Skip to content

Commit 80dc27e

Browse files
committed
!fixup fix isDereferenceableAndAlignedInLoop.
1 parent 734f683 commit 80dc27e

File tree

5 files changed

+93
-33
lines changed

5 files changed

+93
-33
lines changed

llvm/include/llvm/Analysis/LoopAccessAnalysis.h

Lines changed: 5 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -857,7 +857,10 @@ bool sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy, const DataLayout &DL,
857857
bool isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
858858
ScalarEvolution &SE, bool CheckType = true);
859859

860-
/// Calculate Start and End points of memory access.
860+
/// Calculate Start and End points of memory access using exact backedge taken
861+
/// count \p BTC if computable or maximum backedge taken count \p MaxBTC
862+
/// otherwise.
863+
///
861864
/// Let's assume A is the first access and B is a memory access on N-th loop
862865
/// iteration. Then B is calculated as:
863866
/// B = A + Step*N .
@@ -872,7 +875,7 @@ bool isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
872875
/// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
873876
std::pair<const SCEV *, const SCEV *> getStartAndEndForAccess(
874877
const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy, const SCEV *BTC,
875-
const SCEV *SymbolicMaxBTC, ScalarEvolution *SE,
878+
const SCEV *MaxBTC, ScalarEvolution *SE,
876879
DenseMap<std::pair<const SCEV *, Type *>,
877880
std::pair<const SCEV *, const SCEV *>> *PointerBounds);
878881

llvm/lib/Analysis/Loads.cpp

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -319,14 +319,14 @@ bool llvm::isDereferenceableAndAlignedInLoop(
319319
const SCEV *MaxBECount =
320320
Predicates ? SE.getPredicatedConstantMaxBackedgeTakenCount(L, *Predicates)
321321
: SE.getConstantMaxBackedgeTakenCount(L);
322-
const SCEV *SymbolicMaxBECount =
323-
Predicates ? SE.getPredicatedConstantMaxBackedgeTakenCount(L, *Predicates)
324-
: SE.getConstantMaxBackedgeTakenCount(L);
322+
const SCEV *BECount = Predicates
323+
? SE.getPredicatedBackedgeTakenCount(L, *Predicates)
324+
: SE.getBackedgeTakenCount(L);
325325
if (isa<SCEVCouldNotCompute>(MaxBECount))
326326
return false;
327327

328328
const auto &[AccessStart, AccessEnd] = getStartAndEndForAccess(
329-
L, PtrScev, LI->getType(), MaxBECount, SymbolicMaxBECount, &SE, nullptr);
329+
L, PtrScev, LI->getType(), BECount, MaxBECount, &SE, nullptr);
330330
if (isa<SCEVCouldNotCompute>(AccessStart) ||
331331
isa<SCEVCouldNotCompute>(AccessEnd))
332332
return false;

llvm/lib/Analysis/LoopAccessAnalysis.cpp

Lines changed: 4 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -190,7 +190,7 @@ RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup(
190190

191191
std::pair<const SCEV *, const SCEV *> llvm::getStartAndEndForAccess(
192192
const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy, const SCEV *BTC,
193-
const SCEV *SymbolicMaxBTC, ScalarEvolution *SE,
193+
const SCEV *MaxBTC, ScalarEvolution *SE,
194194
DenseMap<std::pair<const SCEV *, Type *>,
195195
std::pair<const SCEV *, const SCEV *>> *PointerBounds) {
196196
std::pair<const SCEV *, const SCEV *> *PtrBoundsPair;
@@ -226,7 +226,9 @@ std::pair<const SCEV *, const SCEV *> llvm::getStartAndEndForAccess(
226226
// will get incremented by EltSize before returning, so this effectively
227227
// sets ScEnd to unsigned max. Note that LAA separately checks that
228228
// accesses cannot not wrap, so unsigned max represents an upper bound.
229-
ScEnd = AR->evaluateAtIteration(SymbolicMaxBTC, *SE);
229+
// TODO: Use additional information to determine no-wrap including
230+
// size/dereferencability info from the accessed object.
231+
ScEnd = AR->evaluateAtIteration(MaxBTC, *SE);
230232
if (!SE->isKnownNonNegative(SE->getMinusSCEV(ScEnd, ScStart)))
231233
ScEnd = SE->getNegativeSCEV(
232234
SE->getAddExpr(EltSizeSCEV, SE->getOne(EltSizeSCEV->getType())));

llvm/test/Transforms/LoopVectorize/dereferenceable-info-from-assumption-variable-size.ll

