Skip to content

[LV] Fix the cost of min/max reductions. #98453

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
Jul 12, 2024
Merged

[LV] Fix the cost of min/max reductions. #98453

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
23169a4
Add test case
Mel-Chen Jul 11, 2024
a00f725
[LV] Fix the cost of in-loop minmax reduction.
Mel-Chen Jul 11, 2024
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
14 changes: 11 additions & 3 deletions llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6138,12 +6138,20 @@ LoopVectorizationCostModel::getReductionPatternCost(
const RecurrenceDescriptor &RdxDesc =
Legal->getReductionVars().find(cast<PHINode>(ReductionPhi))->second;

InstructionCost BaseCost = TTI.getArithmeticReductionCost(
RdxDesc.getOpcode(), VectorTy, RdxDesc.getFastMathFlags(), CostKind);
InstructionCost BaseCost;
RecurKind RK = RdxDesc.getRecurrenceKind();
if (RecurrenceDescriptor::isMinMaxRecurrenceKind(RK)) {
Intrinsic::ID MinMaxID = getMinMaxReductionIntrinsicOp(RK);
BaseCost = TTI.getMinMaxReductionCost(MinMaxID, VectorTy,
RdxDesc.getFastMathFlags(), CostKind);
} else {
BaseCost = TTI.getArithmeticReductionCost(
RdxDesc.getOpcode(), VectorTy, RdxDesc.getFastMathFlags(), CostKind);
}

// For a call to the llvm.fmuladd intrinsic we need to add the cost of a
// normal fmul instruction to the cost of the fadd reduction.
if (RdxDesc.getRecurrenceKind() == RecurKind::FMulAdd)
if (RK == RecurKind::FMulAdd)
BaseCost +=
TTI.getArithmeticInstrCost(Instruction::FMul, VectorTy, CostKind);

Expand Down
178 changes: 178 additions & 0 deletions llvm/test/Transforms/LoopVectorize/RISCV/inloop-reduction.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,3 +198,181 @@ for.cond.cleanup: ; preds = %for.body, %entry
%r.0.lcssa = phi i32 [ 0, %entry ], [ %add, %for.body ]
ret i32 %r.0.lcssa
}

