Skip to content

Commit

Permalink
[LV] Ignore live-out users in cost model if scalar epilogue is required.
Browse files Browse the repository at this point in the history 
Follow-up to ba8126b.

If a scalar epilogue is required, users outside the loop won't use
live-outs from the vector loop but from the scalar epilogue. Ignore them if
that is the case.

This fixes another case where the VPlan-based cost-model more accurately
computes cost.

Fixes llvm#100464.
  • Loading branch information
@fhahn
fhahn committed Jul 25, 2024
1 parent ba461f8 commit b72689a
Show file tree
Hide file tree
Showing 3 changed files with 119 additions and 4 deletions.
24 changes: 20 additions & 4 deletions llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6678,6 +6678,15 @@ void LoopVectorizationCostModel::collectValuesToIgnore() {

SmallVector<Value *, 4> DeadInterleavePointerOps;
SmallVector<Value *, 4> DeadOps;

// If a scalar epilogue is required, users outside the loop won't use
// live-outs from the vector loop but from the scalar epilogue. Ignore them if
// that is the case.
bool RequiresScalarEpilogue = requiresScalarEpilogue(true);
auto IsLiveOutDead = [this, RequiresScalarEpilogue](User *U) {
return RequiresScalarEpilogue &&
!TheLoop->contains(cast<Instruction>(U)->getParent());
};
for (BasicBlock *BB : TheLoop->blocks())
for (Instruction &I : *BB) {
// Find all stores to invariant variables. Since they are going to sink
Expand All @@ -6693,8 +6702,9 @@ void LoopVectorizationCostModel::collectValuesToIgnore() {
// Add instructions that would be trivially dead and are only used by
// values already ignored to DeadOps to seed worklist.
if (wouldInstructionBeTriviallyDead(&I, TLI) &&
all_of(I.users(), [this](User *U) {
return VecValuesToIgnore.contains(U) || ValuesToIgnore.contains(U);
all_of(I.users(), [this, IsLiveOutDead](User *U) {
return VecValuesToIgnore.contains(U) ||
ValuesToIgnore.contains(U) || IsLiveOutDead(U);
}))
DeadOps.push_back(&I);

Expand Down Expand Up @@ -6727,16 +6737,22 @@ void LoopVectorizationCostModel::collectValuesToIgnore() {

// Mark ops that would be trivially dead and are only used by ignored
// instructions as free.
BasicBlock *Header = TheLoop->getHeader();
for (unsigned I = 0; I != DeadOps.size(); ++I) {
auto *Op = dyn_cast<Instruction>(DeadOps[I]);
// Skip any op that shouldn't be considered dead.
if (!Op || !TheLoop->contains(Op) ||
(isa<PHINode>(Op) && Op->getParent() == Header) ||
!wouldInstructionBeTriviallyDead(Op, TLI) ||
any_of(Op->users(), [this](User *U) {
return !VecValuesToIgnore.contains(U) && !ValuesToIgnore.contains(U);
any_of(Op->users(), [this, IsLiveOutDead](User *U) {
return !VecValuesToIgnore.contains(U) && ValuesToIgnore.contains(U) &&
!IsLiveOutDead(U);
}))
continue;

if (!TheLoop->contains(Op->getParent()))
continue;

// If all of Op's users are in ValuesToIgnore, add it to ValuesToIgnore
// which applies for both scalar and vector versions. Otherwise it is only
// dead in vector versions, so only add it to VecValuesToIgnore.
Expand Down
97 changes: 97 additions & 0 deletions llvm/test/Transforms/LoopVectorize/RISCV/dead-ops-cost.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -86,9 +86,106 @@ loop:
exit:
ret void
}

