ARM: optimizeSelect has to consider the previous register class

optimizeSelect folds (predicated) copy instructions, it must not ignore
the original register class of the operand when replacing the register
with the copies dest register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191963 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 32 insertions, 4 deletions

diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index b4266693c1..251171daad 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -1706,7 +1706,7 @@ MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
                                                bool PreferFalse) const {
   assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
          "Unknown select instruction");
-  const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
+  MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
   MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
   bool Invert = !DefMI;
   if (!DefMI)
@@ -1714,11 +1714,17 @@ MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
   if (!DefMI)
     return 0;
 
+  // Find new register class to use.
+  MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
+  unsigned       DestReg  = MI->getOperand(0).getReg();
+  const TargetRegisterClass *PreviousClass = MRI.getRegClass(FalseReg.getReg());
+  if (!MRI.constrainRegClass(DestReg, PreviousClass))
+    return 0;
+
   // Create a new predicated version of DefMI.
   // Rfalse is the first use.
   MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
-                                      DefMI->getDesc(),
-                                      MI->getOperand(0).getReg());
+                                      DefMI->getDesc(), DestReg);
 
   // Copy all the DefMI operands, excluding its (null) predicate.
   const MCInstrDesc &DefDesc = DefMI->getDesc();
@@ -1741,7 +1747,6 @@ MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
   // register operand tied to the first def.
   // The tie makes the register allocator ensure the FalseReg is allocated the
   // same register as operand 0.
-  MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
   FalseReg.setImplicit();
   NewMI.addOperand(FalseReg);
   NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
diff --git a/test/CodeGen/ARM/optselect-regclass.ll b/test/CodeGen/ARM/optselect-regclass.ll
new file mode 100644
index 0000000000..1aa4520896
--- /dev/null
+++ b/test/CodeGen/ARM/optselect-regclass.ll
@@ -0,0 +1,23 @@
+; RUN: llc < %s -march=arm -mcpu=swift -verify-machineinstrs
+%union.opcode.0.2.5.8.15.28 = type { i32 }
+
+@opcode = external global %union.opcode.0.2.5.8.15.28, align 4
+@operands = external hidden global [50 x i8], align 4
+@.str86 = external hidden unnamed_addr constant [13 x i8], align 1
+
+; Function Attrs: nounwind ssp
+define void @xfr() {
+entry:
+  %bf.load4 = load i32* getelementptr inbounds (%union.opcode.0.2.5.8.15.28* @opcode, i32 0, i32 0), align 4
+  %bf.clear10 = and i32 %bf.load4, 65535
+  %and11 = and i32 %bf.load4, 32768
+  %tobool12 = icmp ne i32 %and11, 0
+  %cond13 = select i1 %tobool12, i32 1073676288, i32 0
+  %or = or i32 %cond13, %bf.clear10
+  %shl = shl nuw i32 %or, 2
+  %add = add i32 0, %shl
+  tail call void (i8*, i32, i32, i8*, ...)* @__sprintf_chk(i8* getelementptr inbounds ([50 x i8]* @operands, i32 0, i32 0), i32 0, i32 50, i8* getelementptr inbounds ([13 x i8]* @.str86, i32 0, i32 0), i32 undef, i32 undef, i32 %add)
+  ret void
+}
+
+declare void @__sprintf_chk(i8*, i32, i32, i8*, ...)