Subzero ARM: lower alloca instruction.

src/IceTargetLoweringARM32.cpp

 //===- subzero/src/IceTargetLoweringARM32.cpp - ARM32 lowering ------------===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the TargetLoweringARM32 class, which consists almost
@@ skipping 105 matching lines @@
   static_assert(                                                               \
       _table1_##tag == _table2_##tag,                                          \
       "Inconsistency between ICMPARM32_TABLE and ICEINSTICMP_TABLE");
 ICEINSTICMP_TABLE
 #undef X
 } // end of namespace dummy1
 
 // The maximum number of arguments to pass in GPR registers.
 const uint32_t ARM32_MAX_GPR_ARG = 4;
 
+// Stack alignment
+const uint32_t ARM32_STACK_ALIGNMENT_BYTES = 16;
+
 } // end of anonymous namespace
 
 TargetARM32::TargetARM32(Cfg *Func)
     : TargetLowering(Func), UsesFramePointer(false) {
   // TODO: Don't initialize IntegerRegisters and friends every time.
   // Instead, initialize in some sort of static initializer for the
   // class.
   llvm::SmallBitVector IntegerRegisters(RegARM32::Reg_NUM);
   llvm::SmallBitVector FloatRegisters(RegARM32::Reg_NUM);
   llvm::SmallBitVector VectorRegisters(RegARM32::Reg_NUM);
@@ skipping 464 matching lines @@
 }
 
 void TargetARM32::lowerAlloca(const InstAlloca *Inst) {
   UsesFramePointer = true;
   // Conservatively require the stack to be aligned.  Some stack
   // adjustment operations implemented below assume that the stack is
   // aligned before the alloca.  All the alloca code ensures that the
   // stack alignment is preserved after the alloca.  The stack alignment
   // restriction can be relaxed in some cases.
   NeedsStackAlignment = true;
-  (void)Inst;
-  UnimplementedError(Func->getContext()->getFlags());
+
+  // TODO(sehr,stichnot): minimize the number of adjustments of SP, etc.

[Jim Stichnoth, 2015/05/26 16:58:55]

Can probably remove sehr... :)

[jvoung (off chromium), 2015/05/26 20:53:46]

Done -- darn =)

+  Variable *SP = getPhysicalRegister(RegARM32::Reg_sp);
+  Variable *Dest = Inst->getDest();
+  uint32_t AlignmentParam = Inst->getAlignInBytes();
+  // For default align=0, set it to the real value 1, to avoid any
+  // bit-manipulation problems below.
+  AlignmentParam = std::max(AlignmentParam, 1u);
+
+  // LLVM enforces power of 2 alignment.
+  assert(llvm::isPowerOf2_32(AlignmentParam));
+  assert(llvm::isPowerOf2_32(ARM32_STACK_ALIGNMENT_BYTES));
+
+  uint32_t Alignment = std::max(AlignmentParam, ARM32_STACK_ALIGNMENT_BYTES);
+  if (Alignment > ARM32_STACK_ALIGNMENT_BYTES) {
+    alignRegisterPow2(SP, Alignment);
+  }
+  Operand *TotalSize = Inst->getSizeInBytes();
+  if (ConstantInteger32 *ConstantTotalSize =

[Jim Stichnoth, 2015/05/26 16:58:55]

const auto *ConstantTotalSize
?

[jvoung (off chromium), 2015/05/26 20:53:46]

Done.

+          llvm::dyn_cast<ConstantInteger32>(TotalSize)) {
+    uint32_t Value = ConstantTotalSize->getValue();
+    Value = Utils::applyAlignment(Value, Alignment);
+    Operand *SubAmount = legalize(Ctx->getConstantInt32(Value));
+    _sub(SP, SP, SubAmount);
+  } else {
+    // Non-constant sizes need to be adjusted to the next highest
+    // multiple of the required alignment at runtime.
+    TotalSize = legalize(TotalSize);
+    Variable *T = makeReg(IceType_i32);
+    _mov(T, TotalSize);
+    Operand *AddAmount = legalize(Ctx->getConstantInt32(Alignment - 1));
+    _add(T, T, AddAmount);
+    alignRegisterPow2(T, Alignment);
+    _sub(SP, SP, T);
+  }
+  _mov(Dest, SP);
 }
 
 void TargetARM32::lowerArithmetic(const InstArithmetic *Inst) {
   Variable *Dest = Inst->getDest();
   // TODO(jvoung): Should be able to flip Src0 and Src1 if it is easier
   // to legalize Src0 to flex or Src1 to flex and there is a reversible
   // instruction. E.g., reverse subtract with immediate, register vs
   // register, immediate.
   // Or it may be the case that the operands aren't swapped, but the
   // bits can be flipped and a different operation applied.
@@ skipping 899 matching lines @@
   // There aren't any 64-bit integer registers for ARM32.
   assert(Type != IceType_i64);
   Variable *Reg = Func->makeVariable(Type);
   if (RegNum == Variable::NoRegister)
     Reg->setWeightInfinite();
   else
     Reg->setRegNum(RegNum);
   return Reg;
 }
 
+void TargetARM32::alignRegisterPow2(Variable *Reg, uint32_t Align) {
+  assert(llvm::isPowerOf2_32(Align));
+  uint32_t RotateAmt = 0;
+  uint32_t Immed_8;
+  Operand *Mask;
+  // Use AND or BIC to mask off the bits, depending on which immediate fits
+  // (if it fits at all). Assume Align is usually small, in which case BIC
+  // works better.
+  if (OperandARM32FlexImm::canHoldImm(Align - 1, &RotateAmt, &Immed_8)) {
+    Mask = legalize(Ctx->getConstantInt32(Align - 1), Legal_Reg | Legal_Flex);
+    _bic(Reg, Reg, Mask);
+  } else {
+    Mask = legalize(Ctx->getConstantInt32(-Align), Legal_Reg | Legal_Flex);
+    _and(Reg, Reg, Mask);
+  }
+}
+
 void TargetARM32::postLower() {
   if (Ctx->getFlags().getOptLevel() == Opt_m1)
     return;
   inferTwoAddress();
 }
 
 void TargetARM32::makeRandomRegisterPermutation(
     llvm::SmallVectorImpl<int32_t> &Permutation,
     const llvm::SmallBitVector &ExcludeRegisters) const {
   (void)Permutation;
@@ skipping 54 matching lines @@
   }
 }
 
 void TargetDataARM32::lowerConstants() const {
   if (Ctx->getFlags().getDisableTranslation())
     return;
   UnimplementedError(Ctx->getFlags());
 }
 
 } // end of namespace Ice