Skip to content

[Matrix] Propagate shape information through Select insts #141876

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 8 commits into from
Jun 12, 2025
Merged

[Matrix] Propagate shape information through Select insts #141876

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
4040d3f
[Matrix] Propagate shape information through Select insts
jroelofs May 28, 2025
e952bcb
Merge remote-tracking branch 'origin/main' into jroelofs/lower-matrix…
jroelofs Jun 5, 2025
c0c63f3
select with mismatched shape
jroelofs Jun 5, 2025
5a4a6a5
no return value
jroelofs Jun 5, 2025
6693f56
clean up condition lookup
jroelofs Jun 5, 2025
1a4faa4
review feedback: don't take shape info from select conditions
jroelofs Jun 10, 2025
48bed98
Merge remote-tracking branch 'origin/main' into jroelofs/lower-matrix…
jroelofs Jun 10, 2025
dc97261
Merge remote-tracking branch 'origin/main' into jroelofs/lower-matrix…
jroelofs Jun 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
56 changes: 55 additions & 1 deletion llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@
#include "llvm/IR/CFG.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
Expand Down Expand Up @@ -269,6 +270,16 @@ computeShapeInfoForInst(Instruction *I,
return OpShape->second;
}

if (auto *Select = dyn_cast<SelectInst>(I)) {
Type *CondTy = Select->getCondition()->getType();
for (Use &Op : CondTy->isVectorTy() ? Select->operands()
: drop_begin(Select->operands())) {
auto OpShape = ShapeMap.find(Op);
if (OpShape != ShapeMap.end())
return OpShape->second;
}
}

if (isUniformShape(I)) {
// Find the first operand that has a known shape and use that.
for (auto &Op : I->operands()) {
Expand Down Expand Up @@ -620,7 +631,8 @@ class LowerMatrixIntrinsics {
default:
return false;
}
return isUniformShape(V) || isa<StoreInst>(V) || isa<LoadInst>(V);
return isUniformShape(V) || isa<StoreInst>(V) || isa<LoadInst>(V) ||
isa<SelectInst>(V);
}

/// Propagate the shape information of instructions to their users.
Expand Down Expand Up @@ -707,6 +719,14 @@ class LowerMatrixIntrinsics {
} else if (isa<StoreInst>(V)) {
// Nothing to do. We forward-propagated to this so we would just
// backward propagate to an instruction with an already known shape.
} else if (auto *Select = dyn_cast<SelectInst>(V)) {
ShapeInfo Shape = ShapeMap[V];
Type *CondTy = Select->getCondition()->getType();
for (Use &Op : CondTy->isVectorTy() ? Select->operands()
: drop_begin(Select->operands())) {
if (setShapeInfo(Op, Shape))
pushInstruction(Select, WorkList);
}
} else if (isUniformShape(V)) {
// Propagate to all operands.
ShapeInfo Shape = ShapeMap[V];
Expand Down Expand Up @@ -1066,6 +1086,8 @@ class LowerMatrixIntrinsics {
VisitUnaryOperator(UnOp, SI);
else if (CallInst *CInst = dyn_cast<CallInst>(Inst))
VisitCallInst(CInst);
else if (auto *Select = dyn_cast<SelectInst>(Inst))
VisitSelectInst(Select, SI);
else if (match(Inst, m_Load(m_Value(Op1))))
VisitLoad(cast<LoadInst>(Inst), SI, Op1, Builder);
else if (match(Inst, m_Store(m_Value(Op1), m_Value(Op2))))
Expand Down Expand Up @@ -2177,6 +2199,38 @@ class LowerMatrixIntrinsics {
Builder);
}

/// Lower selects.
void VisitSelectInst(SelectInst *Inst, const ShapeInfo &Shape) {
Value *Cond = Inst->getOperand(0);
Value *OpA = Inst->getOperand(1);
Value *OpB = Inst->getOperand(2);

IRBuilder<> Builder(Inst);

MatrixTy Result;
MatrixTy A = getMatrix(OpA, Shape, Builder);
MatrixTy B = getMatrix(OpB, Shape, Builder);

Value *CondV[2];
if (isa<FixedVectorType>(Cond->getType())) {
MatrixTy C = getMatrix(Cond, Shape, Builder);
CondV[0] = C.getVector(0);
CondV[1] = C.getVector(1);
} else {
CondV[0] = Cond;
CondV[1] = Cond;
}

for (unsigned I = 0, E = Shape.getNumVectors(); I != E; ++I)
Result.addVector(
Builder.CreateSelect(CondV[I], A.getVector(I), B.getVector(I)));

finalizeLowering(Inst,
Result.addNumComputeOps(getNumOps(Result.getVectorTy()) *
Result.getNumVectors()),
Builder);
}

