Skip to content

[X86] checkBitcastSrcVectorSize - early return when reach to MaxRecursionDepth. #130226

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

Merged
merged 2 commits into from
Mar 10, 2025
Merged

[X86] checkBitcastSrcVectorSize - early return when reach to MaxRecursionDepth. #130226

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4574283
[X86] Fix duplicated compute in recursive search.
haonanya1 Mar 7, 2025
b5aeeee
Add a recursive depth limit.
haonanya1 Mar 10, 2025
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
74 changes: 53 additions & 21 deletions llvm/lib/Target/X86/X86ISelLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,6 +63,7 @@
#include <bitset>
#include <cctype>
#include <numeric>
#include <tuple>
using namespace llvm;

#define DEBUG_TYPE "x86-isel"
Expand Down Expand Up @@ -44745,31 +44746,59 @@ bool X86TargetLowering::isSplatValueForTargetNode(SDValue Op,

// Helper to peek through bitops/trunc/setcc to determine size of source vector.
// Allows combineBitcastvxi1 to determine what size vector generated a <X x i1>.
static bool checkBitcastSrcVectorSize(SDValue Src, unsigned Size,
bool AllowTruncate) {
static bool
checkBitcastSrcVectorSize(SDValue Src, unsigned Size, bool AllowTruncate,
std::map<std::tuple<SDValue, unsigned, bool>, bool>
&BitcastSrcVectorSizeMap) {
auto Tp = std::make_tuple(Src, Size, AllowTruncate);
if (BitcastSrcVectorSizeMap.count(Tp))
return BitcastSrcVectorSizeMap[Tp];
switch (Src.getOpcode()) {
case ISD::TRUNCATE:
if (!AllowTruncate)
if (!AllowTruncate) {
BitcastSrcVectorSizeMap[Tp] = false;
Copy link
Contributor

@phoebewang phoebewang Mar 7, 2025

Choose a reason for hiding this comment

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

How about define a bool Result = false; to avoid repeat BitcastSrcVectorSizeMap[Tp] =?

Copy link
Contributor Author

@haonanya1 haonanya1 Mar 10, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

Update pr to Early return when reach to MaxRecursionDepth. It's same as current other implementation, such as SimplifyMultipleUseDemandedBits, matchIndexRecursively and so on.

return false;
}
[[fallthrough]];
case ISD::SETCC:
return Src.getOperand(0).getValueSizeInBits() == Size;
case ISD::FREEZE:
return checkBitcastSrcVectorSize(Src.getOperand(0), Size, AllowTruncate);
case ISD::SETCC: {
auto Ret = Src.getOperand(0).getValueSizeInBits() == Size;
BitcastSrcVectorSizeMap[Tp] = Ret;
return Ret;
}
case ISD::FREEZE: {
auto Ret = checkBitcastSrcVectorSize(Src.getOperand(0), Size, AllowTruncate,
BitcastSrcVectorSizeMap);
BitcastSrcVectorSizeMap[Tp] = Ret;
return Ret;
}
case ISD::AND:
case ISD::XOR:
case ISD::OR:
return checkBitcastSrcVectorSize(Src.getOperand(0), Size, AllowTruncate) &&
checkBitcastSrcVectorSize(Src.getOperand(1), Size, AllowTruncate);
case ISD::OR: {
auto Ret1 = checkBitcastSrcVectorSize(
Src.getOperand(0), Size, AllowTruncate, BitcastSrcVectorSizeMap);
auto Ret2 = checkBitcastSrcVectorSize(
Src.getOperand(1), Size, AllowTruncate, BitcastSrcVectorSizeMap);
BitcastSrcVectorSizeMap[Tp] = Ret1 && Ret2;
return Ret1 && Ret2;
}
case ISD::SELECT:
case ISD::VSELECT:
return Src.getOperand(0).getScalarValueSizeInBits() == 1 &&
checkBitcastSrcVectorSize(Src.getOperand(1), Size, AllowTruncate) &&
checkBitcastSrcVectorSize(Src.getOperand(2), Size, AllowTruncate);
case ISD::BUILD_VECTOR:
return ISD::isBuildVectorAllZeros(Src.getNode()) ||
ISD::isBuildVectorAllOnes(Src.getNode());
case ISD::VSELECT: {
auto Ret1 = checkBitcastSrcVectorSize(
Src.getOperand(1), Size, AllowTruncate, BitcastSrcVectorSizeMap);
auto Ret2 = checkBitcastSrcVectorSize(
Src.getOperand(2), Size, AllowTruncate, BitcastSrcVectorSizeMap);
auto Ret3 = Src.getOperand(0).getScalarValueSizeInBits() == 1;
BitcastSrcVectorSizeMap[Tp] = Ret1 && Ret2 && Ret3;
return Ret1 && Ret2 && Ret3;
}
case ISD::BUILD_VECTOR: {
auto Ret = ISD::isBuildVectorAllZeros(Src.getNode()) ||
ISD::isBuildVectorAllOnes(Src.getNode());
BitcastSrcVectorSizeMap[Tp] = Ret;
return Ret;
}
}
BitcastSrcVectorSizeMap[Tp] = false;
return false;
}

