// sanity checks that apply to all kinds of IR
void Rationalizer::SanityCheck()
{
// TODO: assert(!IsLIR());
BasicBlock* block;
foreach_block(comp, block)
{
for (GenTree* statement = block->bbTreeList; statement != nullptr; statement = statement->gtNext)
{
ValidateStatement(statement, block);
for (GenTree* tree = statement->gtStmt.gtStmtList; tree; tree = tree->gtNext)
{
// QMARK nodes should have been removed before this phase.
assert(tree->OperGet() != GT_QMARK);
if (tree->OperGet() == GT_ASG)
{
if (tree->gtGetOp1()->OperGet() == GT_LCL_VAR)
{
assert(tree->gtGetOp1()->gtFlags & GTF_VAR_DEF);
}
else if (tree->gtGetOp2()->OperGet() == GT_LCL_VAR)
{
assert(!(tree->gtGetOp2()->gtFlags & GTF_VAR_DEF));
}
}
}
}
}
}
//------------------------------------------------------------------------
// DecomposeNeg: Decompose GT_NEG.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeNeg(LIR::Use& use)
{
assert(use.IsInitialized());
assert(use.Def()->OperGet() == GT_NEG);
GenTree* tree = use.Def();
GenTree* gtLong = tree->gtGetOp1();
noway_assert(gtLong->OperGet() == GT_LONG);
LIR::Use op1(Range(), >Long->gtOp.gtOp1, gtLong);
op1.ReplaceWithLclVar(m_compiler, m_blockWeight);
LIR::Use op2(Range(), >Long->gtOp.gtOp2, gtLong);
op2.ReplaceWithLclVar(m_compiler, m_blockWeight);
// Neither GT_NEG nor the introduced temporaries have side effects.
tree->gtFlags &= ~GTF_ALL_EFFECT;
GenTree* loOp1 = gtLong->gtGetOp1();
GenTree* hiOp1 = gtLong->gtGetOp2();
Range().Remove(gtLong);
GenTree* loResult = tree;
loResult->gtType = TYP_INT;
loResult->gtOp.gtOp1 = loOp1;
GenTree* zero = m_compiler->gtNewZeroConNode(TYP_INT);
GenTree* hiAdjust = m_compiler->gtNewOperNode(GT_ADD_HI, TYP_INT, hiOp1, zero);
GenTree* hiResult = m_compiler->gtNewOperNode(GT_NEG, TYP_INT, hiAdjust);
hiResult->gtFlags = tree->gtFlags;
Range().InsertAfter(loResult, zero, hiAdjust, hiResult);
return FinalizeDecomposition(use, loResult, hiResult);
}
//------------------------------------------------------------------------
// StoreNodeToVar: Check if the user is a STORE_LCL_VAR, and if it isn't,
// store the node to a var. Then decompose the new LclVar.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::StoreNodeToVar(LIR::Use& use)
{
if (use.IsDummyUse())
return use.Def()->gtNext;
GenTree* tree = use.Def();
GenTree* user = use.User();
if (user->OperGet() == GT_STORE_LCL_VAR)
{
// If parent is already a STORE_LCL_VAR, we can skip it if
// it is already marked as lvIsMultiRegRet.
unsigned varNum = user->AsLclVarCommon()->gtLclNum;
if (m_compiler->lvaTable[varNum].lvIsMultiRegRet)
{
return tree->gtNext;
}
else if (!m_compiler->lvaTable[varNum].lvPromoted)
{
// If var wasn't promoted, we can just set lvIsMultiRegRet.
m_compiler->lvaTable[varNum].lvIsMultiRegRet = true;
return tree->gtNext;
}
}
// Otherwise, we need to force var = call()
unsigned varNum = use.ReplaceWithLclVar(m_compiler, m_blockWeight);
m_compiler->lvaTable[varNum].lvIsMultiRegRet = true;
// Decompose the new LclVar use
return DecomposeLclVar(use);
}
//------------------------------------------------------------------------
// DecomposeMul: Decompose GT_MUL. The only GT_MULs that make it to decompose are
// those with the GTF_MUL_64RSLT flag set. These muls result in a mul instruction that
// returns its result in two registers like GT_CALLs do. Additionally, these muls are
// guaranteed to be in the form long = (long)int * (long)int. Therefore, to decompose
// these nodes, we convert them into GT_MUL_LONGs, undo the cast from int to long by
// stripping out the lo ops, and force them into the form var = mul, as we do for
// GT_CALLs. In codegen, we then produce a mul instruction that produces the result
// in edx:eax, and store those registers on the stack in genStoreLongLclVar.
//
// All other GT_MULs have been converted to helper calls in morph.cpp
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeMul(LIR::Use& use)
{
assert(use.IsInitialized());
GenTree* tree = use.Def();
genTreeOps oper = tree->OperGet();
assert(oper == GT_MUL);
assert((tree->gtFlags & GTF_MUL_64RSLT) != 0);
GenTree* op1 = tree->gtGetOp1();
GenTree* op2 = tree->gtGetOp2();
GenTree* loOp1 = op1->gtGetOp1();
GenTree* hiOp1 = op1->gtGetOp2();
GenTree* loOp2 = op2->gtGetOp1();
GenTree* hiOp2 = op2->gtGetOp2();
Range().Remove(hiOp1);
Range().Remove(hiOp2);
Range().Remove(op1);
Range().Remove(op2);
// Get rid of the hi ops. We don't need them.
tree->gtOp.gtOp1 = loOp1;
tree->gtOp.gtOp2 = loOp2;
tree->gtOper = GT_MUL_LONG;
return StoreNodeToVar(use);
}
void Rationalizer::RewriteAssignment(LIR::Use& use)
{
assert(use.IsInitialized());
GenTreeOp* assignment = use.Def()->AsOp();
assert(assignment->OperGet() == GT_ASG);
GenTree* location = assignment->gtGetOp1();
GenTree* value = assignment->gtGetOp2();
genTreeOps locationOp = location->OperGet();
switch (locationOp)
{
case GT_LCL_VAR:
case GT_LCL_FLD:
case GT_REG_VAR:
case GT_PHI_ARG:
RewriteAssignmentIntoStoreLclCore(assignment, location, value, locationOp);
BlockRange().Remove(location);
break;
case GT_IND:
{
GenTreeStoreInd* store =
new (comp, GT_STOREIND) GenTreeStoreInd(location->TypeGet(), location->gtGetOp1(), value);
copyFlags(store, assignment, GTF_ALL_EFFECT);
copyFlags(store, location, GTF_IND_FLAGS);
if (assignment->IsReverseOp())
{
store->gtFlags |= GTF_REVERSE_OPS;
}
// TODO: JIT dump
// Remove the GT_IND node and replace the assignment node with the store
BlockRange().Remove(location);
BlockRange().InsertBefore(assignment, store);
use.ReplaceWith(comp, store);
BlockRange().Remove(assignment);
}
break;
case GT_CLS_VAR:
{
location->SetOper(GT_CLS_VAR_ADDR);
location->gtType = TYP_BYREF;
assignment->SetOper(GT_STOREIND);
// TODO: JIT dump
}
break;
default:
unreached();
break;
}
}
GenTree* isNodeCallArg(ArrayStack<GenTree*>* parentStack)
{
for (int i = 1; // 0 is current node, so start at 1
i < parentStack->Height(); i++)
{
GenTree* node = parentStack->Index(i);
switch (node->OperGet())
{
case GT_LIST:
case GT_ARGPLACE:
break;
case GT_NOP:
// Currently there's an issue when the rationalizer performs
// the fixup of a call argument: the case is when we remove an
// inserted NOP as a parent of a call introduced by fgMorph;
// when then the rationalizer removes it, the tree stack in the
// walk is not consistent with the node it was just deleted, so the
// solution is just to go 1 level deeper.
// TODO-Cleanup: This has to be fixed in a proper way: make the rationalizer
// correctly modify the evaluation stack when removing treenodes.
if (node->gtOp.gtOp1->gtOper == GT_CALL)
{
return node->gtOp.gtOp1;
}
break;
case GT_CALL:
return node;
default:
return nullptr;
}
}
return nullptr;
}
// Rewrite a SIMD indirection as GT_IND(GT_LEA(obj.op1)), or as a simple
// lclVar if possible.
//
// Arguments:
// use - A use reference for a block node
// keepBlk - True if this should remain a block node if it is not a lclVar
//
// Return Value:
// None.
//
// TODO-1stClassStructs: These should be eliminated earlier, once we can handle
// lclVars in all the places that used to have GT_OBJ.
//
void Rationalizer::RewriteSIMDOperand(LIR::Use& use, bool keepBlk)
{
#ifdef FEATURE_SIMD
// No lowering is needed for non-SIMD nodes, so early out if featureSIMD is not enabled.
if (!comp->featureSIMD)
{
return;
}
GenTree* tree = use.Def();
if (!tree->OperIsIndir())
{
return;
}
var_types simdType = tree->TypeGet();
if (!varTypeIsSIMD(simdType))
{
return;
}
// If we have GT_IND(GT_LCL_VAR_ADDR) and the GT_LCL_VAR_ADDR is TYP_BYREF/TYP_I_IMPL,
// and the var is a SIMD type, replace the expression by GT_LCL_VAR.
