Skip to content

Commit 050b588

Browse files
ElvisWang123ElvisWang123
committed
[LV][VPlan] Implement VPlan-based cost for exit condition.
This patch tried to model the cost of exit conditions through vplan-based cost model. * `BranchOnCount` will generate icmp + br. The branch instruction is already implemented by the VPRegionBlock so we only need to calculate the cost of icmp. If the VF is same as the trip count of the loop, the cost of the BranchOnCount is free. This patch is not quite NFC for following reasons. * Some of the BranchOnCount could be optimized to BranchOnCond, which is free. * Some of the instructions calculated in the exit condition in legacy cost model will be optimized out.
1 parent fc4210f commit 050b588

14 files changed

+292
-284
lines changed

‎llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

Lines changed: 0 additions & 39 deletions
Original file line numberDiff line numberDiff line change
@@ -7264,45 +7264,6 @@ LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,
72647264
}
72657265
}
72667266

7267-
/// Compute the cost of all exiting conditions of the loop using the legacy
7268-
/// cost model. This is to match the legacy behavior, which adds the cost of
7269-
/// all exit conditions. Note that this over-estimates the cost, as there will
7270-
/// be a single condition to control the vector loop.
7271-
SmallVector<BasicBlock *> Exiting;
7272-
CM.TheLoop->getExitingBlocks(Exiting);
7273-
SetVector<Instruction *> ExitInstrs;
7274-
// Collect all exit conditions.
7275-
for (BasicBlock *EB : Exiting) {
7276-
auto *Term = dyn_cast<BranchInst>(EB->getTerminator());
7277-
if (!Term)
7278-
continue;
7279-
if (auto *CondI = dyn_cast<Instruction>(Term->getOperand(0))) {
7280-
ExitInstrs.insert(CondI);
7281-
}
7282-
}
7283-
// Compute the cost of all instructions only feeding the exit conditions.
7284-
for (unsigned I = 0; I != ExitInstrs.size(); ++I) {
7285-
Instruction *CondI = ExitInstrs[I];
7286-
if (!OrigLoop->contains(CondI) ||
7287-
!CostCtx.SkipCostComputation.insert(CondI).second)
7288-
continue;
7289-
InstructionCost CondICost = CostCtx.getLegacyCost(CondI, VF);
7290-
LLVM_DEBUG({
7291-
dbgs() << "Cost of " << CondICost << " for VF " << VF
7292-
<< ": exit condition instruction " << *CondI << "\n";
7293-
});
7294-
Cost += CondICost;
7295-
for (Value *Op : CondI->operands()) {
7296-
auto *OpI = dyn_cast<Instruction>(Op);
7297-
if (!OpI || any_of(OpI->users(), [&ExitInstrs, this](User *U) {
7298-
return OrigLoop->contains(cast<Instruction>(U)->getParent()) &&
7299-
!ExitInstrs.contains(cast<Instruction>(U));
7300-
}))
7301-
continue;
7302-
ExitInstrs.insert(OpI);
7303-
}
7304-
}
7305-
73067267
// The legacy cost model has special logic to compute the cost of in-loop
73077268
// reductions, which may be smaller than the sum of all instructions involved
73087269
// in the reduction.

‎llvm/lib/Transforms/Vectorize/VPlan.h

Lines changed: 1 addition & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -972,10 +972,7 @@ class VPInstruction : public VPRecipeWithIRFlags,
972972

973973
/// Return the cost of this VPInstruction.
974974
InstructionCost computeCost(ElementCount VF,
975-
VPCostContext &Ctx) const override {
976-
// TODO: Compute accurate cost after retiring the legacy cost model.
977-
return 0;
978-
}
975+
VPCostContext &Ctx) const override;
979976

980977
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
981978
/// Print the VPInstruction to \p O.