define i32 @smin(ptr %a, i64 %n, i32 %start) {
; OUTLOOP-LABEL: @smin(
; OUTLOOP-NEXT: entry:
; OUTLOOP-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
; OUTLOOP-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 4
; OUTLOOP-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]]
; OUTLOOP-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
; OUTLOOP: vector.ph:
; OUTLOOP-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
; OUTLOOP-NEXT: [[TMP3:%.*]] = mul i64 [[TMP2]], 4
; OUTLOOP-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]
; OUTLOOP-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
; OUTLOOP-NEXT: [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
; OUTLOOP-NEXT: [[TMP5:%.*]] = mul i64 [[TMP4]], 4
; OUTLOOP-NEXT: [[MINMAX_IDENT_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[START:%.*]], i64 0
; OUTLOOP-NEXT: [[MINMAX_IDENT_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[MINMAX_IDENT_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
; OUTLOOP-NEXT: br label [[VECTOR_BODY:%.*]]
; OUTLOOP: vector.body:
; OUTLOOP-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; OUTLOOP-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ [[MINMAX_IDENT_SPLAT]], [[VECTOR_PH]] ], [ [[TMP10:%.*]], [[VECTOR_BODY]] ]
; OUTLOOP-NEXT: [[TMP6:%.*]] = add i64 [[INDEX]], 0
; OUTLOOP-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP6]]
; OUTLOOP-NEXT: [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[TMP7]], i32 0
; OUTLOOP-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x i32>, ptr [[TMP8]], align 4
; OUTLOOP-NEXT: [[TMP9:%.*]] = icmp slt <vscale x 4 x i32> [[WIDE_LOAD]], [[VEC_PHI]]
; OUTLOOP-NEXT: [[TMP10]] = select <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i32> [[WIDE_LOAD]], <vscale x 4 x i32> [[VEC_PHI]]
; OUTLOOP-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]]
; OUTLOOP-NEXT: [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; OUTLOOP-NEXT: br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; OUTLOOP: middle.block:
; OUTLOOP-NEXT: [[TMP12:%.*]] = call i32 @llvm.vector.reduce.smin.nxv4i32(<vscale x 4 x i32> [[TMP10]])
; OUTLOOP-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
; OUTLOOP-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
; OUTLOOP: scalar.ph:
; OUTLOOP-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
; OUTLOOP-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP12]], [[MIDDLE_BLOCK]] ], [ [[START]], [[ENTRY]] ]
; OUTLOOP-NEXT: br label [[FOR_BODY:%.*]]
; OUTLOOP: for.body:
; OUTLOOP-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY]] ]
; OUTLOOP-NEXT: [[RDX:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[SMIN:%.*]], [[FOR_BODY]] ]
; OUTLOOP-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV]]
; OUTLOOP-NEXT: [[TMP13:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; OUTLOOP-NEXT: [[CMP_I:%.*]] = icmp slt i32 [[TMP13]], [[RDX]]
; OUTLOOP-NEXT: [[SMIN]] = select i1 [[CMP_I]], i32 [[TMP13]], i32 [[RDX]]
; OUTLOOP-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
; OUTLOOP-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
; OUTLOOP-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
; OUTLOOP: for.end:
; OUTLOOP-NEXT: [[SMIN_LCSSA:%.*]] = phi i32 [ [[SMIN]], [[FOR_BODY]] ], [ [[TMP12]], [[MIDDLE_BLOCK]] ]
; OUTLOOP-NEXT: ret i32 [[SMIN_LCSSA]]
;
; INLOOP-LABEL: @smin(
; INLOOP-NEXT: entry:
; INLOOP-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
; INLOOP-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 4
; INLOOP-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]]
; INLOOP-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
; INLOOP: vector.ph:
; INLOOP-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
; INLOOP-NEXT: [[TMP3:%.*]] = mul i64 [[TMP2]], 4
; INLOOP-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]
; INLOOP-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
; INLOOP-NEXT: [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
; INLOOP-NEXT: [[TMP5:%.*]] = mul i64 [[TMP4]], 4
; INLOOP-NEXT: br label [[VECTOR_BODY:%.*]]
; INLOOP: vector.body:
; INLOOP-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; INLOOP-NEXT: [[VEC_PHI:%.*]] = phi i32 [ [[START:%.*]], [[VECTOR_PH]] ], [ [[RDX_MINMAX:%.*]], [[VECTOR_BODY]] ]
; INLOOP-NEXT: [[TMP6:%.*]] = add i64 [[INDEX]], 0
; INLOOP-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP6]]
; INLOOP-NEXT: [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[TMP7]], i32 0
; INLOOP-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x i32>, ptr [[TMP8]], align 4
; INLOOP-NEXT: [[TMP9:%.*]] = call i32 @llvm.vector.reduce.smin.nxv4i32(<vscale x 4 x i32> [[WIDE_LOAD]])
; INLOOP-NEXT: [[RDX_MINMAX]] = call i32 @llvm.smin.i32(i32 [[TMP9]], i32 [[VEC_PHI]])
; INLOOP-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]]
; INLOOP-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; INLOOP-NEXT: br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; INLOOP: middle.block:
; INLOOP-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
; INLOOP-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
; INLOOP: scalar.ph:
; INLOOP-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
; INLOOP-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[RDX_MINMAX]], [[MIDDLE_BLOCK]] ], [ [[START]], [[ENTRY]] ]
; INLOOP-NEXT: br label [[FOR_BODY:%.*]]
; INLOOP: for.body:
; INLOOP-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY]] ]