GenTree* addr = tree->AsIndir()->Addr();
if (addr->OperIsLocalAddr() && comp->isAddrOfSIMDType(addr))
{
BlockRange().Remove(tree);
addr->SetOper(loadForm(addr->OperGet()));
addr->gtType = simdType;
use.ReplaceWith(comp, addr);
}
else if ((addr->OperGet() == GT_ADDR) && (addr->gtGetOp1()->OperGet() == GT_SIMD))
{
// if we have GT_IND(GT_ADDR(GT_SIMD)), remove the GT_IND(GT_ADDR()), leaving just the GT_SIMD.
BlockRange().Remove(tree);
BlockRange().Remove(addr);
use.ReplaceWith(comp, addr->gtGetOp1());
}
else if (!keepBlk)
{
tree->SetOper(GT_IND);
tree->gtType = simdType;
}
#endif // FEATURE_SIMD
}
//------------------------------------------------------------------------
// ContainCheckIndir: Determine whether operands of an indir should be contained.
//
// Arguments:
// indirNode - The indirection node of interest
//
// Notes:
// This is called for both store and load indirections.
//
// Return Value:
// None.
//
void Lowering::ContainCheckIndir(GenTreeIndir* indirNode)
{
// If this is the rhs of a block copy it will be handled when we handle the store.
if (indirNode->TypeGet() == TYP_STRUCT)
{
return;
}
#ifdef FEATURE_SIMD
// If indirTree is of TYP_SIMD12, don't mark addr as contained
// so that it always get computed to a register. This would
// mean codegen side logic doesn't need to handle all possible
// addr expressions that could be contained.
//
// TODO-ARM64-CQ: handle other addr mode expressions that could be marked
// as contained.
if (indirNode->TypeGet() == TYP_SIMD12)
{
return;
}
#endif // FEATURE_SIMD
GenTree* addr = indirNode->Addr();
bool makeContained = true;
if ((addr->OperGet() == GT_LEA) && IsSafeToContainMem(indirNode, addr))
{
GenTreeAddrMode* lea = addr->AsAddrMode();
GenTree* base = lea->Base();
GenTree* index = lea->Index();
int cns = lea->Offset();
#ifdef _TARGET_ARM_
// ARM floating-point load/store doesn't support a form similar to integer
// ldr Rdst, [Rbase + Roffset] with offset in a register. The only supported
// form is vldr Rdst, [Rbase + imm] with a more limited constraint on the imm.
if (lea->HasIndex() || !emitter::emitIns_valid_imm_for_vldst_offset(cns))
{
if (indirNode->OperGet() == GT_STOREIND)
{
if (varTypeIsFloating(indirNode->AsStoreInd()->Data()))
{
makeContained = false;
}
}
else if (indirNode->OperGet() == GT_IND)
{
if (varTypeIsFloating(indirNode))
{
makeContained = false;
}
}
}
#endif
if (makeContained)
{
MakeSrcContained(indirNode, addr);
}
}
}
//------------------------------------------------------------------------
// AssertWhenAllocObjFoundVisitor: Look for a GT_ALLOCOBJ node and assert
// when found one.
Compiler::fgWalkResult ObjectAllocator::AssertWhenAllocObjFoundVisitor(GenTree** pTree, Compiler::fgWalkData* data)
{
GenTree* tree = *pTree;
assert(tree != nullptr);
assert(tree->OperGet() != GT_ALLOCOBJ);
return Compiler::fgWalkResult::WALK_CONTINUE;
}
void Rationalizer::RewriteAddress(LIR::Use& use)
{
assert(use.IsInitialized());
GenTreeUnOp* address = use.Def()->AsUnOp();
assert(address->OperGet() == GT_ADDR);
GenTree* location = address->gtGetOp1();
genTreeOps locationOp = location->OperGet();
if (location->IsLocal())
{
// We are changing the child from GT_LCL_VAR TO GT_LCL_VAR_ADDR.
// Therefore gtType of the child needs to be changed to a TYP_BYREF
#ifdef DEBUG
if (locationOp == GT_LCL_VAR)
{
JITDUMP("Rewriting GT_ADDR(GT_LCL_VAR) to GT_LCL_VAR_ADDR:\n");
}
else
{
assert(locationOp == GT_LCL_FLD);
JITDUMP("Rewriting GT_ADDR(GT_LCL_FLD) to GT_LCL_FLD_ADDR:\n");
}
#endif // DEBUG
location->SetOper(addrForm(locationOp));
location->gtType = TYP_BYREF;
copyFlags(location, address, GTF_ALL_EFFECT);
use.ReplaceWith(comp, location);
BlockRange().Remove(address);
}
else if (locationOp == GT_CLS_VAR)
{
location->SetOper(GT_CLS_VAR_ADDR);
location->gtType = TYP_BYREF;
copyFlags(location, address, GTF_ALL_EFFECT);
use.ReplaceWith(comp, location);
BlockRange().Remove(address);
JITDUMP("Rewriting GT_ADDR(GT_CLS_VAR) to GT_CLS_VAR_ADDR:\n");
}
else if (location->OperIsIndir())
{
use.ReplaceWith(comp, location->gtGetOp1());
BlockRange().Remove(location);
BlockRange().Remove(address);
JITDUMP("Rewriting GT_ADDR(GT_IND(X)) to X:\n");
}
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
void Rationalizer::RewriteAssignmentIntoStoreLcl(GenTreeOp* assignment)
{
assert(assignment != nullptr);
assert(assignment->OperGet() == GT_ASG);
GenTree* location = assignment->gtGetOp1();
GenTree* value = assignment->gtGetOp2();
RewriteAssignmentIntoStoreLclCore(assignment, location, value, location->OperGet());
}
//------------------------------------------------------------------------
// DecomposeArith: Decompose GT_ADD, GT_SUB, GT_OR, GT_XOR, GT_AND.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeArith(LIR::Use& use)
{
assert(use.IsInitialized());
GenTree* tree = use.Def();
genTreeOps oper = tree->OperGet();
assert((oper == GT_ADD) || (oper == GT_SUB) || (oper == GT_OR) || (oper == GT_XOR) || (oper == GT_AND));
GenTree* op1 = tree->gtGetOp1();
GenTree* op2 = tree->gtGetOp2();
// Both operands must have already been decomposed into GT_LONG operators.
noway_assert((op1->OperGet() == GT_LONG) && (op2->OperGet() == GT_LONG));
// Capture the lo and hi halves of op1 and op2.
GenTree* loOp1 = op1->gtGetOp1();
GenTree* hiOp1 = op1->gtGetOp2();
GenTree* loOp2 = op2->gtGetOp1();
GenTree* hiOp2 = op2->gtGetOp2();
// Now, remove op1 and op2 from the node list.
Range().Remove(op1);
Range().Remove(op2);
// We will reuse "tree" for the loResult, which will now be of TYP_INT, and its operands
// will be the lo halves of op1 from above.
GenTree* loResult = tree;
loResult->SetOper(GetLoOper(oper));
loResult->gtType = TYP_INT;
loResult->gtOp.gtOp1 = loOp1;
loResult->gtOp.gtOp2 = loOp2;
GenTree* hiResult = new (m_compiler, oper) GenTreeOp(GetHiOper(oper), TYP_INT, hiOp1, hiOp2);
Range().InsertAfter(loResult, hiResult);
if ((oper == GT_ADD) || (oper == GT_SUB))
{
if (loResult->gtOverflow())
{
hiResult->gtFlags |= GTF_OVERFLOW;
loResult->gtFlags &= ~GTF_OVERFLOW;
}
if (loResult->gtFlags & GTF_UNSIGNED)
{
hiResult->gtFlags |= GTF_UNSIGNED;
}
}
return FinalizeDecomposition(use, loResult, hiResult);
}
//------------------------------------------------------------------------
// ContainCheckCast: determine whether the source of a CAST node should be contained.
//
// Arguments:
// node - pointer to the node
//
void Lowering::ContainCheckCast(GenTreeCast* node)
{
#ifdef _TARGET_ARM_
GenTree* castOp = node->CastOp();
var_types castToType = node->CastToType();
var_types srcType = castOp->TypeGet();
if (varTypeIsLong(castOp))
{
assert(castOp->OperGet() == GT_LONG);
MakeSrcContained(node, castOp);
}
#endif // _TARGET_ARM_
}
//------------------------------------------------------------------------
// ContainCheckIndir: Determine whether operands of an indir should be contained.
//
// Arguments:
// indirNode - The indirection node of interest
//
// Notes:
// This is called for both store and load indirections.
//
// Return Value:
// None.
