X86: Custom lower <2 x i64> eq and ne when SSE41 is not available.

pcmpeqd, pshufd, pshufd, pand is cheaper than unpack + cmpq, sbbq, cmpq, sbbq + pack.
Small speedup on loop-vectorized viterbi (-march=core2).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@171063 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 50 insertions, 2 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 5f7f915097..a173712b5b 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -9171,8 +9171,30 @@ SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
   if (VT == MVT::v2i64) {
     if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42())
       return SDValue();
-    if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41())
-      return SDValue();
+    if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
+      // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
+      // pcmpeqd + 2 shuffles + pand.
+      assert(Subtarget->hasSSE2() && !FlipSigns && "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);
+
+      // Do the compare.
+      SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
+
+      // Make sure the lower and upper halves are both all-ones.
+      const int Mask1[] = { 0, 0, 2, 2 };
+      SDValue S1 = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask1);
+      const int Mask2[] = { 1, 1, 3, 3 };
+      SDValue S2 = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask2);
+      Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, S1, S2);
+
+      if (Invert)
+        Result = DAG.getNOT(dl, Result, MVT::v4i32);
+
+      return DAG.getNode(ISD::BITCAST, dl, VT, Result);
+    }
   }
 
   // Since SSE has no unsigned integer comparisons, we need to flip  the sign
diff --git a/test/CodeGen/X86/vec_compare.ll b/test/CodeGen/X86/vec_compare.ll
index 367dd27f30..cf86c737c7 100644
--- a/test/CodeGen/X86/vec_compare.ll
+++ b/test/CodeGen/X86/vec_compare.ll
@@ -41,3 +41,29 @@ define <4 x i32> @test4(<4 x i32> %A, <4 x i32> %B) nounwind {
         %D = sext <4 x i1> %C to <4 x i32>
 	ret <4 x i32> %D
 }
+
+define <2 x i64> @test5(<2 x i64> %A, <2 x i64> %B) nounwind {
+; CHECK: test5:
+; CHECK: pcmpeqd
+; CHECK: pshufd $-11
+; CHECK: pshufd $-96
+; CHECK: pand
+; CHECK: ret
+	%C = icmp eq <2 x i64> %A, %B
+	%D = sext <2 x i1> %C to <2 x i64>
+	ret <2 x i64> %D
+}
+
+define <2 x i64> @test6(<2 x i64> %A, <2 x i64> %B) nounwind {
+; CHECK: test6:
+; CHECK: pcmpeqd
+; CHECK: pshufd $-11
+; CHECK: pshufd $-96
+; CHECK: pand
+; CHECK: pcmpeqd
+; CHECK: pxor
+; CHECK: ret
+	%C = icmp ne <2 x i64> %A, %B
+	%D = sext <2 x i1> %C to <2 x i64>
+	ret <2 x i64> %D
+}