[X86] Don't emit KTEST instructions unless only the Z flag is being used

Summary:
KTEST has weird flag behavior. The Z flag is set for all bits in the AND of the k-registers being 0, and the C flag is set for all bits being 1. All other flags are cleared.

We currently emit this instruction in EmitTEST and don't check the condition code. This can lead to strange things like using the S flag after a KTEST for a signed compare.

The domain reassignment pass can also transform TEST instructions into KTEST and is not protected against the flag usage either. For now I've disabled this part of the domain reassignment pass. I tried to comment out the checks in the mir test so that we could recover them later, but I couldn't figure out how to get that to work.

This patch moves the KTEST handling into LowerSETCC and now creates a ktest+x86setcc. I've chosen this approach because I'd like to add support for the C flag for all ones in a followup patch. To do that requires that I can rewrite the condition code going in the x86setcc to be different than the original SETCC condition code.

This fixes PR36182. I'll file a PR to fix domain reassignment once this goes in. Should this be merged to 6.0?

Reviewers: spatel, guyblank, RKSimon, zvi

Reviewed By: guyblank

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D42770

llvm-svn: 324576

1 files changed, 32 insertions, 25 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index bbd04249734..6cc185be1c8 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -17114,24 +17114,6 @@ static bool hasNonFlagsUse(SDValue Op) {
   return false;
 }
 
-// Emit KTEST instruction for bit vectors on AVX-512
-static SDValue EmitKTEST(SDValue Op, SelectionDAG &DAG,
-                         const X86Subtarget &Subtarget) {
-  if (Op.getOpcode() == ISD::BITCAST) {
-    auto hasKTEST = [&](MVT VT) {
-      unsigned SizeInBits = VT.getSizeInBits();
-      return (Subtarget.hasDQI() && (SizeInBits == 8 || SizeInBits == 16)) ||
-        (Subtarget.hasBWI() && (SizeInBits == 32 || SizeInBits == 64));
-    };
-    SDValue Op0 = Op.getOperand(0);
-    MVT Op0VT = Op0.getValueType().getSimpleVT();
-    if (Op0VT.isVector() && Op0VT.getVectorElementType() == MVT::i1 &&
-        hasKTEST(Op0VT))
-      return DAG.getNode(X86ISD::KTEST, SDLoc(Op), Op0VT, Op0, Op0);
-  }
-  return SDValue();
-}
-
 /// Emit nodes that will be selected as "test Op0,Op0", or something
 /// equivalent.
 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, const SDLoc &dl,
@@ -17171,9 +17153,6 @@ SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, const SDLoc &dl,
   // doing a separate TEST. TEST always sets OF and CF to 0, so unless
   // we prove that the arithmetic won't overflow, we can't use OF or CF.
   if (Op.getResNo() != 0 || NeedOF || NeedCF) {
-    // Emit KTEST for bit vectors
-    if (auto Node = EmitKTEST(Op, DAG, Subtarget))
-      return Node;
     // Emit a CMP with 0, which is the TEST pattern.
     return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
                        DAG.getConstant(0, dl, Op.getValueType()));
@@ -17396,10 +17375,6 @@ SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, const SDLoc &dl,
   }
 
   if (Opcode == 0) {
-    // Emit KTEST for bit vectors
-    if (auto Node = EmitKTEST(Op, DAG, Subtarget))
-      return Node;
-
     // Emit a CMP with 0, which is the TEST pattern.
     return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
                        DAG.getConstant(0, dl, Op.getValueType()));
@@ -18146,6 +18121,34 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget &Subtarget,
   return Result;
 }
 
+// Try to select this as a KTEST+SETCC if possible.
+static SDValue EmitKTEST(SDValue Op0, SDValue Op1, ISD::CondCode CC,
+                         const SDLoc &dl, SelectionDAG &DAG,
+                         const X86Subtarget &Subtarget) {
+  // Only support equality comparisons.
+  if (CC != ISD::SETEQ && CC != ISD::SETNE)
+    return SDValue();
+
+  // Must be a bitcast from vXi1.
+  if (Op0.getOpcode() != ISD::BITCAST)
+    return SDValue();
+
+  Op0 = Op0.getOperand(0);
+  MVT VT = Op0.getSimpleValueType();
+  if (!(Subtarget.hasDQI() && (VT == MVT::v8i1  || VT == MVT::v16i1)) &&
+      !(Subtarget.hasBWI() && (VT == MVT::v32i1 || VT == MVT::v64i1)))
+    return SDValue();
+
+  X86::CondCode X86CC;
+  if (isNullConstant(Op1)) {
+    X86CC = CC == ISD::SETEQ ? X86::COND_E : X86::COND_NE;
+  } else
+    return SDValue();
+
+  SDValue KTEST = DAG.getNode(X86ISD::KTEST, dl, MVT::i32, Op0, Op0);
+  return getSETCC(X86CC, KTEST, dl, DAG);
+}
+
 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
 
   MVT VT = Op.getSimpleValueType();
@@ -18176,6 +18179,10 @@ SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
       return NewSetCC;
   }
 
+  // Try to lower using KTEST.
+  if (SDValue NewSetCC = EmitKTEST(Op0, Op1, CC, dl, DAG, Subtarget))
+    return NewSetCC;
+
   // Look for X == 0, X == 1, X != 0, or X != 1.  We can simplify some forms of
   // these.
   if ((isOneConstant(Op1) || isNullConstant(Op1)) &&