For functions that use vector registers, save VRSAVE, mark used
registers, and update it on entry to each function, then restore it on exit.

This compiles:

void func(vfloat *a, vfloat *b, vfloat *c) {
        *a = *b * *c + *c;
}

to this:

_func:
        mfspr r2, 256
        oris r6, r2, 49152
        mtspr 256, r6
        lvx v0, 0, r5
        lvx v1, 0, r4
        vmaddfp v0, v1, v0, v0
        stvx v0, 0, r3
        mtspr 256, r2
        blr

GCC produces this (which has additional stack accesses):

_func:
        mfspr r0,256
        stw r0,-4(r1)
        oris r0,r0,0xc000
        mtspr 256,r0
        lvx v0,0,r5
        lvx v1,0,r4
        lwz r12,-4(r1)
        vmaddfp v0,v0,v1,v0
        stvx v0,0,r3
        mtspr 256,r12
        blr


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26733 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
index 09bbfc6..0288143 100644
--- a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
+++ b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
@@ -196,8 +196,65 @@
   CodeGenMap.clear();
   DAG.RemoveDeadNodes();
   
-  // Emit machine code to BB. 
+  // Emit machine code to BB.
   ScheduleAndEmitDAG(DAG);
+  
+  // Check to see if this function uses vector registers, which means we have to
+  // save and restore the VRSAVE register and update it with the regs we use.  
+  //
+  // In this case, there will be virtual registers of vector type type created
+  // by the scheduler.  Detect them now.
+  SSARegMap *RegMap = DAG.getMachineFunction().getSSARegMap();
+  bool HasVectorVReg = false;
+  for (unsigned i = MRegisterInfo::FirstVirtualRegister, 
+       e = RegMap->getLastVirtReg(); i != e; ++i)
+    if (RegMap->getRegClass(i) == &PPC::VRRCRegClass) {
+      HasVectorVReg = true;
+      break;
+    }
+  
+  // If we have a vector register, we want to emit code into the entry and exit
+  // blocks to save and restore the VRSAVE register.  We do this here (instead
+  // of marking all vector instructions as clobbering VRSAVE) for two reasons:
+  //
+  // 1. This (trivially) reduces the load on the register allocator, by not
+  //    having to represent the live range of the VRSAVE register.
+  // 2. This (more significantly) allows us to create a temporary virtual
+  //    register to hold the saved VRSAVE value, allowing this temporary to be
+  //    register allocated, instead of forcing it to be spilled to the stack.
+  if (HasVectorVReg) {
+    // Create two vregs - one to hold the VRSAVE register that is live-in to the
+    // function and one for the value after having bits or'd into it.
+    unsigned InVRSAVE = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
+    unsigned UpdatedVRSAVE = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
+    
+    MachineFunction &MF = DAG.getMachineFunction();
+    MachineBasicBlock &EntryBB = *MF.begin();
+    // Emit the following code into the entry block:
+    // InVRSAVE = MFVRSAVE
+    // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
+    // MTVRSAVE UpdatedVRSAVE
+    MachineBasicBlock::iterator IP = EntryBB.begin();  // Insert Point
+    BuildMI(EntryBB, IP, PPC::MFVRSAVE, 0, InVRSAVE);
+    BuildMI(EntryBB, IP, PPC::UPDATE_VRSAVE, 1, UpdatedVRSAVE).addReg(InVRSAVE);
+    BuildMI(EntryBB, IP, PPC::MTVRSAVE, 1).addReg(UpdatedVRSAVE);
+    
+    // Find all return blocks, outputting a restore in each epilog.
+    const TargetInstrInfo &TII = *DAG.getTarget().getInstrInfo();
+    for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB)
+      if (!BB->empty() && TII.isReturn(BB->back().getOpcode())) {
+        IP = BB->end(); --IP;
+        
+        // Skip over all terminator instructions, which are part of the return
+        // sequence.
+        MachineBasicBlock::iterator I2 = IP;
+        while (I2 != BB->begin() && TII.isTerminatorInstr((--I2)->getOpcode()))
+          IP = I2;
+        
+        // Emit: MTVRSAVE InVRSave
+        BuildMI(*BB, IP, PPC::MTVRSAVE, 1).addReg(InVRSAVE);
+      }        
+  }
 }
 
