Use default lowering of DYNAMIC_STACKALLOC. As far as I can tell, ARM isle is doing the right thing and codegen looks correct for both Thumb and Thumb2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@101410 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARMISelDAGToDAG.cpp

@@ -121,9 +121,6 @@ private:
   SDNode *SelectARMIndexedLoad(SDNode *N);
   SDNode *SelectT2IndexedLoad(SDNode *N);
 
-  /// SelectDYN_ALLOC - Select dynamic alloc for Thumb.
-  SDNode *SelectDYN_ALLOC(SDNode *N);
-
   /// SelectVLD - Select NEON load intrinsics.  NumVecs should be
   /// 1, 2, 3 or 4.  The opcode arrays specify the instructions used for
   /// loads of D registers and even subregs and odd subregs of Q registers.
@@ -939,59 +936,6 @@ SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
   return NULL;
 }
 
-SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDNode *N) {
-  DebugLoc dl = N->getDebugLoc();
-  EVT VT = N->getValueType(0);
-  SDValue Chain = N->getOperand(0);
-  SDValue Size = N->getOperand(1);
-  SDValue Align = N->getOperand(2);
-  SDValue SP = CurDAG->getCopyFromReg(Chain, dl, ARM::SP, MVT::i32);
-  int32_t AlignVal = cast<ConstantSDNode>(Align)->getSExtValue();
-  if (AlignVal < 0)
-    // We need to align the stack. Use Thumb1 tAND which is the only thumb
-    // instruction that can read and write SP. This matches to a pseudo
-    // instruction that has a chain to ensure the result is written back to
-    // the stack pointer.
-    SP = SDValue(CurDAG->getMachineNode(ARM::tANDsp, dl, VT, SP, Align), 0);
-
-  bool isC = isa<ConstantSDNode>(Size);
-  uint32_t C = isC ? cast<ConstantSDNode>(Size)->getZExtValue() : ~0UL;
-  // Handle the most common case for both Thumb1 and Thumb2:
-  // tSUBspi - immediate is between 0 ... 508 inclusive.
-  if (C <= 508 && ((C & 3) == 0))
-    // FIXME: tSUBspi encode scale 4 implicitly.
-    return CurDAG->SelectNodeTo(N, ARM::tSUBspi_, VT, MVT::Other, SP,
-                                CurDAG->getTargetConstant(C/4, MVT::i32),
-                                Chain);
-
-  if (Subtarget->isThumb1Only()) {
-    // Use tADDspr since Thumb1 does not have a sub r, sp, r. ARMISelLowering
-    // should have negated the size operand already. FIXME: We can't insert
-    // new target independent node at this stage so we are forced to negate
-    // it earlier. Is there a better solution?
-    return CurDAG->SelectNodeTo(N, ARM::tADDspr_, VT, MVT::Other, SP, Size,
-                                Chain);
-  } else if (Subtarget->isThumb2()) {
-    if (isC && Predicate_t2_so_imm(Size.getNode())) {
-      // t2SUBrSPi
-      SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
-      return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi_, VT, MVT::Other, Ops, 3);
-    } else if (isC && Predicate_imm0_4095(Size.getNode())) {
-      // t2SUBrSPi12
-      SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
-      return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi12_, VT, MVT::Other, Ops, 3);
-    } else {
-      // t2SUBrSPs
-      SDValue Ops[] = { SP, Size,
-                        getI32Imm(ARM_AM::getSORegOpc(ARM_AM::lsl,0)), Chain };
-      return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPs_, VT, MVT::Other, Ops, 4);
-    }
-  }
-
-  // FIXME: Add ADD / SUB sp instructions for ARM.
-  return 0;
-}
-
 /// PairDRegs - Insert a pair of double registers into an implicit def to
 /// form a quad register.
 SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
@@ -1571,8 +1515,6 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
       return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
     }
   }
-  case ARMISD::DYN_ALLOC:
-    return SelectDYN_ALLOC(N);
   case ISD::SRL:
     if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
                       Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX))

lib/Target/ARM/ARMISelLowering.cpp

@@ -384,10 +384,7 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
   // FIXME: Shouldn't need this, since no register is used, but the legalizer
   // doesn't yet know how to not do that for SjLj.
   setExceptionSelectorRegister(ARM::R0);
-  if (Subtarget->isThumb())
+  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
-    setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
-  else
-    setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
   setOperationAction(ISD::MEMBARRIER,         MVT::Other, Custom);
 
   if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {

test/CodeGen/Thumb2/2010-04-15-DynAllocBug.ll

@@ -7,15 +7,15 @@
 define arm_apcscc void @t() nounwind ssp {
 entry:
 ; CHECK: t:
-; CHECK: mvn r0, #7
+; CHECK: mov r0, sp
-; CHECK: ands sp, r0
+; CHECK: bfc r0, #0, #3
-; CHECK: mov r1, sp
+; CHECK: subs r0, #16
-; CHECK: mov sp, r1
+; CHECK: mov sp, r0
 ; Yes, this is stupid codegen, but it's correct.
-; CHECK: sub sp, #16
+; CHECK: mov r0, sp
-; CHECK: mov r1, sp
+; CHECK: bfc r0, #0, #3
-; CHECK: mov sp, r1
+; CHECK: subs r0, #16
-; CHECK: ands sp, r0
+; CHECK: mov sp, r0
   %size = mul i32 8, 2
   %vla_a = alloca i8, i32 %size, align 8
   %vla_b = alloca i8, i32 %size, align 8