Add a target hook to allow changing the tail duplication limit based on the

contents of the block to be duplicated.  Use this for ARM Cortex A8/9 to
be more aggressive tail duplicating indirect branches, since it makes it
much more likely that they will be predicted in the branch target buffer.
Testcase coming soon.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@89187 91177308-0d34-0410-b5e6-96231b3b80d8

6 files changed, 32 insertions, 4 deletions

diff --git a/include/llvm/Target/TargetInstrInfo.h b/include/llvm/Target/TargetInstrInfo.h
index 40b7780dfd..43fd54e183 100644
--- a/include/llvm/Target/TargetInstrInfo.h
+++ b/include/llvm/Target/TargetInstrInfo.h
@@ -536,6 +536,13 @@ public:
   /// length.
   virtual unsigned getInlineAsmLength(const char *Str,
                                       const MCAsmInfo &MAI) const;
+
+  /// TailDuplicationLimit - Returns the limit on the number of instructions
+  /// in basic block MBB beyond which it will not be tail-duplicated.
+  virtual unsigned TailDuplicationLimit(const MachineBasicBlock &MBB,
+                                        unsigned DefaultLimit) const {
+    return DefaultLimit;
+  }
 };
 
 /// TargetInstrInfoImpl - This is the default implementation of
diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp
index d344af06fd..94bfb7204b 100644
--- a/lib/CodeGen/BranchFolding.cpp
+++ b/lib/CodeGen/BranchFolding.cpp
@@ -1033,12 +1033,13 @@ bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB,
   if (TailBB->isSuccessor(TailBB))
     return false;
 
-  // Duplicate up to one less than the tail-merge threshold. When optimizing
-  // for size, duplicate only one, because one branch instruction can be
-  // eliminated to compensate for the duplication.
+  // Set the limit on the number of instructions to duplicate, with a default
+  // of one less than the tail-merge threshold. When optimizing for size,
+  // duplicate only one, because one branch instruction can be eliminated to
+  // compensate for the duplication.
   unsigned MaxDuplicateCount =
     MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ?
-      1 : (TailMergeSize - 1);
+    1 : TII->TailDuplicationLimit(*TailBB, TailMergeSize - 1);
 
   // Check the instructions in the block to determine whether tail-duplication
   // is invalid or unlikely to be profitable.
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 828041c64e..b50b6098dd 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -1005,6 +1005,16 @@ bool ARMBaseInstrInfo::isIdentical(const MachineInstr *MI0,
   return TargetInstrInfoImpl::isIdentical(MI0, MI1, MRI);
 }
 
+unsigned ARMBaseInstrInfo::TailDuplicationLimit(const MachineBasicBlock &MBB,
+                                                unsigned DefaultLimit) const {
+  // If the target processor can predict indirect branches, it is highly
+  // desirable to duplicate them, since it can often make them predictable.
+  if (!MBB.empty() && isIndirectBranchOpcode(MBB.back().getOpcode()) &&
+      getSubtarget().hasBranchTargetBuffer())
+    return DefaultLimit + 2;
+  return DefaultLimit;
+}
+
 /// getInstrPredicate - If instruction is predicated, returns its predicate
 /// condition, otherwise returns AL. It also returns the condition code
 /// register by reference.
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.h b/lib/Target/ARM/ARMBaseInstrInfo.h
index 29855e2c61..73e854faf2 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.h
+++ b/lib/Target/ARM/ARMBaseInstrInfo.h
@@ -272,6 +272,9 @@ public:
 
   virtual bool isIdentical(const MachineInstr *MI, const MachineInstr *Other,
                            const MachineRegisterInfo *MRI) const;
+
+  virtual unsigned TailDuplicationLimit(const MachineBasicBlock &MBB,
+                                        unsigned DefaultLimit) const;
 };
 
 static inline
diff --git a/lib/Target/ARM/ARMSubtarget.cpp b/lib/Target/ARM/ARMSubtarget.cpp
index 87b78a9ee5..432ed78c19 100644
--- a/lib/Target/ARM/ARMSubtarget.cpp
+++ b/lib/Target/ARM/ARMSubtarget.cpp
@@ -109,6 +109,8 @@ ARMSubtarget::ARMSubtarget(const std::string &TT, const std::string &FS,
     if (UseNEONFP.getPosition() == 0)
       UseNEONForSinglePrecisionFP = true;
   }
+  HasBranchTargetBuffer = (CPUString == "cortex-a8" ||
+                           CPUString == "cortex-a9");
 }
 
 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol.
diff --git a/lib/Target/ARM/ARMSubtarget.h b/lib/Target/ARM/ARMSubtarget.h
index a95c8a5ccf..3d0e01e99b 100644
--- a/lib/Target/ARM/ARMSubtarget.h
+++ b/lib/Target/ARM/ARMSubtarget.h
@@ -50,6 +50,9 @@ protected:
   /// determine if NEON should actually be used.
   bool UseNEONForSinglePrecisionFP;
 
+  /// HasBranchTargetBuffer - True if processor can predict indirect branches.
+  bool HasBranchTargetBuffer;
+
   /// IsThumb - True if we are in thumb mode, false if in ARM mode.
   bool IsThumb;
 
@@ -123,6 +126,8 @@ protected:
   bool isThumb2() const { return IsThumb && (ThumbMode == Thumb2); }
   bool hasThumb2() const { return ThumbMode >= Thumb2; }
 
+  bool hasBranchTargetBuffer() const { return HasBranchTargetBuffer; }
+
   bool isR9Reserved() const { return IsR9Reserved; }
 
   const std::string & getCPUString() const { return CPUString; }