Merge r100559 from mainline to fix PR6696.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_27@100851 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 84 insertions, 14 deletions

diff --git a/lib/Target/X86/X86.h b/lib/Target/X86/X86.h
index ba0ee6cedd..1fdcce0cd2 100644
--- a/lib/Target/X86/X86.h
+++ b/lib/Target/X86/X86.h
@@ -66,6 +66,12 @@ TargetAsmBackend *createX86_64AsmBackend(const Target &, MCAssembler &);
 ///
 FunctionPass *createEmitX86CodeToMemory();
 
+/// createX86MaxStackAlignmentHeuristicPass - This function returns a pass
+/// which determines whether the frame pointer register should be
+/// reserved in case dynamic stack alignment is later required.
+///
+FunctionPass *createX86MaxStackAlignmentHeuristicPass();
+
 extern Target TheX86_32Target, TheX86_64Target;
 
 } // End llvm namespace
diff --git a/lib/Target/X86/X86MachineFunctionInfo.h b/lib/Target/X86/X86MachineFunctionInfo.h
index 4b2529bccf..a916c6372e 100644
--- a/lib/Target/X86/X86MachineFunctionInfo.h
+++ b/lib/Target/X86/X86MachineFunctionInfo.h
@@ -52,6 +52,10 @@ class X86MachineFunctionInfo : public MachineFunctionInfo {
   /// relocation models.
   unsigned GlobalBaseReg;
 
+  /// ReserveFP - whether the function should reserve the frame pointer
+  /// when allocating, even if there may not actually be a frame pointer used.
+  bool ReserveFP;
+
 public:
   X86MachineFunctionInfo() : ForceFramePointer(false),
                              CalleeSavedFrameSize(0),
@@ -68,7 +72,8 @@ public:
       ReturnAddrIndex(0),
       TailCallReturnAddrDelta(0),
       SRetReturnReg(0),
-      GlobalBaseReg(0) {}
+      GlobalBaseReg(0),
+      ReserveFP(false) {}
   
   bool getForceFramePointer() const { return ForceFramePointer;} 
   void setForceFramePointer(bool forceFP) { ForceFramePointer = forceFP; }
@@ -90,6 +95,9 @@ public:
 
   unsigned getGlobalBaseReg() const { return GlobalBaseReg; }
   void setGlobalBaseReg(unsigned Reg) { GlobalBaseReg = Reg; }
+
+  bool getReserveFP() const { return ReserveFP; }
+  void setReserveFP(bool reserveFP) { ReserveFP = reserveFP; }
 };
 
 } // End llvm namespace
diff --git a/lib/Target/X86/X86RegisterInfo.cpp b/lib/Target/X86/X86RegisterInfo.cpp
index 4aeeecb4d3..4fe4d0a525 100644
--- a/lib/Target/X86/X86RegisterInfo.cpp
+++ b/lib/Target/X86/X86RegisterInfo.cpp
@@ -1489,3 +1489,46 @@ unsigned getX86SubSuperRegister(unsigned Reg, EVT VT, bool High) {
 }
 