/// Helper to linearize a matrix expression tree into a string. Currently
/// matrix expressions are linarized by starting at an expression leaf and
/// linearizing bottom up.
Expand Down
146 changes: 146 additions & 0 deletions llvm/test/Transforms/LowerMatrixIntrinsics/select.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,146 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -passes='lower-matrix-intrinsics' -S < %s | FileCheck %s

define void @select_2x2_bot(i1 %cond, ptr %lhs, ptr %rhs, ptr %out) {
; CHECK-LABEL: @select_2x2_bot(
; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x float>, ptr [[LHS:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr float, ptr [[LHS]], i64 2
; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x float>, ptr [[RHS:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr float, ptr [[RHS]], i64 2
; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x float>, ptr [[VEC_GEP3]], align 8
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[COND:%.*]], <2 x float> [[COL_LOAD]], <2 x float> [[COL_LOAD2]]
; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[COND]], <2 x float> [[COL_LOAD1]], <2 x float> [[COL_LOAD4]]
; CHECK-NEXT: store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 4
; CHECK-NEXT: [[VEC_GEP5:%.*]] = getelementptr float, ptr [[OUT]], i64 2
; CHECK-NEXT: store <2 x float> [[TMP2]], ptr [[VEC_GEP5]], align 4
; CHECK-NEXT: ret void
;
%lhsv = load <4 x float>, ptr %lhs
%rhsv = load <4 x float>, ptr %rhs
%op = select i1 %cond, <4 x float> %lhsv, <4 x float> %rhsv
call void @llvm.matrix.column.major.store(<4 x float> %op, ptr %out, i64 2, i1 false, i32 2, i32 2)
ret void
}

define void @select_2x2_lhs(i1 %cond, ptr %lhs, ptr %rhs, ptr %out) {
; CHECK-LABEL: @select_2x2_lhs(
; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x float>, ptr [[LHS:%.*]], align 4
; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr float, ptr [[LHS]], i64 2
; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 4
; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x float>, ptr [[RHS:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr float, ptr [[RHS]], i64 2
; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x float>, ptr [[VEC_GEP3]], align 8
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[COND:%.*]], <2 x float> [[COL_LOAD]], <2 x float> [[COL_LOAD2]]
; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[COND]], <2 x float> [[COL_LOAD1]], <2 x float> [[COL_LOAD4]]
; CHECK-NEXT: store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP5:%.*]] = getelementptr float, ptr [[OUT]], i64 2
; CHECK-NEXT: store <2 x float> [[TMP2]], ptr [[VEC_GEP5]], align 8
; CHECK-NEXT: ret void
;
%lhsv = call <4 x float> @llvm.matrix.column.major.load(ptr %lhs, i64 2, i1 false, i32 2, i32 2)
%rhsv = load <4 x float>, ptr %rhs
%op = select i1 %cond, <4 x float> %lhsv, <4 x float> %rhsv
store <4 x float> %op, ptr %out
ret void
}

define void @select_2x2_rhs(i1 %cond, ptr %lhs, ptr %rhs, ptr %out) {
; CHECK-LABEL: @select_2x2_rhs(
; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x float>, ptr [[RHS:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr float, ptr [[RHS]], i64 2
; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x float>, ptr [[RHS1:%.*]], align 4
; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr float, ptr [[RHS1]], i64 2
; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x float>, ptr [[VEC_GEP3]], align 4
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[COND:%.*]], <2 x float> [[COL_LOAD]], <2 x float> [[COL_LOAD2]]
; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[COND]], <2 x float> [[COL_LOAD1]], <2 x float> [[COL_LOAD4]]
; CHECK-NEXT: store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr float, ptr [[OUT]], i64 2
; CHECK-NEXT: store <2 x float> [[TMP2]], ptr [[VEC_GEP2]], align 8
; CHECK-NEXT: ret void
;
%lhsv = load <4 x float>, ptr %lhs
%rhsv = call <4 x float> @llvm.matrix.column.major.load(ptr %rhs, i64 2, i1 false, i32 2, i32 2)
%op = select i1 %cond, <4 x float> %lhsv, <4 x float> %rhsv
store <4 x float> %op, ptr %out
ret void
}

define void @select_2x2_vcond_shape1(ptr %cond, ptr %lhs, ptr %rhs, ptr %out) {
; CHECK-LABEL: @select_2x2_vcond_shape1(
; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x float>, ptr [[LHS:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr float, ptr [[LHS]], i64 2
; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x i1>, ptr [[RHS:%.*]], align 1
; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr i1, ptr [[RHS]], i64 2
; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x i1>, ptr [[VEC_GEP3]], align 1
; CHECK-NEXT: [[COL_LOAD5:%.*]] = load <2 x float>, ptr [[RHS1:%.*]], align 4
; CHECK-NEXT: [[VEC_GEP6:%.*]] = getelementptr float, ptr [[RHS1]], i64 2
; CHECK-NEXT: [[COL_LOAD7:%.*]] = load <2 x float>, ptr [[VEC_GEP6]], align 4
; CHECK-NEXT: [[TMP1:%.*]] = select <2 x i1> [[COL_LOAD2]], <2 x float> [[COL_LOAD]], <2 x float> [[COL_LOAD5]]
; CHECK-NEXT: [[TMP2:%.*]] = select <2 x i1> [[COL_LOAD4]], <2 x float> [[COL_LOAD1]], <2 x float> [[COL_LOAD7]]
; CHECK-NEXT: store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP8:%.*]] = getelementptr float, ptr [[OUT]], i64 2
; CHECK-NEXT: store <2 x float> [[TMP2]], ptr [[VEC_GEP8]], align 8
; CHECK-NEXT: ret void
;
%lhsv = load <4 x float>, ptr %lhs
%condv = load <4 x i1>, ptr %cond
%rhsv = call <4 x float> @llvm.matrix.column.major.load(ptr %rhs, i64 2, i1 false, i32 2, i32 2)
%op = select <4 x i1> %condv, <4 x float> %lhsv, <4 x float> %rhsv
store <4 x float> %op, ptr %out
ret void
}

