MachineLICM CSE should match destination register classes; avoid hoisting implicit_def's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@65592 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/MachineLICM.cpp b/lib/CodeGen/MachineLICM.cpp
index dd32977..aaa4de4 100644 (file)
--- a/lib/CodeGen/MachineLICM.cpp
+++ b/lib/CodeGen/MachineLICM.cpp
@@ -288,6 +288,9 @@ static bool HasPHIUses(unsigned Reg, MachineRegisterInfo *RegInfo) {
 /// IsProfitableToHoist - Return true if it is potentially profitable to hoist
 /// the given loop invariant.
 bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
+  if (MI.getOpcode() == TargetInstrInfo::IMPLICIT_DEF)
+    return false;
+
   const TargetInstrDesc &TID = MI.getDesc();
 
   // FIXME: For now, only hoist re-materilizable instructions. LICM will
@@ -312,7 +315,8 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
 }
 
 static const MachineInstr *LookForDuplicate(const MachineInstr *MI,
-                                      std::vector<const MachineInstr*> &PrevMIs) {
+                                      std::vector<const MachineInstr*> &PrevMIs,
+                                      MachineRegisterInfo *RegInfo) {
   unsigned NumOps = MI->getNumOperands();
   for (unsigned i = 0, e = PrevMIs.size(); i != e; ++i) {
     const MachineInstr *PrevMI = PrevMIs[i];
@@ -322,8 +326,14 @@ static const MachineInstr *LookForDuplicate(const MachineInstr *MI,
     bool IsSame = true;
     for (unsigned j = 0; j != NumOps; ++j) {
       const MachineOperand &MO = MI->getOperand(j);
-      if (MO.isReg() && MO.isDef()) 
+      if (MO.isReg() && MO.isDef()) {
+        if (RegInfo->getRegClass(MO.getReg()) !=
+            RegInfo->getRegClass(PrevMI->getOperand(j).getReg())) {
+          IsSame = false;
+          break;
+        }
         continue;
+      }
       if (!MO.isIdenticalTo(PrevMI->getOperand(j))) {
         IsSame = false;
         break;
@@ -362,7 +372,7 @@ void MachineLICM::Hoist(MachineInstr &MI) {
     std::vector<const MachineInstr*> >::iterator CI = CSEMap.find(BBOpcPair);
   bool DoneCSE = false;
   if (CI != CSEMap.end()) {
-    const MachineInstr *Dup = LookForDuplicate(&MI, CI->second);
+    const MachineInstr *Dup = LookForDuplicate(&MI, CI->second, RegInfo);
     if (Dup) {
       DOUT << "CSEing " << MI;
       DOUT << " with " << *Dup;

diff --git a/test/CodeGen/X86/2009-02-26-MachineLICMBug.ll b/test/CodeGen/X86/2009-02-26-MachineLICMBug.ll
new file mode 100644 (file)
index 0000000..8bf6c23
--- /dev/null
+++ b/test/CodeGen/X86/2009-02-26-MachineLICMBug.ll
@@ -0,0 +1,47 @@
+; RUN: llvm-as < %s | llc -march=x86-64 -mattr=+sse3 -stats |& not grep {machine-licm}
+; rdar://6627786
+
+target triple = "x86_64-apple-darwin10.0"
+       %struct.Key = type { i64 }
+       %struct.__Rec = type opaque
+       %struct.__vv = type {  }
+
+define %struct.__vv* @t(%struct.Key* %desc) nounwind ssp {
+entry:
+       br label %bb4
+
+bb4:           ; preds = %bb.i, %bb26, %bb4, %entry
+       %0 = call i32 (...)* @xxGetOffsetForCode(i32 undef) nounwind            ; <i32> [#uses=0]
+       %ins = or i64 0, 0              ; <i64> [#uses=1]
+       %1 = call i32 (...)* @xxCalculateMidType(%struct.Key* %desc, i32 0) nounwind            ; <i32> [#uses=1]
+       %cond = icmp eq i32 %1, 1               ; <i1> [#uses=1]
+       br i1 %cond, label %bb26, label %bb4
+
+bb26:          ; preds = %bb4
+       %2 = and i64 %ins, 15728640             ; <i64> [#uses=1]
+       %cond.i = icmp eq i64 %2, 1048576               ; <i1> [#uses=1]
+       br i1 %cond.i, label %bb.i, label %bb4
+
+bb.i:          ; preds = %bb26
+       %3 = load i32* null, align 4            ; <i32> [#uses=1]
+       %4 = uitofp i32 %3 to float             ; <float> [#uses=1]
+       %.sum13.i = add i64 0, 4                ; <i64> [#uses=1]
+       %5 = getelementptr i8* null, i64 %.sum13.i              ; <i8*> [#uses=1]
+       %6 = bitcast i8* %5 to i32*             ; <i32*> [#uses=1]
+       %7 = load i32* %6, align 4              ; <i32> [#uses=1]
+       %8 = uitofp i32 %7 to float             ; <float> [#uses=1]
+       %.sum.i = add i64 0, 8          ; <i64> [#uses=1]
+       %9 = getelementptr i8* null, i64 %.sum.i                ; <i8*> [#uses=1]
+       %10 = bitcast i8* %9 to i32*            ; <i32*> [#uses=1]
+       %11 = load i32* %10, align 4            ; <i32> [#uses=1]
+       %12 = uitofp i32 %11 to float           ; <float> [#uses=1]
+       %13 = insertelement <4 x float> undef, float %4, i32 0          ; <<4 x float>> [#uses=1]
+       %14 = insertelement <4 x float> %13, float %8, i32 1            ; <<4 x float>> [#uses=1]
+       %15 = insertelement <4 x float> %14, float %12, i32 2           ; <<4 x float>> [#uses=1]
+       store <4 x float> %15, <4 x float>* null, align 16
+       br label %bb4
+}
+
+declare i32 @xxGetOffsetForCode(...)
+
+declare i32 @xxCalculateMidType(...)