Enable generating PPC pre-increment (r+imm) instructions by default.

[oota-llvm.git] / lib / Target / PowerPC / PPCInstrInfo.cpp

diff --git a/lib/Target/PowerPC/PPCInstrInfo.cpp b/lib/Target/PowerPC/PPCInstrInfo.cpp
index fe103c657718dbb56f1ac06fa786f4ce081133b3..c0b9e92c2d0a83c3dd8f9bd589a26a0106f2b545 100644 (file)
--- a/lib/Target/PowerPC/PPCInstrInfo.cpp
+++ b/lib/Target/PowerPC/PPCInstrInfo.cpp
@@ -1,4 +1,4 @@
-//===- PPCInstrInfo.cpp - PowerPC32 Instruction Information -----*- C++ -*-===//
+//===-- PPCInstrInfo.cpp - PowerPC Instruction Information ----------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
@@ -22,6 +22,7 @@
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineMemOperand.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/PseudoSourceValue.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/ErrorHandling.h"
@@ -49,9 +50,9 @@ ScheduleHazardRecognizer *PPCInstrInfo::CreateTargetHazardRecognizer(
   const TargetMachine *TM,
   const ScheduleDAG *DAG) const {
   unsigned Directive = TM->getSubtarget<PPCSubtarget>().getDarwinDirective();
-  if (Directive == PPC::DIR_440) {
+  if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2) {
     const InstrItineraryData *II = TM->getInstrItineraryData();
-    return new PPCHazardRecognizer440(II, DAG);
+    return new PPCScoreboardHazardRecognizer(II, DAG);
   }
 
   return TargetInstrInfoImpl::CreateTargetHazardRecognizer(TM, DAG);
@@ -65,14 +66,14 @@ ScheduleHazardRecognizer *PPCInstrInfo::CreateTargetPostRAHazardRecognizer(
   unsigned Directive = TM.getSubtarget<PPCSubtarget>().getDarwinDirective();
 
   // Most subtargets use a PPC970 recognizer.
-  if (Directive != PPC::DIR_440) {
+  if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2) {
     const TargetInstrInfo *TII = TM.getInstrInfo();
     assert(TII && "No InstrInfo?");
 
     return new PPCHazardRecognizer970(*TII);
   }
 
-  return TargetInstrInfoImpl::CreateTargetPostRAHazardRecognizer(II, DAG);
+  return new PPCScoreboardHazardRecognizer(II, DAG);
 }
 unsigned PPCInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
                                            int &FrameIndex) const {
@@ -353,7 +354,7 @@ PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF,
                                   const TargetRegisterClass *RC,
                                   SmallVectorImpl<MachineInstr*> &NewMIs) const{
   DebugLoc DL;
-  if (PPC::GPRCRegisterClass->hasSubClassEq(RC)) {
+  if (PPC::GPRCRegClass.hasSubClassEq(RC)) {
     if (SrcReg != PPC::LR) {
       NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STW))
                                          .addReg(SrcReg,
@@ -369,7 +370,7 @@ PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF,
                                                  getKillRegState(isKill)),
                                          FrameIdx));
     }
-  } else if (PPC::G8RCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::G8RCRegClass.hasSubClassEq(RC)) {
     if (SrcReg != PPC::LR8) {
       NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STD))
                                          .addReg(SrcReg,
@@ -377,7 +378,7 @@ PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF,
                                          FrameIdx));
     } else {
       // FIXME: this spills LR immediately to memory in one step.  To do this,
-      // we use R11, which we know cannot be used in the prolog/epilog.  This is
+      // we use X11, which we know cannot be used in the prolog/epilog.  This is
       // a hack.
       NewMIs.push_back(BuildMI(MF, DL, get(PPC::MFLR8), PPC::X11));
       NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STD))
@@ -385,17 +386,17 @@ PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF,
                                                  getKillRegState(isKill)),
                                          FrameIdx));
     }
-  } else if (PPC::F8RCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
     NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFD))
                                        .addReg(SrcReg,
                                                getKillRegState(isKill)),
                                        FrameIdx));
-  } else if (PPC::F4RCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) {
     NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFS))
                                        .addReg(SrcReg,
                                                getKillRegState(isKill)),
                                        FrameIdx));
-  } else if (PPC::CRRCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) {
     if ((!DisablePPC32RS && !TM.getSubtargetImpl()->isPPC64()) ||
         (!DisablePPC64RS && TM.getSubtargetImpl()->isPPC64())) {
       NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CR))
@@ -410,11 +411,14 @@ PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF,
       // We hack this on Darwin by reserving R2.  It's probably broken on Linux
       // at the moment.
 
