Add materialization of virtual base registers for frame indices allocated into

the local block. Resolve references to those indices to a new base register.
For simplification and testing purposes, a new virtual base register is
allocated for each frame index being resolved. The result is truly horrible,
but correct, code that's good for exercising the new code paths.

Next up is adding thumb1 support, which should be very simple. Following that
will be adding base register re-use and implementing a reasonable ARM
heuristic for when a virtual base register should be generated at all.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@111315 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 74 insertions, 3 deletions

diff --git a/lib/CodeGen/LocalStackSlotAllocation.cpp b/lib/CodeGen/LocalStackSlotAllocation.cpp
index 065cf80ca4..3927beec11 100644
--- a/lib/CodeGen/LocalStackSlotAllocation.cpp
+++ b/lib/CodeGen/LocalStackSlotAllocation.cpp
@@ -27,6 +27,7 @@
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
@@ -160,6 +161,14 @@ void LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
          E = Fn.end(); BB != E; ++BB) {
     for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
       MachineInstr *MI = I;
+      // Debug value instructions can't be out of range, so they don't need
+      // any updates.
+      // FIXME: When we extend this stuff to handle functions with both
+      // VLAs and dynamic realignment, we should update the debug values
+      // to reference the new base pointer when possible.
+      if (MI->isDebugValue())
+        continue;
+
       // For now, allocate the base register(s) within the basic block
       // where they're used, and don't try to keep them around outside
       // of that. It may be beneficial to try sharing them more broadly
@@ -169,8 +178,12 @@ void LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
       for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
         // Consider replacing all frame index operands that reference
         // an object allocated in the local block.
-        if (MI->getOperand(i).isFI() &&
-            MFI->isObjectPreAllocated(MI->getOperand(i).getIndex())) {
+        if (MI->getOperand(i).isFI()) {
+          int FrameIdx = MI->getOperand(i).getIndex();
+          // Don't try this with values not in the local block.
+          if (!MFI->isObjectPreAllocated(FrameIdx))
+            continue;
+
           DEBUG(dbgs() << "Considering: " << *MI);
           if (TRI->needsFrameBaseReg(MI, i)) {
             DEBUG(dbgs() << "  Replacing FI in: " << *MI);
@@ -182,6 +195,17 @@ void LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
             // create a new one.
             // FIXME: For the moment, just always create a new one.
 
+            const TargetRegisterClass *RC = TRI->getPointerRegClass();
+            unsigned BaseReg = Fn.getRegInfo().createVirtualRegister(RC);
+
+            // Tell the target to insert the instruction to initialize
+            // the base register.
+            TRI->materializeFrameBaseRegister(I, BaseReg, FrameIdx);
+
+            // Modify the instruction to use the new base register rather
+            // than the frame index operand.
+            TRI->resolveFrameIndex(I, BaseReg, 0);
+
             ++NumBaseRegisters;
             ++NumReplacements;
           }
diff --git a/lib/CodeGen/PrologEpilogInserter.cpp b/lib/CodeGen/PrologEpilogInserter.cpp
index 7b67473636..569faceffd 100644
--- a/lib/CodeGen/PrologEpilogInserter.cpp
+++ b/lib/CodeGen/PrologEpilogInserter.cpp
@@ -560,7 +560,7 @@ void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) {
   // check for whether the frame is large enough to want to use virtual
   // frame index registers. Functions which don't want/need this optimization
   // will continue to use the existing code path.
-  if (EnableLocalStackAlloc) {
+  if (EnableLocalStackAlloc && MFI->getLocalFrameSize()) {
     unsigned Align = MFI->getLocalFrameMaxAlign();
 
     // Adjust to alignment boundary.
diff --git a/lib/Target/ARM/ARMBaseRegisterInfo.cpp b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
index 5612f04c5f..23fe6216f6 100644
--- a/lib/Target/ARM/ARMBaseRegisterInfo.cpp
+++ b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
@@ -1407,6 +1407,49 @@ needsFrameBaseReg(MachineInstr *MI, unsigned operand) const {
   }
 }
 
+/// materializeFrameBaseRegister - Insert defining instruction(s) for
+/// BaseReg to be a pointer to FrameIdx before insertion point I.
+void ARMBaseRegisterInfo::
+materializeFrameBaseRegister(MachineBasicBlock::iterator I,
+                             unsigned BaseReg, int FrameIdx) const {
+  ARMFunctionInfo *AFI =
+    I->getParent()->getParent()->getInfo<ARMFunctionInfo>();
+  unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri : ARM::t2ADDri;
+  assert(!AFI->isThumb1OnlyFunction() &&
+         "This materializeFrameBaseRegister does not support Thumb1!");
+
+  MachineInstrBuilder MIB =
+    BuildMI(*I->getParent(), I, I->getDebugLoc(), TII.get(ADDriOpc), BaseReg)
+    .addFrameIndex(FrameIdx).addImm(0);
+  AddDefaultCC(AddDefaultPred(MIB));
+}
+
+void
+ARMBaseRegisterInfo::resolveFrameIndex(MachineBasicBlock::iterator I,
+                                       unsigned BaseReg, int64_t Offset) const {
+  MachineInstr &MI = *I;
+  MachineBasicBlock &MBB = *MI.getParent();
+  MachineFunction &MF = *MBB.getParent();
+  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+  int Off = Offset; // ARM doesn't need the general 64-bit offsets
+  unsigned i = 0;
+
+  assert(!AFI->isThumb1OnlyFunction() &&
+         "This resolveFrameIndex does not support Thumb1!");
+
+  while (!MI.getOperand(i).isFI()) {
+    ++i;
+    assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!");
+  }
+  bool Done = false;
+  if (!AFI->isThumbFunction())
+    Done = rewriteARMFrameIndex(MI, i, BaseReg, Off, TII);
+  else {
+    assert(AFI->isThumb2Function());
+    Done = rewriteT2FrameIndex(MI, i, BaseReg, Off, TII);
+  }
+  assert (Done && "Unable to resolve frame index!");
+}
 
 unsigned
 ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
diff --git a/lib/Target/ARM/ARMBaseRegisterInfo.h b/lib/Target/ARM/ARMBaseRegisterInfo.h
index f3ccfb4cc4..7570d1fd16 100644
--- a/lib/Target/ARM/ARMBaseRegisterInfo.h
+++ b/lib/Target/ARM/ARMBaseRegisterInfo.h
@@ -106,6 +106,10 @@ public:
   bool canRealignStack(const MachineFunction &MF) const;
   bool needsStackRealignment(const MachineFunction &MF) const;
   bool needsFrameBaseReg(MachineInstr *MI, unsigned operand) const;
+  void materializeFrameBaseRegister(MachineBasicBlock::iterator I,
+                                    unsigned BaseReg, int FrameIdx) const;
+  void resolveFrameIndex(MachineBasicBlock::iterator I,
+                         unsigned BaseReg, int64_t Offset) const;
 
   bool cannotEliminateFrame(const MachineFunction &MF) const;