; INLOOP-NEXT: [[RDX:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[SMIN:%.*]], [[FOR_BODY]] ]
; INLOOP-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV]]
; INLOOP-NEXT: [[TMP11:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; INLOOP-NEXT: [[CMP_I:%.*]] = icmp slt i32 [[TMP11]], [[RDX]]
; INLOOP-NEXT: [[SMIN]] = select i1 [[CMP_I]], i32 [[TMP11]], i32 [[RDX]]
; INLOOP-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
; INLOOP-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
; INLOOP-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
; INLOOP: for.end:
; INLOOP-NEXT: [[SMIN_LCSSA:%.*]] = phi i32 [ [[SMIN]], [[FOR_BODY]] ], [ [[RDX_MINMAX]], [[MIDDLE_BLOCK]] ]
; INLOOP-NEXT: ret i32 [[SMIN_LCSSA]]
;
; IF-EVL-LABEL: @smin(
; IF-EVL-NEXT: entry:
; IF-EVL-NEXT: [[TMP0:%.*]] = sub i64 -1, [[N:%.*]]
; IF-EVL-NEXT: [[TMP1:%.*]] = call i64 @llvm.vscale.i64()
; IF-EVL-NEXT: [[TMP2:%.*]] = mul i64 [[TMP1]], 4
; IF-EVL-NEXT: [[TMP3:%.*]] = icmp ult i64 [[TMP0]], [[TMP2]]
; IF-EVL-NEXT: br i1 [[TMP3]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
; IF-EVL: vector.ph:
; IF-EVL-NEXT: [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
; IF-EVL-NEXT: [[TMP5:%.*]] = mul i64 [[TMP4]], 4
; IF-EVL-NEXT: [[TMP6:%.*]] = sub i64 [[TMP5]], 1
; IF-EVL-NEXT: [[N_RND_UP:%.*]] = add i64 [[N]], [[TMP6]]
; IF-EVL-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP5]]
; IF-EVL-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
; IF-EVL-NEXT: [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[N]], 1
; IF-EVL-NEXT: [[TMP7:%.*]] = call i64 @llvm.vscale.i64()
; IF-EVL-NEXT: [[TMP8:%.*]] = mul i64 [[TMP7]], 4
; IF-EVL-NEXT: [[MINMAX_IDENT_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[START:%.*]], i64 0
; IF-EVL-NEXT: [[MINMAX_IDENT_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[MINMAX_IDENT_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
; IF-EVL-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i64 0
; IF-EVL-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
; IF-EVL-NEXT: br label [[VECTOR_BODY:%.*]]
; IF-EVL: vector.body:
; IF-EVL-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; IF-EVL-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ [[MINMAX_IDENT_SPLAT]], [[VECTOR_PH]] ], [ [[TMP16:%.*]], [[VECTOR_BODY]] ]
; IF-EVL-NEXT: [[TMP9:%.*]] = add i64 [[INDEX]], 0
; IF-EVL-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX]], i64 0
; IF-EVL-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
; IF-EVL-NEXT: [[TMP10:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
; IF-EVL-NEXT: [[TMP11:%.*]] = add <vscale x 4 x i64> zeroinitializer, [[TMP10]]
; IF-EVL-NEXT: [[VEC_IV:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT]], [[TMP11]]
; IF-EVL-NEXT: [[TMP12:%.*]] = icmp ule <vscale x 4 x i64> [[VEC_IV]], [[BROADCAST_SPLAT2]]
; IF-EVL-NEXT: [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP9]]
; IF-EVL-NEXT: [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[TMP13]], i32 0
; IF-EVL-NEXT: [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 4 x i32> @llvm.masked.load.nxv4i32.p0(ptr [[TMP14]], i32 4, <vscale x 4 x i1> [[TMP12]], <vscale x 4 x i32> poison)
; IF-EVL-NEXT: [[TMP15:%.*]] = icmp slt <vscale x 4 x i32> [[WIDE_MASKED_LOAD]], [[VEC_PHI]]
; IF-EVL-NEXT: [[TMP16]] = select <vscale x 4 x i1> [[TMP15]], <vscale x 4 x i32> [[WIDE_MASKED_LOAD]], <vscale x 4 x i32> [[VEC_PHI]]
; IF-EVL-NEXT: [[TMP17:%.*]] = select <vscale x 4 x i1> [[TMP12]], <vscale x 4 x i32> [[TMP16]], <vscale x 4 x i32> [[VEC_PHI]]
; IF-EVL-NEXT: [[INDEX_NEXT]] = add i64 [[INDEX]], [[TMP8]]
; IF-EVL-NEXT: [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; IF-EVL-NEXT: br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; IF-EVL: middle.block:
; IF-EVL-NEXT: [[TMP19:%.*]] = call i32 @llvm.vector.reduce.smin.nxv4i32(<vscale x 4 x i32> [[TMP17]])
; IF-EVL-NEXT: br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
; IF-EVL: scalar.ph:
; IF-EVL-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
; IF-EVL-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP19]], [[MIDDLE_BLOCK]] ], [ [[START]], [[ENTRY]] ]
; IF-EVL-NEXT: br label [[FOR_BODY:%.*]]
; IF-EVL: for.body:
; IF-EVL-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY]] ]
; IF-EVL-NEXT: [[RDX:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[SMIN:%.*]], [[FOR_BODY]] ]
; IF-EVL-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV]]
; IF-EVL-NEXT: [[TMP20:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; IF-EVL-NEXT: [[CMP_I:%.*]] = icmp slt i32 [[TMP20]], [[RDX]]
; IF-EVL-NEXT: [[SMIN]] = select i1 [[CMP_I]], i32 [[TMP20]], i32 [[RDX]]
; IF-EVL-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
; IF-EVL-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
; IF-EVL-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
; IF-EVL: for.end:
; IF-EVL-NEXT: [[SMIN_LCSSA:%.*]] = phi i32 [ [[SMIN]], [[FOR_BODY]] ], [ [[TMP19]], [[MIDDLE_BLOCK]] ]
; IF-EVL-NEXT: ret i32 [[SMIN_LCSSA]]
;
entry:
br label %for.body

for.body:
%iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
%rdx = phi i32 [ %start, %entry ], [ %smin, %for.body ]
%arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
%0 = load i32, ptr %arrayidx, align 4
%cmp.i = icmp slt i32 %0, %rdx
%smin = select i1 %cmp.i, i32 %0, i32 %rdx
%iv.next = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %iv.next, %n
br i1 %exitcond.not, label %for.end, label %for.body

for.end:
ret i32 %smin
}
Loading