 /// getGlobalBaseReg - Output the instructions required to put the
diff --git a/lib/Target/PowerPC/PPCInstrInfo.td b/lib/Target/PowerPC/PPCInstrInfo.td
index b41a77a..1aa43ba 100644
--- a/lib/Target/PowerPC/PPCInstrInfo.td
+++ b/lib/Target/PowerPC/PPCInstrInfo.td
@@ -210,6 +210,9 @@
 def ADJCALLSTACKUP   : Pseudo<(ops u16imm:$amt),
                               "; ADJCALLSTACKUP",
                               [(callseq_end imm:$amt)]>;
+
+def UPDATE_VRSAVE    : Pseudo<(ops GPRC:$rD, GPRC:$rS),
+                              "UPDATE_VRSAVE $rD, $rS", []>;
 }
 def IMPLICIT_DEF_GPR : Pseudo<(ops GPRC:$rD), "; $rD = IMPLICIT_DEF_GPRC",
                               [(set GPRC:$rD, (undef))]>;
@@ -694,8 +697,24 @@
 //
 def MFCTR : XFXForm_1_ext<31, 339, 9, (ops GPRC:$rT), "mfctr $rT", SprMFSPR>,
             PPC970_DGroup_First, PPC970_Unit_FXU;
+def MTCTR : XFXForm_7_ext<31, 467, 9, (ops GPRC:$rS), "mtctr $rS", SprMTSPR>,
+            PPC970_DGroup_First, PPC970_Unit_FXU;
+
+def MTLR  : XFXForm_7_ext<31, 467, 8, (ops GPRC:$rS), "mtlr $rS", SprMTSPR>,
+            PPC970_DGroup_First, PPC970_Unit_FXU;
 def MFLR  : XFXForm_1_ext<31, 339, 8, (ops GPRC:$rT), "mflr $rT",  SprMFSPR>,
             PPC970_DGroup_First, PPC970_Unit_FXU;
+
+// Move to/from VRSAVE: despite being a SPR, the VRSAVE register is renamed like
+// a GPR on the PPC970.  As such, copies in and out have the same performance
+// characteristics as an OR instruction.
+def MTVRSAVE : XFXForm_7_ext<31, 467, 256, (ops GPRC:$rS),
+                             "mtspr 256, $rS", IntGeneral>,
+               PPC970_Unit_FXU;
+def MFVRSAVE : XFXForm_1_ext<31, 339, 256, (ops GPRC:$rT),
+                             "mfspr $rT, 256", IntGeneral>,
+               PPC970_Unit_FXU;
+
 def MFCR  : XFXForm_3<31, 19, (ops GPRC:$rT), "mfcr $rT", SprMFCR>,
             PPC970_MicroCode, PPC970_Unit_CRU;
 def MTCRF : XFXForm_5<31, 144, (ops crbitm:$FXM, GPRC:$rS),
@@ -704,13 +723,6 @@
 def MFOCRF: XFXForm_5a<31, 19, (ops GPRC:$rT, crbitm:$FXM),
                        "mfcr $rT, $FXM", SprMFCR>,
             PPC970_DGroup_First, PPC970_Unit_CRU;
-def MTCTR : XFXForm_7_ext<31, 467, 9, (ops GPRC:$rS), "mtctr $rS", SprMTSPR>,
-            PPC970_DGroup_First, PPC970_Unit_FXU;
-def MTLR  : XFXForm_7_ext<31, 467, 8, (ops GPRC:$rS), "mtlr $rS", SprMTSPR>,
-            PPC970_DGroup_First, PPC970_Unit_FXU;
-def MTSPR : XFXForm_7<31, 467, (ops GPRC:$rS, u16imm:$UIMM), "mtspr $UIMM, $rS",
-                      SprMTSPR>,
-            PPC970_DGroup_Single, PPC970_Unit_FXU;
 
 // XS-Form instructions.  Just 'sradi'
 //
diff --git a/lib/Target/PowerPC/PPCRegisterInfo.cpp b/lib/Target/PowerPC/PPCRegisterInfo.cpp
index 8d2037c..d55bdc7 100644
--- a/lib/Target/PowerPC/PPCRegisterInfo.cpp
+++ b/lib/Target/PowerPC/PPCRegisterInfo.cpp
@@ -266,12 +266,63 @@
   }
 }
 