//
void Lowering::ContainCheckIndir(GenTreeIndir* indirNode)
{
// If this is the rhs of a block copy it will be handled when we handle the store.
if (indirNode->TypeGet() == TYP_STRUCT)
{
return;
}
GenTree* addr = indirNode->Addr();
bool makeContained = true;
if ((addr->OperGet() == GT_LEA) && IsSafeToContainMem(indirNode, addr))
{
GenTreeAddrMode* lea = addr->AsAddrMode();
GenTree* base = lea->Base();
GenTree* index = lea->Index();
int cns = lea->Offset();
#ifdef _TARGET_ARM_
// ARM floating-point load/store doesn't support a form similar to integer
// ldr Rdst, [Rbase + Roffset] with offset in a register. The only supported
// form is vldr Rdst, [Rbase + imm] with a more limited constraint on the imm.
if (lea->HasIndex() || !emitter::emitIns_valid_imm_for_vldst_offset(cns))
{
if (indirNode->OperGet() == GT_STOREIND)
{
if (varTypeIsFloating(indirNode->AsStoreInd()->Data()))
{
makeContained = false;
}
}
else if (indirNode->OperGet() == GT_IND)
{
if (varTypeIsFloating(indirNode))
{
makeContained = false;
}
}
}
#endif
if (makeContained)
{
MakeSrcContained(indirNode, addr);
}
}
}
//------------------------------------------------------------------------
// ContainCheckShiftRotate: Determine whether a mul op's operands should be contained.
//
// Arguments:
// node - the node we care about
//
void Lowering::ContainCheckShiftRotate(GenTreeOp* node)
{
GenTree* shiftBy = node->gtOp2;
assert(node->OperIsShiftOrRotate());
#ifdef _TARGET_ARM_
GenTree* source = node->gtOp1;
if (node->OperIs(GT_LSH_HI, GT_RSH_LO))
{
assert(source->OperGet() == GT_LONG);
MakeSrcContained(node, source);
}
#endif // _TARGET_ARM_
if (shiftBy->IsCnsIntOrI())
{
MakeSrcContained(node, shiftBy);
}
}
// Rewrite GT_OBJ of SIMD Vector as GT_IND(GT_LEA(obj.op1)) of a SIMD type.
//
// Arguments:
// ppTree - A pointer-to-a-pointer for the GT_OBJ
// fgWalkData - A pointer to tree walk data providing the context
//
// Return Value:
// None.
//
// TODO-Cleanup: Once SIMD types are plumbed through the frontend, this will no longer
// be required.
//
void Rationalizer::RewriteObj(LIR::Use& use)
{
#ifdef FEATURE_SIMD
GenTreeObj* obj = use.Def()->AsObj();
// For UNIX struct passing, we can have Obj nodes for arguments.
// For other cases, we should never see a non-SIMD type here.
#ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
if (!varTypeIsSIMD(obj))
{
return;
}
#endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
// Should come here only if featureSIMD is enabled
noway_assert(comp->featureSIMD);
// We should only call this with a SIMD type.
noway_assert(varTypeIsSIMD(obj));
var_types simdType = obj->TypeGet();
// If the operand of obj is a GT_ADDR(GT_LCL_VAR) and LclVar is known to be a SIMD type,
// replace obj by GT_LCL_VAR.
GenTree* srcAddr = obj->gtGetOp1();
if (srcAddr->OperIsLocalAddr() && comp->isAddrOfSIMDType(srcAddr))
{
BlockRange().Remove(obj);
srcAddr->SetOper(loadForm(srcAddr->OperGet()));
srcAddr->gtType = simdType;
use.ReplaceWith(comp, srcAddr);
}
else
{
obj->SetOper(GT_IND);
obj->gtType = simdType;
}
#else
// we should never reach without feature SIMD
assert(!"Unexpected obj during rationalization\n");
unreached();
#endif
}
//------------------------------------------------------------------------
// DecomposeNot: Decompose GT_NOT.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeNot(LIR::Use& use)
{
assert(use.IsInitialized());
assert(use.Def()->OperGet() == GT_NOT);
GenTree* tree = use.Def();
GenTree* gtLong = tree->gtGetOp1();
noway_assert(gtLong->OperGet() == GT_LONG);
GenTree* loOp1 = gtLong->gtGetOp1();
GenTree* hiOp1 = gtLong->gtGetOp2();
Range().Remove(gtLong);
GenTree* loResult = tree;
loResult->gtType = TYP_INT;
loResult->gtOp.gtOp1 = loOp1;
GenTree* hiResult = new (m_compiler, GT_NOT) GenTreeOp(GT_NOT, TYP_INT, hiOp1, nullptr);
Range().InsertAfter(loResult, hiResult);
return FinalizeDecomposition(use, loResult, hiResult);
}
// Rewrite a SIMD indirection as GT_IND(GT_LEA(obj.op1)), or as a simple
// lclVar if possible.
//
// Arguments:
// use - A use reference for a block node
// keepBlk - True if this should remain a block node if it is not a lclVar
//
// Return Value:
// None.
//
// TODO-1stClassStructs: These should be eliminated earlier, once we can handle
// lclVars in all the places that used to have GT_OBJ.
//
void Rationalizer::RewriteSIMDOperand(LIR::Use& use, bool keepBlk)
{
#ifdef FEATURE_SIMD
// No lowering is needed for non-SIMD nodes, so early out if featureSIMD is not enabled.
if (!comp->featureSIMD)
{
return;
}
GenTree* tree = use.Def();
if (!tree->OperIsIndir())
{
return;
}
var_types simdType = tree->TypeGet();
if (!varTypeIsSIMD(simdType))
{
return;
}
// If the operand of is a GT_ADDR(GT_LCL_VAR) and LclVar is known to be of simdType,
// replace obj by GT_LCL_VAR.
GenTree* addr = tree->AsIndir()->Addr();
if (addr->OperIsLocalAddr() && comp->isAddrOfSIMDType(addr))
{
BlockRange().Remove(tree);
addr->SetOper(loadForm(addr->OperGet()));
addr->gtType = simdType;
use.ReplaceWith(comp, addr);
}
else if (!keepBlk)
{
tree->SetOper(GT_IND);
tree->gtType = simdType;
}
#endif // FEATURE_SIMD
}
//------------------------------------------------------------------------
// DecomposeCall: Decompose GT_CALL.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeCall(LIR::Use& use)
{
assert(use.IsInitialized());
assert(use.Def()->OperGet() == GT_CALL);
// We only need to force var = call() if the call's result is used.
if (use.IsDummyUse())
return use.Def()->gtNext;
GenTree* user = use.User();
if (user->OperGet() == GT_STORE_LCL_VAR)
{
// If parent is already a STORE_LCL_VAR, we can skip it if
// it is already marked as lvIsMultiRegRet.
unsigned varNum = user->AsLclVarCommon()->gtLclNum;
if (m_compiler->lvaTable[varNum].lvIsMultiRegRet)
{
return use.Def()->gtNext;
}
else if (!m_compiler->lvaTable[varNum].lvPromoted)
{
// If var wasn't promoted, we can just set lvIsMultiRegRet.
m_compiler->lvaTable[varNum].lvIsMultiRegRet = true;
return use.Def()->gtNext;
}
}
GenTree* originalNode = use.Def();
// Otherwise, we need to force var = call()
unsigned varNum = use.ReplaceWithLclVar(m_compiler, m_blockWeight);
m_compiler->lvaTable[varNum].lvIsMultiRegRet = true;
// Decompose the new LclVar use
return DecomposeLclVar(use);
}
//------------------------------------------------------------------------
// TreeNodeInfoInitIndir: Specify register requirements for address expression
// of an indirection operation.
//
// Arguments:
// indirTree - GT_IND, GT_STOREIND, block node or GT_NULLCHECK gentree node
//
void Lowering::TreeNodeInfoInitIndir(GenTreeIndir* indirTree)
{
ContainCheckIndir(indirTree);
// If this is the rhs of a block copy (i.e. non-enregisterable struct),
// it has no register requirements.
if (indirTree->TypeGet() == TYP_STRUCT)
{
return;
}
TreeNodeInfo* info = &(indirTree->gtLsraInfo);
bool isStore = (indirTree->gtOper == GT_STOREIND);
info->srcCount = GetIndirSourceCount(indirTree);
GenTree* addr = indirTree->Addr();
GenTree* index = nullptr;
unsigned cns = 0;
#ifdef _TARGET_ARM_
// Unaligned loads/stores for floating point values must first be loaded into integer register(s)
if (indirTree->gtFlags & GTF_IND_UNALIGNED)
{
var_types type = TYP_UNDEF;
if (indirTree->OperGet() == GT_STOREIND)
{
type = indirTree->AsStoreInd()->Data()->TypeGet();
}
else if (indirTree->OperGet() == GT_IND)
{
type = indirTree->TypeGet();
}
if (type == TYP_FLOAT)
{
info->internalIntCount = 1;
}
else if (type == TYP_DOUBLE)
{
info->internalIntCount = 2;
}
}
#endif
if (addr->isContained())
{
assert(addr->OperGet() == GT_LEA);
GenTreeAddrMode* lea = addr->AsAddrMode();
index = lea->Index();
cns = lea->gtOffset;
// On ARM we may need a single internal register
// (when both conditions are true then we still only need a single internal register)
if ((index != nullptr) && (cns != 0))
{
// ARM does not support both Index and offset so we need an internal register
info->internalIntCount++;
}
else if (!emitter::emitIns_valid_imm_for_ldst_offset(cns, emitTypeSize(indirTree)))
{
// This offset can't be contained in the ldr/str instruction, so we need an internal register
info->internalIntCount++;
}
}
}
//------------------------------------------------------------------------
// DecomposeStoreLclVar: Decompose GT_STORE_LCL_VAR.
