Set mayLoad or mayStore flags for SC and LL in order to prevent LICM from

moving them out of the loop. Previously, stores and loads to a stack frame
object were inserted to accomplish this. Remove the code that was needed to do
this. Patch by Sasa Stankovic.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@135415 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 41 insertions, 109 deletions

diff --git a/lib/Target/Mips/MipsISelLowering.cpp b/lib/Target/Mips/MipsISelLowering.cpp
index 873c99a1f4..dfde8c9718 100644
--- a/lib/Target/Mips/MipsISelLowering.cpp
+++ b/lib/Target/Mips/MipsISelLowering.cpp
@@ -733,11 +733,10 @@ MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
   const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
   DebugLoc dl = MI->getDebugLoc();
 
-  unsigned Dest = MI->getOperand(0).getReg();
+  unsigned Oldval = MI->getOperand(0).getReg();
   unsigned Ptr = MI->getOperand(1).getReg();
   unsigned Incr = MI->getOperand(2).getReg();
 
-  unsigned Oldval = RegInfo.createVirtualRegister(RC);
   unsigned Tmp1 = RegInfo.createVirtualRegister(RC);
   unsigned Tmp2 = RegInfo.createVirtualRegister(RC);
   unsigned Tmp3 = RegInfo.createVirtualRegister(RC);
@@ -759,38 +758,16 @@ MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
 
   //  thisMBB:
   //    ...
-  //    sw incr, fi(sp)           // store incr to stack (when BinOpcode == 0)
   //    fallthrough --> loopMBB
-
-  // Note: for atomic.swap (when BinOpcode == 0), storing incr to stack before
-  // the loop and then loading it from stack in block loopMBB is necessary to
-  // prevent MachineLICM pass to hoist "or" instruction out of the block
-  // loopMBB.
-
-  int fi = 0;
-  if (BinOpcode == 0 && !Nand) {
-    // Get or create a temporary stack location.
-    MipsFunctionInfo *MipsFI = MF->getInfo<MipsFunctionInfo>();
-    fi = MipsFI->getAtomicFrameIndex();
-    if (fi == -1) {
-      fi = MF->getFrameInfo()->CreateStackObject(Size, Size, false);
-      MipsFI->setAtomicFrameIndex(fi);
-    }
-
-    BuildMI(BB, dl, TII->get(Mips::SW))
-        .addReg(Incr).addFrameIndex(fi).addImm(0);
-  }
   BB->addSuccessor(loopMBB);
 
   //  loopMBB:
   //    ll oldval, 0(ptr)
-  //    or dest, $0, oldval
   //    <binop> tmp1, oldval, incr
   //    sc tmp1, 0(ptr)
   //    beq tmp1, $0, loopMBB
   BB = loopMBB;
   BuildMI(BB, dl, TII->get(Mips::LL), Oldval).addReg(Ptr).addImm(0);
-  BuildMI(BB, dl, TII->get(Mips::OR), Dest).addReg(Mips::ZERO).addReg(Oldval);
   if (Nand) {
     //  and tmp2, oldval, incr
     //  nor tmp1, $0, tmp2
@@ -800,10 +777,7 @@ MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
     //  <binop> tmp1, oldval, incr
     BuildMI(BB, dl, TII->get(BinOpcode), Tmp1).addReg(Oldval).addReg(Incr);
   } else {
-    //  lw tmp2, fi(sp)              // load incr from stack
-    //  or tmp1, $zero, tmp2
-    BuildMI(BB, dl, TII->get(Mips::LW), Tmp2).addFrameIndex(fi).addImm(0);
-    BuildMI(BB, dl, TII->get(Mips::OR), Tmp1).addReg(Mips::ZERO).addReg(Tmp2);
+    Tmp1 = Incr;
   }
   BuildMI(BB, dl, TII->get(Mips::SC), Tmp3).addReg(Tmp1).addReg(Ptr).addImm(0);
   BuildMI(BB, dl, TII->get(Mips::BEQ))
@@ -880,12 +854,6 @@ MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
   //    nor     mask2,$0,mask
   //    andi    tmp4,incr,255
   //    sll     incr2,tmp4,shift
-  //    sw      incr2, fi(sp)      // store incr2 to stack (when BinOpcode == 0)
-
-  // Note: for atomic.swap (when BinOpcode == 0), storing incr2 to stack before
-  // the loop and then loading it from stack in block loopMBB is necessary to
-  // prevent MachineLICM pass to hoist "or" instruction out of the block
-  // loopMBB.
 