+// HandleVRSaveUpdate - MI is the UPDATE_VRSAVE instruction introduced by the
+// instruction selector.  Based on the vector registers that have been used,
+// transform this into the appropriate ORI instruction.
+static void HandleVRSaveUpdate(MachineInstr *MI, const bool *UsedRegs) {
+  unsigned UsedRegMask = 0;
+#define HANDLEREG(N) if (UsedRegs[PPC::V##N]) UsedRegMask |= 1 << (31-N)
+  HANDLEREG( 0); HANDLEREG( 1); HANDLEREG( 2); HANDLEREG( 3);
+  HANDLEREG( 4); HANDLEREG( 5); HANDLEREG( 6); HANDLEREG( 7);
+  HANDLEREG( 8); HANDLEREG( 9); HANDLEREG(10); HANDLEREG(11);
+  HANDLEREG(12); HANDLEREG(13); HANDLEREG(14); HANDLEREG(15);
+  HANDLEREG(16); HANDLEREG(17); HANDLEREG(18); HANDLEREG(19);
+  HANDLEREG(20); HANDLEREG(21); HANDLEREG(22); HANDLEREG(23);
+  HANDLEREG(24); HANDLEREG(25); HANDLEREG(26); HANDLEREG(27);
+  HANDLEREG(28); HANDLEREG(29); HANDLEREG(30); HANDLEREG(31);
+#undef HANDLEREG
+  unsigned SrcReg = MI->getOperand(1).getReg();
+  unsigned DstReg = MI->getOperand(0).getReg();
+  // If no registers are used, turn this into a copy.
+  if (UsedRegMask == 0) {
+    if (SrcReg != DstReg)
+      BuildMI(*MI->getParent(), MI, PPC::OR4, 2, DstReg)
+        .addReg(SrcReg).addReg(SrcReg);
+  } else if ((UsedRegMask & 0xFFFF) == UsedRegMask) {
+    BuildMI(*MI->getParent(), MI, PPC::ORI, 2, DstReg)
+        .addReg(SrcReg).addImm(UsedRegMask);
+  } else if ((UsedRegMask & 0xFFFF0000) == UsedRegMask) {
+    BuildMI(*MI->getParent(), MI, PPC::ORIS, 2, DstReg)
+        .addReg(SrcReg).addImm(UsedRegMask >> 16);
+  } else {
+    BuildMI(*MI->getParent(), MI, PPC::ORIS, 2, DstReg)
+       .addReg(SrcReg).addImm(UsedRegMask >> 16);
+    BuildMI(*MI->getParent(), MI, PPC::ORI, 2, DstReg)
+      .addReg(DstReg).addImm(UsedRegMask & 0xFFFF);
+  }
+  
+  // Remove the old UPDATE_VRSAVE instruction.
+  MI->getParent()->erase(MI);
+}
+
 
 void PPCRegisterInfo::emitPrologue(MachineFunction &MF) const {
   MachineBasicBlock &MBB = MF.front();   // Prolog goes in entry BB
   MachineBasicBlock::iterator MBBI = MBB.begin();
   MachineFrameInfo *MFI = MF.getFrameInfo();
 
+  // Scan the first few instructions of the prolog, looking for an UPDATE_VRSAVE
+  // instruction.  If we find it, process it.
+  for (unsigned i = 0; MBBI != MBB.end() && i < 5; ++i, ++MBBI) {
+    if (MBBI->getOpcode() == PPC::UPDATE_VRSAVE) {
+      HandleVRSaveUpdate(MBBI, MF.getUsedPhysregs());
+      break;
+    }
+  }
+  
+  // Move MBBI back to the beginning of the function.
+  MBBI = MBB.begin();
+  
   // Get the number of bytes to allocate from the FrameInfo
   unsigned NumBytes = MFI->getStackSize();
   
diff --git a/lib/Target/PowerPC/PPCRegisterInfo.td b/lib/Target/PowerPC/PPCRegisterInfo.td
index f349500..68c73bc 100644
--- a/lib/Target/PowerPC/PPCRegisterInfo.td
+++ b/lib/Target/PowerPC/PPCRegisterInfo.td
@@ -152,9 +152,9 @@
     GPRCClass::iterator
     GPRCClass::allocation_order_end(MachineFunction &MF) const {
       if (hasFP(MF))
-        return end()-4;
+        return end()-4;  // don't allocate R31, R0, R1, LR
       else
-        return end()-3;
+        return end()-3;  // don't allocate R0, R1, LR
     }
   }];
 }