[X86] SimplifyDemandedBitsForTargetNode - add TESTPS/TESTPD support

We only need the sign bits from these nodes

Another step towards Issue llvm#60007

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -43742,6 +43742,20 @@ bool X86TargetLowering::SimplifyDemandedBitsForTargetNode(
       return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, SDLoc(Op), VT, NewSrc));
     return false;
   }
+  case X86ISD::TESTP: {
+    SDValue Op0 = Op.getOperand(0);
+    SDValue Op1 = Op.getOperand(1);
+    MVT OpVT = Op0.getSimpleValueType();
+    assert((OpVT.getVectorElementType() == MVT::f32 ||
+            OpVT.getVectorElementType() == MVT::f64) &&
+           "Illegal vector type for X86ISD::TESTP");
+
+    // TESTPS/TESTPD only demands the sign bits of ALL the elements.
+    KnownBits KnownSrc;
+    APInt SignMask = APInt::getSignMask(OpVT.getScalarSizeInBits());
+    return SimplifyDemandedBits(Op0, SignMask, KnownSrc, TLO, Depth + 1) ||
+           SimplifyDemandedBits(Op1, SignMask, KnownSrc, TLO, Depth + 1);
+  }
   case X86ISD::BEXTR:
   case X86ISD::BEXTRI: {
     SDValue Op0 = Op.getOperand(0);
@@ -54658,6 +54672,21 @@ static SDValue combineMOVMSK(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+static SDValue combineTESTP(SDNode *N, SelectionDAG &DAG,
+                            TargetLowering::DAGCombinerInfo &DCI,
+                            const X86Subtarget &Subtarget) {
+  MVT VT = N->getSimpleValueType(0);
+  unsigned NumBits = VT.getScalarSizeInBits();
+
+  // Simplify the inputs.
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  APInt DemandedMask(APInt::getAllOnes(NumBits));
+  if (TLI.SimplifyDemandedBits(SDValue(N, 0), DemandedMask, DCI))
+    return SDValue(N, 0);
+
+  return SDValue();
+}
+
 static SDValue combineX86GatherScatter(SDNode *N, SelectionDAG &DAG,
                                        TargetLowering::DAGCombinerInfo &DCI,
                                        const X86Subtarget &Subtarget) {
@@ -57545,6 +57574,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case X86ISD::FMADDSUB:
   case X86ISD::FMSUBADD:    return combineFMADDSUB(N, DAG, DCI);
   case X86ISD::MOVMSK:      return combineMOVMSK(N, DAG, DCI, Subtarget);
+  case X86ISD::TESTP:       return combineTESTP(N, DAG, DCI, Subtarget);
   case X86ISD::MGATHER:
   case X86ISD::MSCATTER:
     return combineX86GatherScatter(N, DAG, DCI, Subtarget);

llvm/test/CodeGen/X86/combine-testpd.ll

@@ -157,8 +157,6 @@ define i32 @testpdc_128_signbit(<2 x double> %c, <2 x double> %d, i32 %a, i32 %b
 ; CHECK-LABEL: testpdc_128_signbit:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movl %edi, %eax
-; CHECK-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; CHECK-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm0
 ; CHECK-NEXT:    vtestpd %xmm1, %xmm0
 ; CHECK-NEXT:    cmovael %esi, %eax
 ; CHECK-NEXT:    retq
@@ -175,11 +173,6 @@ define i32 @testpdz_256_signbit(<4 x double> %c, <4 x double> %d, i32 %a, i32 %b
 ; CHECK-LABEL: testpdz_256_signbit:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movl %edi, %eax
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; CHECK-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; CHECK-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
-; CHECK-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
 ; CHECK-NEXT:    vtestpd %ymm1, %ymm0
 ; CHECK-NEXT:    cmovnel %esi, %eax
 ; CHECK-NEXT:    vzeroupper

llvm/test/CodeGen/X86/combine-testps.ll

@@ -157,7 +157,6 @@ define i32 @testpsz_128_signbit(<4 x float> %c, <4 x float> %d, i32 %a, i32 %b)
 ; CHECK-LABEL: testpsz_128_signbit:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movl %edi, %eax
-; CHECK-NEXT:    vpsrad $31, %xmm0, %xmm0
 ; CHECK-NEXT:    vtestps %xmm1, %xmm0
 ; CHECK-NEXT:    cmovnel %esi, %eax
 ; CHECK-NEXT:    retq
@@ -174,11 +173,6 @@ define i32 @testpsnzc_256_signbit(<8 x float> %c, <8 x float> %d, i32 %a, i32 %b
 ; CHECK-LABEL: testpsnzc_256_signbit:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movl %edi, %eax
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; CHECK-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; CHECK-NEXT:    vpcmpgtd %xmm2, %xmm3, %xmm2
-; CHECK-NEXT:    vpcmpgtd %xmm0, %xmm3, %xmm0
-; CHECK-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
 ; CHECK-NEXT:    vtestps %ymm1, %ymm0
 ; CHECK-NEXT:    cmovnel %esi, %eax
 ; CHECK-NEXT:    vzeroupper