   int64_t MaskImm = (Size == 1) ? 255 : 65535;
   BuildMI(BB, dl, TII->get(Mips::ADDiu), Tmp1).addReg(Mips::ZERO).addImm(-4);
@@ -904,21 +872,9 @@ MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
     BuildMI(BB, dl, TII->get(Mips::SLL), Incr2).addReg(Tmp5).addReg(Shift);
   }
 
-  int fi = 0;
-  if (BinOpcode == 0 && !Nand) {
-    // Get or create a temporary stack location.
-    MipsFunctionInfo *MipsFI = MF->getInfo<MipsFunctionInfo>();
-    fi = MipsFI->getAtomicFrameIndex();
-    if (fi == -1) {
-      fi = MF->getFrameInfo()->CreateStackObject(Size, Size, false);
-      MipsFI->setAtomicFrameIndex(fi);
-    }
-
-    BuildMI(BB, dl, TII->get(Mips::SW))
-        .addReg(Incr2).addFrameIndex(fi).addImm(0);
-  }
   BB->addSuccessor(loopMBB);
 
+  // atomic.load.binop
   // loopMBB:
   //   ll      oldval,0(addr)
   //   binop   tmp7,oldval,incr2
@@ -927,6 +883,15 @@ MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
   //   or      tmp9,tmp8,newval
   //   sc      tmp9,0(addr)
   //   beq     tmp9,$0,loopMBB
+  
+  // atomic.swap
+  // loopMBB:
+  //   ll      oldval,0(addr)
+  //   and     tmp8,oldval,mask2
+  //   or      tmp9,tmp8,incr2
+  //   sc      tmp9,0(addr)
+  //   beq     tmp9,$0,loopMBB
+
   BB = loopMBB;
   BuildMI(BB, dl, TII->get(Mips::LL), Oldval).addReg(Addr).addImm(0);
   if (Nand) {
@@ -940,15 +905,14 @@ MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
   } else if (BinOpcode) {
     //  <binop> tmp7, oldval, incr2
     BuildMI(BB, dl, TII->get(BinOpcode), Tmp7).addReg(Oldval).addReg(Incr2);
-  } else {
-    //  lw tmp6, fi(sp)              // load incr2 from stack
-    //  or tmp7, $zero, tmp6
-    BuildMI(BB, dl, TII->get(Mips::LW), Tmp6).addFrameIndex(fi).addImm(0);
-    BuildMI(BB, dl, TII->get(Mips::OR), Tmp7).addReg(Mips::ZERO).addReg(Tmp6);
   }
-  BuildMI(BB, dl, TII->get(Mips::AND), Newval).addReg(Tmp7).addReg(Mask);
+  if (BinOpcode != 0 || Nand)
+    BuildMI(BB, dl, TII->get(Mips::AND), Newval).addReg(Tmp7).addReg(Mask);
   BuildMI(BB, dl, TII->get(Mips::AND), Tmp8).addReg(Oldval).addReg(Mask2);
-  BuildMI(BB, dl, TII->get(Mips::OR), Tmp9).addReg(Tmp8).addReg(Newval);
+  if (BinOpcode != 0 || Nand)
+    BuildMI(BB, dl, TII->get(Mips::OR), Tmp9).addReg(Tmp8).addReg(Newval);
+  else
+    BuildMI(BB, dl, TII->get(Mips::OR), Tmp9).addReg(Tmp8).addReg(Incr2);
   BuildMI(BB, dl, TII->get(Mips::SC), Tmp13)
     .addReg(Tmp9).addReg(Addr).addImm(0);
   BuildMI(BB, dl, TII->get(Mips::BEQ))
@@ -996,7 +960,6 @@ MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
   unsigned Newval  = MI->getOperand(3).getReg();
 
   unsigned Tmp1 = RegInfo.createVirtualRegister(RC);
-  unsigned Tmp2 = RegInfo.createVirtualRegister(RC);
   unsigned Tmp3 = RegInfo.createVirtualRegister(RC);
 
