From 17418964f8133a7200ab708338c5faefae36da6d Mon Sep 17 00:00:00 2001
From: Nadav Rotem <nrotem@apple.com>
Date: Wed, 10 Oct 2012 21:31:55 +0000
Subject: [PATCH] Patch by Shuxin Yang <shuxin.llvm@gmail.com>.

Original message:

The attached is the fix to radar://11663049. The optimization can be outlined by following rules:

   (select (x != c), e, c) -> select (x != c), e, x),
   (select (x == c), c, e) -> select (x == c), x, e)
where the <c> is an integer constant.

 The reason for this change is that : on x86, conditional-move-from-constant needs two instructions;
however, conditional-move-from-register need only one instruction.

  While the LowerSELECT() sounds to be the most convenient place for this optimization, it turns out to be a bad place. The reason is that by replacing the constant <c> with a symbolic value, it obscure some instruction-combining opportunities which would otherwise be very easy to spot. For that reason, I have to postpone the change to last instruction-combining phase.

  The change passes the test of "make check-all -C <build-root/test" and "make -C project/test-suite/SingleSource".

llvm-svn: 165661
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 40 +++++++++++++++++++++++++++++++++
 llvm/test/CodeGen/X86/select_const.ll   | 16 +++++++++++++
 2 files changed, 56 insertions(+)
 create mode 100644 llvm/test/CodeGen/X86/select_const.ll

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 7df1098..2fbc6bd 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -14418,6 +14418,7 @@ static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
       if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
         CC = X86::GetOppositeBranchCondition(CC);
         std::swap(TrueC, FalseC);
+        std::swap(TrueOp, FalseOp);
       }
 
       // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3.  Likewise for any pow2/0.
@@ -14500,6 +14501,45 @@ static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
       }
     }
   }
+
+  // Handle these cases:
+  //   (select (x != c), e, c) -> select (x != c), e, x),
+  //   (select (x == c), c, e) -> select (x == c), x, e) 
+  // where the c is an integer constant, and the "select" is the combination 
+  // of CMOV and CMP.
+  //
+  // The rationale for this change is that the conditional-move from a constant
+  // needs two instructions, however, conditional-move from a register needs 
+  // only one instruction.
+  // 
+  // CAVEAT: By replacing a constant with a symbolic value, it may obscure 
+  //  some instruction-combining opportunities. This opt needs to be 
+  //  postponed as late as possible.
+  //
+  if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
+    // the DCI.xxxx conditions are provided to postpone the optimization as 
+    // late as possible.
+
+    ConstantSDNode *CmpAgainst = 0;
+    if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) && 
+        (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) && 
+        dyn_cast<ConstantSDNode>(Cond.getOperand(0)) == 0) {
+
+      if (CC == X86::COND_NE && 
+          CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
+        CC = X86::GetOppositeBranchCondition(CC);
+        std::swap(TrueOp, FalseOp);
+      }
+
+      if (CC == X86::COND_E && 
+          CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
+        SDValue Ops[] = { FalseOp, Cond.getOperand(0), N->getOperand(2), Cond };
+        return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops, 
+                           array_lengthof(Ops));
+      }
+    }
+  }
+
   return SDValue();
 }
 
diff --git a/llvm/test/CodeGen/X86/select_const.ll b/llvm/test/CodeGen/X86/select_const.ll
new file mode 100644
index 0000000..5b2409d
--- /dev/null
+++ b/llvm/test/CodeGen/X86/select_const.ll
@@ -0,0 +1,16 @@
+; RUN: llc < %s -mtriple=x86_64-apple-darwin10 -mcpu=corei7 | FileCheck %s
+
+define i64 @test1(i64 %x) nounwind {
+entry:
+  %cmp = icmp eq i64 %x, 2
+  %add = add i64 %x, 1
+  %retval.0 = select i1 %cmp, i64 2, i64 %add
+  ret i64 %retval.0
+
+; CHECK: test1:
+; CHECK: leaq 1(%rdi), %rax
+; CHECK: cmpq $2, %rdi
+; CHECK: cmoveq %rdi, %rax
+; CHECK: ret
+
+}
-- 
2.7.4