[llvm] r353615 - [TargetLowering] avoid miscompile in setcc transform (PR40657)

[Sanjay Patel via llvm-commits]

Author: spatel
Date: Sat Feb  9 07:59:02 2019
New Revision: 353615

URL: http://llvm.org/viewvc/llvm-project?rev=353615&view=rev
Log:
[TargetLowering] avoid miscompile in setcc transform (PR40657)

Modified:
    llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
    llvm/trunk/test/CodeGen/X86/setcc-combine.ll

Modified: llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp?rev=353615&r1=353614&r2=353615&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp Sat Feb  9 07:59:02 2019
@@ -3074,7 +3074,9 @@ SDValue TargetLowering::SimplifySetCC(EV
             return DAG.getSetCC(dl, VT, N0.getOperand(0),
                                 DAG.getConstant(0, dl, N0.getValueType()),
                                 Cond);
-          if (N0.getNode()->hasOneUse()) {
+          // The shift is not valid if this is a bool (i1).
+          // TODO: This transform needs evidence to justify its existence.
+          if (N0.getNode()->hasOneUse() && OpVT.getScalarSizeInBits() != 1) {
             assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
             auto &DL = DAG.getDataLayout();
             // (Z-X) == X  --> Z == X<<1

Modified: llvm/trunk/test/CodeGen/X86/setcc-combine.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/setcc-combine.ll?rev=353615&r1=353614&r2=353615&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/setcc-combine.ll (original)
+++ llvm/trunk/test/CodeGen/X86/setcc-combine.ll Sat Feb  9 07:59:02 2019
@@ -262,12 +262,15 @@ define void @test_i1_uge(i1 *%A2) {
   ret void
 }
 
-; FIXME: This should not get folded to 0.
+; This should not get folded to 0.
 
 define i64 @PR40657(i8 %var2, i8 %var9) {
 ; CHECK-LABEL: PR40657:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xorl %eax, %eax
+; CHECK-NEXT:    notb %sil
+; CHECK-NEXT:    addb %dil, %sil
+; CHECK-NEXT:    movzbl %sil, %eax
+; CHECK-NEXT:    andl $1, %eax
 ; CHECK-NEXT:    retq
   %var6 = trunc i8 %var9 to i1
   %var7 = trunc i8 175 to i1