 #include "X86GenRegisterInfo.inc"
+
+namespace {
+  struct MSAH : public MachineFunctionPass {
+    static char ID;
+    MSAH() : MachineFunctionPass(&ID) {}
+
+    virtual bool runOnMachineFunction(MachineFunction &MF) {
+      const X86TargetMachine *TM =
+        static_cast<const X86TargetMachine *>(&MF.getTarget());
+      const X86RegisterInfo *X86RI = TM->getRegisterInfo();
+      MachineRegisterInfo &RI = MF.getRegInfo();
+      X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
+      unsigned StackAlignment = X86RI->getStackAlignment();
+
+      // Be over-conservative: scan over all vreg defs and find whether vector
+      // registers are used. If yes, there is a possibility that vector register
+      // will be spilled and thus require dynamic stack realignment.
+      for (unsigned RegNum = TargetRegisterInfo::FirstVirtualRegister;
+           RegNum < RI.getLastVirtReg(); ++RegNum)
+        if (RI.getRegClass(RegNum)->getAlignment() > StackAlignment) {
+          FuncInfo->setReserveFP(true);
+          return true;
+        }
+
+      // Nothing to do
+      return false;
+    }
+
+    virtual const char *getPassName() const {
+      return "X86 Maximal Stack Alignment Check";
+    }
+
+    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+      AU.setPreservesCFG();
+      MachineFunctionPass::getAnalysisUsage(AU);
+    }
+  };
+
+  char MSAH::ID = 0;
+}
+
+FunctionPass*
+llvm::createX86MaxStackAlignmentHeuristicPass() { return new MSAH(); }
diff --git a/lib/Target/X86/X86RegisterInfo.td b/lib/Target/X86/X86RegisterInfo.td
index ed2ce6c353..0a6f6579fa 100644
--- a/lib/Target/X86/X86RegisterInfo.td
+++ b/lib/Target/X86/X86RegisterInfo.td
@@ -352,11 +352,12 @@ def GR8 : RegisterClass<"X86", [i8],  8,
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
       const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       // Does the function dedicate RBP / EBP to being a frame ptr?
       if (!Subtarget.is64Bit())
         // In 32-mode, none of the 8-bit registers aliases EBP or ESP.
         return begin() + 8;
-      else if (RI->hasFP(MF))
+      else if (RI->hasFP(MF) || MFI->getReserveFP())
         // If so, don't allocate SPL or BPL.
         return array_endof(X86_GR8_AO_64) - 1;
       else
@@ -396,9 +397,10 @@ def GR16 : RegisterClass<"X86", [i16], 16,
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
       const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       if (Subtarget.is64Bit()) {
         // Does the function dedicate RBP to being a frame ptr?
-        if (RI->hasFP(MF))
+        if (RI->hasFP(MF) || MFI->getReserveFP())
           // If so, don't allocate SP or BP.
           return array_endof(X86_GR16_AO_64) - 1;
         else
@@ -406,7 +408,7 @@ def GR16 : RegisterClass<"X86", [i16], 16,
           return array_endof(X86_GR16_AO_64);
       } else {
         // Does the function dedicate EBP to being a frame ptr?
-        if (RI->hasFP(MF))
+        if (RI->hasFP(MF) || MFI->getReserveFP())
           // If so, don't allocate SP or BP.
           return begin() + 6;
         else
@@ -447,9 +449,10 @@ def GR32 : RegisterClass<"X86", [i32], 32,
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
       const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       if (Subtarget.is64Bit()) {
         // Does the function dedicate RBP to being a frame ptr?
-        if (RI->hasFP(MF))
+        if (RI->hasFP(MF) || MFI->getReserveFP())
           // If so, don't allocate ESP or EBP.
           return array_endof(X86_GR32_AO_64) - 1;
         else
@@ -457,7 +460,7 @@ def GR32 : RegisterClass<"X86", [i32], 32,
           return array_endof(X86_GR32_AO_64);
       } else {
         // Does the function dedicate EBP to being a frame ptr?
-        if (RI->hasFP(MF))
+        if (RI->hasFP(MF) || MFI->getReserveFP())
           // If so, don't allocate ESP or EBP.
           return begin() + 6;
         else
@@ -484,9 +487,11 @@ def GR64 : RegisterClass<"X86", [i64], 64,
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
       const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       if (!Subtarget.is64Bit())
