ARM scheduler model: Swift has varying latencies, uops for simple ALU ops

[oota-llvm.git] / lib / Target / ARM / ARMBaseInstrInfo.cpp

diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index b5fa2763bf606b868e9f8a1e4ccd8689d255b31b..9e68ff44890e1d9cb455eb8a2ca22617f95071bb 100644 (file)
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -18,9 +18,7 @@
 #include "ARMHazardRecognizer.h"
 #include "ARMMachineFunctionInfo.h"
 #include "MCTargetDesc/ARMAddressingModes.h"
-#include "llvm/Constants.h"
-#include "llvm/Function.h"
-#include "llvm/GlobalValue.h"
+#include "llvm/ADT/STLExtras.h"
 #include "llvm/CodeGen/LiveVariables.h"
 #include "llvm/CodeGen/MachineConstantPool.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
@@ -29,12 +27,14 @@
 #include "llvm/CodeGen/MachineMemOperand.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/SelectionDAGNodes.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/GlobalValue.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/Support/BranchProbability.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
-#include "llvm/ADT/STLExtras.h"
 
 #define GET_INSTRINFO_CTOR
 #include "ARMGenInstrInfo.inc"
@@ -464,8 +464,9 @@ PredicateInstruction(MachineInstr *MI,
   unsigned Opc = MI->getOpcode();
   if (isUncondBranchOpcode(Opc)) {
     MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
-    MI->addOperand(MachineOperand::CreateImm(Pred[0].getImm()));
-    MI->addOperand(MachineOperand::CreateReg(Pred[1].getReg(), false));
+    MachineInstrBuilder(*MI->getParent()->getParent(), MI)
+      .addImm(Pred[0].getImm())
+      .addReg(Pred[1].getReg());
     return true;
   }
 
@@ -1124,7 +1125,7 @@ bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
   // copyPhysReg() calls.  Look for VMOVS instructions that can legally be
   // widened to VMOVD.  We prefer the VMOVD when possible because it may be
   // changed into a VORR that can go down the NEON pipeline.
-  if (!WidenVMOVS || !MI->isCopy())
+  if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15())
     return false;
 
   // Look for a copy between even S-registers.  That is where we keep floats
@@ -1154,6 +1155,7 @@ bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
 
   // All clear, widen the COPY.
   DEBUG(dbgs() << "widening:    " << *MI);
+  MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
 
   // Get rid of the old <imp-def> of DstRegD.  Leave it if it defines a Q-reg
   // or some other super-register.
@@ -1165,14 +1167,14 @@ bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
   MI->setDesc(get(ARM::VMOVD));
   MI->getOperand(0).setReg(DstRegD);
   MI->getOperand(1).setReg(SrcRegD);
-  AddDefaultPred(MachineInstrBuilder(MI));
+  AddDefaultPred(MIB);
 
   // We are now reading SrcRegD instead of SrcRegS.  This may upset the
   // register scavenger and machine verifier, so we need to indicate that we
   // are reading an undefined value from SrcRegD, but a proper value from
   // SrcRegS.
   MI->getOperand(1).setIsUndef();
-  MachineInstrBuilder(MI).addReg(SrcRegS, RegState::Implicit);
+  MIB.addReg(SrcRegS, RegState::Implicit);
 
   // SrcRegD may actually contain an unrelated value in the ssub_1
   // sub-register.  Don't kill it.  Only kill the ssub_0 sub-register.
@@ -1716,7 +1718,7 @@ MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
   // same register as operand 0.
   MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
   FalseReg.setImplicit();
-  NewMI->addOperand(FalseReg);
+  NewMI.addOperand(FalseReg);
   NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
 
   // The caller will erase MI, but not DefMI.
@@ -2717,7 +2719,6 @@ ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
   case ARM::t2STMDB_UPD: {
     unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
     if (Subtarget.isSwift()) {
-      // rdar://8402126
       int UOps = 1 + NumRegs;  // One for address computation, one for each ld / st.
       switch (Opc) {
       default: break;
@@ -3327,8 +3328,9 @@ ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
     // instructions).
     if (Latency > 0 && Subtarget.isThumb2()) {
       const MachineFunction *MF = DefMI->getParent()->getParent();
-      if (MF->getFunction()->getFnAttributes().
-            hasAttribute(Attributes::OptimizeForSize))
+      if (MF->getFunction()->getAttributes().
+            hasAttribute(AttributeSet::FunctionIndex,
+                         Attribute::OptimizeForSize))
         --Latency;
     }
     return Latency;
@@ -3732,9 +3734,9 @@ ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
   if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
     return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
 
-  // A9-like cores are particularly picky about mixing the two and want these
+  // CortexA9 is particularly picky about mixing the two and wants these
   // converted.
-  if (Subtarget.isLikeA9() && !isPredicated(MI) &&
+  if (Subtarget.isCortexA9() && !isPredicated(MI) &&
       (MI->getOpcode() == ARM::VMOVRS ||
        MI->getOpcode() == ARM::VMOVSR ||
        MI->getOpcode() == ARM::VMOVS))
@@ -3819,7 +3821,7 @@ void
 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
   unsigned DstReg, SrcReg, DReg;
   unsigned Lane;
-  MachineInstrBuilder MIB(MI);
+  MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
   const TargetRegisterInfo *TRI = &getRegisterInfo();
   switch (MI->getOpcode()) {
     default:
@@ -4021,14 +4023,12 @@ ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
 // VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
 //
 // FCONSTD can be used as a dependency-breaking instruction.
-
-
 unsigned ARMBaseInstrInfo::
 getPartialRegUpdateClearance(const MachineInstr *MI,
                              unsigned OpNum,
                              const TargetRegisterInfo *TRI) const {
-  // Only Swift has partial register update problems.
-  if (!SwiftPartialUpdateClearance || !Subtarget.isSwift())
+  if (!SwiftPartialUpdateClearance ||
+      !(Subtarget.isSwift() || Subtarget.isCortexA15()))
     return 0;
 
   assert(TRI && "Need TRI instance");
@@ -4044,7 +4044,6 @@ getPartialRegUpdateClearance(const MachineInstr *MI,
   case ARM::VLDRS:
   case ARM::FCONSTS:
   case ARM::VMOVSR:
-    // rdar://problem/8791586
   case ARM::VMOVv8i8:
   case ARM::VMOVv4i16:
   case ARM::VMOVv2i32:
@@ -4055,7 +4054,7 @@ getPartialRegUpdateClearance(const MachineInstr *MI,
 
     // Explicitly reads the dependency.
   case ARM::VLD1LNd32:
-    UseOp = 1;
+    UseOp = 3;
     break;
   default:
     return 0;
@@ -4124,3 +4123,15 @@ breakPartialRegDependency(MachineBasicBlock::iterator MI,
 bool ARMBaseInstrInfo::hasNOP() const {
   return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
 }
+
+bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {
+  unsigned ShOpVal = MI->getOperand(3).getImm();
+  unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal);
+  // Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1.
+  if ((ShImm == 1 && ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsr) ||
+      ((ShImm == 1 || ShImm == 2) &&
+       ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsl))
+    return true;
+
+  return false;
+}