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

Jump to bottom

ARM64 - Emitting msub instruction #66621

Merged
merged 7 commits into from
Apr 12, 2022
Merged

ARM64 - Emitting msub instruction #66621

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
09dc7e3
Emitting MSUB for ARM64
TIHan Mar 14, 2022
0247254
Update codegenarm64.cpp
TIHan Mar 14, 2022
ff8ee6f
Updated comments
TIHan Mar 14, 2022
39d098c
Formatting
TIHan Mar 15, 2022
f006f8e
Combining some logic for GT_MADD and GT_MSUB
TIHan Mar 15, 2022
1acab50
Merge remote-tracking branch 'upstream/main' into arm64-msub
TIHan Apr 9, 2022
4655c4f
Can only use gtOverflow after checking to see if it's a valid operato…
TIHan Apr 9, 2022
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
Next Next commit
Emitting MSUB for ARM64
  • Loading branch information
@TIHan
TIHan committed Mar 14, 2022
commit 09dc7e38315f1bc65a23e6d6862aa906c8b624d8
1 change: 1 addition & 0 deletions src/coreclr/jit/codegen.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1336,6 +1336,7 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
#ifdef TARGET_ARM64
void genCodeForJumpCompare(GenTreeOp* tree);
void genCodeForMadd(GenTreeOp* tree);
void genCodeForMsub(GenTreeOp* tree);
void genCodeForBfiz(GenTreeOp* tree);
void genCodeForAddEx(GenTreeOp* tree);
#endif // TARGET_ARM64
Expand Down
27 changes: 26 additions & 1 deletion src/coreclr/jit/codegenarm64.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10023,7 +10023,7 @@ void CodeGen::instGen_MemoryBarrier(BarrierKind barrierKind)
}

//-----------------------------------------------------------------------------------
// genCodeForMadd: Emit a madd/msub (Multiply-Add) instruction
// genCodeForMadd: Emit a madd (Multiply-Add) instruction
//
// Arguments:
// tree - GT_MADD tree where op1 or op2 is GT_ADD
Expand Down Expand Up @@ -10067,6 +10067,31 @@ void CodeGen::genCodeForMadd(GenTreeOp* tree)
genProduceReg(tree);
}

//-----------------------------------------------------------------------------------
// genCodeForMsub: Emit a msub (Multiply-Subtract) instruction
//
// Arguments:
// tree - GT_MSUB tree where op2 is GT_SUB
//
void CodeGen::genCodeForMsub(GenTreeOp* tree)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Have you considered to just handle it as part of GT_MADD? I bet it'd be much less lines of code

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I did consider it - I guess it's a design choice whether or not to use GT_MADD or introduce GT_MSUB.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm actually conflicted on it. On the one hand, just using GT_MADD is sufficient, but on the other hand, the codegen for GT_MADD is a bit complicated since we are making GT_NEG nodes as contained.

The goal of introducing GT_MSUB and its code-gen was to make it easy to understand.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The name GT_MADD as a lowering-only op that is specific for ARM64 which reflects the actual instruction 'madd' - at least for me, I wouldn't it to expect to emit 'msub', but I totally understand why it does though.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we consider actually just making GT_MADD only emit 'madd' ? Then you could do the decision to use GT_MADD or GT_MSUB in lowering rather than in code-gen. It would make containment less complicated and code-gen simpler.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we consider actually just making GT_MADD only emit 'madd'

I like this and it makes it have "parity" with GT_ADD and GT_SUB.

{
assert(tree->OperIs(GT_MSUB) && varTypeIsIntegral(tree) && !(tree->gtFlags & GTF_SET_FLAGS));
genConsumeOperands(tree);

assert(tree->gtGetOp2()->OperIs(GT_MUL));
assert(tree->gtGetOp2()->isContained());

GenTree* a = tree->gtGetOp1();
GenTree* b = tree->gtGetOp2()->gtGetOp1();
GenTree* c = tree->gtGetOp2()->gtGetOp2();

// d = a - b * c
// MADD d, b, c, a
GetEmitter()->emitIns_R_R_R_R(INS_msub, emitActualTypeSize(tree), tree->GetRegNum(),
b->GetRegNum(), c->GetRegNum(), a->GetRegNum());
genProduceReg(tree);
}

//------------------------------------------------------------------------
// genCodeForBfiz: Generates the code sequence for a GenTree node that
// represents a bitfield insert in zero with sign/zero extension.
Expand Down
4 changes: 4 additions & 0 deletions src/coreclr/jit/codegenarmarch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -306,6 +306,10 @@ void CodeGen::genCodeForTreeNode(GenTree* treeNode)
genCodeForMadd(treeNode->AsOp());
break;

case GT_MSUB:
genCodeForMsub(treeNode->AsOp());
break;

case GT_INC_SATURATE:
genCodeForIncSaturate(treeNode);
break;
Expand Down
4 changes: 3 additions & 1 deletion src/coreclr/jit/gtlist.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -220,8 +220,10 @@ GTNODE(MUL_LONG , GenTreeOp ,1,GTK_BINOP|DBK_NOTHIR)
GTNODE(AND_NOT , GenTreeOp ,0,GTK_BINOP|DBK_NOTHIR)

#ifdef TARGET_ARM64
GTNODE(MADD , GenTreeOp ,0,GTK_BINOP|DBK_NOTHIR) // Generates the Multiply-Add instruction (madd/msub) In the future, we might consider
GTNODE(MADD , GenTreeOp ,0,GTK_BINOP|DBK_NOTHIR) // Generates the Multiply-Add instruction. In the future, we might consider
// enabling it for both armarch and xarch for floating-point MADD "unsafe" math.
GTNODE(MSUB , GenTreeOp ,0,GTK_BINOP|DBK_NOTHIR) // Generates the Multiply-Subtract instruction. In the future, we might consider
// enabling it for both armarch and xarch for floating-point MSUB "unsafe" math.
GTNODE(ADDEX, GenTreeOp ,0,GTK_BINOP|DBK_NOTHIR) // Add with sign/zero extension.
GTNODE(BFIZ , GenTreeOp ,0,GTK_BINOP|DBK_NOTHIR) // Bitfield Insert in Zero.
#endif
Expand Down
15 changes: 15 additions & 0 deletions src/coreclr/jit/lowerarmarch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1673,6 +1673,21 @@ void Lowering::ContainCheckBinary(GenTreeOp* node)
}
}

// Find "a - b * c" in order to emit MSUB
if (comp->opts.OptimizationEnabled() && varTypeIsIntegral(node) && !node->isContained() && node->OperIs(GT_SUB) &&
TIHan marked this conversation as resolved.
Show resolved Hide resolved
!node->gtOverflow() && op2->OperIs(GT_MUL) && !op2->isContained() && !op2->gtOverflow() && varTypeIsIntegral(op2))
{
GenTree* a = op1;
GenTree* b = op2->gtGetOp1();
GenTree* c = op2->gtGetOp2();

if (!a->isContained() && !b->isContained() && !c->isContained())
{
node->ChangeOper(GT_MSUB);
MakeSrcContained(node, op2);
}
}

// Change ADD TO ADDEX for ADD(X, CAST(Y)) or ADD(CAST(X), Y) where CAST is int->long
// or for ADD(LSH(X, CNS), X) or ADD(X, LSH(X, CNS)) where CNS is in the (0..typeWidth) range
if (node->OperIs(GT_ADD) && !op1->isContained() && !op2->isContained() && varTypeIsIntegral(node) &&
Expand Down