Lines changed: 80 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -22,8 +22,26 @@ define void @deref_assumption_in_preheader_non_constant_trip_count_access_i8(ptr
2222
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[TMP2]], i32 0
2323
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i8>, ptr [[TMP3]], align 1
2424
; CHECK-NEXT: [[TMP4:%.*]] = icmp sge <2 x i8> [[WIDE_LOAD]], zeroinitializer
25-
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[TMP1]], i32 0
26-
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x i8>, ptr [[TMP5]], align 1
25+
; CHECK-NEXT: [[TMP15:%.*]] = xor <2 x i1> [[TMP4]], splat (i1 true)
26+
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i1> [[TMP15]], i32 0
27+
; CHECK-NEXT: br i1 [[TMP5]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
28+
; CHECK: [[PRED_LOAD_IF]]:
29+
; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP0]]
30+
; CHECK-NEXT: [[TMP17:%.*]] = load i8, ptr [[TMP16]], align 1
31+
; CHECK-NEXT: [[TMP18:%.*]] = insertelement <2 x i8> poison, i8 [[TMP17]], i32 0
32+
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE]]
33+
; CHECK: [[PRED_LOAD_CONTINUE]]:
34+
; CHECK-NEXT: [[TMP9:%.*]] = phi <2 x i8> [ poison, %[[VECTOR_BODY]] ], [ [[TMP18]], %[[PRED_LOAD_IF]] ]
35+
; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x i1> [[TMP15]], i32 1
36+
; CHECK-NEXT: br i1 [[TMP10]], label %[[PRED_LOAD_IF1:.*]], label %[[PRED_LOAD_CONTINUE2]]
37+
; CHECK: [[PRED_LOAD_IF1]]:
38+
; CHECK-NEXT: [[TMP11:%.*]] = add i64 [[INDEX]], 1
39+
; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP11]]
40+
; CHECK-NEXT: [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 1
41+
; CHECK-NEXT: [[TMP14:%.*]] = insertelement <2 x i8> [[TMP9]], i8 [[TMP13]], i32 1
42+
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE2]]
43+
; CHECK: [[PRED_LOAD_CONTINUE2]]:
44+
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = phi <2 x i8> [ [[TMP9]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP14]], %[[PRED_LOAD_IF1]] ]
2745
; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP4]], <2 x i8> [[WIDE_LOAD]], <2 x i8> [[WIDE_LOAD1]]
2846
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8, ptr [[C]], i64 [[TMP0]]
2947
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i8, ptr [[TMP6]], i32 0
@@ -105,8 +123,26 @@ define void @deref_assumption_in_preheader_non_constant_trip_count_access_i32(pt
105123
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TMP2]], i32 0
106124
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP3]], align 1
107125
; CHECK-NEXT: [[TMP4:%.*]] = icmp sge <2 x i32> [[WIDE_LOAD]], zeroinitializer
108-
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i32, ptr [[TMP1]], i32 0
109-
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x i32>, ptr [[TMP5]], align 1
126+
; CHECK-NEXT: [[TMP15:%.*]] = xor <2 x i1> [[TMP4]], splat (i1 true)
127+
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i1> [[TMP15]], i32 0
128+
; CHECK-NEXT: br i1 [[TMP5]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
129+
; CHECK: [[PRED_LOAD_IF]]:
130+
; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP0]]
131+
; CHECK-NEXT: [[TMP17:%.*]] = load i32, ptr [[TMP16]], align 1
132+
; CHECK-NEXT: [[TMP18:%.*]] = insertelement <2 x i32> poison, i32 [[TMP17]], i32 0
133+
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE]]
134+
; CHECK: [[PRED_LOAD_CONTINUE]]:
135+
; CHECK-NEXT: [[TMP9:%.*]] = phi <2 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP18]], %[[PRED_LOAD_IF]] ]
136+
; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x i1> [[TMP15]], i32 1
137+
; CHECK-NEXT: br i1 [[TMP10]], label %[[PRED_LOAD_IF1:.*]], label %[[PRED_LOAD_CONTINUE2]]
138+
; CHECK: [[PRED_LOAD_IF1]]:
139+
; CHECK-NEXT: [[TMP11:%.*]] = add i64 [[INDEX]], 1
140+
; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP11]]
141+
; CHECK-NEXT: [[TMP13:%.*]] = load i32, ptr [[TMP12]], align 1
142+
; CHECK-NEXT: [[TMP14:%.*]] = insertelement <2 x i32> [[TMP9]], i32 [[TMP13]], i32 1
143+
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE2]]
144+
; CHECK: [[PRED_LOAD_CONTINUE2]]:
145+
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = phi <2 x i32> [ [[TMP9]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP14]], %[[PRED_LOAD_IF1]] ]
110146
; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP4]], <2 x i32> [[WIDE_LOAD]], <2 x i32> [[WIDE_LOAD1]]
111147
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[TMP0]]
112148
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[TMP6]], i32 0