//
// Arguments:
// ppTree - the tree to decompose
// data - tree walk context
//
// Return Value:
// None.
//
void DecomposeLongs::DecomposeStoreLclVar(GenTree** ppTree, Compiler::fgWalkData* data)
{
assert(ppTree != nullptr);
assert(*ppTree != nullptr);
assert(data != nullptr);
assert((*ppTree)->OperGet() == GT_STORE_LCL_VAR);
assert(m_compiler->compCurStmt != nullptr);
GenTreeStmt* curStmt = m_compiler->compCurStmt->AsStmt();
GenTree* tree = *ppTree;
GenTree* nextTree = tree->gtNext;
GenTree* rhs = tree->gtGetOp1();
if ((rhs->OperGet() == GT_PHI) || (rhs->OperGet() == GT_CALL))
{
// GT_CALLs are not decomposed, so will not be converted to GT_LONG
// GT_STORE_LCL_VAR = GT_CALL are handled in genMultiRegCallStoreToLocal
return;
}
noway_assert(rhs->OperGet() == GT_LONG);
unsigned varNum = tree->AsLclVarCommon()->gtLclNum;
LclVarDsc* varDsc = m_compiler->lvaTable + varNum;
m_compiler->lvaDecRefCnts(tree);
GenTree* loRhs = rhs->gtGetOp1();
GenTree* hiRhs = rhs->gtGetOp2();
GenTree* hiStore = m_compiler->gtNewLclLNode(varNum, TYP_INT);
if (varDsc->lvPromoted)
{
assert(varDsc->lvFieldCnt == 2);
unsigned loVarNum = varDsc->lvFieldLclStart;
unsigned hiVarNum = loVarNum + 1;
tree->AsLclVarCommon()->SetLclNum(loVarNum);
hiStore->SetOper(GT_STORE_LCL_VAR);
hiStore->AsLclVarCommon()->SetLclNum(hiVarNum);
}
else
{
noway_assert(varDsc->lvLRACandidate == false);
tree->SetOper(GT_STORE_LCL_FLD);
tree->AsLclFld()->gtLclOffs = 0;
tree->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
hiStore->SetOper(GT_STORE_LCL_FLD);
hiStore->AsLclFld()->gtLclOffs = 4;
hiStore->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
}
tree->gtOp.gtOp1 = loRhs;
tree->gtType = TYP_INT;
loRhs->gtNext = tree;
tree->gtPrev = loRhs;
hiStore->gtOp.gtOp1 = hiRhs;
hiStore->CopyCosts(tree);
hiStore->gtFlags |= GTF_VAR_DEF;
m_compiler->lvaIncRefCnts(tree);
m_compiler->lvaIncRefCnts(hiStore);
tree->gtNext = hiRhs;
hiRhs->gtPrev = tree;
hiRhs->gtNext = hiStore;
hiStore->gtPrev = hiRhs;
hiStore->gtNext = nextTree;
if (nextTree != nullptr)
{
nextTree->gtPrev = hiStore;
}
nextTree = hiRhs;
bool isEmbeddedStmt = !curStmt->gtStmtIsTopLevel();
if (!isEmbeddedStmt)
{
tree->gtNext = nullptr;
hiRhs->gtPrev = nullptr;
}
InsertNodeAsStmt(hiStore);
}
//------------------------------------------------------------------------
// DecomposeNode: Decompose long-type trees into lower and upper halves.
//
// Arguments:
// *ppTree - A node that may or may not require decomposition.
// data - The tree-walk data that provides the context.
//
// Return Value:
// None. It the tree at *ppTree is of TYP_LONG, it will generally be replaced.
//
void DecomposeLongs::DecomposeNode(GenTree** ppTree, Compiler::fgWalkData* data)
{
GenTree* tree = *ppTree;
// Handle the case where we are implicitly using the lower half of a long lclVar.
if ((tree->TypeGet() == TYP_INT) && tree->OperIsLocal())
{
LclVarDsc* varDsc = m_compiler->lvaTable + tree->AsLclVarCommon()->gtLclNum;
if (varTypeIsLong(varDsc) && varDsc->lvPromoted)
{
#ifdef DEBUG
if (m_compiler->verbose)
{
printf("Changing implicit reference to lo half of long lclVar to an explicit reference of its promoted half:\n");
m_compiler->gtDispTree(tree);
}
#endif // DEBUG
m_compiler->lvaDecRefCnts(tree);
unsigned loVarNum = varDsc->lvFieldLclStart;
tree->AsLclVarCommon()->SetLclNum(loVarNum);
m_compiler->lvaIncRefCnts(tree);
return;
}
}
if (tree->TypeGet() != TYP_LONG)
{
return;
}
#ifdef DEBUG
if (m_compiler->verbose)
{
printf("Decomposing TYP_LONG tree. BEFORE:\n");
m_compiler->gtDispTree(tree);
}
#endif // DEBUG
switch (tree->OperGet())
{
case GT_PHI:
case GT_PHI_ARG:
break;
case GT_LCL_VAR:
DecomposeLclVar(ppTree, data);
break;
case GT_LCL_FLD:
DecomposeLclFld(ppTree, data);
break;
case GT_STORE_LCL_VAR:
DecomposeStoreLclVar(ppTree, data);
break;
case GT_CAST:
DecomposeCast(ppTree, data);
break;
case GT_CNS_LNG:
DecomposeCnsLng(ppTree, data);
break;
case GT_CALL:
DecomposeCall(ppTree, data);
break;
case GT_RETURN:
assert(tree->gtOp.gtOp1->OperGet() == GT_LONG);
break;
case GT_STOREIND:
DecomposeStoreInd(ppTree, data);
break;
case GT_STORE_LCL_FLD:
assert(tree->gtOp.gtOp1->OperGet() == GT_LONG);
NYI("st.lclFld of of TYP_LONG");
break;
case GT_IND:
DecomposeInd(ppTree, data);
break;
case GT_NOT:
DecomposeNot(ppTree, data);
break;
case GT_NEG:
DecomposeNeg(ppTree, data);
break;
// Binary operators. Those that require different computation for upper and lower half are
// handled by the use of GetHiOper().
case GT_ADD:
case GT_SUB:
case GT_OR:
case GT_XOR:
case GT_AND:
DecomposeArith(ppTree, data);
break;
case GT_MUL:
NYI("Arithmetic binary operators on TYP_LONG - GT_MUL");
break;
case GT_DIV:
NYI("Arithmetic binary operators on TYP_LONG - GT_DIV");
break;
case GT_MOD:
NYI("Arithmetic binary operators on TYP_LONG - GT_MOD");
break;
case GT_UDIV:
NYI("Arithmetic binary operators on TYP_LONG - GT_UDIV");
break;
case GT_UMOD:
NYI("Arithmetic binary operators on TYP_LONG - GT_UMOD");
break;
case GT_LSH:
case GT_RSH:
case GT_RSZ:
NYI("Arithmetic binary operators on TYP_LONG - SHIFT");
break;
case GT_ROL:
case GT_ROR:
NYI("Arithmetic binary operators on TYP_LONG - ROTATE");
break;
case GT_MULHI:
NYI("Arithmetic binary operators on TYP_LONG - MULHI");
break;
case GT_LOCKADD:
case GT_XADD:
case GT_XCHG:
case GT_CMPXCHG:
NYI("Interlocked operations on TYP_LONG");
break;
default:
{
JITDUMP("Illegal TYP_LONG node %s in Decomposition.", GenTree::NodeName(tree->OperGet()));
noway_assert(!"Illegal TYP_LONG node in Decomposition.");
break;
}
}
#ifdef DEBUG
if (m_compiler->verbose)
{
printf(" AFTER:\n");
m_compiler->gtDispTree(*ppTree);
}
#endif
}
//------------------------------------------------------------------------
// DecomposeArith: Decompose GT_ADD, GT_SUB, GT_OR, GT_XOR, GT_AND.
//
// Arguments:
// ppTree - the tree to decompose
// data - tree walk context
//
// Return Value:
// None.
//
void DecomposeLongs::DecomposeArith(GenTree** ppTree, Compiler::fgWalkData* data)
{
assert(ppTree != nullptr);
assert(*ppTree != nullptr);
assert(data != nullptr);
assert(m_compiler->compCurStmt != nullptr);
GenTreeStmt* curStmt = m_compiler->compCurStmt->AsStmt();
GenTree* tree = *ppTree;
genTreeOps oper = tree->OperGet();
assert((oper == GT_ADD) ||
(oper == GT_SUB) ||
(oper == GT_OR) ||
(oper == GT_XOR) ||
(oper == GT_AND));
NYI_IF((tree->gtFlags & GTF_REVERSE_OPS) != 0, "Binary operator with GTF_REVERSE_OPS");
GenTree* op1 = tree->gtGetOp1();
GenTree* op2 = tree->gtGetOp2();
// Both operands must have already been decomposed into GT_LONG operators.
noway_assert((op1->OperGet() == GT_LONG) && (op2->OperGet() == GT_LONG));
// Capture the lo and hi halves of op1 and op2.
GenTree* loOp1 = op1->gtGetOp1();
GenTree* hiOp1 = op1->gtGetOp2();
GenTree* loOp2 = op2->gtGetOp1();
GenTree* hiOp2 = op2->gtGetOp2();
// We don't have support to decompose a TYP_LONG node that already has a child that has
// been decomposed into parts, where the high part depends on the value generated by the
// low part (via the flags register). For example, if we have:
// +(gt_long(+(lo3, lo4), +Hi(hi3, hi4)), gt_long(lo2, hi2))
// We would decompose it here to:
// gt_long(+(+(lo3, lo4), lo2), +Hi(+Hi(hi3, hi4), hi2))
// But this would generate incorrect code, because the "+Hi(hi3, hi4)" code generation
// needs to immediately follow the "+(lo3, lo4)" part. Also, if this node is one that
// requires a unique high operator, and the child nodes are not simple locals (e.g.,
// they are decomposed nodes), then we also can't decompose the node, as we aren't
// guaranteed the high and low parts will be executed immediately after each other.
NYI_IF(hiOp1->OperIsHigh() ||
hiOp2->OperIsHigh() ||
(GenTree::OperIsHigh(GetHiOper(oper)) &&
(!loOp1->OperIsLeaf() ||
!hiOp1->OperIsLeaf() ||
!loOp1->OperIsLeaf() ||
!hiOp2->OperIsLeaf())),
"Can't decompose expression tree TYP_LONG node");
// Now, remove op1 and op2 from the node list.
m_compiler->fgSnipNode(curStmt, op1);
m_compiler->fgSnipNode(curStmt, op2);
// We will reuse "tree" for the loResult, which will now be of TYP_INT, and its operands
// will be the lo halves of op1 from above.
GenTree* loResult = tree;
loResult->SetOper(GetLoOper(loResult->OperGet()));
loResult->gtType = TYP_INT;
loResult->gtOp.gtOp1 = loOp1;
loResult->gtOp.gtOp2 = loOp2;
// The various halves will be correctly threaded internally. We simply need to
// relink them into the proper order, i.e. loOp1 is followed by loOp2, and then
// the loResult node.
// (This rethreading, and that below, are where we need to address the reverse ops case).
// The current order is (after snipping op1 and op2):
// ... loOp1-> ... hiOp1->loOp2First ... loOp2->hiOp2First ... hiOp2
// The order we want is:
// ... loOp1->loOp2First ... loOp2->loResult
// ... hiOp1->hiOp2First ... hiOp2->hiResult
// i.e. we swap hiOp1 and loOp2, and create (for now) separate loResult and hiResult trees
GenTree* loOp2First = hiOp1->gtNext;
GenTree* hiOp2First = loOp2->gtNext;
// First, we will NYI if both hiOp1 and loOp2 have side effects.
NYI_IF(((loOp2->gtFlags & GTF_ALL_EFFECT) != 0) && ((hiOp1->gtFlags & GTF_ALL_EFFECT) != 0),
"Binary long operator with non-reorderable sub expressions");
// Now, we reorder the loOps and the loResult.
loOp1->gtNext = loOp2First;
loOp2First->gtPrev = loOp1;
loOp2->gtNext = loResult;
loResult->gtPrev = loOp2;
// Next, reorder the hiOps and the hiResult.
GenTree* hiResult = new (m_compiler, oper) GenTreeOp(GetHiOper(oper), TYP_INT, hiOp1, hiOp2);
hiOp1->gtNext = hiOp2First;
hiOp2First->gtPrev = hiOp1;
hiOp2->gtNext = hiResult;
hiResult->gtPrev = hiOp2;
if ((oper == GT_ADD) || (oper == GT_SUB))
{
if (loResult->gtOverflow())
{
hiResult->gtFlags |= GTF_OVERFLOW;
loResult->gtFlags &= ~GTF_OVERFLOW;
}
if (loResult->gtFlags & GTF_UNSIGNED)
{
hiResult->gtFlags |= GTF_UNSIGNED;
}
}
FinalizeDecomposition(ppTree, data, loResult, hiResult);
}
//------------------------------------------------------------------------
// DecomposeStoreLclVar: Decompose GT_STORE_LCL_VAR.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeStoreLclVar(LIR::Use& use)
{
assert(use.IsInitialized());
assert(use.Def()->OperGet() == GT_STORE_LCL_VAR);
GenTree* tree = use.Def();
GenTree* rhs = tree->gtGetOp1();
if ((rhs->OperGet() == GT_PHI) || (rhs->OperGet() == GT_CALL) ||
((rhs->OperGet() == GT_MUL_LONG) && (rhs->gtFlags & GTF_MUL_64RSLT) != 0))
{
// GT_CALLs are not decomposed, so will not be converted to GT_LONG
// GT_STORE_LCL_VAR = GT_CALL are handled in genMultiRegCallStoreToLocal
// GT_MULs are not decomposed, so will not be converted to GT_LONG
return tree->gtNext;
}
noway_assert(rhs->OperGet() == GT_LONG);
unsigned varNum = tree->AsLclVarCommon()->gtLclNum;
LclVarDsc* varDsc = m_compiler->lvaTable + varNum;
m_compiler->lvaDecRefCnts(tree);
GenTree* loRhs = rhs->gtGetOp1();
GenTree* hiRhs = rhs->gtGetOp2();
GenTree* hiStore = m_compiler->gtNewLclLNode(varNum, TYP_INT);
if (varDsc->lvPromoted)
{
assert(varDsc->lvFieldCnt == 2);
unsigned loVarNum = varDsc->lvFieldLclStart;
unsigned hiVarNum = loVarNum + 1;
tree->AsLclVarCommon()->SetLclNum(loVarNum);
hiStore->SetOper(GT_STORE_LCL_VAR);
hiStore->AsLclVarCommon()->SetLclNum(hiVarNum);
}
else
{
noway_assert(varDsc->lvLRACandidate == false);
tree->SetOper(GT_STORE_LCL_FLD);
tree->AsLclFld()->gtLclOffs = 0;
tree->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
hiStore->SetOper(GT_STORE_LCL_FLD);
hiStore->AsLclFld()->gtLclOffs = 4;
hiStore->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
}
// 'tree' is going to steal the loRhs node for itself, so we need to remove the
// GT_LONG node from the threading.
Range().Remove(rhs);
tree->gtOp.gtOp1 = loRhs;
tree->gtType = TYP_INT;
hiStore->gtOp.gtOp1 = hiRhs;
hiStore->gtFlags |= GTF_VAR_DEF;
m_compiler->lvaIncRefCnts(tree);
m_compiler->lvaIncRefCnts(hiStore);
Range().InsertAfter(tree, hiStore);
return hiStore->gtNext;
}
//------------------------------------------------------------------------
// DecomposeCast: Decompose GT_CAST.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeCast(LIR::Use& use)
{
assert(use.IsInitialized());
assert(use.Def()->OperGet() == GT_CAST);
GenTree* cast = use.Def()->AsCast();
GenTree* loResult = nullptr;
GenTree* hiResult = nullptr;
var_types srcType = cast->CastFromType();
var_types dstType = cast->CastToType();
if ((cast->gtFlags & GTF_UNSIGNED) != 0)
{
srcType = genUnsignedType(srcType);
}
if (varTypeIsLong(srcType))
{
if (cast->gtOverflow() && (varTypeIsUnsigned(srcType) != varTypeIsUnsigned(dstType)))
{
GenTree* srcOp = cast->gtGetOp1();
noway_assert(srcOp->OperGet() == GT_LONG);
GenTree* loSrcOp = srcOp->gtGetOp1();
GenTree* hiSrcOp = srcOp->gtGetOp2();
//
// When casting between long types an overflow check is needed only if the types
// have different signedness. In both cases (long->ulong and ulong->long) we only
// need to check if the high part is negative or not. Use the existing cast node
// to perform a int->uint cast of the high part to take advantage of the overflow
// check provided by codegen.
//
loResult = loSrcOp;
hiResult = cast;
hiResult->gtType = TYP_INT;
hiResult->AsCast()->gtCastType = TYP_UINT;
hiResult->gtFlags &= ~GTF_UNSIGNED;
hiResult->gtOp.gtOp1 = hiSrcOp;
Range().Remove(cast);
Range().Remove(srcOp);
Range().InsertAfter(hiSrcOp, hiResult);
}
else
{
NYI("Unimplemented long->long no-op cast decomposition");
}
}
else if (varTypeIsIntegralOrI(srcType))
{
if (cast->gtOverflow() && !varTypeIsUnsigned(srcType) && varTypeIsUnsigned(dstType))
{
//
// An overflow check is needed only when casting from a signed type to ulong.
// Change the cast type to uint to take advantage of the overflow check provided
// by codegen and then zero extend the resulting uint to ulong.
