000000000000000
-; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    ldr z4, [x0]
+; CHECK-NEXT:    ldr z5, [x0, #2, mul vl]
+; CHECK-NEXT:    ldr z6, [x0, #3, mul vl]
 ; CHECK-NEXT:    umov w8, v0.b[8]
 ; CHECK-NEXT:    mov v1.b[1], v0.b[1]
 ; CHECK-NEXT:    fmov s2, w8
@@ -60,31 +62,20 @@ define void @crash_when_lowering_extract_shuffle(ptr %dst, i1 %cond) vscale_rang
 ; CHECK-NEXT:    asr z1.s, z1.s, #31
 ; CHECK-NEXT:    uunpklo z3.s, z3.h
 ; CHECK-NEXT:    lsl z0.s, z0.s, #31
-; CHECK-NEXT:    and z1.s, z1.s, #0x1
+; CHECK-NEXT:    bic z1.d, z4.d, z1.d
 ; CHECK-NEXT:    lsl z2.s, z2.s, #31
+; CHECK-NEXT:    ldr z4, [x0, #1, mul vl]
 ; CHECK-NEXT:    asr z0.s, z0.s, #31
-; CHECK-NEXT:    cmpne p1.s, p0/z, z1.s, #0
-; CHECK-NEXT:    ldr z1, [x0]
+; CHECK-NEXT:    str z1, [x0]
 ; CHECK-NEXT:    lsl z3.s, z3.s, #31
 ; CHECK-NEXT:    asr z2.s, z2.s, #31
-; CHECK-NEXT:    and z0.s, z0.s, #0x1
+; CHECK-NEXT:    bic z0.d, z5.d, z0.d
 ; CHECK-NEXT:    asr z3.s, z3.s, #31
-; CHECK-NEXT:    and z2.s, z2.s, #0x1
-; CHECK-NEXT:    mov z1.s, p1/m, #0 // =0x0
-; CHECK-NEXT:    cmpne p2.s, p0/z, z0.s, #0
-; CHECK-NEXT:    ldr z0, [x0, #2, mul vl]
-; CHECK-NEXT:    and z3.s, z3.s, #0x1
-; CHECK-NEXT:    str z1, [x0]
-; CHECK-NEXT:    cmpne p3.s, p0/z, z3.s, #0
-; CHECK-NEXT:    cmpne p0.s, p0/z, z2.s, #0
-; CHECK-NEXT:    ldr z3, [x0, #3, mul vl]
-; CHECK-NEXT:    ldr z2, [x0, #1, mul vl]
-; CHECK-NEXT:    mov z0.s, p2/m, #0 // =0x0
-; CHECK-NEXT:    mov z3.s, p3/m, #0 // =0x0
-; CHECK-NEXT:    mov z2.s, p0/m, #0 // =0x0
+; CHECK-NEXT:    bic z1.d, z4.d, z2.d
 ; CHECK-NEXT:    str z0, [x0, #2, mul vl]
+; CHECK-NEXT:    bic z3.d, z6.d, z3.d
+; CHECK-NEXT:    str z1, [x0, #1, mul vl]
 ; CHECK-NEXT:    str z3, [x0, #3, mul vl]
-; CHECK-NEXT:    str z2, [x0, #1, mul vl]
 ; CHECK-NEXT:  .LBB1_2: // %exit
 ; CHECK-NEXT:    ret
   %broadcast.splat = shufflevector <32 x i1> zeroinitializer, <32 x i1> zeroinitializer, <32 x i32> zeroinitializer

llvm/test/CodeGen/AArch64/vselect-constants.ll

@@ -169,11 +169,9 @@ define <4 x i32> @cmp_sel_1_or_0_vec(<4 x i32> %x, <4 x i32> %y) {
 define <4 x i32> @sel_0_or_1_vec(<4 x i1> %cond) {
 ; CHECK-LABEL: sel_0_or_1_vec:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ushll v0.4s, v0.4h, #0
 ; CHECK-NEXT:    movi v1.4s, #1
-; CHECK-NEXT:    shl v0.4s, v0.4s, #31
-; CHECK-NEXT:    cmge v0.4s, v0.4s, #0
-; CHECK-NEXT:    and v0.16b, v0.16b, v1.16b
+; CHECK-NEXT:    ushll v0.4s, v0.4h, #0
+; CHECK-NEXT:    bic v0.16b, v1.16b, v0.16b
 ; CHECK-NEXT:    ret
   %add = select <4 x i1> %cond, <4 x i32> <i32 0, i32 0, i32 0, i32 0>, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
   ret <4 x i32> %add