指令选择期间的结果重新洗牌

时间:2017-06-04 04:22:26

标签: tuples llvm llvm-ir machine-instruction multiple-return-values

在LLVM后端,在指令选择期间,我的输入看起来像这样:

t17: i16,ch = load t16:1, t2, undef:i16

我想选择一个有一些额外结果的操作码,即用

之类的内容替换上面的操作码
t17: i16, _: i16, ch = LDW t2, undef:i16

当然,如果我只是尝试用此load替换LDW,则会失败,因为返回类型不同:

llc: lib/CodeGen/SelectionDAG/SelectionDAG.cpp:6525: 
void llvm::SelectionDAG::ReplaceAllUsesWith(llvm::SDNode*, llvm::SDNode*): 
Assertion `(!From->hasAnyUseOfValue(i) || From->getValueType(i) == To->getValueType(i)) && 
"Cannot use this version of ReplaceAllUsesWith!"' failed.

那是因为i == 1From->getValueType(i)chTo->getValueType(i)i16

我的问题是,如何在指令选择期间将元组(x: i16, y: i16, ch)重新转换为(x : i16, ch)?即我该怎么办

t17 : i16, ch = project [0, 2] (LDW t2, undef:i16)

其中project [0, 2]是一个神奇的元组重组伪指令,只需要处理掉坐标1?

编辑:发现看似可能的解决方案,但仍然无法正常工作

我发现MERGE_VALUES节点类型应该是我尝试做的事情,因为它允许接受一些输入并将它们打包成多值输出。所以我试过

SDNode* LDW = CurDAG->getMachineNode(
  AVR::LDWRdPtr, SDLoc(N), VT, PtrVT, MVT::Other,
  LD->getBasePtr(), LD->getChain());
SDValue Unpack[] = { SDValue(LDW, 0), SDValue(LDW, 2) };
SDNode* NN = CurDAG->getMergeValues(Unpack, SDLoc(N)).getNode();

ReplaceNode(N, NN);

如果我用--debug运行指令选择,这看起来很有希望,因为它转为:

SelectionDAG has 16 nodes:
  t0: ch = EntryToken
    t2: i16,ch = CopyFromReg t0, Register:i16 %vreg0
  t5: i16,ch = load<LD2[%v25](align=1)(dereferenceable)> t0, t2, undef:i16
    t9: ch,glue = callseq_start t5:1, TargetConstant:i16<0>
  t11: ch,glue = CopyToReg t9, Register:i16 %R25R24, t5
  t13: ch,glue = CALL t11, TargetGlobalAddress:i16<i8 (i16)* @read_ram> 0, Register:i16 %R25R24, RegisterMask:Untyped, t11:1
  t14: ch,glue = callseq_end t13, TargetConstant:i16<0>, TargetConstant:i16<0>, t13:1
    t16: i8,ch,glue = CopyFromReg t14, Register:i8 %R24, t14:1
  t17: ch = RET_FLAG t16:1

进入这个:

SelectionDAG has 16 nodes:
  t0: ch = EntryToken
    t9: i16,ch,glue = ADJCALLSTACKDOWN TargetConstant:i16<0>, t19:1
  t11: ch,glue = CopyToReg t9:1, Register:i16 %R25R24, t19
  t13: ch,glue = CALLk TargetGlobalAddress:i16<i8 (i16)* @read_ram> 0, Register:i16 %R25R24, RegisterMask:Untyped, t11, t11:1
  t14: i16,ch,glue = ADJCALLSTACKUP TargetConstant:i16<0>, TargetConstant:i16<0>, t13, t13:1
    t2: i16,ch = CopyFromReg t0, Register:i16 %vreg0
  t18: i16,i16,ch = LDWRdPtr t2, t0
  t19: i16,ch = merge_values t18, t18:2
    t16: i8,ch,glue = CopyFromReg t14:1, Register:i8 %R24, t14:2
  t17: ch = RET t16:1

对我来说很好:t18,_,ch是我们存储LDW结果的地方,然后t19,ch将其解包。

但是,这会在以后启动指令调度:

llc: CodeGen/SelectionDAG/InstrEmitter.cpp:305: 
unsigned int llvm::InstrEmitter::getVR(llvm::SDValue, llvm::DenseMap<llvm::SDValue, unsigned int>&): 
Assertion `I != VRBaseMap.end() && "Node emitted out of order - late"' failed.

#9 0x0000000001a7eae8 llvm::InstrEmitter::getVR(llvm::SDValue, llvm::DenseMap<llvm::SDValue, unsigned int, llvm::DenseMapInfo<llvm::SDValue>, llvm::detail::DenseMapPair<llvm::SDValue, unsigned int> >&) CodeGen/SelectionDAG/InstrEmitter.cpp:305:0
#10 0x0000000001a819c2 llvm::InstrEmitter::EmitSpecialNode(llvm::SDNode*, bool, bool, llvm::DenseMap<llvm::SDValue, unsigned int, llvm::DenseMapInfo<llvm::SDValue>, llvm::detail::DenseMapPair<llvm::SDValue, unsigned int> >&) CodeGen/SelectionDAG/InstrEmitter.cpp:928:0
#11 0x0000000001971c9c llvm::InstrEmitter::EmitNode(llvm::SDNode*, bool, bool, llvm::DenseMap<llvm::SDValue, unsigned int, llvm::DenseMapInfo<llvm::SDValue>, llvm::detail::DenseMapPair<llvm::SDValue, unsigned int> >&) CodeGen/SelectionDAG/InstrEmitter.h:124:0
#12 0x0000000001988bdb llvm::ScheduleDAGSDNodes::EmitSchedule(llvm::MachineInstrBundleIterator<llvm::MachineInstr, false>&) CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp:841:0
#13 0x0000000001a33cdc llvm::SelectionDAGISel::CodeGenAndEmitDAG() CodeGen/SelectionDAG/SelectionDAGISel.cpp:890:0

0 个答案:

没有答案