//
loResult = cast;
loResult->AsCast()->gtCastType = TYP_UINT;
loResult->gtType = TYP_INT;
hiResult = m_compiler->gtNewZeroConNode(TYP_INT);
Range().InsertAfter(loResult, hiResult);
}
else
{
if (varTypeIsUnsigned(srcType))
{
loResult = cast->gtGetOp1();
hiResult = m_compiler->gtNewZeroConNode(TYP_INT);
Range().Remove(cast);
Range().InsertAfter(loResult, hiResult);
}
else
{
LIR::Use src(Range(), &(cast->gtOp.gtOp1), cast);
unsigned lclNum = src.ReplaceWithLclVar(m_compiler, m_blockWeight);
loResult = src.Def();
GenTree* loCopy = m_compiler->gtNewLclvNode(lclNum, TYP_INT);
GenTree* shiftBy = m_compiler->gtNewIconNode(31, TYP_INT);
hiResult = m_compiler->gtNewOperNode(GT_RSH, TYP_INT, loCopy, shiftBy);
Range().Remove(cast);
Range().InsertAfter(loResult, loCopy, shiftBy, hiResult);
m_compiler->lvaIncRefCnts(loCopy);
}
}
}
else
{
NYI("Unimplemented cast decomposition");
}
return FinalizeDecomposition(use, loResult, hiResult);
}
//------------------------------------------------------------------------
// DecomposeShift: Decompose GT_LSH, GT_RSH, GT_RSZ. For shift nodes, we need to use
// the shift helper functions, so we here convert the shift into a helper call by
// pulling its arguments out of linear order and making them the args to a call, then
// replacing the original node with the new call.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeShift(LIR::Use& use)
{
assert(use.IsInitialized());
GenTree* tree = use.Def();
GenTree* gtLong = tree->gtGetOp1();
genTreeOps oper = tree->OperGet();
assert((oper == GT_LSH) || (oper == GT_RSH) || (oper == GT_RSZ));
LIR::Use loOp1Use(Range(), >Long->gtOp.gtOp1, gtLong);
loOp1Use.ReplaceWithLclVar(m_compiler, m_blockWeight);
LIR::Use hiOp1Use(Range(), >Long->gtOp.gtOp2, gtLong);
hiOp1Use.ReplaceWithLclVar(m_compiler, m_blockWeight);
LIR::Use shiftWidthUse(Range(), &tree->gtOp.gtOp2, tree);
shiftWidthUse.ReplaceWithLclVar(m_compiler, m_blockWeight);
GenTree* loOp1 = gtLong->gtGetOp1();
GenTree* hiOp1 = gtLong->gtGetOp2();
GenTree* shiftWidthOp = tree->gtGetOp2();
Range().Remove(gtLong);
Range().Remove(loOp1);
Range().Remove(hiOp1);
Range().Remove(shiftWidthOp);
// TODO-X86-CQ: If the shift operand is a GT_CNS_INT, we should pipe the instructions through to codegen
// and generate the shift instructions ourselves there, rather than replacing it with a helper call.
unsigned helper;
switch (oper)
{
case GT_LSH:
helper = CORINFO_HELP_LLSH;
break;
case GT_RSH:
helper = CORINFO_HELP_LRSH;
break;
case GT_RSZ:
helper = CORINFO_HELP_LRSZ;
break;
default:
unreached();
}
GenTreeArgList* argList = m_compiler->gtNewArgList(loOp1, hiOp1, shiftWidthOp);
GenTree* call = m_compiler->gtNewHelperCallNode(helper, TYP_LONG, 0, argList);
GenTreeCall* callNode = call->AsCall();
ReturnTypeDesc* retTypeDesc = callNode->GetReturnTypeDesc();
retTypeDesc->InitializeLongReturnType(m_compiler);
call = m_compiler->fgMorphArgs(callNode);
Range().InsertAfter(tree, LIR::SeqTree(m_compiler, call));
Range().Remove(tree);
use.ReplaceWith(m_compiler, call);
return call;
}
void Rationalizer::DoPhase()
{
DBEXEC(TRUE, SanityCheck());
comp->compCurBB = nullptr;
comp->fgOrder = Compiler::FGOrderLinear;
BasicBlock* firstBlock = comp->fgFirstBB;
for (BasicBlock* block = comp->fgFirstBB; block != nullptr; block = block->bbNext)
{
comp->compCurBB = block;
m_block = block;
// Establish the first and last nodes for the block. This is necessary in order for the LIR
// utilities that hang off the BasicBlock type to work correctly.
GenTreeStmt* firstStatement = block->firstStmt();
if (firstStatement == nullptr)
{
// No statements in this block; skip it.
block->MakeLIR(nullptr, nullptr);
continue;
}
GenTreeStmt* lastStatement = block->lastStmt();
// Rewrite intrinsics that are not supported by the target back into user calls.
// This needs to be done before the transition to LIR because it relies on the use
// of fgMorphArgs, which is designed to operate on HIR. Once this is done for a
// particular statement, link that statement's nodes into the current basic block.
//
// This walk also clears the GTF_VAR_USEDEF bit on locals, which is not necessary
// in the backend.
GenTree* lastNodeInPreviousStatement = nullptr;
for (GenTreeStmt* statement = firstStatement; statement != nullptr; statement = statement->getNextStmt())
{
assert(statement->gtStmtList != nullptr);
assert(statement->gtStmtList->gtPrev == nullptr);
assert(statement->gtStmtExpr != nullptr);
assert(statement->gtStmtExpr->gtNext == nullptr);
SplitData splitData;
splitData.root = statement;
splitData.block = block;
splitData.thisPhase = this;
comp->fgWalkTreePost(&statement->gtStmtExpr,
[](GenTree** use, Compiler::fgWalkData* walkData) -> Compiler::fgWalkResult {
GenTree* node = *use;
if (node->OperGet() == GT_INTRINSIC &&
Compiler::IsIntrinsicImplementedByUserCall(node->gtIntrinsic.gtIntrinsicId))
{
RewriteIntrinsicAsUserCall(use, walkData);
}
else if (node->OperIsLocal())
{
node->gtFlags &= ~GTF_VAR_USEDEF;
}
return Compiler::WALK_CONTINUE;
},
&splitData, true);
GenTree* firstNodeInStatement = statement->gtStmtList;
if (lastNodeInPreviousStatement != nullptr)
{
lastNodeInPreviousStatement->gtNext = firstNodeInStatement;
}
firstNodeInStatement->gtPrev = lastNodeInPreviousStatement;
lastNodeInPreviousStatement = statement->gtStmtExpr;
}
block->MakeLIR(firstStatement->gtStmtList, lastStatement->gtStmtExpr);
// Rewrite HIR nodes into LIR nodes.
for (GenTreeStmt *statement = firstStatement, *nextStatement; statement != nullptr; statement = nextStatement)
{
nextStatement = statement->getNextStmt();
// If this statement has correct offset information, change it into an IL offset
// node and insert it into the LIR.
if (statement->gtStmtILoffsx != BAD_IL_OFFSET)
{
assert(!statement->IsPhiDefnStmt());
statement->SetOper(GT_IL_OFFSET);
statement->gtNext = nullptr;
statement->gtPrev = nullptr;
BlockRange().InsertBefore(statement->gtStmtList, statement);
}
m_statement = statement;
comp->fgWalkTreePost(&statement->gtStmtExpr,
[](GenTree** use, Compiler::fgWalkData* walkData) -> Compiler::fgWalkResult {
return reinterpret_cast<Rationalizer*>(walkData->pCallbackData)
->RewriteNode(use, *walkData->parentStack);
},
this, true);
}
assert(BlockRange().CheckLIR(comp));
}
comp->compRationalIRForm = true;
}
void Rationalizer::RewriteAssignment(LIR::Use& use)
{
assert(use.IsInitialized());
GenTreeOp* assignment = use.Def()->AsOp();
assert(assignment->OperGet() == GT_ASG);
GenTree* location = assignment->gtGetOp1();
GenTree* value = assignment->gtGetOp2();
genTreeOps locationOp = location->OperGet();
#ifdef FEATURE_SIMD
if (varTypeIsSIMD(location) && assignment->OperIsInitBlkOp())
{
if (location->OperGet() == GT_LCL_VAR)
{
var_types simdType = location->TypeGet();
GenTree* initVal = assignment->gtOp.gtOp2;
var_types baseType = comp->getBaseTypeOfSIMDLocal(location);
if (baseType != TYP_UNKNOWN)
{
GenTreeSIMD* simdTree = new (comp, GT_SIMD)
GenTreeSIMD(simdType, initVal, SIMDIntrinsicInit, baseType, genTypeSize(simdType));
assignment->gtOp.gtOp2 = simdTree;
value = simdTree;
initVal->gtNext = simdTree;
simdTree->gtPrev = initVal;
simdTree->gtNext = location;
location->gtPrev = simdTree;
}
}
else
{
assert(location->OperIsBlk());
}
}
#endif // FEATURE_SIMD
switch (locationOp)
{
case GT_LCL_VAR:
case GT_LCL_FLD:
case GT_REG_VAR:
case GT_PHI_ARG:
RewriteAssignmentIntoStoreLclCore(assignment, location, value, locationOp);
BlockRange().Remove(location);
break;
case GT_IND:
{
GenTreeStoreInd* store =
new (comp, GT_STOREIND) GenTreeStoreInd(location->TypeGet(), location->gtGetOp1(), value);
copyFlags(store, assignment, GTF_ALL_EFFECT);
copyFlags(store, location, GTF_IND_FLAGS);
if (assignment->IsReverseOp())
{
store->gtFlags |= GTF_REVERSE_OPS;
}
// TODO: JIT dump
// Remove the GT_IND node and replace the assignment node with the store
BlockRange().Remove(location);
BlockRange().InsertBefore(assignment, store);
use.ReplaceWith(comp, store);
BlockRange().Remove(assignment);
}
break;
case GT_CLS_VAR:
{
location->SetOper(GT_CLS_VAR_ADDR);
location->gtType = TYP_BYREF;
assignment->SetOper(GT_STOREIND);
// TODO: JIT dump
}
break;
case GT_BLK:
case GT_OBJ:
case GT_DYN_BLK:
{
assert(varTypeIsStruct(location));
GenTreeBlk* storeBlk = location->AsBlk();
genTreeOps storeOper;
switch (location->gtOper)
{
case GT_BLK:
storeOper = GT_STORE_BLK;
break;
case GT_OBJ:
storeOper = GT_STORE_OBJ;
break;
case GT_DYN_BLK:
storeOper = GT_STORE_DYN_BLK;
break;
default:
unreached();
}
JITDUMP("Rewriting GT_ASG(%s(X), Y) to %s(X,Y):\n", GenTree::NodeName(location->gtOper),
GenTree::NodeName(storeOper));
storeBlk->SetOperRaw(storeOper);
storeBlk->gtFlags &= ~GTF_DONT_CSE;
storeBlk->gtFlags |= (assignment->gtFlags & (GTF_ALL_EFFECT | GTF_REVERSE_OPS | GTF_BLK_VOLATILE |
GTF_BLK_UNALIGNED | GTF_BLK_INIT | GTF_DONT_CSE));
storeBlk->gtBlk.Data() = value;
// Replace the assignment node with the store
use.ReplaceWith(comp, storeBlk);
BlockRange().Remove(assignment);
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
break;
default:
unreached();
break;
}
}
Compiler::fgWalkResult Rationalizer::RewriteNode(GenTree** useEdge, ArrayStack<GenTree*>& parentStack)
{
assert(useEdge != nullptr);
GenTree* node = *useEdge;
assert(node != nullptr);
#ifdef DEBUG
const bool isLateArg = (node->gtFlags & GTF_LATE_ARG) != 0;
#endif
// First, remove any preceeding list nodes, which are not otherwise visited by the tree walk.
