Use multiple virtual registers in PPC CR spilling

Now that the register scavenger can support multiple spill slots, and PEI can
use virtual-register-based scavenging for multiple simultaneous registers, we
can use a virtual register for the transfer register in the CR spilling code.

This should eliminate the last place (outside of the prologue/epilogue) where
we depend on the unconditional availability of the r0 register. We will soon be
able to allocate it (in a somewhat restricted sense) as a GPR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178060 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 58 insertions, 42 deletions

diff --git a/lib/Target/PowerPC/PPCFrameLowering.cpp b/lib/Target/PowerPC/PPCFrameLowering.cpp
index f6c00f0d32..6a3951cf61 100644
--- a/lib/Target/PowerPC/PPCFrameLowering.cpp
+++ b/lib/Target/PowerPC/PPCFrameLowering.cpp
@@ -1094,6 +1094,13 @@ PPCFrameLowering::addScavengingSpillSlot(MachineFunction &MF,
     RS->addScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
                                                        RC->getAlignment(),
                                                        false));
+
+    // These kinds of spills might need two registers.
+    if (spillsCR(MF) || spillsVRSAVE(MF))
+      RS->addScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
+                                                         RC->getAlignment(),
+                                                         false));
+
   }
 }
 
diff --git a/lib/Target/PowerPC/PPCRegisterInfo.cpp b/lib/Target/PowerPC/PPCRegisterInfo.cpp
index 3d1b83767f..1fa0f67137 100644
--- a/lib/Target/PowerPC/PPCRegisterInfo.cpp
+++ b/lib/Target/PowerPC/PPCRegisterInfo.cpp
@@ -304,14 +304,14 @@ void PPCRegisterInfo::lowerCRSpilling(MachineBasicBlock::iterator II,
   MachineInstr &MI = *II;       // ; SPILL_CR <SrcReg>, <offset>
   // Get the instruction's basic block.
   MachineBasicBlock &MBB = *MI.getParent();
+  MachineFunction &MF = *MBB.getParent();
   DebugLoc dl = MI.getDebugLoc();
 
-  // FIXME: Once LLVM supports creating virtual registers here, or the register
-  // scavenger can return multiple registers, stop using reserved registers
-  // here.
-
   bool LP64 = Subtarget.isPPC64();
-  unsigned Reg = LP64 ? PPC::X0 : PPC::R0;
+  const TargetRegisterClass *G8RC = &PPC::G8RCRegClass;
+  const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
+
+  unsigned Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC);
   unsigned SrcReg = MI.getOperand(0).getReg();
 
   // We need to store the CR in the low 4-bits of the saved value. First, issue
@@ -321,13 +321,17 @@ void PPCRegisterInfo::lowerCRSpilling(MachineBasicBlock::iterator II,
     
   // If the saved register wasn't CR0, shift the bits left so that they are in
   // CR0's slot.
-  if (SrcReg != PPC::CR0)
+  if (SrcReg != PPC::CR0) {
+    unsigned Reg1 = Reg;
+    Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC);
+
     // rlwinm rA, rA, ShiftBits, 0, 31.
     BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg)
-      .addReg(Reg, RegState::Kill)
+      .addReg(Reg1, RegState::Kill)
       .addImm(getPPCRegisterNumbering(SrcReg) * 4)
       .addImm(0)
       .addImm(31);
+  }
 
   addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::STW8 : PPC::STW))
                     .addReg(Reg, getKillRegState(MI.getOperand(1).getImm())),
@@ -343,14 +347,14 @@ void PPCRegisterInfo::lowerCRRestore(MachineBasicBlock::iterator II,
   MachineInstr &MI = *II;       // ; <DestReg> = RESTORE_CR <offset>
   // Get the instruction's basic block.
   MachineBasicBlock &MBB = *MI.getParent();
+  MachineFunction &MF = *MBB.getParent();
   DebugLoc dl = MI.getDebugLoc();
 
-  // FIXME: Once LLVM supports creating virtual registers here, or the register
-  // scavenger can return multiple registers, stop using reserved registers
-  // here.
-
   bool LP64 = Subtarget.isPPC64();
-  unsigned Reg = LP64 ? PPC::X0 : PPC::R0;
+  const TargetRegisterClass *G8RC = &PPC::G8RCRegClass;
+  const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
+
+  unsigned Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC);
   unsigned DestReg = MI.getOperand(0).getReg();
   assert(MI.definesRegister(DestReg) &&
     "RESTORE_CR does not define its destination");
@@ -361,15 +365,18 @@ void PPCRegisterInfo::lowerCRRestore(MachineBasicBlock::iterator II,
   // If the reloaded register isn't CR0, shift the bits right so that they are
   // in the right CR's slot.
   if (DestReg != PPC::CR0) {
+    unsigned Reg1 = Reg;
+    Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC);
+
     unsigned ShiftBits = getPPCRegisterNumbering(DestReg)*4;
     // rlwinm r11, r11, 32-ShiftBits, 0, 31.
     BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg)
-             .addReg(Reg).addImm(32-ShiftBits).addImm(0)
+             .addReg(Reg1, RegState::Kill).addImm(32-ShiftBits).addImm(0)
              .addImm(31);
   }
 
   BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MTCRF8 : PPC::MTCRF), DestReg)
-             .addReg(Reg);
+             .addReg(Reg, RegState::Kill);
 