   // insert new blocks after the current block
@@ -1016,25 +979,9 @@ MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
                   BB->end());
   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
 
-  // Get or create a temporary stack location.
-  MipsFunctionInfo *MipsFI = MF->getInfo<MipsFunctionInfo>();
-  int fi = MipsFI->getAtomicFrameIndex();
-  if (fi == -1) {
-    fi = MF->getFrameInfo()->CreateStackObject(Size, Size, false);
-    MipsFI->setAtomicFrameIndex(fi);
-  }
-
   //  thisMBB:
   //    ...
-  //    sw newval, fi(sp)           // store newval to stack
   //    fallthrough --> loop1MBB
-
-  // Note: storing newval to stack before the loop and then loading it from
-  // stack in block loop2MBB is necessary to prevent MachineLICM pass to
-  // hoist "or" instruction out of the block loop2MBB.
-
-  BuildMI(BB, dl, TII->get(Mips::SW))
-      .addReg(Newval).addFrameIndex(fi).addImm(0);
   BB->addSuccessor(loop1MBB);
 
   // loop1MBB:
@@ -1048,13 +995,11 @@ MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
   BB->addSuccessor(loop2MBB);
 
   // loop2MBB:
-  //   lw tmp2, fi(sp)              // load newval from stack
-  //   or tmp1, $0, tmp2
+  //   or tmp1, $0, newval
   //   sc tmp1, 0(ptr)
   //   beq tmp1, $0, loop1MBB
   BB = loop2MBB;
-  BuildMI(BB, dl, TII->get(Mips::LW), Tmp2).addFrameIndex(fi).addImm(0);
-  BuildMI(BB, dl, TII->get(Mips::OR), Tmp1).addReg(Mips::ZERO).addReg(Tmp2);
+  BuildMI(BB, dl, TII->get(Mips::OR), Tmp1).addReg(Mips::ZERO).addReg(Newval);
   BuildMI(BB, dl, TII->get(Mips::SC), Tmp3).addReg(Tmp1).addReg(Ptr).addImm(0);
   BuildMI(BB, dl, TII->get(Mips::BEQ))
     .addReg(Tmp3).addReg(Mips::ZERO).addMBB(loop1MBB);
diff --git a/lib/Target/Mips/MipsInstrInfo.td b/lib/Target/Mips/MipsInstrInfo.td
index d1a0587124..72265d0f22 100644
--- a/lib/Target/Mips/MipsInstrInfo.td
+++ b/lib/Target/Mips/MipsInstrInfo.td
@@ -590,10 +590,10 @@ def SH      : StoreM<0x29, "sh", truncstorei16>;
 def SW      : StoreM<0x2b, "sw", store>;
 
 /// Load-linked, Store-conditional
-let hasDelaySlot = 1 in
+let mayLoad = 1, hasDelaySlot = 1 in
   def LL    : FI<0x30, (outs CPURegs:$dst), (ins mem:$addr),
               "ll\t$dst, $addr", [], IILoad>;
-let Constraints = "$src = $dst" in
+let mayStore = 1, Constraints = "$src = $dst" in
   def SC    : FI<0x38, (outs CPURegs:$dst), (ins CPURegs:$src, mem:$addr),
               "sc\t$src, $addr", [], IIStore>;
 
diff --git a/lib/Target/Mips/MipsMachineFunction.h b/lib/Target/Mips/MipsMachineFunction.h
index dbb7a67442..bc30b6b242 100644
--- a/lib/Target/Mips/MipsMachineFunction.h
+++ b/lib/Target/Mips/MipsMachineFunction.h
@@ -51,16 +51,12 @@ private:
   mutable int DynAllocFI; // Frame index of dynamically allocated stack area.   
   unsigned MaxCallFrameSize;
 
