Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

JIT: Fix LowerHWIntrinsicCreate LIR ordering issues #93210

Merged
merged 4 commits into from
Oct 9, 2023
Merged

JIT: Fix LowerHWIntrinsicCreate LIR ordering issues #93210

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
6e40556
JIT: Fix LowerHWIntrinsicCreate LIR ordering issues
jakobbotsch Oct 9, 2023
e05ecd3
Optimize
jakobbotsch Oct 9, 2023
35eb0d4
Revert unnecessary changes
jakobbotsch Oct 9, 2023
866393d
Revert a change, fix accidental removal
jakobbotsch Oct 9, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
64 changes: 64 additions & 0 deletions src/coreclr/jit/lir.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1794,6 +1794,70 @@ void LIR::InsertBeforeTerminator(BasicBlock* block, LIR::Range&& range)
blockRange.InsertBefore(insertionPoint, std::move(range));
}

//------------------------------------------------------------------------
// LIR::LastNode:
// Given two nodes in the same block range, find which node appears last.
//
// Arguments:
// node1 - The first node
// node2 - The second node
//
// Returns:
// Node that appears last.
//
GenTree* LIR::LastNode(GenTree* node1, GenTree* node2)
{
assert(node1 != nullptr);
assert(node2 != nullptr);

if (node1 == node2)
{
return node1;
}

GenTree* cursor1 = node1->gtNext;
GenTree* cursor2 = node2->gtNext;

while (true)
{
if ((cursor1 == node2) || (cursor2 == nullptr))
{
return node2;
}

if ((cursor2 == node1) || (cursor1 == nullptr))
{
return node1;
}

cursor1 = cursor1->gtNext;
cursor2 = cursor2->gtNext;
}
}

//------------------------------------------------------------------------
// LIR::LastNode:
// Given an array of nodes in a block range, find the last one.
//
// Arguments:
// nodes - Pointer to nodes
// numNodes - Number of nodes
//
// Returns:
// Node that appears last.
//
GenTree* LIR::LastNode(GenTree** nodes, size_t numNodes)
{
assert(numNodes > 0);
GenTree* lastNode = nodes[0];
for (size_t i = 1; i < numNodes; i++)
{
lastNode = LastNode(lastNode, nodes[i]);
}

return lastNode;
}

#ifdef DEBUG
void GenTree::dumpLIRFlags()
{
Expand Down
3 changes: 3 additions & 0 deletions src/coreclr/jit/lir.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -310,6 +310,9 @@ class LIR final
static Range SeqTree(Compiler* compiler, GenTree* tree);

static void InsertBeforeTerminator(BasicBlock* block, LIR::Range&& range);

static GenTree* LastNode(GenTree* node1, GenTree* node2);
static GenTree* LastNode(GenTree** nodes, size_t numNodes);
};

inline void GenTree::SetUnusedValue()
Expand Down
7 changes: 4 additions & 3 deletions src/coreclr/jit/lowerarmarch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1610,11 +1610,12 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
for (N = 1; N < argCnt - 1; N++)
{
opN = node->Op(N + 1);
idx = comp->gtNewIconNode(N);
BlockRange().InsertBefore(opN, idx);

// Place the insert as early as possible to avoid creating a lot of long lifetimes.
GenTree* insertionPoint = LIR::LastNode(tmp1, opN);
idx = comp->gtNewIconNode(N);
tmp1 = comp->gtNewSimdHWIntrinsicNode(simdType, tmp1, idx, opN, NI_AdvSimd_Insert, simdBaseJitType, simdSize);
BlockRange().InsertAfter(opN, tmp1);
BlockRange().InsertAfter(insertionPoint, idx, tmp1);
LowerNode(tmp1);
}

Expand Down
40 changes: 25 additions & 15 deletions src/coreclr/jit/lowerxarch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3210,13 +3210,16 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
size_t halfArgCnt = argCnt / 2;
assert((halfArgCnt * 2) == argCnt);

GenTree* loInsertionPoint = LIR::LastNode(node->GetOperandArray(), halfArgCnt);

GenTree* lo = comp->gtNewSimdHWIntrinsicNode(TYP_SIMD32, node->GetOperandArray(), halfArgCnt,
NI_Vector256_Create, simdBaseJitType, 32);
BlockRange().InsertAfter(node->Op(halfArgCnt), lo);
BlockRange().InsertAfter(loInsertionPoint, lo);

GenTree* hiInsertionPoint = LIR::LastNode(node->GetOperandArray(halfArgCnt), halfArgCnt);
GenTree* hi = comp->gtNewSimdHWIntrinsicNode(TYP_SIMD32, node->GetOperandArray(halfArgCnt), halfArgCnt,
NI_Vector256_Create, simdBaseJitType, 32);
BlockRange().InsertAfter(node->Op(argCnt), hi);
BlockRange().InsertAfter(hiInsertionPoint, hi);

assert(argCnt >= 7);
node->ResetHWIntrinsicId(NI_Vector512_WithUpper, comp, lo, hi);
Expand Down Expand Up @@ -3263,13 +3266,17 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
size_t halfArgCnt = argCnt / 2;
assert((halfArgCnt * 2) == argCnt);

GenTree* loInsertionPoint = LIR::LastNode(node->GetOperandArray(), halfArgCnt);

