Don't redundantly copy implicit operands when rematerializing.

While we're at it - don't copy vreg implicit operands while rematerializing.
This fixes PR12138.

llvm-svn: 151779

Author: lhames

llvm/lib/CodeGen/RegisterCoalescer.cpp

@@ -814,7 +814,8 @@ bool RegisterCoalescer::ReMaterializeTrivialDef(LiveInterval &SrcInt,
          e = NewMI->getNumOperands(); i != e; ++i) {
     MachineOperand &MO = NewMI->getOperand(i);
     if (MO.isReg()) {
-      assert(MO.isDef() && MO.isImplicit() && MO.isDead());
+      assert(MO.isDef() && MO.isImplicit() && MO.isDead() &&
+             TargetRegisterInfo::isPhysicalRegister(MO.getReg()));
       NewMIImplDefs.push_back(MO.getReg());
     }
   }
@@ -824,8 +825,13 @@ bool RegisterCoalescer::ReMaterializeTrivialDef(LiveInterval &SrcInt,
   for (unsigned i = CopyMI->getDesc().getNumOperands(),
          e = CopyMI->getNumOperands(); i != e; ++i) {
     MachineOperand &MO = CopyMI->getOperand(i);
-    if (MO.isReg() && MO.isImplicit())
-      NewMI->addOperand(MO);
+    if (MO.isReg()) {
+      assert(MO.isImplicit() && "No explicit operands after implict operands.");
+      // Discard VReg implicit defs.
+      if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
+        NewMI->addOperand(MO);
+      }
+    }
   }
 
   LIS->ReplaceMachineInstrInMaps(CopyMI, NewMI);
@@ -840,7 +846,6 @@ bool RegisterCoalescer::ReMaterializeTrivialDef(LiveInterval &SrcInt,
     li.addRange(lr);
   }
 
-  NewMI->copyImplicitOps(CopyMI);
   CopyMI->eraseFromParent();
   ReMatCopies.insert(CopyMI);
   ReMatDefs.insert(DefMI);

llvm/test/CodeGen/X86/2012-02-29-CoalescerBug.ll

@@ -0,0 +1,58 @@
+; RUN: llc -O1 <%s
+; PR12138
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128"
+target triple = "i386-apple-macosx10.7.0"
+
+%struct.S0 = type { i8, i32 }
+
+@d = external global [2 x [2 x %struct.S0]], align 4
+@c = external global i32, align 4
+@e = external global i32, align 4
+@b = external global i32, align 4
+@a = external global i32, align 4
+
+define void @fn2() nounwind optsize ssp {
+entry:
+  store i64 0, i64* bitcast ([2 x [2 x %struct.S0]]* @d to i64*), align 4
+  %0 = load i32* @c, align 4
+  %tobool2 = icmp eq i32 %0, 0
+  %1 = load i32* @a, align 4
+  %tobool4 = icmp eq i32 %1, 0
+  br label %for.cond
+
+for.cond:                                         ; preds = %if.end, %entry
+  %f.1.0 = phi i32 [ undef, %entry ], [ %sub, %if.end ]
+  %g.0 = phi i64 [ 0, %entry ], [ %ins, %if.end ]
+  %tobool = icmp eq i32 %f.1.0, 0
+  br i1 %tobool, label %for.end, label %for.body
+
+for.body:                                         ; preds = %for.cond
+  %2 = lshr i64 %g.0, 32
+  %conv = trunc i64 %2 to i16
+  br i1 %tobool2, label %lor.rhs, label %lor.end
+
+lor.rhs:                                          ; preds = %for.body
+  store i32 1, i32* @e, align 4
+  br label %lor.end
+
+lor.end:                                          ; preds = %lor.rhs, %for.body
+  %xor.i = xor i16 %conv, 1
+  %p1.lobit.i8 = lshr i64 %g.0, 47
+  %p1.lobit.i8.tr = trunc i64 %p1.lobit.i8 to i16
+  %p1.lobit.i = and i16 %p1.lobit.i8.tr, 1
+  %and.i = and i16 %p1.lobit.i, %xor.i
+  %3 = xor i16 %and.i, 1
+  %sub.conv.i = sub i16 %conv, %3
+  %conv3 = sext i16 %sub.conv.i to i32
+  store i32 %conv3, i32* @b, align 4
+  br i1 %tobool4, label %if.end, label %for.end
+
+if.end:                                           ; preds = %lor.end
+  %mask = and i64 %g.0, -256
+  %ins = or i64 %mask, 1
+  %sub = add nsw i32 %f.1.0, -1
+  br label %for.cond
+
+for.end:                                          ; preds = %lor.end, %for.cond
+  ret void
+}