[Hexagon] Enable the post-RA scheduler

The aggressive anti-dependency breaker can rename the restored callee-
saved registers. To prevent this, mark these registers are live on all
paths to the return/tail-call instructions, and add implicit use operands
for them to these instructions.

llvm-svn: 270898

lib/Target/Hexagon/HexagonFrameLowering.cpp

@@ -292,6 +292,26 @@ namespace {
     return false;
   }
 
+  /// Returns the "return" instruction from this block, or nullptr if there
+  /// isn't any.
+  MachineInstr *getReturn(MachineBasicBlock &MBB) {
+    for (auto &I : MBB)
+      if (I.isReturn())
+        return &I;
+    return nullptr;
+  }
+
+  bool isRestoreCall(unsigned Opc) {
+    switch (Opc) {
+      case Hexagon::RESTORE_DEALLOC_RET_JMP_V4:
+      case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC:
+      case Hexagon::RESTORE_DEALLOC_BEFORE_TAILCALL_V4:
+      case Hexagon::RESTORE_DEALLOC_BEFORE_TAILCALL_V4_PIC:
+        return true;
+    }
+    return false;
+  }
+
   inline bool isOptNone(const MachineFunction &MF) {
     return MF.getFunction()->hasFnAttribute(Attribute::OptimizeNone) ||
            MF.getTarget().getOptLevel() == CodeGenOpt::None;
@@ -445,6 +465,28 @@ void HexagonFrameLowering::emitPrologue(MachineFunction &MF,
     for (auto &B : MF)
       if (B.isReturnBlock())
         insertEpilogueInBlock(B);
+
+    for (auto &B : MF) {
+      if (B.empty())
+        continue;
+      MachineInstr *RetI = getReturn(B);
+      if (!RetI || isRestoreCall(RetI->getOpcode()))
+        continue;
+      for (auto &R : CSI)
+        RetI->addOperand(MachineOperand::CreateReg(R.getReg(), false, true));
+    }
+  }
+
+  if (EpilogB) {
+    // If there is an epilog block, it may not have a return instruction.
+    // In such case, we need to add the callee-saved registers as live-ins
+    // in all blocks on all paths from the epilog to any return block.
+    unsigned MaxBN = 0;
+    for (auto &B : MF)
+      if (B.getNumber() >= 0)
+        MaxBN = std::max(MaxBN, unsigned(B.getNumber()));
+    BitVector DoneT(MaxBN+1), DoneF(MaxBN+1), Path(MaxBN+1);
+    updateExitPaths(*EpilogB, EpilogB, DoneT, DoneF, Path);
   }
 }
 
@@ -544,13 +586,7 @@ void HexagonFrameLowering::insertEpilogueInBlock(MachineBasicBlock &MBB) const {
   auto &HRI = *HST.getRegisterInfo();
   unsigned SP = HRI.getStackRegister();
 
-  MachineInstr *RetI = nullptr;
-  for (auto &I : MBB) {
-    if (!I.isReturn())
-      continue;
-    RetI = &I;
-    break;
-  }
+  MachineInstr *RetI = getReturn(MBB);
   unsigned RetOpc = RetI ? RetI->getOpcode() : 0;
 
   MachineBasicBlock::iterator InsertPt = MBB.getFirstTerminator();
@@ -614,6 +650,51 @@ void HexagonFrameLowering::insertEpilogueInBlock(MachineBasicBlock &MBB) const {
 }
 
 
+bool HexagonFrameLowering::updateExitPaths(MachineBasicBlock &MBB,
+      MachineBasicBlock *RestoreB, BitVector &DoneT, BitVector &DoneF,
+      BitVector &Path) const {
+  assert(MBB.getNumber() >= 0);
+  unsigned BN = MBB.getNumber();
+  if (Path[BN] || DoneF[BN])
+    return false;
+  if (DoneT[BN])
+    return true;
+
+  auto &CSI = MBB.getParent()->getFrameInfo()->getCalleeSavedInfo();
+
+  Path[BN] = true;
+  bool ReachedExit = false;
+  for (auto &SB : MBB.successors())
+    ReachedExit |= updateExitPaths(*SB, RestoreB, DoneT, DoneF, Path);
+
+  if (!MBB.empty() && MBB.back().isReturn()) {
+    // Add implicit uses of all callee-saved registers to the reached
+    // return instructions. This is to prevent the anti-dependency breaker
+    // from renaming these registers.
+    MachineInstr &RetI = MBB.back();
+    if (!isRestoreCall(RetI.getOpcode()))
+      for (auto &R : CSI)
+        RetI.addOperand(MachineOperand::CreateReg(R.getReg(), false, true));
+    ReachedExit = true;
+  }
+
+  // We don't want to add unnecessary live-ins to the restore block: since
+  // the callee-saved registers are being defined in it, the entry of the
+  // restore block cannot be on the path from the definitions to any exit.
+  if (ReachedExit && &MBB != RestoreB) {
+    for (auto &R : CSI)
+      if (!MBB.isLiveIn(R.getReg()))
+        MBB.addLiveIn(R.getReg());
+    DoneT[BN] = true;
+  }
+  if (!ReachedExit)
+    DoneF[BN] = true;
+
+  Path[BN] = false;
+  return ReachedExit;
+}
+
+
 namespace {
   bool IsAllocFrame(MachineBasicBlock::const_iterator It) {
     if (!It->isBundle())

