[X86] Improve lowering of (v2i64 (setgt X, -1)) on pre-SSE2 targets. Enable v2i64 in foldVectorXorShiftIntoCmp.

Similar to D72302 but for the canonical form for the opposite case. I've changed foldVectorXorShiftIntoCmp to form a target independent setcc node instead of PCMPGT now and enabled its for v2i64 on pre-SSE4.2 targets. The setcc should eventually get lowered to PCMPGT or the new v2i64 sequence.

Differential Revision: https://reviews.llvm.org/D72318

Author: topperc

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -21597,6 +21597,17 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget &Subtarget,
         return DAG.getBitcast(VT, Result);
       }
 
+      if (!FlipSigns && !Invert && ISD::isBuildVectorAllOnes(Op1.getNode())) {
+        Op0 = DAG.getBitcast(MVT::v4i32, Op0);
+        Op1 = DAG.getConstant(-1, dl, MVT::v4i32);
+
+        SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
+        static const int MaskHi[] = { 1, 1, 3, 3 };
+        SDValue Result = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
+
+        return DAG.getBitcast(VT, Result);
+      }
+
       // Since SSE has no unsigned integer comparisons, we need to flip the sign
       // bits of the inputs before performing those operations. The lower
       // compare is always unsigned.
@@ -40814,8 +40825,8 @@ static SDValue foldVectorXorShiftIntoCmp(SDNode *N, SelectionDAG &DAG,
   default: return SDValue();
   case MVT::v16i8:
   case MVT::v8i16:
-  case MVT::v4i32: if (!Subtarget.hasSSE2()) return SDValue(); break;
-  case MVT::v2i64: if (!Subtarget.hasSSE42()) return SDValue(); break;
+  case MVT::v4i32:
+  case MVT::v2i64: if (!Subtarget.hasSSE2()) return SDValue(); break;
   case MVT::v32i8:
   case MVT::v16i16:
   case MVT::v8i32:
@@ -40839,7 +40850,7 @@ static SDValue foldVectorXorShiftIntoCmp(SDNode *N, SelectionDAG &DAG,
 
   // Create a greater-than comparison against -1. We don't use the more obvious
   // greater-than-or-equal-to-zero because SSE/AVX don't have that instruction.
-  return DAG.getNode(X86ISD::PCMPGT, SDLoc(N), VT, Shift.getOperand(0), Ones);
+  return DAG.getSetCC(SDLoc(N), VT, Shift.getOperand(0), Ones, ISD::SETGT);
 }
 
 /// Detect patterns of truncation with unsigned saturation:

llvm/test/CodeGen/X86/vector-pcmp.ll

@@ -61,10 +61,9 @@ define <4 x i32> @test_pcmpgtd(<4 x i32> %x) {
 define <2 x i64> @test_pcmpgtq(<2 x i64> %x) {
 ; SSE2-LABEL: test_pcmpgtq:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE2-NEXT:    pxor %xmm1, %xmm0
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; SSE2-NEXT:    retq
 ;
 ; SSE42-LABEL: test_pcmpgtq:
@@ -187,13 +186,11 @@ define <8 x i32> @test_pcmpgtd_256(<8 x i32> %x) {
 define <4 x i64> @test_pcmpgtq_256(<4 x i64> %x) {
 ; SSE2-LABEL: test_pcmpgtq_256:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    psrad $31, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd %xmm2, %xmm2
-; SSE2-NEXT:    pxor %xmm2, %xmm0
-; SSE2-NEXT:    pxor %xmm2, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE2-NEXT:    retq
 ;
 ; SSE42-LABEL: test_pcmpgtq_256: