Change if-conversion block size limit checks to add some flexibility.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106901 91177308-0d34-0410-b5e6-96231b3b80d8

10 files changed, 101 insertions, 66 deletions

diff --git a/include/llvm/Target/TargetInstrInfo.h b/include/llvm/Target/TargetInstrInfo.h
index e586a06..2c6080b 100644
--- a/include/llvm/Target/TargetInstrInfo.h
+++ b/include/llvm/Target/TargetInstrInfo.h
@@ -331,6 +331,31 @@ public:
                                    MachineBasicBlock::iterator MBBI) const {
     return true;
   }
+
+  /// isProfitableToIfCvt - Return true if it's profitable to first "NumInstrs"
+  /// of the specified basic block.
+  virtual
+  bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumInstrs) const {
+    return false;
+  }
+  
+  /// isProfitableToIfCvt - Second variant of isProfitableToIfCvt, this one
+  /// checks for the case where two basic blocks from true and false path
+  /// of a if-then-else (diamond) are predicated on mutally exclusive
+  /// predicates.
+  virtual bool
+  isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumTInstrs,
+                      MachineBasicBlock &FMBB, unsigned NumFInstrs) const {
+    return false;
+  }
+
+  /// isProfitableToDupForIfCvt - Return true if it's profitable for
+  /// if-converter to duplicate a specific number of instructions in the
+  /// specified MBB to enable if-conversion.
+  virtual bool
+  isProfitableToDupForIfCvt(MachineBasicBlock &MBB,unsigned NumInstrs) const {
+    return false;
+  }
   
   /// copyRegToReg - Emit instructions to copy between a pair of registers. It
   /// returns false if the target does not how to copy between the specified
diff --git a/include/llvm/Target/TargetLowering.h b/include/llvm/Target/TargetLowering.h
index 4aa8a4d..eb048f6 100644
--- a/include/llvm/Target/TargetLowering.h
+++ b/include/llvm/Target/TargetLowering.h
@@ -683,19 +683,6 @@ public:
     return JumpBufAlignment;
   }
 
-  /// getIfCvtBlockLimit - returns the target specific if-conversion block size
-  /// limit. Any block whose size is greater should not be predicated.
-  unsigned getIfCvtBlockSizeLimit() const {
-    return IfCvtBlockSizeLimit;
-  }
-
-  /// getIfCvtDupBlockLimit - returns the target specific size limit for a
-  /// block to be considered for duplication. Any block whose size is greater
-  /// should not be duplicated to facilitate its predication.
-  unsigned getIfCvtDupBlockSizeLimit() const {
-    return IfCvtDupBlockSizeLimit;
-  }
-
   /// getPrefLoopAlignment - return the preferred loop alignment.
   ///
   unsigned getPrefLoopAlignment() const {
@@ -1078,19 +1065,6 @@ protected:
     JumpBufAlignment = Align;
   }
 
-  /// setIfCvtBlockSizeLimit - Set the target's if-conversion block size
-  /// limit (in number of instructions); default is 2.
-  void setIfCvtBlockSizeLimit(unsigned Limit) {
-    IfCvtBlockSizeLimit = Limit;
-  }
-
-  /// setIfCvtDupBlockSizeLimit - Set the target's block size limit (in number
-  /// of instructions) to be considered for code duplication during
-  /// if-conversion; default is 2.
-  void setIfCvtDupBlockSizeLimit(unsigned Limit) {
-    IfCvtDupBlockSizeLimit = Limit;
-  }
-
   /// setPrefLoopAlignment - Set the target's preferred loop alignment. Default
   /// alignment is zero, it means the target does not care about loop alignment.
   void setPrefLoopAlignment(unsigned Align) {
@@ -1538,14 +1512,6 @@ private:
   /// buffers
   unsigned JumpBufAlignment;
 
-  /// IfCvtBlockSizeLimit - The maximum allowed size for a block to be
-  /// if-converted.
-  unsigned IfCvtBlockSizeLimit;
-
-  /// IfCvtDupBlockSizeLimit - The maximum allowed size for a block to be
-  /// duplicated during if-conversion.
-  unsigned IfCvtDupBlockSizeLimit;
-
   /// PrefLoopAlignment - The perferred loop alignment.
   ///
   unsigned PrefLoopAlignment;
diff --git a/lib/CodeGen/IfConversion.cpp b/lib/CodeGen/IfConversion.cpp
index 0b40d8e..ea4e68d 100644
--- a/lib/CodeGen/IfConversion.cpp
+++ b/lib/CodeGen/IfConversion.cpp
@@ -188,8 +188,14 @@ namespace {
                                bool IgnoreBr = false);
     void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI);
 
-    bool MeetIfcvtSizeLimit(unsigned Size) const {
-      return Size > 0 && Size <= TLI->getIfCvtBlockSizeLimit();
+    bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, unsigned Size) const {
+      return Size > 0 && TII->isProfitableToIfCvt(BB, Size);
+    }
+
+    bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, unsigned TSize,
+                            MachineBasicBlock &FBB, unsigned FSize) const {
+      return TSize > 0 && FSize > 0 &&
+        TII->isProfitableToIfCvt(TBB, TSize, FBB, FSize);
     }
 
     // blockAlwaysFallThrough - Block ends without a terminator.
@@ -436,7 +442,7 @@ bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const {
 
   if (TrueBBI.BB->pred_size() > 1) {
     if (TrueBBI.CannotBeCopied ||
-        TrueBBI.NonPredSize > TLI->getIfCvtDupBlockSizeLimit())
+        !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize))
       return false;
     Dups = TrueBBI.NonPredSize;
   }
@@ -473,7 +479,7 @@ bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI,
           ++Size;
       }
     }
-    if (Size > TLI->getIfCvtDupBlockSizeLimit())
+    if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size))
       return false;
     Dups = Size;
   }
@@ -761,8 +767,8 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
   bool FNeedSub = FalseBBI.Predicate.size() > 0;
   bool Enqueued = false;
   if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) &&
-      MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) &&
-      MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) &&
+      MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize - (Dups + Dups2),
+                         *FalseBBI.BB, FalseBBI.NonPredSize - (Dups + Dups2)) &&
       FeasibilityAnalysis(TrueBBI, BBI.BrCond) &&
       FeasibilityAnalysis(FalseBBI, RevCond)) {
     // Diamond:
@@ -779,7 +785,7 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
   }
 
   if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) &&
-      MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
+      MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize) &&
       FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
     // Triangle:
     //   EBB
@@ -793,14 +799,14 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
   }
 
   if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) &&
-      MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
+      MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize) &&
       FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
     Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups));
     Enqueued = true;
   }
 
   if (ValidSimple(TrueBBI, Dups) &&
-      MeetIfcvtSizeLimit(TrueBBI.NonPredSize) &&
+      MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize) &&
       FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
     // Simple (split, no rejoin):
     //   EBB
@@ -816,21 +822,21 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
   if (CanRevCond) {
     // Try the other path...
     if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) &&
-        MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
+        MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize) &&
         FeasibilityAnalysis(FalseBBI, RevCond, true)) {
       Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups));
       Enqueued = true;
     }
 
     if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) &&
-        MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
+        MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize) &&
         FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
       Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups));
       Enqueued = true;
     }
 
     if (ValidSimple(FalseBBI, Dups) &&
-        MeetIfcvtSizeLimit(FalseBBI.NonPredSize) &&
+        MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize) &&
         FeasibilityAnalysis(FalseBBI, RevCond)) {
       Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups));
       Enqueued = true;
diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 7ad0d38..8019116 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -578,8 +578,6 @@ TargetLowering::TargetLowering(const TargetMachine &tm,
   SchedPreferenceInfo = Sched::Latency;
   JumpBufSize = 0;
   JumpBufAlignment = 0;
-  IfCvtBlockSizeLimit = 2;
-  IfCvtDupBlockSizeLimit = 0;
   PrefLoopAlignment = 0;
   ShouldFoldAtomicFences = false;
 
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 7ef7fd6..c0d0509 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -1447,6 +1447,24 @@ bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
   return false;
 }
 
+bool ARMBaseInstrInfo::
+isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumInstrs) const {
+  if (!NumInstrs)
+    return false;
+  if (Subtarget.getCPUString() == "generic")
+    // Generic (and overly aggressive) if-conversion limits for testing.
+    return NumInstrs <= 10;
+  else if (Subtarget.hasV7Ops())
+    return NumInstrs <= 3;
+  return NumInstrs <= 2;
+}
+  
+bool ARMBaseInstrInfo::
+isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
+                    MachineBasicBlock &FMBB, unsigned NumF) const {
+  return NumT && NumF && NumT <= 2 && NumF <= 2;
+}
+
 /// 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 d7d9f52..f673c99 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.h
+++ b/lib/Target/ARM/ARMBaseInstrInfo.h
@@ -343,6 +343,17 @@ public:
   virtual bool isSchedulingBoundary(const MachineInstr *MI,
                                     const MachineBasicBlock *MBB,
                                     const MachineFunction &MF) const;
+
+  virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB,
+                                   unsigned NumInstrs) const;
+
+  virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB,unsigned NumT,
+                                   MachineBasicBlock &FMBB,unsigned NumF) const;
+
+  virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
+                                         unsigned NumInstrs) const {
+    return NumInstrs && NumInstrs == 1;
+  }
 };
 
 static inline
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 07de1dd..d4332f0 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -531,23 +531,6 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
   else
     setSchedulingPreference(Sched::Hybrid);
 
-  // FIXME: If-converter should use instruction latency to determine
-  // profitability rather than relying on fixed limits.
-  if (Subtarget->getCPUString() == "generic") {
-    // Generic (and overly aggressive) if-conversion limits.
-    setIfCvtBlockSizeLimit(10);
-    setIfCvtDupBlockSizeLimit(2);
-  } else if (Subtarget->hasV7Ops()) {
-    setIfCvtBlockSizeLimit(3);
-    setIfCvtDupBlockSizeLimit(1);
-  } else if (Subtarget->hasV6Ops()) {
-    setIfCvtBlockSizeLimit(2);
-    setIfCvtDupBlockSizeLimit(1);
-  } else {
-    setIfCvtBlockSizeLimit(3);
-    setIfCvtDupBlockSizeLimit(2);
-  }
-
   maxStoresPerMemcpy = 1;   //// temporary - rewrite interface to use type
   // Do not enable CodePlacementOpt for now: it currently runs after the
   // ARMConstantIslandPass and messes up branch relaxation and placement
diff --git a/lib/Target/ARM/Thumb2InstrInfo.cpp b/lib/Target/ARM/Thumb2InstrInfo.cpp
index 42fe509..d7c86c6 100644
--- a/lib/Target/ARM/Thumb2InstrInfo.cpp
+++ b/lib/Target/ARM/Thumb2InstrInfo.cpp
@@ -24,9 +24,18 @@
 #include "llvm/CodeGen/MachineMemOperand.h"
 #include "llvm/CodeGen/PseudoSourceValue.h"
 #include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/CommandLine.h"
 
 using namespace llvm;
 
+static cl::opt<unsigned>
+IfCvtLimit("thumb2-ifcvt-limit (default 3)",
+           cl::Hidden, cl::init(3));
+
+static cl::opt<unsigned>
+IfCvtDiamondLimit("thumb2-ifcvt-diamond-limit (default 3)",
+                  cl::Hidden, cl::init(3));
+
 Thumb2InstrInfo::Thumb2InstrInfo(const ARMSubtarget &STI)
   : ARMBaseInstrInfo(STI), RI(*this, STI) {
 }
@@ -94,6 +103,20 @@ Thumb2InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
   return llvm::getITInstrPredicate(MBBI, PredReg) == ARMCC::AL;
 }
 
+bool Thumb2InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
+                                          unsigned NumInstrs) const {
+  return NumInstrs && NumInstrs <= IfCvtLimit;
+}
+  
+bool Thumb2InstrInfo::
+isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
+                    MachineBasicBlock &FMBB, unsigned NumF) const {
+  // FIXME: Catch optimization such as:
+  //        r0 = movne
+  //        r0 = moveq
+  return NumT && NumF &&
+    NumT <= (IfCvtDiamondLimit) && NumF <= (IfCvtDiamondLimit);
+}
 
 bool
 Thumb2InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
diff --git a/lib/Target/ARM/Thumb2InstrInfo.h b/lib/Target/ARM/Thumb2InstrInfo.h
index 8b01bcb..13b801a 100644
--- a/lib/Target/ARM/Thumb2InstrInfo.h
+++ b/lib/Target/ARM/Thumb2InstrInfo.h
@@ -38,6 +38,11 @@ public:
   bool isLegalToSplitMBBAt(MachineBasicBlock &MBB,
                            MachineBasicBlock::iterator MBBI) const;
 
+  bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumInstrs) const;
+  
+  bool isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumTInstrs,
+                           MachineBasicBlock &FMBB, unsigned NumFInstrs) const;
+
   bool copyRegToReg(MachineBasicBlock &MBB,
                     MachineBasicBlock::iterator I,
                     unsigned DestReg, unsigned SrcReg,
diff --git a/test/CodeGen/ARM/lsr-code-insertion.ll b/test/CodeGen/ARM/lsr-code-insertion.ll
index 1bbb96d..b8c543b 100644
--- a/test/CodeGen/ARM/lsr-code-insertion.ll
+++ b/test/CodeGen/ARM/lsr-code-insertion.ll
@@ -1,4 +1,4 @@
-; RUN: llc < %s -stats |& grep {39.*Number of machine instrs printed}
+; RUN: llc < %s -stats |& grep {38.*Number of machine instrs printed}
 ; RUN: llc < %s -stats |& not grep {.*Number of re-materialization}
 ; This test really wants to check that the resultant "cond_true" block only 
 ; has a single store in it, and that cond_true55 only has code to materialize