define void @select_2x2_vcond_shape2(ptr %cond, ptr %lhs, ptr %rhs, ptr %out) {
; CHECK-LABEL: @select_2x2_vcond_shape2(
; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x float>, ptr [[LHS:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr float, ptr [[LHS]], i64 2
; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x i1>, ptr [[COND:%.*]], align 1
; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr i1, ptr [[COND]], i64 2
; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x i1>, ptr [[VEC_GEP3]], align 1
; CHECK-NEXT: [[COL_LOAD5:%.*]] = load <2 x float>, ptr [[RHS:%.*]], align 4
; CHECK-NEXT: [[VEC_GEP6:%.*]] = getelementptr float, ptr [[RHS]], i64 2
; CHECK-NEXT: [[COL_LOAD7:%.*]] = load <2 x float>, ptr [[VEC_GEP6]], align 4
; CHECK-NEXT: [[TMP1:%.*]] = select <2 x i1> [[COL_LOAD2]], <2 x float> [[COL_LOAD]], <2 x float> [[COL_LOAD5]]
; CHECK-NEXT: [[TMP2:%.*]] = select <2 x i1> [[COL_LOAD4]], <2 x float> [[COL_LOAD1]], <2 x float> [[COL_LOAD7]]
; CHECK-NEXT: store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP8:%.*]] = getelementptr float, ptr [[OUT]], i64 2
; CHECK-NEXT: store <2 x float> [[TMP2]], ptr [[VEC_GEP8]], align 8
; CHECK-NEXT: ret void
;
%lhsv = load <4 x float>, ptr %lhs
%condv = call <4 x i1> @llvm.matrix.column.major.load(ptr %cond, i64 2, i1 false, i32 2, i32 2)
%rhsv = call <4 x float> @llvm.matrix.column.major.load(ptr %rhs, i64 2, i1 false, i32 2, i32 2)
%op = select <4 x i1> %condv, <4 x float> %lhsv, <4 x float> %rhsv
store <4 x float> %op, ptr %out
ret void
}

define void @select_2x2_vcond_shape3(ptr %cond, ptr %lhs, ptr %rhs, ptr %out) {
; CHECK-LABEL: @select_2x2_vcond_shape3(
; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x float>, ptr [[LHS:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr float, ptr [[LHS]], i64 2
; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <4 x i1>, ptr [[COND:%.*]], align 1
; CHECK-NEXT: [[COL_LOAD3:%.*]] = load <2 x float>, ptr [[RHS:%.*]], align 4
; CHECK-NEXT: [[VEC_GEP4:%.*]] = getelementptr float, ptr [[RHS]], i64 2
; CHECK-NEXT: [[COL_LOAD5:%.*]] = load <2 x float>, ptr [[VEC_GEP4]], align 4
; CHECK-NEXT: [[SPLIT:%.*]] = shufflevector <4 x i1> [[COL_LOAD2]], <4 x i1> poison, <2 x i32> <i32 0, i32 1>
; CHECK-NEXT: [[SPLIT6:%.*]] = shufflevector <4 x i1> [[COL_LOAD2]], <4 x i1> poison, <2 x i32> <i32 2, i32 3>
; CHECK-NEXT: [[TMP1:%.*]] = select <2 x i1> [[SPLIT]], <2 x float> [[COL_LOAD]], <2 x float> [[COL_LOAD3]]
; CHECK-NEXT: [[TMP2:%.*]] = select <2 x i1> [[SPLIT6]], <2 x float> [[COL_LOAD1]], <2 x float> [[COL_LOAD5]]
; CHECK-NEXT: store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 16
; CHECK-NEXT: [[VEC_GEP7:%.*]] = getelementptr float, ptr [[OUT]], i64 2
; CHECK-NEXT: store <2 x float> [[TMP2]], ptr [[VEC_GEP7]], align 8
; CHECK-NEXT: ret void
;
%lhsv = load <4 x float>, ptr %lhs
%condv = call <4 x i1> @llvm.matrix.column.major.load(ptr %cond, i64 4, i1 false, i32 4, i32 1)
%rhsv = call <4 x float> @llvm.matrix.column.major.load(ptr %rhs, i64 2, i1 false, i32 2, i32 2)
%op = select <4 x i1> %condv, <4 x float> %lhsv, <4 x float> %rhsv
store <4 x float> %op, ptr %out
ret void
}
Loading