@@ -189,8 +225,26 @@ define void @deref_assumption_in_preheader_too_small_non_constant_trip_count_acc
189225
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TMP2]], i32 0
190226
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP3]], align 1
191227
; CHECK-NEXT: [[TMP4:%.*]] = icmp sge <2 x i32> [[WIDE_LOAD]], zeroinitializer
192-
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i32, ptr [[TMP1]], i32 0
193-
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x i32>, ptr [[TMP5]], align 1
228+
; CHECK-NEXT: [[TMP15:%.*]] = xor <2 x i1> [[TMP4]], splat (i1 true)
229+
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i1> [[TMP15]], i32 0
230+
; CHECK-NEXT: br i1 [[TMP5]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
231+
; CHECK: [[PRED_LOAD_IF]]:
232+
; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP0]]
233+
; CHECK-NEXT: [[TMP17:%.*]] = load i32, ptr [[TMP16]], align 1
234+
; CHECK-NEXT: [[TMP18:%.*]] = insertelement <2 x i32> poison, i32 [[TMP17]], i32 0
235+
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE]]
236+
; CHECK: [[PRED_LOAD_CONTINUE]]:
237+
; CHECK-NEXT: [[TMP9:%.*]] = phi <2 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP18]], %[[PRED_LOAD_IF]] ]
238+
; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x i1> [[TMP15]], i32 1
239+
; CHECK-NEXT: br i1 [[TMP10]], label %[[PRED_LOAD_IF1:.*]], label %[[PRED_LOAD_CONTINUE2]]
240+
; CHECK: [[PRED_LOAD_IF1]]:
241+
; CHECK-NEXT: [[TMP11:%.*]] = add i64 [[INDEX]], 1
242+
; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP11]]
243+
; CHECK-NEXT: [[TMP13:%.*]] = load i32, ptr [[TMP12]], align 1
244+
; CHECK-NEXT: [[TMP14:%.*]] = insertelement <2 x i32> [[TMP9]], i32 [[TMP13]], i32 1
245+
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE2]]
246+
; CHECK: [[PRED_LOAD_CONTINUE2]]:
247+
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = phi <2 x i32> [ [[TMP9]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP14]], %[[PRED_LOAD_IF1]] ]
194248
; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP4]], <2 x i32> [[WIDE_LOAD]], <2 x i32> [[WIDE_LOAD1]]
195249
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[TMP0]]
196250
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[TMP6]], i32 0
@@ -271,8 +325,26 @@ define void @deref_assumption_in_preheader_too_small2_non_constant_trip_count_ac
271325
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TMP2]], i32 0
272326
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP3]], align 1
273327
; CHECK-NEXT: [[TMP4:%.*]] = icmp sge <2 x i32> [[WIDE_LOAD]], zeroinitializer
274-
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i32, ptr [[TMP1]], i32 0
275-
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x i32>, ptr [[TMP5]], align 1
328+
; CHECK-NEXT: [[TMP15:%.*]] = xor <2 x i1> [[TMP4]], splat (i1 true)
329+
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i1> [[TMP15]], i32 0
330+
; CHECK-NEXT: br i1 [[TMP5]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
331+
; CHECK: [[PRED_LOAD_IF]]:
332+
; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP0]]
333+
; CHECK-NEXT: [[TMP17:%.*]] = load i32, ptr [[TMP16]], align 1
334+
; CHECK-NEXT: [[TMP18:%.*]] = insertelement <2 x i32> poison, i32 [[TMP17]], i32 0
335+
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE]]
336+
; CHECK: [[PRED_LOAD_CONTINUE]]:
337+
; CHECK-NEXT: [[TMP9:%.*]] = phi <2 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP18]], %[[PRED_LOAD_IF]] ]
338+
; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x i1> [[TMP15]], i32 1
339+
; CHECK-NEXT: br i1 [[TMP10]], label %[[PRED_LOAD_IF1:.*]], label %[[PRED_LOAD_CONTINUE2]]
340+
; CHECK: [[PRED_LOAD_IF1]]:
341+
; CHECK-NEXT: [[TMP11:%.*]] = add i64 [[INDEX]], 1
342+
; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP11]]
343+
; CHECK-NEXT: [[TMP13:%.*]] = load i32, ptr [[TMP12]], align 1
344+
; CHECK-NEXT: [[TMP14:%.*]] = insertelement <2 x i32> [[TMP9]], i32 [[TMP13]], i32 1
345+
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE2]]
346+
; CHECK: [[PRED_LOAD_CONTINUE2]]:
347+
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = phi <2 x i32> [ [[TMP9]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP14]], %[[PRED_LOAD_IF1]] ]
276348
; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP4]], <2 x i32> [[WIDE_LOAD]], <2 x i32> [[WIDE_LOAD1]]
277349
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[TMP0]]
278350
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[TMP6]], i32 0