         return begin();  // None of these are allocatable in 32-bit.
-      if (RI->hasFP(MF)) // Does the function dedicate RBP to being a frame ptr?
+      // Does the function dedicate RBP to being a frame ptr?
+      if (RI->hasFP(MF) || MFI->getReserveFP())
         return end()-3;  // If so, don't allocate RIP, RSP or RBP
       else
         return end()-2;  // If not, just don't allocate RIP or RSP
@@ -582,8 +587,9 @@ def GR16_NOREX : RegisterClass<"X86", [i16], 16,
     GR16_NOREXClass::allocation_order_end(const MachineFunction &MF) const {
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       // Does the function dedicate RBP / EBP to being a frame ptr?
-      if (RI->hasFP(MF))
+      if (RI->hasFP(MF) || MFI->getReserveFP())
         // If so, don't allocate SP or BP.
         return end() - 2;
       else
@@ -604,8 +610,9 @@ def GR32_NOREX : RegisterClass<"X86", [i32], 32,
     GR32_NOREXClass::allocation_order_end(const MachineFunction &MF) const {
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       // Does the function dedicate RBP / EBP to being a frame ptr?
-      if (RI->hasFP(MF))
+      if (RI->hasFP(MF) || MFI->getReserveFP())
         // If so, don't allocate ESP or EBP.
         return end() - 2;
       else
@@ -626,8 +633,9 @@ def GR64_NOREX : RegisterClass<"X86", [i64], 64,
     GR64_NOREXClass::allocation_order_end(const MachineFunction &MF) const {
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       // Does the function dedicate RBP to being a frame ptr?
-      if (RI->hasFP(MF))
+      if (RI->hasFP(MF) || MFI->getReserveFP())
         // If so, don't allocate RIP, RSP or RBP.
         return end() - 3;
       else
@@ -668,9 +676,10 @@ def GR32_NOSP : RegisterClass<"X86", [i32], 32,
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
       const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       if (Subtarget.is64Bit()) {
         // Does the function dedicate RBP to being a frame ptr?
-        if (RI->hasFP(MF))
+        if (RI->hasFP(MF) || MFI->getReserveFP())
           // If so, don't allocate EBP.
           return array_endof(X86_GR32_NOSP_AO_64) - 1;
         else
@@ -678,7 +687,7 @@ def GR32_NOSP : RegisterClass<"X86", [i32], 32,
           return array_endof(X86_GR32_NOSP_AO_64);
       } else {
         // Does the function dedicate EBP to being a frame ptr?
-        if (RI->hasFP(MF))
+        if (RI->hasFP(MF) || MFI->getReserveFP())
           // If so, don't allocate EBP.
           return begin() + 6;
         else
@@ -703,9 +712,11 @@ def GR64_NOSP : RegisterClass<"X86", [i64], 64,
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
       const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       if (!Subtarget.is64Bit())
         return begin();  // None of these are allocatable in 32-bit.
-      if (RI->hasFP(MF)) // Does the function dedicate RBP to being a frame ptr?
+      // Does the function dedicate RBP to being a frame ptr?
+      if (RI->hasFP(MF) || MFI->getReserveFP())
         return end()-1;  // If so, don't allocate RBP
       else
         return end();  // If not, any reg in this class is ok.
@@ -726,8 +737,9 @@ def GR64_NOREX_NOSP : RegisterClass<"X86", [i64], 64,
   {
       const TargetMachine &TM = MF.getTarget();
       const TargetRegisterInfo *RI = TM.getRegisterInfo();
+      const X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
       // Does the function dedicate RBP to being a frame ptr?
-      if (RI->hasFP(MF))
+      if (RI->hasFP(MF) || MFI->getReserveFP())
         // If so, don't allocate RBP.
         return end() - 1;
       else
diff --git a/lib/Target/X86/X86TargetMachine.cpp b/lib/Target/X86/X86TargetMachine.cpp
index 56ddaf858a..98f0dc0c48 100644
--- a/lib/Target/X86/X86TargetMachine.cpp
+++ b/lib/Target/X86/X86TargetMachine.cpp
@@ -160,6 +160,7 @@ bool X86TargetMachine::addInstSelector(PassManagerBase &PM,
 
 bool X86TargetMachine::addPreRegAlloc(PassManagerBase &PM,
                                       CodeGenOpt::Level OptLevel) {
+  PM.add(createX86MaxStackAlignmentHeuristicPass());
   return false;  // -print-machineinstr shouldn't print after this.
 }