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;
}
//------------------------------------------------------------------------
// 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
}
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 GT_LIST nodes, which are not otherwise visited by the tree walk.
//
// NOTE: GT_LIST nodes that are used by block ops and phi nodes will in fact be visited.
for (GenTree* prev = node->gtPrev; prev != nullptr && prev->OperGet() == GT_LIST; prev = node->gtPrev)
{
BlockRange().Remove(prev);
}
// In addition, remove the current node if it is a GT_LIST node.
if ((*useEdge)->OperGet() == GT_LIST)
{
BlockRange().Remove(*useEdge);
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_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;
#ifdef _TARGET_XARCH_
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_INITBLK:
RewriteInitBlk(use);
break;
case GT_COPYBLK:
RewriteCopyBlk(use);
break;
case GT_OBJ:
RewriteObj(use);
break;
case GT_LCL_FLD:
case GT_STORE_LCL_FLD:
// TODO-1stClassStructs: Eliminate this.
FixupIfSIMDLocal(node->AsLclVarCommon());
break;
case GT_STOREIND:
case GT_IND:
if (node->gtType == TYP_STRUCT)
{
GenTree* addr = node->AsIndir()->Addr();
assert(addr->TypeGet() == TYP_BYREF);
if (addr->OperIsLocal())
{
LclVarDsc* varDsc = &(comp->lvaTable[addr->AsLclVarCommon()->gtLclNum]);
assert(varDsc->lvSIMDType);
unsigned simdSize = (unsigned int)roundUp(varDsc->lvExactSize, TARGET_POINTER_SIZE);
node->gtType = comp->getSIMDTypeForSize(simdSize);
}
#if DEBUG
else
{
// If the address is not a local var, assert that the user of this IND is an ADDR node.
assert((use.User()->OperGet() == GT_ADDR) || use.User()->OperIsLocalAddr());
}
#endif
}
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;
}
else if (simdNode->gtType == TYP_STRUCT || varTypeIsSIMD(simdNode))
{
node->gtType = simdType;
}
// 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->gtGetOp2();
if (op2 != nullptr && op2->gtType == TYP_STRUCT)
{
op2->gtType = simdType;
}
}
}
break;
#endif // FEATURE_SIMD
default:
break;
}
// Do some extra processing on top-level nodes to remove unused local reads.
if (use.IsDummyUse() && node->OperIsLocalRead())
{
assert((node->gtFlags & GTF_ALL_EFFECT) == 0);
comp->lvaDecRefCnts(node);
BlockRange().Remove(node);
}
assert(isLateArg == ((node->gtFlags & GTF_LATE_ARG) != 0));
return Compiler::WALK_CONTINUE;
}