Expand Down Expand Up @@ -44925,6 +44954,7 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, EVT VT, SDValue Src,
// (v16i8 shuffle <0,2,4,6,8,10,12,14,u,u,...,u> (v16i8 bitcast t0), undef)
MVT SExtVT;
bool PropagateSExt = false;
std::map<std::tuple<SDValue, unsigned, bool>, bool> BitcastSrcVectorSizeMap;
Copy link
Contributor

@phoebewang phoebewang Mar 7, 2025

Choose a reason for hiding this comment

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

DenseMap?

Copy link
Contributor Author

@haonanya1 haonanya1 Mar 10, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

Update pr to Early return when reach to MaxRecursionDepth. It's same as current other implementation, such as SimplifyMultipleUseDemandedBits, matchIndexRecursively and so on.

switch (SrcVT.getSimpleVT().SimpleTy) {
default:
return SDValue();
Expand All @@ -44936,7 +44966,8 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, EVT VT, SDValue Src,
// For cases such as (i4 bitcast (v4i1 setcc v4i64 v1, v2))
// sign-extend to a 256-bit operation to avoid truncation.
if (Subtarget.hasAVX() &&
checkBitcastSrcVectorSize(Src, 256, Subtarget.hasAVX2())) {
checkBitcastSrcVectorSize(Src, 256, Subtarget.hasAVX2(),
BitcastSrcVectorSizeMap)) {
SExtVT = MVT::v4i64;
PropagateSExt = true;
}
Expand All @@ -44948,8 +44979,9 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, EVT VT, SDValue Src,
// If the setcc operand is 128-bit, prefer sign-extending to 128-bit over
// 256-bit because the shuffle is cheaper than sign extending the result of
// the compare.
if (Subtarget.hasAVX() && (checkBitcastSrcVectorSize(Src, 256, true) ||
checkBitcastSrcVectorSize(Src, 512, true))) {
if (Subtarget.hasAVX() &&
(checkBitcastSrcVectorSize(Src, 256, true, BitcastSrcVectorSizeMap) ||
checkBitcastSrcVectorSize(Src, 512, true, BitcastSrcVectorSizeMap))) {
SExtVT = MVT::v8i32;
PropagateSExt = true;
}
Expand All @@ -44974,7 +45006,7 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, EVT VT, SDValue Src,
break;
}
// Split if this is a <64 x i8> comparison result.
if (checkBitcastSrcVectorSize(Src, 512, false)) {
if (checkBitcastSrcVectorSize(Src, 512, false, BitcastSrcVectorSizeMap)) {
SExtVT = MVT::v64i8;
break;
}
Expand Down