GenTree* lo = comp->gtNewSimdHWIntrinsicNode(TYP_SIMD16, node->GetOperandArray(), halfArgCnt,
NI_Vector128_Create, simdBaseJitType, 16);
BlockRange().InsertAfter(node->Op(halfArgCnt), lo);
BlockRange().InsertAfter(loInsertionPoint, lo);

GenTree* hiInsertionPoint = LIR::LastNode(node->GetOperandArray(halfArgCnt), halfArgCnt);

GenTree* hi = comp->gtNewSimdHWIntrinsicNode(TYP_SIMD16, node->GetOperandArray(halfArgCnt), halfArgCnt,
NI_Vector128_Create, simdBaseJitType, 16);
BlockRange().InsertAfter(node->Op(argCnt), hi);
BlockRange().InsertAfter(hiInsertionPoint, hi);

assert(argCnt >= 3);
node->ResetHWIntrinsicId(NI_Vector256_WithUpper, comp, lo, hi);
Expand Down Expand Up @@ -3336,11 +3343,12 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
opN = node->Op(N + 1);

idx = comp->gtNewIconNode(N, TYP_INT);
BlockRange().InsertAfter(opN, idx);
// Place the insert as early as possible to avoid creating a lot of long lifetimes.
GenTree* insertionPoint = LIR::LastNode(tmp1, opN);

tmp1 = comp->gtNewSimdHWIntrinsicNode(simdType, tmp1, opN, idx, insIntrinsic, simdBaseJitType,
simdSize);
BlockRange().InsertAfter(idx, tmp1);
BlockRange().InsertAfter(insertionPoint, idx, tmp1);
LowerNode(tmp1);
}

Expand Down Expand Up @@ -3419,17 +3427,17 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)

tmp1 = comp->gtNewSimdHWIntrinsicNode(simdType, op[N], op[O], NI_SSE2_UnpackLow, CORINFO_TYPE_UBYTE,
simdSize);
BlockRange().InsertAfter(op[O], tmp1);
BlockRange().InsertAfter(LIR::LastNode(op[N], op[O]), tmp1);
LowerNode(tmp1);

tmp2 = comp->gtNewSimdHWIntrinsicNode(simdType, op[P], op[Q], NI_SSE2_UnpackLow, CORINFO_TYPE_UBYTE,
simdSize);
BlockRange().InsertAfter(op[Q], tmp2);
BlockRange().InsertAfter(LIR::LastNode(op[P], op[Q]), tmp2);
LowerNode(tmp2);

tmp3 = comp->gtNewSimdHWIntrinsicNode(simdType, tmp1, tmp2, NI_SSE2_UnpackLow, CORINFO_TYPE_USHORT,
simdSize);
BlockRange().InsertAfter(tmp2, tmp3);
BlockRange().InsertAfter(LIR::LastNode(tmp1, tmp2), tmp3);
LowerNode(tmp3);

// This caches the result in index 0 through 3, depending on which
Expand Down Expand Up @@ -3468,12 +3476,12 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)

tmp1 =
comp->gtNewSimdHWIntrinsicNode(simdType, op[0], op[1], NI_SSE2_UnpackLow, CORINFO_TYPE_UINT, simdSize);
BlockRange().InsertAfter(op[1], tmp1);
BlockRange().InsertAfter(LIR::LastNode(op[0], op[1]), tmp1);
LowerNode(tmp1);

tmp2 =
comp->gtNewSimdHWIntrinsicNode(simdType, op[2], op[3], NI_SSE2_UnpackLow, CORINFO_TYPE_UINT, simdSize);
BlockRange().InsertAfter(op[3], tmp2);
BlockRange().InsertAfter(LIR::LastNode(op[2], op[3]), tmp2);
LowerNode(tmp2);

node->ResetHWIntrinsicId(NI_SSE2_UnpackLow, tmp1, tmp2);
Expand Down Expand Up @@ -3565,11 +3573,13 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
LowerNode(tmp2);

idx = comp->gtNewIconNode(N << 4, TYP_INT);
BlockRange().InsertAfter(tmp2, idx);

// Place the insert as early as possible to avoid creating a lot of long lifetimes.
GenTree* insertionPoint = LIR::LastNode(tmp1, tmp2);

tmp3 = comp->gtNewSimdHWIntrinsicNode(simdType, tmp1, tmp2, idx, NI_SSE41_Insert, simdBaseJitType,
simdSize);
BlockRange().InsertAfter(idx, tmp3);
BlockRange().InsertAfter(insertionPoint, idx, tmp3);

insertedNodes[N] = tmp3;
tmp1 = tmp3;
Expand Down Expand Up @@ -3655,11 +3665,11 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
}

tmp1 = comp->gtNewSimdHWIntrinsicNode(simdType, op[0], op[1], NI_SSE_UnpackLow, simdBaseJitType, simdSize);
BlockRange().InsertAfter(op[1], tmp1);
BlockRange().InsertAfter(LIR::LastNode(op[0], op[1]), tmp1);
LowerNode(tmp1);

tmp2 = comp->gtNewSimdHWIntrinsicNode(simdType, op[2], op[3], NI_SSE_UnpackLow, simdBaseJitType, simdSize);
BlockRange().InsertAfter(op[3], tmp2);
BlockRange().InsertAfter(LIR::LastNode(op[2], op[3]), tmp2);
LowerNode(tmp2);

node->ResetHWIntrinsicId(NI_SSE_MoveLowToHigh, tmp1, tmp2);
Expand Down