   // Discard the pseudo instruction.
   MBB.erase(II);
@@ -381,13 +388,11 @@ void PPCRegisterInfo::lowerVRSAVESpilling(MachineBasicBlock::iterator II,
   MachineInstr &MI = *II;       // ; SPILL_VRSAVE <SrcReg>, <offset>
   // Get the instruction's basic block.
   MachineBasicBlock &MBB = *MI.getParent();
+  MachineFunction &MF = *MBB.getParent();
   DebugLoc dl = MI.getDebugLoc();
 
-  // FIXME: Once LLVM supports creating virtual registers here, or the register
-  // scavenger can return multiple registers, stop using reserved registers
-  // here.
-
-  unsigned Reg = PPC::R0;
+  const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
+  unsigned Reg = MF.getRegInfo().createVirtualRegister(GPRC);
   unsigned SrcReg = MI.getOperand(0).getReg();
 
   BuildMI(MBB, II, dl, TII.get(PPC::MFVRSAVEv), Reg)
@@ -407,13 +412,11 @@ void PPCRegisterInfo::lowerVRSAVERestore(MachineBasicBlock::iterator II,
   MachineInstr &MI = *II;       // ; <DestReg> = RESTORE_VRSAVE <offset>
   // Get the instruction's basic block.
   MachineBasicBlock &MBB = *MI.getParent();
+  MachineFunction &MF = *MBB.getParent();
   DebugLoc dl = MI.getDebugLoc();
 
-  // FIXME: Once LLVM supports creating virtual registers here, or the register
-  // scavenger can return multiple registers, stop using reserved registers
-  // here.
-
-  unsigned Reg = PPC::R0;
+  const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
+  unsigned Reg = MF.getRegInfo().createVirtualRegister(GPRC);
   unsigned DestReg = MI.getOperand(0).getReg();
   assert(MI.definesRegister(DestReg) &&
     "RESTORE_VRSAVE does not define its destination");
@@ -422,7 +425,7 @@ void PPCRegisterInfo::lowerVRSAVERestore(MachineBasicBlock::iterator II,
                               Reg), FrameIndex);
 
   BuildMI(MBB, II, dl, TII.get(PPC::MTVRSAVEv), DestReg)
-             .addReg(Reg);
+             .addReg(Reg, RegState::Kill);
 
   // Discard the pseudo instruction.
   MBB.erase(II);
diff --git a/test/CodeGen/PowerPC/2010-02-12-saveCR.ll b/test/CodeGen/PowerPC/2010-02-12-saveCR.ll
index 0da6e4351b..79e8f96b39 100644
--- a/test/CodeGen/PowerPC/2010-02-12-saveCR.ll
+++ b/test/CodeGen/PowerPC/2010-02-12-saveCR.ll
@@ -2,21 +2,22 @@
 ; ModuleID = 'hh.c'
 target datalayout = "E-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f128:64:128-n32"
 target triple = "powerpc-apple-darwin9.6"
-; This formerly used R0 for both the stack address and CR.
 
 define void @foo() nounwind {
 entry:
-;CHECK:  mfcr r0
-;CHECK:  lis r2, 1
-;CHECK:  rlwinm r0, r0, 8, 0, 31
-;CHECK:  ori r2, r2, 34540
-;CHECK:  stwx r0, r1, r2
-; Make sure that the register scavenger returns the same temporary register.
-;CHECK:  lis r2, 1
-;CHECK:  mfcr r0
-;CHECK:  ori r2, r2, 34536
-;CHECK:  rlwinm r0, r0, 12, 0, 31
-;CHECK:  stwx r0, r1, r2
+; Note that part of what is being checked here is proper register reuse.
+; CHECK: mfcr [[T1:r[0-9]+]]                         ; cr2
+; CHECK: lis [[T2:r[0-9]+]], 1
+; FIXME: There should only be one lis needed here!
+; CHECK: lis [[T3:r[0-9]+]], 1
+; CHECK: addi r3, r1, 72
+; CHECK: rlwinm [[T1]], [[T1]], 8, 0, 31
+; CHECK: ori [[T2]], [[T2]], 34540
+; CHECK: ori [[T3]], [[T3]], 34536
+; CHECK: stwx [[T1]], r1, [[T2]]
+; CHECK: mfcr [[T4:r[0-9]+]]                         ; cr3
+; CHECK: rlwinm [[T4]], [[T4]], 12, 0, 31
+; CHECK: stwx r4, r1, [[T3]]
   %x = alloca [100000 x i8]                       ; <[100000 x i8]*> [#uses=1]
   %"alloca point" = bitcast i32 0 to i32          ; <i32> [#uses=0]
   %x1 = bitcast [100000 x i8]* %x to i8*          ; <i8*> [#uses=1]
@@ -25,11 +26,16 @@ entry:
   br label %return
 
 return:                                           ; preds = %entry
-;CHECK:  lis r2, 1
-;CHECK:  ori r2, r2, 34540
-;CHECK:  lwzx r0, r1, r2
-;CHECK:  rlwinm r0, r0, 24, 0, 31
-;CHECK:  mtcrf 32, r0
+; CHECK: lis [[T1:r[0-9]+]], 1
+; CHECK: ori [[T1]], [[T1]], 34536
+; CHECK: lwzx [[T1]], r1, [[T1]]
+; CHECK: rlwinm [[T1]], [[T1]], 20, 0, 31
+; CHECK: mtcrf 16, [[T1]]
+; CHECK: lis [[T1]], 1
+; CHECK: ori [[T1]], [[T1]], 34540
+; CHECK: lwzx [[T1]], r1, [[T1]]
+; CHECK: rlwinm [[T1]], [[T1]], 24, 0, 31
+; CHECK: mtcrf 32, [[T1]]
   ret void
 }