X86: Add an SSE2 lowering for 64 bit compares when pcmpgtq (SSE4.2) isn't available.

This pattern started popping up in vectorized min/max reductions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179797 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 40 insertions, 15 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 3564ce39206..296bcf6ed33 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -9350,11 +9350,49 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
   if (Swap)
     std::swap(Op0, Op1);
 
+  // Since SSE has no unsigned integer comparisons, we need to flip  the sign
+  // bits of the inputs before performing those operations.
+  if (FlipSigns) {
+    EVT EltVT = VT.getVectorElementType();
+    SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
+                                      EltVT);
+    std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
+    SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
+                                    SignBits.size());
+    Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
+    Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
+  }
+
   // Check that the operation in question is available (most are plain SSE2,
   // but PCMPGTQ and PCMPEQQ have different requirements).
   if (VT == MVT::v2i64) {
-    if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42())
-      return SDValue();
+    if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
+      assert(Subtarget->hasSSE2() && "Don't know how to lower!");
+
+      // First cast everything to the right type,
+      Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
+      Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
+
+      // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
+      SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
+      SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
+
+      // Create masks for only the low parts/high parts of the 64 bit integers.
+      const int MaskHi[] = { 1, 1, 3, 3 };
+      const int MaskLo[] = { 0, 0, 2, 2 };
+      SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
+      SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
+      SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
+
+      SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
+      Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
+
+      if (Invert)
+        Result = DAG.getNOT(dl, Result, MVT::v4i32);
+
+      return DAG.getNode(ISD::BITCAST, dl, VT, Result);
+    }
+
     if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
       // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
       // pcmpeqd + pshufd + pand.
@@ -9379,19 +9417,6 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
     }
   }
 
-  // Since SSE has no unsigned integer comparisons, we need to flip  the sign
-  // bits of the inputs before performing those operations.
-  if (FlipSigns) {
-    EVT EltVT = VT.getVectorElementType();
-    SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
-                                      EltVT);
-    std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
-    SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
-                                    SignBits.size());
-    Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
-    Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
-  }
-
   SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
 
   // If the logical-not of the result is required, perform that now.