add X86-specific DAG optimization to simplify boolean test

- if a boolean test (X86ISD::CMP or X86ISD:SUB) checks a boolean value
  generated from X86ISD::SETCC, try to simplify the boolean value
  generation and checking by reusing the original EFLAGS with proper
  condition code
- add hooks to X86 specific SETCC/BRCOND/CMOV, the major 3 places
  consuming EFLAGS

part of patches fixing PR12312



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@161687 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 155 insertions, 4 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index cc02393..c2a81c9 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -13782,6 +13782,72 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+// Check whether a boolean test is testing a boolean value generated by
+// X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
+// code.
+//
+// Simplify the following patterns:
+// (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
+// (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
+// to (Op EFLAGS Cond)
+//
+// (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
+// (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
+// to (Op EFLAGS !Cond)
+//
+// where Op could be BRCOND or CMOV.
+//
+static SDValue BoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
+  // Quit if not CMP and SUB with its value result used.
+  if (Cmp.getOpcode() != X86ISD::CMP &&
+      (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
+      return SDValue();
+
+  // Quit if not used as a boolean value.
+  if (CC != X86::COND_E && CC != X86::COND_NE)
+    return SDValue();
+
+  // Check CMP operands. One of them should be 0 or 1 and the other should be
+  // an SetCC or extended from it.
+  SDValue Op1 = Cmp.getOperand(0);
+  SDValue Op2 = Cmp.getOperand(1);
+
+  SDValue SetCC;
+  const ConstantSDNode* C = 0;
+  bool needOppositeCond = (CC == X86::COND_E);
+
+  if ((C = dyn_cast<ConstantSDNode>(Op1)))
+    SetCC = Op2;
+  else if ((C = dyn_cast<ConstantSDNode>(Op2)))
+    SetCC = Op1;
+  else // Quit if all operands are not constants.
+    return SDValue();
+
+  if (C->getZExtValue() == 1)
+    needOppositeCond = !needOppositeCond;
+  else if (C->getZExtValue() != 0)
+    // Quit if the constant is neither 0 or 1.
+    return SDValue();
+
+  // Skip 'zext' node.
+  if (SetCC.getOpcode() == ISD::ZERO_EXTEND)
+    SetCC = SetCC.getOperand(0);
+
+  // Quit if not SETCC.
+  // FIXME: So far we only handle the boolean value generated from SETCC. If
+  // there is other ways to generate boolean values, we need handle them here
+  // as well.
+  if (SetCC.getOpcode() != X86ISD::SETCC)
+    return SDValue();
+
+  // Set the condition code or opposite one if necessary.
+  CC = X86::CondCode(SetCC.getConstantOperandVal(0));
+  if (needOppositeCond)
+    CC = X86::GetOppositeBranchCondition(CC);
+
+  return SetCC.getOperand(1);
+}
+
 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
                                   TargetLowering::DAGCombinerInfo &DCI) {
@@ -13795,6 +13861,7 @@ static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
   SDValue TrueOp = N->getOperand(1);
   X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
   SDValue Cond = N->getOperand(3);
+
   if (CC == X86::COND_E || CC == X86::COND_NE) {
     switch (Cond.getOpcode()) {
     default: break;
@@ -13806,6 +13873,16 @@ static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  SDValue Flags;
+
+  Flags = BoolTestSetCCCombine(Cond, CC);
+  if (Flags.getNode()) {
+    SDValue Ops[] = { FalseOp, TrueOp,
+                      DAG.getConstant(CC, MVT::i8), Flags };
+    return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(),
+                       Ops, array_lengthof(Ops));
+  }
+
   // If this is a select between two integer constants, try to do some
   // optimizations.  Note that the operands are ordered the opposite of SELECT
   // operands.
@@ -15285,19 +15362,50 @@ static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG) {
 
 // Optimize  RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
-  unsigned X86CC = N->getConstantOperandVal(0);
-  SDValue EFLAG = N->getOperand(1);
   DebugLoc DL = N->getDebugLoc();
+  X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
+  SDValue EFLAGS = N->getOperand(1);
 
   // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
   // a zext and produces an all-ones bit which is more useful than 0/1 in some
   // cases.
-  if (X86CC == X86::COND_B)
+  if (CC == X86::COND_B)
     return DAG.getNode(ISD::AND, DL, MVT::i8,
                        DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
-                                   DAG.getConstant(X86CC, MVT::i8), EFLAG),
+                                   DAG.getConstant(CC, MVT::i8), EFLAGS),
                        DAG.getConstant(1, MVT::i8));
 
+  SDValue Flags;
+
+  Flags = BoolTestSetCCCombine(EFLAGS, CC);
+  if (Flags.getNode()) {
+    SDValue Cond = DAG.getConstant(CC, MVT::i8);
+    return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
+  }
+
+  return SDValue();
+}
+
+// Optimize branch condition evaluation.
+//
+static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
+                                    TargetLowering::DAGCombinerInfo &DCI,
+                                    const X86Subtarget *Subtarget) {
+  DebugLoc DL = N->getDebugLoc();
+  SDValue Chain = N->getOperand(0);
+  SDValue Dest = N->getOperand(1);
+  SDValue EFLAGS = N->getOperand(3);
+  X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
+
+  SDValue Flags;
+
+  Flags = BoolTestSetCCCombine(EFLAGS, CC);
+  if (Flags.getNode()) {
+    SDValue Cond = DAG.getConstant(CC, MVT::i8);
+    return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
+                       Flags);
+  }
+
   return SDValue();
 }
 
@@ -15515,6 +15623,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::TRUNCATE:       return PerformTruncateCombine(N, DAG, DCI);
   case ISD::SETCC:          return PerformISDSETCCCombine(N, DAG);
   case X86ISD::SETCC:       return PerformSETCCCombine(N, DAG);
+  case X86ISD::BRCOND:      return PerformBrCondCombine(N, DAG, DCI, Subtarget);
   case X86ISD::SHUFP:       // Handle all target specific shuffles
   case X86ISD::PALIGN:
   case X86ISD::UNPCKH:
diff --git a/test/CodeGen/X86/bool-simplify.ll b/test/CodeGen/X86/bool-simplify.ll
new file mode 100644
index 0000000..0cb9fd9
--- /dev/null
+++ b/test/CodeGen/X86/bool-simplify.ll
@@ -0,0 +1,42 @@
+; RUN: llc < %s -march=x86-64 -mattr=+sse41,-avx | FileCheck %s
+
+define i32 @foo(<2 x i64> %c, i32 %a, i32 %b) {
+  %t1 = call i32 @llvm.x86.sse41.ptestz(<2 x i64> %c, <2 x i64> %c)
+  %t2 = icmp ne i32 %t1, 0
+  %t3 = select i1 %t2, i32 %a, i32 %b
+  ret i32 %t3
+; CHECK: foo
+; CHECK: ptest
+; CHECK-NOT: testl
+; CHECK: cmov
+; CHECK: ret
+}
+
+define i32 @bar(<2 x i64> %c) {
+entry:
+  %0 = call i32 @llvm.x86.sse41.ptestz(<2 x i64> %c, <2 x i64> %c)
+  %1 = icmp ne i32 %0, 0
+  br i1 %1, label %if-true-block, label %endif-block
+if-true-block:                                    ; preds = %entry
+  ret i32 0
+endif-block:                                      ; preds = %entry,
+  ret i32 1
+; CHECK: bar
+; CHECK: ptest
+; CHECK-NOT: testl
+; CHECK: jne
+; CHECK: ret
+}
+
+define i32 @bax(<2 x i64> %c) {
+  %t1 = call i32 @llvm.x86.sse41.ptestz(<2 x i64> %c, <2 x i64> %c)
+  %t2 = icmp eq i32 %t1, 1
+  %t3 = zext i1 %t2 to i32
+  ret i32 %t3
+; CHECK: bax
+; CHECK: ptest
+; CHECK-NOT: cmpl
+; CHECK: ret
+}
+
+declare i32 @llvm.x86.sse41.ptestz(<2 x i64>, <2 x i64>) nounwind readnone