-  /// AtomicFrameIndex - To implement atomic.swap and atomic.cmp.swap
-  /// intrinsics, it is necessary to use a temporary stack location.
-  /// This field holds the frame index of this location.
-  int AtomicFrameIndex;
 public:
   MipsFunctionInfo(MachineFunction& MF)
   : MF(MF), SRetReturnReg(0), GlobalBaseReg(0),
     VarArgsFrameIndex(0), InArgFIRange(std::make_pair(-1, 0)),
     OutArgFIRange(std::make_pair(-1, 0)), GPFI(0), DynAllocFI(0),
-    MaxCallFrameSize(0), AtomicFrameIndex(-1)
+    MaxCallFrameSize(0)
   {}
 
   bool isInArgFI(int FI) const {
@@ -104,9 +100,6 @@ public:
 
   unsigned getMaxCallFrameSize() const { return MaxCallFrameSize; }
   void setMaxCallFrameSize(unsigned S) { MaxCallFrameSize = S; }
-
-  int getAtomicFrameIndex() const { return AtomicFrameIndex; }
-  void setAtomicFrameIndex(int Index) { AtomicFrameIndex = Index; }
 };
 
 } // end of namespace llvm
diff --git a/test/CodeGen/Mips/atomic.ll b/test/CodeGen/Mips/atomic.ll
index 2d5555bd2c..71e39285db 100644
--- a/test/CodeGen/Mips/atomic.ll
+++ b/test/CodeGen/Mips/atomic.ll
@@ -24,7 +24,6 @@ entry:
 ; CHECK:   lw      $[[R0:[0-9]+]], %got(x)($gp)
 ; CHECK:   $[[BB0:[A-Z_0-9]+]]:
 ; CHECK:   ll      $[[R1:[0-9]+]], 0($[[R0]])
-; CHECK:   or      $2, $zero, $[[R1]]
 ; CHECK:   addu    $[[R2:[0-9]+]], $[[R1]], $4
 ; CHECK:   sc      $[[R2]], 0($[[R0]])
 ; CHECK:   beq     $[[R2]], $zero, $[[BB0]]
@@ -39,43 +38,42 @@ entry:
 ; CHECK:   lw      $[[R0:[0-9]+]], %got(x)($gp)
 ; CHECK:   $[[BB0:[A-Z_0-9]+]]:
 ; CHECK:   ll      $[[R1:[0-9]+]], 0($[[R0]])
-; CHECK:   or      $2, $zero, $[[R1]]
-; CHECK:   and     $[[R1]], $[[R1]], $4
-; CHECK:   nor     $[[R2:[0-9]+]], $zero, $[[R1]]
+; CHECK:   and     $[[R3:[0-9]+]], $[[R1]], $4
+; CHECK:   nor     $[[R2:[0-9]+]], $zero, $[[R3]]
 ; CHECK:   sc      $[[R2]], 0($[[R0]])
 ; CHECK:   beq     $[[R2]], $zero, $[[BB0]]
 }
 
-define i32 @AtomicSwap32(i32 %oldval) nounwind {
+define i32 @AtomicSwap32(i32 %newval) nounwind {
 entry:
-  %0 = call i32 @llvm.atomic.swap.i32.p0i32(i32* @x, i32 %oldval)
+  %newval.addr = alloca i32, align 4
+  store i32 %newval, i32* %newval.addr, align 4
+  %tmp = load i32* %newval.addr, align 4
+  %0 = call i32 @llvm.atomic.swap.i32.p0i32(i32* @x, i32 %tmp)
   ret i32 %0
 