//
// NOTE: GT_FIELD_LIST head nodes, and GT_LIST nodes used by phi nodes will in fact be visited.
for (GenTree* prev = node->gtPrev; prev != nullptr && prev->OperIsAnyList() && !(prev->OperIsFieldListHead());
prev = node->gtPrev)
{
BlockRange().Remove(prev);
}
// In addition, remove the current node if it is a GT_LIST node that is not an aggregate.
if (node->OperIsAnyList())
{
GenTreeArgList* list = node->AsArgList();
if (!list->OperIsFieldListHead())
{
BlockRange().Remove(list);
}
return Compiler::WALK_CONTINUE;
}
LIR::Use use;
if (parentStack.Height() < 2)
{
use = LIR::Use::GetDummyUse(BlockRange(), *useEdge);
}
else
{
use = LIR::Use(BlockRange(), useEdge, parentStack.Index(1));
}
assert(node == use.Def());
switch (node->OperGet())
{
case GT_ASG:
RewriteAssignment(use);
break;
case GT_BOX:
// GT_BOX at this level just passes through so get rid of it
use.ReplaceWith(comp, node->gtGetOp1());
BlockRange().Remove(node);
break;
case GT_ADDR:
RewriteAddress(use);
break;
case GT_IND:
// Clear the `GTF_IND_ASG_LHS` flag, which overlaps with `GTF_IND_REQ_ADDR_IN_REG`.
node->gtFlags &= ~GTF_IND_ASG_LHS;
if (varTypeIsSIMD(node))
{
RewriteSIMDOperand(use, false);
}
else
{
// Due to promotion of structs containing fields of type struct with a
// single scalar type field, we could potentially see IR nodes of the
// form GT_IND(GT_ADD(lclvarAddr, 0)) where 0 is an offset representing
// a field-seq. These get folded here.
//
// TODO: This code can be removed once JIT implements recursive struct
// promotion instead of lying about the type of struct field as the type
// of its single scalar field.
GenTree* addr = node->AsIndir()->Addr();
if (addr->OperGet() == GT_ADD && addr->gtGetOp1()->OperGet() == GT_LCL_VAR_ADDR &&
addr->gtGetOp2()->IsIntegralConst(0))
{
GenTreeLclVarCommon* lclVarNode = addr->gtGetOp1()->AsLclVarCommon();
unsigned lclNum = lclVarNode->GetLclNum();
LclVarDsc* varDsc = comp->lvaTable + lclNum;
if (node->TypeGet() == varDsc->TypeGet())
{
JITDUMP("Rewriting GT_IND(GT_ADD(LCL_VAR_ADDR,0)) to LCL_VAR\n");
lclVarNode->SetOper(GT_LCL_VAR);
lclVarNode->gtType = node->TypeGet();
use.ReplaceWith(comp, lclVarNode);
BlockRange().Remove(addr);
BlockRange().Remove(addr->gtGetOp2());
BlockRange().Remove(node);
}
}
}
break;
case GT_NOP:
// fgMorph sometimes inserts NOP nodes between defs and uses
// supposedly 'to prevent constant folding'. In this case, remove the
// NOP.
if (node->gtGetOp1() != nullptr)
{
use.ReplaceWith(comp, node->gtGetOp1());
BlockRange().Remove(node);
}
break;
case GT_COMMA:
{
GenTree* op1 = node->gtGetOp1();
if ((op1->gtFlags & GTF_ALL_EFFECT) == 0)
{
// The LHS has no side effects. Remove it.
bool isClosed = false;
unsigned sideEffects = 0;
LIR::ReadOnlyRange lhsRange = BlockRange().GetTreeRange(op1, &isClosed, &sideEffects);
// None of the transforms performed herein violate tree order, so these
// should always be true.
assert(isClosed);
assert((sideEffects & GTF_ALL_EFFECT) == 0);
BlockRange().Delete(comp, m_block, std::move(lhsRange));
}
GenTree* replacement = node->gtGetOp2();
if (!use.IsDummyUse())
{
use.ReplaceWith(comp, replacement);
}
else
{
// This is a top-level comma. If the RHS has no side effects we can remove
// it as well.
if ((replacement->gtFlags & GTF_ALL_EFFECT) == 0)
{
bool isClosed = false;
unsigned sideEffects = 0;
LIR::ReadOnlyRange rhsRange = BlockRange().GetTreeRange(replacement, &isClosed, &sideEffects);
// None of the transforms performed herein violate tree order, so these
// should always be true.
assert(isClosed);
assert((sideEffects & GTF_ALL_EFFECT) == 0);
BlockRange().Delete(comp, m_block, std::move(rhsRange));
}
}
BlockRange().Remove(node);
}
break;
case GT_ARGPLACE:
// Remove argplace and list nodes from the execution order.
//
// TODO: remove phi args and phi nodes as well?
BlockRange().Remove(node);
break;
#if defined(_TARGET_XARCH_) || defined(_TARGET_ARM_)
case GT_CLS_VAR:
{
// Class vars that are the target of an assignment will get rewritten into
// GT_STOREIND(GT_CLS_VAR_ADDR, val) by RewriteAssignment. This check is
// not strictly necessary--the GT_IND(GT_CLS_VAR_ADDR) pattern that would
// otherwise be generated would also be picked up by RewriteAssignment--but
// skipping the rewrite here saves an allocation and a bit of extra work.
const bool isLHSOfAssignment = (use.User()->OperGet() == GT_ASG) && (use.User()->gtGetOp1() == node);
if (!isLHSOfAssignment)
{
GenTree* ind = comp->gtNewOperNode(GT_IND, node->TypeGet(), node);
node->SetOper(GT_CLS_VAR_ADDR);
node->gtType = TYP_BYREF;
BlockRange().InsertAfter(node, ind);
use.ReplaceWith(comp, ind);
// TODO: JIT dump
}
}
break;
#endif // _TARGET_XARCH_
case GT_INTRINSIC:
// Non-target intrinsics should have already been rewritten back into user calls.
assert(Compiler::IsTargetIntrinsic(node->gtIntrinsic.gtIntrinsicId));
break;
#ifdef FEATURE_SIMD
case GT_BLK:
case GT_OBJ:
{
// TODO-1stClassStructs: These should have been transformed to GT_INDs, but in order
// to preserve existing behavior, we will keep this as a block node if this is the
// lhs of a block assignment, and either:
// - It is a "generic" TYP_STRUCT assignment, OR
// - It is an initblk, OR
// - Neither the lhs or rhs are known to be of SIMD type.
