[X86] Turn setne X, signedmax into setgt signedmax, X in LowerVSETCC to avoid an...

We won't be able to fold the constant pool load, but its still better than materialing ones and xoring for the invert if we used PCMPEQ.

This will fix another regression from D42948.

llvm-svn: 325845

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 5dd0b9ea189800bfa55a2b620362543dbef8166e..0f9f9bf1b833a3b8a586d39686a60b0135bf5cb5 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -18100,12 +18100,16 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget &Subtarget,
   }
 
   // If this is a SETNE against the signed minimum value, change it to SETGT.
+  // If this is a SETNE against the signed maximum value, change it to SETLT 
+  // which will be swapped to SETGT.
   // Otherwise we use PCMPEQ+invert.
   APInt ConstValue;
   if (Cond == ISD::SETNE &&
-      ISD::isConstantSplatVector(Op1.getNode(), ConstValue),
-      ConstValue.isMinSignedValue()) {
-    Cond = ISD::SETGT;
+      ISD::isConstantSplatVector(Op1.getNode(), ConstValue)) {
+    if (ConstValue.isMinSignedValue())
+      Cond = ISD::SETGT;
+    else if (ConstValue.isMaxSignedValue())
+      Cond = ISD::SETLT;
   }
 
   // If both operands are known non-negative, then an unsigned compare is the

diff --git a/llvm/test/CodeGen/X86/vector-compare-simplify.ll b/llvm/test/CodeGen/X86/vector-compare-simplify.ll
index f1ac60134eb80d25544e76b168bf27394589839e..e7f1d849b2f7b7a5ac799b886511ff10f7ec4c98 100644 (file)
--- a/llvm/test/CodeGen/X86/vector-compare-simplify.ll
+++ b/llvm/test/CodeGen/X86/vector-compare-simplify.ll
@@ -345,3 +345,18 @@ define <4 x i32> @ne_smin(<4 x i32> %x) {
   ret <4 x i32> %r
 }
 
+; Make sure we can efficiently handle ne smax by turning into sgt. We can't fold
+; the constant pool load, but the alternative is a cmpeq+invert which is 3 instructions.
+; The PCMPGT version is two instructions given sufficient register allocation freedom
+; to avoid the last mov to %xmm0 seen here.
+define <4 x i32> @ne_smax(<4 x i32> %x) {
+; CHECK-LABEL: ne_smax:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movdqa {{.*#+}} xmm1 = [2147483647,2147483647,2147483647,2147483647]
+; CHECK-NEXT:    pcmpgtd %xmm0, %xmm1
+; CHECK-NEXT:    movdqa %xmm1, %xmm0
+; CHECK-NEXT:    retq
+  %cmp = icmp ne <4 x i32> %x, <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647>
+  %r = sext <4 x i1> %cmp to <4 x i32>
+  ret <4 x i32> %r
+}