 ; CHECK:   AtomicSwap32:
 ; CHECK:   lw      $[[R0:[0-9]+]], %got(x)($gp)
-; CHECK:   sw      $4, [[OFFSET:[0-9]+]]($sp)
 ; CHECK:   $[[BB0:[A-Z_0-9]+]]:
-; CHECK:   ll      $[[R1:[0-9]+]], 0($[[R0]])
-; CHECK:   or      $2, $zero, $[[R1]]
-; CHECK:   lw      $[[R2:[0-9]+]], [[OFFSET]]($sp)
-; CHECK:   or      $[[R3:[0-9]+]], $zero, $[[R2]]
-; CHECK:   sc      $[[R3]], 0($[[R0]])
-; CHECK:   beq     $[[R3]], $zero, $[[BB0]]
+; CHECK:   ll      ${{[0-9]+}}, 0($[[R0]])
+; CHECK:   sc      $[[R2:[0-9]+]], 0($[[R0]])
+; CHECK:   beq     $[[R2]], $zero, $[[BB0]]
 }
 
 define i32 @AtomicCmpSwap32(i32 %oldval, i32 %newval) nounwind {
 entry:
-  %0 = call i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* @x, i32 %oldval, i32 %newval)
+  %newval.addr = alloca i32, align 4
+  store i32 %newval, i32* %newval.addr, align 4
+  %tmp = load i32* %newval.addr, align 4
+  %0 = call i32 @llvm.atomic.cmp.swap.i32.p0i32(i32* @x, i32 %oldval, i32 %tmp)
   ret i32 %0
 
 ; CHECK:   AtomicCmpSwap32:
 ; CHECK:   lw      $[[R0:[0-9]+]], %got(x)($gp)
-; CHECK:   sw      $5, [[OFFSET:[0-9]+]]($sp)
 ; CHECK:   $[[BB0:[A-Z_0-9]+]]:
 ; CHECK:   ll      $2, 0($[[R0]])
 ; CHECK:   bne     $2, $4, $[[BB1:[A-Z_0-9]+]]
-; CHECK:   lw      $[[R1:[0-9]+]], [[OFFSET]]($sp)
-; CHECK:   or      $[[R2:[0-9]+]], $zero, $[[R1]]
+; CHECK:   or      $[[R2:[0-9]+]], $zero, $5
 ; CHECK:   sc      $[[R2]], 0($[[R0]])
 ; CHECK:   beq     $[[R2]], $zero, $[[BB0]]
 ; CHECK:   $[[BB1]]:
@@ -183,9 +181,9 @@ entry:
 ; CHECK:   sra     $2, $[[R17]], 24
 }
 
-define signext i8 @AtomicSwap8(i8 signext %oldval) nounwind {
+define signext i8 @AtomicSwap8(i8 signext %newval) nounwind {
 entry:
-  %0 = call i8 @llvm.atomic.swap.i8.p0i8(i8* @y, i8 %oldval)
+  %0 = call i8 @llvm.atomic.swap.i8.p0i8(i8* @y, i8 %newval)
   ret i8 %0
 
 ; CHECK:   AtomicSwap8:
@@ -199,15 +197,11 @@ entry:
 ; CHECK:   nor     $[[R7:[0-9]+]], $zero, $[[R6]]
 ; CHECK:   andi    $[[R8:[0-9]+]], $4, 255
 ; CHECK:   sll     $[[R9:[0-9]+]], $[[R8]], $[[R4]]
-; CHECK:   sw      $[[R9]], [[OFFSET:[0-9]+]]($sp)
 
 ; CHECK:   $[[BB0:[A-Z_0-9]+]]:
 ; CHECK:   ll      $[[R10:[0-9]+]], 0($[[R2]])
-; CHECK:   lw      $[[R18:[0-9]+]], [[OFFSET]]($sp)
-; CHECK:   or      $[[R11:[0-9]+]], $zero, $[[R18]]
-; CHECK:   and     $[[R12:[0-9]+]], $[[R11]], $[[R6]]
 ; CHECK:   and     $[[R13:[0-9]+]], $[[R10]], $[[R7]]
-; CHECK:   or      $[[R14:[0-9]+]], $[[R13]], $[[R12]]
+; CHECK:   or      $[[R14:[0-9]+]], $[[R13]], $[[R9]]
 ; CHECK:   sc      $[[R14]], 0($[[R2]])
 ; CHECK:   beq     $[[R14]], $zero, $[[BB0]]