; Test case for https://github.com/llvm/llvm-project/issues/100464.
; Loop with a live-out %l and scalar epilogue required due to an interleave
; group. As the scalar epilogue is required the live-out is fed from the scalar
; epilogue and dead in the vector loop.
define i8 @dead_live_out_due_to_scalar_epilogue_required(ptr %src, ptr %dst) {
; CHECK-LABEL: define i8 @dead_live_out_due_to_scalar_epilogue_required(
; CHECK-SAME: ptr [[SRC:%.*]], ptr [[DST:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[ENTRY:.*]]:
; CHECK-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()
; CHECK-NEXT: [[TMP1:%.*]] = mul i32 [[TMP0]], 4
; CHECK-NEXT: [[TMP2:%.*]] = call i32 @llvm.umax.i32(i32 8, i32 [[TMP1]])
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i32 252, [[TMP2]]
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_MEMCHECK:.*]]
; CHECK: [[VECTOR_MEMCHECK]]:
; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[DST]], i64 1005
; CHECK-NEXT: [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[SRC]], i64 1005
; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult ptr [[DST]], [[SCEVGEP1]]
; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[SRC]], [[SCEVGEP]]
; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
; CHECK-NEXT: br i1 [[FOUND_CONFLICT]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
; CHECK: [[VECTOR_PH]]:
; CHECK-NEXT: [[TMP3:%.*]] = call i32 @llvm.vscale.i32()
; CHECK-NEXT: [[TMP4:%.*]] = mul i32 [[TMP3]], 4
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 252, [[TMP4]]
; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[N_MOD_VF]], 0
; CHECK-NEXT: [[TMP6:%.*]] = select i1 [[TMP5]], i32 [[TMP4]], i32 [[N_MOD_VF]]
; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 252, [[TMP6]]
; CHECK-NEXT: [[IND_END:%.*]] = mul i32 [[N_VEC]], 4
; CHECK-NEXT: [[TMP7:%.*]] = call i32 @llvm.vscale.i32()
; CHECK-NEXT: [[TMP8:%.*]] = mul i32 [[TMP7]], 4
; CHECK-NEXT: [[TMP9:%.*]] = call <vscale x 4 x i32> @llvm.experimental.stepvector.nxv4i32()
; CHECK-NEXT: [[TMP10:%.*]] = add <vscale x 4 x i32> [[TMP9]], zeroinitializer
; CHECK-NEXT: [[TMP11:%.*]] = mul <vscale x 4 x i32> [[TMP10]], shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 4, i64 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer)
; CHECK-NEXT: [[INDUCTION:%.*]] = add <vscale x 4 x i32> zeroinitializer, [[TMP11]]
; CHECK-NEXT: [[TMP12:%.*]] = call i32 @llvm.vscale.i32()
; CHECK-NEXT: [[TMP13:%.*]] = mul i32 [[TMP12]], 4
; CHECK-NEXT: [[TMP14:%.*]] = mul i32 4, [[TMP13]]
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[TMP14]], i64 0
; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[DOTSPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 4 x i32> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP15:%.*]] = sext <vscale x 4 x i32> [[VEC_IND]] to <vscale x 4 x i64>
; CHECK-NEXT: [[TMP16:%.*]] = getelementptr i8, ptr [[DST]], <vscale x 4 x i64> [[TMP15]]
; CHECK-NEXT: call void @llvm.masked.scatter.nxv4i8.nxv4p0(<vscale x 4 x i8> zeroinitializer, <vscale x 4 x ptr> [[TMP16]], i32 1, <vscale x 4 x i1> shufflevector (<vscale x 4 x i1> insertelement (<vscale x 4 x i1> poison, i1 true, i64 0), <vscale x 4 x i1> poison, <vscale x 4 x i32> zeroinitializer)), !alias.scope [[META4:![0-9]+]], !noalias [[META7:![0-9]+]]
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP8]]
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 4 x i32> [[VEC_IND]], [[DOTSPLAT]]
; CHECK-NEXT: [[TMP17:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP17]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
; CHECK: [[MIDDLE_BLOCK]]:
; CHECK-NEXT: br label %[[SCALAR_PH]]
; CHECK: [[SCALAR_PH]]:
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_MEMCHECK]] ]
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
; CHECK-NEXT: [[IDXPROM:%.*]] = sext i32 [[IV]] to i64
; CHECK-NEXT: [[GEP_SRC:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[IDXPROM]]
; CHECK-NEXT: [[L:%.*]] = load i8, ptr [[GEP_SRC]], align 1
; CHECK-NEXT: [[GEP_DST:%.*]] = getelementptr i8, ptr [[DST]], i64 [[IDXPROM]]
; CHECK-NEXT: store i8 0, ptr [[GEP_DST]], align 1
; CHECK-NEXT: [[IV_NEXT]] = add i32 [[IV]], 4
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[IV]], 1001
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]], !llvm.loop [[LOOP10:![0-9]+]]
; CHECK: [[EXIT]]:
; CHECK-NEXT: [[R:%.*]] = phi i8 [ [[L]], %[[LOOP]] ]
; CHECK-NEXT: ret i8 [[R]]
;
entry:
br label %loop

loop:
%iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
%idxprom = sext i32 %iv to i64
%gep.src = getelementptr i8, ptr %src, i64 %idxprom
%l = load i8, ptr %gep.src, align 1
%gep.dst = getelementptr i8, ptr %dst, i64 %idxprom
store i8 0, ptr %gep.dst, align 1
%iv.next = add i32 %iv, 4
%cmp = icmp ult i32 %iv, 1001
br i1 %cmp, label %loop, label %exit

exit:
%r = phi i8 [ %l, %loop ]
ret i8 %r
}


;.
; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
; CHECK: [[META4]] = !{[[META5:![0-9]+]]}
; CHECK: [[META5]] = distinct !{[[META5]], [[META6:![0-9]+]]}
; CHECK: [[META6]] = distinct !{[[META6]], !"LVerDomain"}
; CHECK: [[META7]] = !{[[META8:![0-9]+]]}
; CHECK: [[META8]] = distinct !{[[META8]], [[META6]]}
; CHECK: [[LOOP9]] = distinct !{[[LOOP9]], [[META1]], [[META2]]}
; CHECK: [[LOOP10]] = distinct !{[[LOOP10]], [[META1]]}
;.
2 changes: 2 additions & 0 deletions llvm/test/Transforms/LoopVectorize/RISCV/riscv-vector-reverse.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@ define void @vector_reverse_i64(ptr nocapture noundef writeonly %A, ptr nocaptur
; CHECK-NEXT: LV: Found an induction variable.
; CHECK-NEXT: LV: Did not find one integer induction var.
; CHECK-NEXT: LV: We can vectorize this loop (with a runtime bound check)!
; CHECK-NEXT: LV: Loop does not require scalar epilogue
; CHECK-NEXT: LV: Found trip count: 0
; CHECK-NEXT: LV: Scalable vectorization is available
; CHECK-NEXT: LV: The max safe fixed VF is: 67108864.
Expand Down Expand Up @@ -222,6 +223,7 @@ define void @vector_reverse_f32(ptr nocapture noundef writeonly %A, ptr nocaptur
; CHECK-NEXT: LV: Found FP op with unsafe algebra.
; CHECK-NEXT: LV: Did not find one integer induction var.
; CHECK-NEXT: LV: We can vectorize this loop (with a runtime bound check)!
; CHECK-NEXT: LV: Loop does not require scalar epilogue
; CHECK-NEXT: LV: Found trip count: 0
; CHECK-NEXT: LV: Scalable vectorization is available
; CHECK-NEXT: LV: The max safe fixed VF is: 67108864.
Expand Down

0 comments on commit b72689a

Please sign in to comment.