‎llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

Lines changed: 23 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -834,6 +834,29 @@ bool VPInstruction::onlyFirstPartUsed(const VPValue *Op) const {
834834
llvm_unreachable("switch should return");
835835
}
836836

837+
InstructionCost VPInstruction::computeCost(ElementCount VF,
838+
VPCostContext &Ctx) const {
839+
Type *ValTy = Ctx.Types.inferScalarType(getOperand(0));
840+
841+
switch (getOpcode()) {
842+
case VPInstruction::BranchOnCount: {
843+
// BranchOnCount will genearte icmp_eq + br instructions and the
844+
// cost of branch will be calculated in VPRegionBlock.
845+
// If the vector loop only executed once, ignore the cost of the cmp.
846+
auto TC = dyn_cast_if_present<ConstantInt>(
847+
getParent()->getPlan()->getTripCount()->getUnderlyingValue());
848+
if (TC && VF.isFixed() && TC->getSExtValue() == VF.getFixedValue())
849+
return 0;
850+
return Ctx.TTI.getCmpSelInstrCost(Instruction::ICmp, ValTy, nullptr,
851+
CmpInst::ICMP_EQ, Ctx.CostKind);
852+
}
853+
default:
854+
// TODO: Compute accurate cost after retiring the legacy cost model.
855+
return 0;
856+
};
857+
llvm_unreachable("switch should return");
858+
}
859+
837860
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
838861
void VPInstruction::dump() const {
839862
VPSlotTracker SlotTracker(getParent()->getPlan());