llvm/test/Transforms/LoopVectorize/single_early_exit_live_outs.ll

Lines changed: 0 additions & 17 deletions
Original file line numberDiff line numberDiff line change
@@ -2233,20 +2233,3 @@ attributes #0 = { "vector-function-abi-variant"="_ZGVsNxv_foo(foo_vec)" }
22332233
; CHECK: [[LOOP35]] = distinct !{[[LOOP35]], [[META2]], [[META1]]}
22342234
; CHECK: [[LOOP36]] = distinct !{[[LOOP36]], [[META1]], [[META2]]}
22352235
; CHECK: [[LOOP37]] = distinct !{[[LOOP37]], [[META2]], [[META1]]}
2236-
; CHECK: [[LOOP38]] = distinct !{[[LOOP38]], [[META1]], [[META2]]}
2237-
; CHECK: [[LOOP39]] = distinct !{[[LOOP39]], [[META2]], [[META1]]}
2238-
; CHECK: [[LOOP40]] = distinct !{[[LOOP40]], [[META1]], [[META2]]}
2239-
; CHECK: [[LOOP41]] = distinct !{[[LOOP41]], [[META2]], [[META1]]}
2240-
; CHECK: [[LOOP42]] = distinct !{[[LOOP42]], [[META1]], [[META2]]}
2241-
; CHECK: [[LOOP43]] = distinct !{[[LOOP43]], [[META2]], [[META1]]}
2242-
; CHECK: [[LOOP44]] = distinct !{[[LOOP44]], [[META1]], [[META2]]}
2243-
; CHECK: [[LOOP45]] = distinct !{[[LOOP45]], [[META2]], [[META1]]}
2244-
; CHECK: [[LOOP46]] = distinct !{[[LOOP46]], [[META1]], [[META2]]}
2245-
; CHECK: [[LOOP47]] = distinct !{[[LOOP47]], [[META2]], [[META1]]}
2246-
; CHECK: [[LOOP48]] = distinct !{[[LOOP48]], [[META1]], [[META2]]}
2247-
; CHECK: [[LOOP49]] = distinct !{[[LOOP49]], [[META2]], [[META1]]}
2248-
; CHECK: [[LOOP50]] = distinct !{[[LOOP50]], [[META1]], [[META2]]}
2249-
; CHECK: [[LOOP51]] = distinct !{[[LOOP51]], [[META2]], [[META1]]}
2250-
; CHECK: [[LOOP52]] = distinct !{[[LOOP52]], [[META1]], [[META2]]}
2251-
; CHECK: [[LOOP53]] = distinct !{[[LOOP53]], [[META2]], [[META1]]}
2252-
;.

0 commit comments

Comments
 (0)