Infrastructure for dynamic stack realignment on ARM. For now, this is off by

default behind a command line option. This will enable better performance for
vectors on NEON enabled processors.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85333 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARMBaseRegisterInfo.cpp

@@ -29,6 +29,7 @@
 #include "llvm/CodeGen/MachineLocation.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/RegisterScavenging.h"
+#include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Target/TargetFrameInfo.h"
@@ -48,6 +49,10 @@ static cl::opt<bool>
 ReuseFrameIndexVals("arm-reuse-frame-index-vals", cl::Hidden, cl::init(false),
           cl::desc("Reuse repeated frame index values"));
 
+static cl::opt<bool>
+ARMDynamicStackAlign("arm-dynamic-stack-alignment", cl::Hidden, cl::init(false),
+          cl::desc("Dynamically re-align the stack as needed"));
+
 unsigned ARMBaseRegisterInfo::getRegisterNumbering(unsigned RegEnum,
                                                    bool *isSPVFP) {
   if (isSPVFP)
@@ -466,6 +471,21 @@ ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg,
   }
 }
 
+static unsigned calculateMaxStackAlignment(const MachineFrameInfo *FFI) {
+  unsigned MaxAlign = 0;
+
+  for (int i = FFI->getObjectIndexBegin(),
+         e = FFI->getObjectIndexEnd(); i != e; ++i) {
+    if (FFI->isDeadObjectIndex(i))
+      continue;
+
+    unsigned Align = FFI->getObjectAlignment(i);
+    MaxAlign = std::max(MaxAlign, Align);
+  }
+
+  return MaxAlign;
+}
+
 /// hasFP - Return true if the specified function should have a dedicated frame
 /// pointer register.  This is true if the function has variable sized allocas
 /// or if frame pointer elimination is disabled.
@@ -473,10 +493,28 @@ ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg,
 bool ARMBaseRegisterInfo::hasFP(const MachineFunction &MF) const {
   const MachineFrameInfo *MFI = MF.getFrameInfo();
   return (NoFramePointerElim ||
+          needsStackRealignment(MF) ||
           MFI->hasVarSizedObjects() ||
           MFI->isFrameAddressTaken());
 }
 
+bool ARMBaseRegisterInfo::
+needsStackRealignment(const MachineFunction &MF) const {
+  // Only do this for ARM if explicitly enabled
+  // FIXME: Once it's passing all the tests, enable by default
+  if (!ARMDynamicStackAlign)
+    return false;
+
+  const MachineFrameInfo *MFI = MF.getFrameInfo();
+  const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+  unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
+  return (RealignStack &&
+          !AFI->isThumb1OnlyFunction() &&
+          (MFI->getMaxAlignment() > StackAlign) &&
+          !MFI->hasVarSizedObjects());
+
+}
+
 bool ARMBaseRegisterInfo::cannotEliminateFrame(const MachineFunction &MF) const {
   const MachineFrameInfo *MFI = MF.getFrameInfo();
   if (NoFramePointerElim && MFI->hasCalls())
@@ -552,6 +590,16 @@ ARMBaseRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
   SmallVector<unsigned, 4> UnspilledCS2GPRs;
   ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
 
+  MachineFrameInfo *MFI = MF.getFrameInfo();
+
+  // Calculate and set max stack object alignment early, so we can decide
+  // whether we will need stack realignment (and thus FP).
+  if (ARMDynamicStackAlign) {
+    unsigned MaxAlign = std::max(MFI->getMaxAlignment(),
+                                 calculateMaxStackAlignment(MFI));
+    MFI->setMaxAlignment(MaxAlign);
+  }
+
   // Don't spill FP if the frame can be eliminated. This is determined
   // by scanning the callee-save registers to see if any is used.
   const unsigned *CSRegs = getCalleeSavedRegs();
@@ -1085,16 +1133,28 @@ ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   int FrameIndex = MI.getOperand(i).getIndex();
   int Offset = MFI->getObjectOffset(FrameIndex) + MFI->getStackSize() + SPAdj;
 
+  // When doing dynamic stack realignment, all of these need to change(?)
   if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
     Offset -= AFI->getGPRCalleeSavedArea1Offset();
   else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
     Offset -= AFI->getGPRCalleeSavedArea2Offset();
   else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
     Offset -= AFI->getDPRCalleeSavedAreaOffset();
-  else if (hasFP(MF) && AFI->hasStackFrame()) {
+  else if (needsStackRealignment(MF)) {
+    // When dynamically realigning the stack, use the frame pointer for
+    // parameters, and the stack pointer for locals.
+    assert (hasFP(MF) && "dynamic stack realignment without a FP!");
+    if (FrameIndex < 0) {
+      FrameReg = getFrameRegister(MF);
+      Offset -= AFI->getFramePtrSpillOffset();
+      // When referencing from the frame pointer, stack pointer adjustments
+      // don't matter.
+      SPAdj = 0;
+    }
+  } else if (hasFP(MF) && AFI->hasStackFrame()) {
     assert(SPAdj == 0 && "Unexpected stack offset!");
     // Use frame pointer to reference fixed objects unless this is a
-    // frameless function,
+    // frameless function.
     FrameReg = getFrameRegister(MF);
     Offset -= AFI->getFramePtrSpillOffset();
   }
@@ -1303,6 +1363,18 @@ emitPrologue(MachineFunction &MF) const {
   AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
   AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
   AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
+
+  // If we need dynamic stack realignment, do it here.
+  if (needsStackRealignment(MF)) {
+    unsigned Opc;
+    unsigned MaxAlign = MFI->getMaxAlignment();
+    assert (!AFI->isThumb1OnlyFunction());
+    Opc = AFI->isThumbFunction() ? ARM::t2BICri : ARM::BICri;
+
+    AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(Opc), ARM::SP)
+                                  .addReg(ARM::SP, RegState::Kill)
+                                  .addImm(MaxAlign-1)));
+  }
 }
 
 static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {

lib/Target/ARM/ARMBaseRegisterInfo.h

@@ -96,6 +96,8 @@ public:
 
   bool hasFP(const MachineFunction &MF) const;
 
+  bool needsStackRealignment(const MachineFunction &MF) const;
+
   bool cannotEliminateFrame(const MachineFunction &MF) const;
 
   void processFunctionBeforeCalleeSavedScan(MachineFunction &MF,