+      bool is64Bit = TM.getSubtargetImpl()->isPPC64();
       // We need to store the CR in the low 4-bits of the saved value.  First,
       // issue a MFCR to save all of the CRBits.
       unsigned ScratchReg = TM.getSubtargetImpl()->isDarwinABI() ?
-                                                           PPC::R2 : PPC::R0;
-      NewMIs.push_back(BuildMI(MF, DL, get(PPC::MFCRpseud), ScratchReg)
+                              (is64Bit ? PPC::X2 : PPC::R2) :
+                              (is64Bit ? PPC::X0 : PPC::R0);
+      NewMIs.push_back(BuildMI(MF, DL, get(is64Bit ? PPC::MFCR8pseud :
+                                             PPC::MFCRpseud), ScratchReg)
                                .addReg(SrcReg, getKillRegState(isKill)));
 
       // If the saved register wasn't CR0, shift the bits left so that they are
@@ -422,17 +426,19 @@ PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF,
       if (SrcReg != PPC::CR0) {
         unsigned ShiftBits = getPPCRegisterNumbering(SrcReg)*4;
         // rlwinm scratch, scratch, ShiftBits, 0, 31.
-        NewMIs.push_back(BuildMI(MF, DL, get(PPC::RLWINM), ScratchReg)
+        NewMIs.push_back(BuildMI(MF, DL, get(is64Bit ? PPC::RLWINM8 :
+                           PPC::RLWINM), ScratchReg)
                        .addReg(ScratchReg).addImm(ShiftBits)
                        .addImm(0).addImm(31));
       }
 
-      NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STW))
+      NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(is64Bit ?
+                                           PPC::STW8 : PPC::STW))
                                          .addReg(ScratchReg,
                                                  getKillRegState(isKill)),
                                          FrameIdx));
     }
-  } else if (PPC::CRBITRCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
     // FIXME: We use CRi here because there is no mtcrf on a bit. Since the
     // backend currently only uses CR1EQ as an individual bit, this should
     // not cause any bug. If we need other uses of CR bits, the following
@@ -464,9 +470,9 @@ PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF,
       Reg = PPC::CR7;
 
     return StoreRegToStackSlot(MF, Reg, isKill, FrameIdx,
-                               PPC::CRRCRegisterClass, NewMIs);
+                               &PPC::CRRCRegClass, NewMIs);
 
-  } else if (PPC::VRRCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
     // We don't have indexed addressing for vector loads.  Emit:
     // R0 = ADDI FI#
     // STVX VAL, 0, R0
@@ -516,7 +522,7 @@ PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL,
                                    unsigned DestReg, int FrameIdx,
                                    const TargetRegisterClass *RC,
                                    SmallVectorImpl<MachineInstr*> &NewMIs)const{
-  if (PPC::GPRCRegisterClass->hasSubClassEq(RC)) {
+  if (PPC::GPRCRegClass.hasSubClassEq(RC)) {
     if (DestReg != PPC::LR) {
       NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ),
                                                  DestReg), FrameIdx));
@@ -525,22 +531,22 @@ PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL,
                                                  PPC::R11), FrameIdx));
       NewMIs.push_back(BuildMI(MF, DL, get(PPC::MTLR)).addReg(PPC::R11));
     }
-  } else if (PPC::G8RCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::G8RCRegClass.hasSubClassEq(RC)) {
     if (DestReg != PPC::LR8) {
       NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LD), DestReg),
                                          FrameIdx));
     } else {
       NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LD),
-                                                 PPC::R11), FrameIdx));
-      NewMIs.push_back(BuildMI(MF, DL, get(PPC::MTLR8)).addReg(PPC::R11));
+                                                 PPC::X11), FrameIdx));
+      NewMIs.push_back(BuildMI(MF, DL, get(PPC::MTLR8)).addReg(PPC::X11));
     }
-  } else if (PPC::F8RCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) {
     NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFD), DestReg),
                                        FrameIdx));
-  } else if (PPC::F4RCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) {
     NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFS), DestReg),
                                        FrameIdx));
-  } else if (PPC::CRRCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) {
     if ((!DisablePPC32RS && !TM.getSubtargetImpl()->isPPC64()) ||
         (!DisablePPC64RS && TM.getSubtargetImpl()->isPPC64())) {
       NewMIs.push_back(addFrameReference(BuildMI(MF, DL,
@@ -568,10 +574,11 @@ PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL,
                       .addImm(31));
       }
   
-      NewMIs.push_back(BuildMI(MF, DL, get(PPC::MTCRF), DestReg)
+      NewMIs.push_back(BuildMI(MF, DL, get(TM.getSubtargetImpl()->isPPC64() ?
+                         PPC::MTCRF8 : PPC::MTCRF), DestReg)
                        .addReg(ScratchReg));
     }
-  } else if (PPC::CRBITRCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) {
 
     unsigned Reg = 0;
     if (DestReg == PPC::CR0LT || DestReg == PPC::CR0GT ||
@@ -600,9 +607,9 @@ PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL,
       Reg = PPC::CR7;
 
     return LoadRegFromStackSlot(MF, DL, Reg, FrameIdx,
-                                PPC::CRRCRegisterClass, NewMIs);
+                                &PPC::CRRCRegClass, NewMIs);
 
-  } else if (PPC::VRRCRegisterClass->hasSubClassEq(RC)) {
+  } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) {
     // We don't have indexed addressing for vector loads.  Emit:
     // R0 = ADDI FI#
     // Dest = LVX 0, R0
@@ -678,6 +685,9 @@ unsigned PPCInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
   case PPC::GC_LABEL:
   case PPC::DBG_VALUE:
     return 0;
+  case PPC::BL8_NOP_ELF:
+  case PPC::BLA8_NOP_ELF:
+    return 8;
   default:
     return 4; // PowerPC instructions are all 4 bytes
   }