lib/Target/Hexagon/HexagonFrameLowering.h

@@ -91,6 +91,8 @@ private:
       const HexagonRegisterInfo &HRI, bool &PrologueStubs) const;
   bool insertCSRRestoresInBlock(MachineBasicBlock &MBB, const CSIVect &CSI,
       const HexagonRegisterInfo &HRI) const;
+  bool updateExitPaths(MachineBasicBlock &MBB, MachineBasicBlock *RestoreB,
+      BitVector &DoneT, BitVector &DoneF, BitVector &Path) const;
   void insertCFIInstructionsAt(MachineBasicBlock &MBB,
       MachineBasicBlock::iterator At) const;
 

lib/Target/Hexagon/HexagonSubtarget.h

@@ -106,6 +106,7 @@ public:
   bool enableMachineSchedDefaultSched() const override { return false; }
 
   AntiDepBreakMode getAntiDepBreakMode() const override { return ANTIDEP_ALL; }
+  bool enablePostRAScheduler() const override { return true; }
 
   const std::string &getCPUString () const { return CPUString; }
 

test/CodeGen/Hexagon/adde.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -disable-hsdr -hexagon-expand-condsets=0 -hexagon-bit=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -disable-hsdr -hexagon-expand-condsets=0 -hexagon-bit=0 -disable-post-ra < %s | FileCheck %s
 
 ; CHECK: r{{[0-9]+:[0-9]+}} = combine(#0, #1)
 ; CHECK: r{{[0-9]+:[0-9]+}} = combine(#0, #0)

test/CodeGen/Hexagon/newvaluestore.ll

@@ -1,22 +1,13 @@
-; RUN: llc -march=hexagon -mcpu=hexagonv4 -disable-hexagon-misched < %s | FileCheck %s
-; Check that we generate new value store packet in V4
+; RUN: llc -march=hexagon < %s | FileCheck %s
+; Check that we generate new value store.
 
 @i = global i32 0, align 4
-@j = global i32 10, align 4
-@k = global i32 100, align 4
 
-define i32 @main() nounwind {
+define i32 @main(i32 %x, i32* %p) nounwind {
 entry:
 ; CHECK: memw(r{{[0-9]+}}+#{{[0-9]+}}) = r{{[0-9]+}}.new
-  %number1 = alloca i32, align 4
-  %number2 = alloca i32, align 4
-  %number3 = alloca i32, align 4
-  %0 = load i32 , i32 * @i, align 4
-  store i32 %0, i32* %number1, align 4
-  %1 = load i32 , i32 * @j, align 4
-  store i32 %1, i32* %number2, align 4
-  %2 = load i32 , i32 * @k, align 4
-  store i32 %2, i32* %number3, align 4
-  ret i32 %0
+  %t0 = load i32, i32* @i, align 4
+  store i32 %t0, i32* %p, align 4
+  ret i32 %x
 }
 

test/CodeGen/Hexagon/sube.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -disable-hsdr -hexagon-expand-condsets=0 -hexagon-bit=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -disable-hsdr -hexagon-expand-condsets=0 -hexagon-bit=0 -disable-post-ra < %s | FileCheck %s
 
 ; CHECK: r{{[0-9]+:[0-9]+}} = combine(#0, #1)
 ; CHECK: r{{[0-9]+:[0-9]+}} = combine(#0, #0)

test/CodeGen/Hexagon/v60Intrins.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -mcpu=hexagonv60 -O2  < %s | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv60 -O2 -disable-post-ra  < %s | FileCheck %s
 
 ; CHECK: q{{[0-3]}} = vand(v{{[0-9]*}},r{{[0-9]*}})
 ; CHECK: q{{[0-3]}} = vsetq(r{{[0-9]*}})