‎llvm/test/Transforms/LoopVectorize/AArch64/clamped-trip-count.ll

Lines changed: 42 additions & 38 deletions
Original file line numberDiff line numberDiff line change
@@ -8,39 +8,41 @@ define void @clamped_tc_8(ptr nocapture %dst, i32 %n, i64 %val) vscale_range(1,1
88
; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
99
; CHECK: vector.ph:
1010
; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
11-
; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 8
11+
; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 2
1212
; CHECK-NEXT: [[TMP4:%.*]] = sub i64 [[TMP1]], 1
1313
; CHECK-NEXT: [[N_RND_UP:%.*]] = add i64 8, [[TMP4]]
1414
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
1515
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
1616
; CHECK-NEXT: [[TMP5:%.*]] = call i64 @llvm.vscale.i64()
17-
; CHECK-NEXT: [[TMP6:%.*]] = mul i64 [[TMP5]], 8
17+
; CHECK-NEXT: [[TMP6:%.*]] = mul i64 [[TMP5]], 2
1818
; CHECK-NEXT: [[IND_END:%.*]] = getelementptr i8, ptr [[DST]], i64 [[N_VEC]]
19-
; CHECK-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 8 x i1> @llvm.get.active.lane.mask.nxv8i1.i64(i64 0, i64 8)
20-
; CHECK-NEXT: [[TMP8:%.*]] = call <vscale x 8 x i64> @llvm.stepvector.nxv8i64()
21-
; CHECK-NEXT: [[TMP7:%.*]] = mul <vscale x 8 x i64> [[TMP8]], splat (i64 1)
22-
; CHECK-NEXT: [[INDUCTION:%.*]] = add <vscale x 8 x i64> zeroinitializer, [[TMP7]]
19+
; CHECK-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 2 x i1> @llvm.get.active.lane.mask.nxv2i1.i64(i64 0, i64 8)
20+
; CHECK-NEXT: [[TMP8:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
21+
; CHECK-NEXT: [[TMP7:%.*]] = mul <vscale x 2 x i64> [[TMP8]], splat (i64 1)
22+
; CHECK-NEXT: [[INDUCTION:%.*]] = add <vscale x 2 x i64> zeroinitializer, [[TMP7]]
2323
; CHECK-NEXT: [[TMP12:%.*]] = mul i64 1, [[TMP6]]
24-
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 8 x i64> poison, i64 [[TMP12]], i64 0
25-
; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 8 x i64> [[DOTSPLATINSERT]], <vscale x 8 x i64> poison, <vscale x 8 x i32> zeroinitializer
26-
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 8 x i64> poison, i64 [[VAL]], i64 0
27-
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 8 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 8 x i64> poison, <vscale x 8 x i32> zeroinitializer
24+
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP12]], i64 0
25+
; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[DOTSPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
26+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[VAL]], i64 0
27+
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
2828
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
2929
; CHECK: vector.body:
3030
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
31-
; CHECK-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 8 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[VECTOR_PH]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]
32-
; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 8 x i64> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
31+
; CHECK-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 2 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[VECTOR_PH]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]
32+
; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
3333
; CHECK-NEXT: [[TMP13:%.*]] = add i64 [[INDEX]], 0
3434
; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[DST]], i64 [[TMP13]]
35-
; CHECK-NEXT: [[TMP10:%.*]] = shl nuw nsw <vscale x 8 x i64> [[VEC_IND]], splat (i64 3)
36-
; CHECK-NEXT: [[TMP11:%.*]] = lshr <vscale x 8 x i64> [[BROADCAST_SPLAT]], [[TMP10]]
37-
; CHECK-NEXT: [[TMP14:%.*]] = trunc <vscale x 8 x i64> [[TMP11]] to <vscale x 8 x i8>
35+
; CHECK-NEXT: [[TMP10:%.*]] = shl nuw nsw <vscale x 2 x i64> [[VEC_IND]], splat (i64 3)
36+
; CHECK-NEXT: [[TMP11:%.*]] = lshr <vscale x 2 x i64> [[BROADCAST_SPLAT]], [[TMP10]]
37+
; CHECK-NEXT: [[TMP16:%.*]] = trunc <vscale x 2 x i64> [[TMP11]] to <vscale x 2 x i8>
3838
; CHECK-NEXT: [[TMP17:%.*]] = getelementptr i8, ptr [[NEXT_GEP]], i32 0
39-
; CHECK-NEXT: call void @llvm.masked.store.nxv8i8.p0(<vscale x 8 x i8> [[TMP14]], ptr [[TMP17]], i32 1, <vscale x 8 x i1> [[ACTIVE_LANE_MASK]])
39+
; CHECK-NEXT: call void @llvm.masked.store.nxv2i8.p0(<vscale x 2 x i8> [[TMP16]], ptr [[TMP17]], i32 1, <vscale x 2 x i1> [[ACTIVE_LANE_MASK]])
4040
; CHECK-NEXT: [[INDEX_NEXT]] = add i64 [[INDEX]], [[TMP6]]
41-
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 8 x i64> [[VEC_IND]], [[DOTSPLAT]]
42-
; CHECK-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 8 x i1> @llvm.get.active.lane.mask.nxv8i1.i64(i64 [[INDEX_NEXT]], i64 8)
43-
; CHECK-NEXT: br i1 true, label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
41+
; CHECK-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 2 x i1> @llvm.get.active.lane.mask.nxv2i1.i64(i64 [[INDEX_NEXT]], i64 8)
42+
; CHECK-NEXT: [[TMP14:%.*]] = xor <vscale x 2 x i1> [[ACTIVE_LANE_MASK_NEXT]], splat (i1 true)
43+
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 2 x i64> [[VEC_IND]], [[DOTSPLAT]]
44+
; CHECK-NEXT: [[TMP15:%.*]] = extractelement <vscale x 2 x i1> [[TMP14]], i32 0
45+
; CHECK-NEXT: br i1 [[TMP15]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
4446
; CHECK: middle.block:
4547
; CHECK-NEXT: br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
4648
; CHECK: scalar.ph:
@@ -94,39 +96,41 @@ define void @clamped_tc_max_8(ptr nocapture %dst, i32 %n, i64 %val) vscale_range
9496
; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
9597
; CHECK: vector.ph:
9698
; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
97-
; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 8
99+
; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 2
98100
; CHECK-NEXT: [[TMP4:%.*]] = sub i64 [[TMP1]], 1
99101
; CHECK-NEXT: [[N_RND_UP:%.*]] = add i64 [[WIDE_TRIP_COUNT]], [[TMP4]]
100102
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
101103
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
102104
; CHECK-NEXT: [[TMP5:%.*]] = call i64 @llvm.vscale.i64()
103-
; CHECK-NEXT: [[TMP6:%.*]] = mul i64 [[TMP5]], 8
105+
; CHECK-NEXT: [[TMP6:%.*]] = mul i64 [[TMP5]], 2
104106
; CHECK-NEXT: [[IND_END:%.*]] = getelementptr i8, ptr [[DST]], i64 [[N_VEC]]
105-
; CHECK-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 8 x i1> @llvm.get.active.lane.mask.nxv8i1.i64(i64 0, i64 [[WIDE_TRIP_COUNT]])
106-
; CHECK-NEXT: [[TMP8:%.*]] = call <vscale x 8 x i64> @llvm.stepvector.nxv8i64()
107-
; CHECK-NEXT: [[TMP7:%.*]] = mul <vscale x 8 x i64> [[TMP8]], splat (i64 1)
108-
; CHECK-NEXT: [[INDUCTION:%.*]] = add <vscale x 8 x i64> zeroinitializer, [[TMP7]]
107+
; CHECK-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 2 x i1> @llvm.get.active.lane.mask.nxv2i1.i64(i64 0, i64 [[WIDE_TRIP_COUNT]])
108+
; CHECK-NEXT: [[TMP8:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
109+
; CHECK-NEXT: [[TMP7:%.*]] = mul <vscale x 2 x i64> [[TMP8]], splat (i64 1)
110+
; CHECK-NEXT: [[INDUCTION:%.*]] = add <vscale x 2 x i64> zeroinitializer, [[TMP7]]
109111
; CHECK-NEXT: [[TMP12:%.*]] = mul i64 1, [[TMP6]]
110-
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 8 x i64> poison, i64 [[TMP12]], i64 0
111-
; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 8 x i64> [[DOTSPLATINSERT]], <vscale x 8 x i64> poison, <vscale x 8 x i32> zeroinitializer
112-
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 8 x i64> poison, i64 [[VAL]], i64 0
113-
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 8 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 8 x i64> poison, <vscale x 8 x i32> zeroinitializer
112+
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP12]], i64 0
113+
; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[DOTSPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
114+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[VAL]], i64 0
115+
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
114116
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
115117
; CHECK: vector.body:
116118
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
117-
; CHECK-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 8 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[VECTOR_PH]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]
118-
; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 8 x i64> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
119+
; CHECK-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 2 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[VECTOR_PH]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]
120+
; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
119121
; CHECK-NEXT: [[TMP13:%.*]] = add i64 [[INDEX]], 0
120122
; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[DST]], i64 [[TMP13]]
121-
; CHECK-NEXT: [[TMP10:%.*]] = shl nuw nsw <vscale x 8 x i64> [[VEC_IND]], splat (i64 3)
122-
; CHECK-NEXT: [[TMP11:%.*]] = lshr <vscale x 8 x i64> [[BROADCAST_SPLAT]], [[TMP10]]
123-
; CHECK-NEXT: [[TMP14:%.*]] = trunc <vscale x 8 x i64> [[TMP11]] to <vscale x 8 x i8>
123+
; CHECK-NEXT: [[TMP10:%.*]] = shl nuw nsw <vscale x 2 x i64> [[VEC_IND]], splat (i64 3)
124+
; CHECK-NEXT: [[TMP11:%.*]] = lshr <vscale x 2 x i64> [[BROADCAST_SPLAT]], [[TMP10]]
125+
; CHECK-NEXT: [[TMP16:%.*]] = trunc <vscale x 2 x i64> [[TMP11]] to <vscale x 2 x i8>
124126
; CHECK-NEXT: [[TMP17:%.*]] = getelementptr i8, ptr [[NEXT_GEP]], i32 0
125-
; CHECK-NEXT: call void @llvm.masked.store.nxv8i8.p0(<vscale x 8 x i8> [[TMP14]], ptr [[TMP17]], i32 1, <vscale x 8 x i1> [[ACTIVE_LANE_MASK]])
127+
; CHECK-NEXT: call void @llvm.masked.store.nxv2i8.p0(<vscale x 2 x i8> [[TMP16]], ptr [[TMP17]], i32 1, <vscale x 2 x i1> [[ACTIVE_LANE_MASK]])
126128
; CHECK-NEXT: [[INDEX_NEXT]] = add i64 [[INDEX]], [[TMP6]]
127-
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 8 x i64> [[VEC_IND]], [[DOTSPLAT]]
128-
; CHECK-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 8 x i1> @llvm.get.active.lane.mask.nxv8i1.i64(i64 [[INDEX_NEXT]], i64 [[WIDE_TRIP_COUNT]])
129-
; CHECK-NEXT: br i1 true, label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
129+
; CHECK-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 2 x i1> @llvm.get.active.lane.mask.nxv2i1.i64(i64 [[INDEX_NEXT]], i64 [[WIDE_TRIP_COUNT]])
130+
; CHECK-NEXT: [[TMP14:%.*]] = xor <vscale x 2 x i1> [[ACTIVE_LANE_MASK_NEXT]], splat (i1 true)
131+
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 2 x i64> [[VEC_IND]], [[DOTSPLAT]]
132+
; CHECK-NEXT: [[TMP15:%.*]] = extractelement <vscale x 2 x i1> [[TMP14]], i32 0
133+
; CHECK-NEXT: br i1 [[TMP15]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
130134
; CHECK: middle.block:
131135
; CHECK-NEXT: br i1 true, label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]], label [[SCALAR_PH]]
132136
; CHECK: scalar.ph:

0 commit comments

Comments
 (0)