GenTree* parent = use.User();
bool keepBlk = false;
if ((parent->OperGet() == GT_ASG) && (node == parent->gtGetOp1()))
{
if ((node->TypeGet() == TYP_STRUCT) || parent->OperIsInitBlkOp())
{
keepBlk = true;
}
else if (!comp->isAddrOfSIMDType(node->AsBlk()->Addr()))
{
GenTree* dataSrc = parent->gtGetOp2();
if (!dataSrc->IsLocal() && (dataSrc->OperGet() != GT_SIMD))
{
noway_assert(dataSrc->OperIsIndir());
keepBlk = !comp->isAddrOfSIMDType(dataSrc->AsIndir()->Addr());
}
}
}
RewriteSIMDOperand(use, keepBlk);
}
break;
case GT_LCL_FLD:
case GT_STORE_LCL_FLD:
// TODO-1stClassStructs: Eliminate this.
FixupIfSIMDLocal(node->AsLclVarCommon());
break;
case GT_SIMD:
{
noway_assert(comp->featureSIMD);
GenTreeSIMD* simdNode = node->AsSIMD();
unsigned simdSize = simdNode->gtSIMDSize;
var_types simdType = comp->getSIMDTypeForSize(simdSize);
// TODO-1stClassStructs: This should be handled more generally for enregistered or promoted
// structs that are passed or returned in a different register type than their enregistered
// type(s).
if (simdNode->gtType == TYP_I_IMPL && simdNode->gtSIMDSize == TARGET_POINTER_SIZE)
{
// This happens when it is consumed by a GT_RET_EXPR.
// It can only be a Vector2f or Vector2i.
assert(genTypeSize(simdNode->gtSIMDBaseType) == 4);
simdNode->gtType = TYP_SIMD8;
}
// Certain SIMD trees require rationalizing.
if (simdNode->gtSIMD.gtSIMDIntrinsicID == SIMDIntrinsicInitArray)
{
// Rewrite this as an explicit load.
JITDUMP("Rewriting GT_SIMD array init as an explicit load:\n");
unsigned int baseTypeSize = genTypeSize(simdNode->gtSIMDBaseType);
GenTree* address = new (comp, GT_LEA) GenTreeAddrMode(TYP_BYREF, simdNode->gtOp1, simdNode->gtOp2,
baseTypeSize, offsetof(CORINFO_Array, u1Elems));
GenTree* ind = comp->gtNewOperNode(GT_IND, simdType, address);
BlockRange().InsertBefore(simdNode, address, ind);
use.ReplaceWith(comp, ind);
BlockRange().Remove(simdNode);
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
else
{
// This code depends on the fact that NONE of the SIMD intrinsics take vector operands
// of a different width. If that assumption changes, we will EITHER have to make these type
// transformations during importation, and plumb the types all the way through the JIT,
// OR add a lot of special handling here.
GenTree* op1 = simdNode->gtGetOp1();
if (op1 != nullptr && op1->gtType == TYP_STRUCT)
{
op1->gtType = simdType;
}
GenTree* op2 = simdNode->gtGetOp2IfPresent();
if (op2 != nullptr && op2->gtType == TYP_STRUCT)
{
op2->gtType = simdType;
}
}
}
break;
#endif // FEATURE_SIMD
default:
// JCC nodes should not be present in HIR.
assert(node->OperGet() != GT_JCC);
break;
}
// Do some extra processing on top-level nodes to remove unused local reads.
if (node->OperIsLocalRead())
{
if (use.IsDummyUse())
{
comp->lvaDecRefCnts(node);
BlockRange().Remove(node);
}
else
{
// Local reads are side-effect-free; clear any flags leftover from frontend transformations.
node->gtFlags &= ~GTF_ALL_EFFECT;
}
}
assert(isLateArg == ((use.Def()->gtFlags & GTF_LATE_ARG) != 0));
return Compiler::WALK_CONTINUE;
}
void Rationalizer::RewriteAssignment(LIR::Use& use)
{
assert(use.IsInitialized());
GenTreeOp* assignment = use.Def()->AsOp();
assert(assignment->OperGet() == GT_ASG);
GenTree* location = assignment->gtGetOp1();
GenTree* value = assignment->gtGetOp2();
genTreeOps locationOp = location->OperGet();
if (assignment->OperIsBlkOp())
{
#ifdef FEATURE_SIMD
if (varTypeIsSIMD(location) && assignment->OperIsInitBlkOp())
{
if (location->OperGet() == GT_LCL_VAR)
{
var_types simdType = location->TypeGet();
GenTree* initVal = assignment->gtOp.gtOp2;
var_types baseType = comp->getBaseTypeOfSIMDLocal(location);
if (baseType != TYP_UNKNOWN)
{
GenTreeSIMD* simdTree = new (comp, GT_SIMD)
GenTreeSIMD(simdType, initVal, SIMDIntrinsicInit, baseType, genTypeSize(simdType));
assignment->gtOp.gtOp2 = simdTree;
value = simdTree;
initVal->gtNext = simdTree;
simdTree->gtPrev = initVal;
simdTree->gtNext = location;
location->gtPrev = simdTree;
}
}
}
#endif // FEATURE_SIMD
if ((location->TypeGet() == TYP_STRUCT) && !assignment->IsPhiDefn() && !value->IsMultiRegCall())
{
if ((location->OperGet() == GT_LCL_VAR))
{
// We need to construct a block node for the location.
// Modify lcl to be the address form.
location->SetOper(addrForm(locationOp));
LclVarDsc* varDsc = &(comp->lvaTable[location->AsLclVarCommon()->gtLclNum]);
location->gtType = TYP_BYREF;
GenTreeBlk* storeBlk = nullptr;
unsigned int size = varDsc->lvExactSize;
if (varDsc->lvStructGcCount != 0)
{
CORINFO_CLASS_HANDLE structHnd = varDsc->lvVerTypeInfo.GetClassHandle();
GenTreeObj* objNode = comp->gtNewObjNode(structHnd, location)->AsObj();
unsigned int slots = (unsigned)(roundUp(size, TARGET_POINTER_SIZE) / TARGET_POINTER_SIZE);
objNode->SetGCInfo(varDsc->lvGcLayout, varDsc->lvStructGcCount, slots);
objNode->ChangeOper(GT_STORE_OBJ);
objNode->SetData(value);
comp->fgMorphUnsafeBlk(objNode);
storeBlk = objNode;
}
else
{
storeBlk = new (comp, GT_STORE_BLK) GenTreeBlk(GT_STORE_BLK, TYP_STRUCT, location, value, size);
}
storeBlk->gtFlags |= (GTF_REVERSE_OPS | GTF_ASG);
storeBlk->gtFlags |= ((location->gtFlags | value->gtFlags) & GTF_ALL_EFFECT);
GenTree* insertionPoint = location->gtNext;
BlockRange().InsertBefore(insertionPoint, storeBlk);
use.ReplaceWith(comp, storeBlk);
BlockRange().Remove(assignment);
JITDUMP("After transforming local struct assignment into a block op:\n");
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
return;
}
else
{
assert(location->OperIsBlk());
}
}
}
switch (locationOp)
{
case GT_LCL_VAR:
case GT_LCL_FLD:
case GT_REG_VAR:
case GT_PHI_ARG:
RewriteAssignmentIntoStoreLclCore(assignment, location, value, locationOp);
BlockRange().Remove(location);
break;
case GT_IND:
{
GenTreeStoreInd* store =
new (comp, GT_STOREIND) GenTreeStoreInd(location->TypeGet(), location->gtGetOp1(), value);
copyFlags(store, assignment, GTF_ALL_EFFECT);
copyFlags(store, location, GTF_IND_FLAGS);
if (assignment->IsReverseOp())
{
store->gtFlags |= GTF_REVERSE_OPS;
}
// TODO: JIT dump
// Remove the GT_IND node and replace the assignment node with the store
BlockRange().Remove(location);
BlockRange().InsertBefore(assignment, store);
use.ReplaceWith(comp, store);
BlockRange().Remove(assignment);
}
break;
case GT_CLS_VAR:
{
location->SetOper(GT_CLS_VAR_ADDR);
location->gtType = TYP_BYREF;
assignment->SetOper(GT_STOREIND);
// TODO: JIT dump
}
break;
case GT_BLK:
case GT_OBJ:
case GT_DYN_BLK:
{
assert(varTypeIsStruct(location));
GenTreeBlk* storeBlk = location->AsBlk();
genTreeOps storeOper;
switch (location->gtOper)
{
case GT_BLK:
storeOper = GT_STORE_BLK;
break;
case GT_OBJ:
storeOper = GT_STORE_OBJ;
break;
case GT_DYN_BLK:
storeOper = GT_STORE_DYN_BLK;
break;
default:
unreached();
}
JITDUMP("Rewriting GT_ASG(%s(X), Y) to %s(X,Y):\n", GenTree::NodeName(location->gtOper),
GenTree::NodeName(storeOper));
storeBlk->SetOperRaw(storeOper);
storeBlk->gtFlags &= ~GTF_DONT_CSE;
storeBlk->gtFlags |= (assignment->gtFlags & (GTF_ALL_EFFECT | GTF_REVERSE_OPS | GTF_BLK_VOLATILE |
GTF_BLK_UNALIGNED | GTF_DONT_CSE));
storeBlk->gtBlk.Data() = value;
// Replace the assignment node with the store
use.ReplaceWith(comp, storeBlk);
BlockRange().Remove(assignment);
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
break;
default:
unreached();
break;
}
}