Skip to content

Commit 102d168

Browse files
david-armdavid-arm
authored
[Analysis] Bail out for negative offsets in isDereferenceableAndAlignedInLoop (#99490)
This patch now bails out explicitly for negative offsets so that it's more consistent with the unsigned remainder and add calculations, and it fixes a genuine bug as shown with the new test.
1 parent f406d83 commit 102d168

File tree

2 files changed

+120
-15
lines changed

2 files changed

+120
-15
lines changed

‎llvm/lib/Analysis/Loads.cpp

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -313,6 +313,13 @@ bool llvm::isDereferenceableAndAlignedInLoop(LoadInst *LI, Loop *L,
313313
const auto *Offset = dyn_cast<SCEVConstant>(StartS->getOperand(0));
314314
const auto *NewBase = dyn_cast<SCEVUnknown>(StartS->getOperand(1));
315315
if (StartS->getNumOperands() == 2 && Offset && NewBase) {
316+
// The following code below assumes the offset is unsigned, but GEP
317+
// offsets are treated as signed so we can end up with a signed value
318+
// here too. For example, suppose the initial PHI value is (i8 255),
319+
// the offset will be treated as (i8 -1) and sign-extended to (i64 -1).
320+
if (Offset->getAPInt().isNegative())
321+
return false;
322+
316323
// For the moment, restrict ourselves to the case where the offset is a
317324
// multiple of the requested alignment and the base is aligned.
318325
// TODO: generalize if a case found which warrants

‎llvm/test/Transforms/LoopVectorize/load-deref-pred-align.ll

Lines changed: 113 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
11
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
22
; RUN: opt -passes=loop-vectorize -force-vector-width=2 -force-vector-interleave=1 -S %s | FileCheck %s
33

4-
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128-ni:1"
4+
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128-ni:1-p:16:16:16:16"
55

66
declare void @init(ptr nocapture nofree)
77

@@ -17,7 +17,7 @@ define i16 @test_access_size_not_multiple_of_align(i64 %len, ptr %test_base) {
1717
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
1818
; CHECK: vector.body:
1919
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_LOAD_CONTINUE2:%.*]] ]
20-
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i16> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP16:%.*]], [[PRED_LOAD_CONTINUE2]] ]
20+
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i16> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[PRED_LOAD_CONTINUE2]] ]
2121
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[INDEX]], 0
2222
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[TEST_BASE:%.*]], i64 [[TMP0]]
2323
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 0
@@ -43,16 +43,16 @@ define i16 @test_access_size_not_multiple_of_align(i64 %len, ptr %test_base) {
4343
; CHECK: pred.load.continue2:
4444
; CHECK-NEXT: [[TMP14:%.*]] = phi <2 x i16> [ [[TMP8]], [[PRED_LOAD_CONTINUE]] ], [ [[TMP13]], [[PRED_LOAD_IF1]] ]
4545
; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP3]], <2 x i16> [[TMP14]], <2 x i16> zeroinitializer
46-
; CHECK-NEXT: [[TMP16]] = add <2 x i16> [[VEC_PHI]], [[PREDPHI]]
46+
; CHECK-NEXT: [[TMP15]] = add <2 x i16> [[VEC_PHI]], [[PREDPHI]]
4747
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
48-
; CHECK-NEXT: [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
49-
; CHECK-NEXT: br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
48+
; CHECK-NEXT: [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
49+
; CHECK-NEXT: br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
5050
; CHECK: middle.block:
51-
; CHECK-NEXT: [[TMP18:%.*]] = call i16 @llvm.vector.reduce.add.v2i16(<2 x i16> [[TMP16]])
51+
; CHECK-NEXT: [[TMP17:%.*]] = call i16 @llvm.vector.reduce.add.v2i16(<2 x i16> [[TMP15]])
5252
; CHECK-NEXT: br i1 true, label [[LOOP_EXIT:%.*]], label [[SCALAR_PH]]
5353
; CHECK: scalar.ph:
5454
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
55-
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i16 [ [[TMP18]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
55+
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i16 [ [[TMP17]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
5656
; CHECK-NEXT: br label [[LOOP:%.*]]
5757
; CHECK: loop:
5858
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LATCH:%.*]] ]
@@ -72,7 +72,7 @@ define i16 @test_access_size_not_multiple_of_align(i64 %len, ptr %test_base) {
7272
; CHECK-NEXT: [[EXIT:%.*]] = icmp eq i64 [[IV]], 4095
7373
; CHECK-NEXT: br i1 [[EXIT]], label [[LOOP_EXIT]], label [[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
7474
; CHECK: loop_exit:
75-
; CHECK-NEXT: [[ACCUM_NEXT_LCSSA:%.*]] = phi i16 [ [[ACCUM_NEXT]], [[LATCH]] ], [ [[TMP18]], [[MIDDLE_BLOCK]] ]
75+
; CHECK-NEXT: [[ACCUM_NEXT_LCSSA:%.*]] = phi i16 [ [[ACCUM_NEXT]], [[LATCH]] ], [ [[TMP17]], [[MIDDLE_BLOCK]] ]
7676
; CHECK-NEXT: ret i16 [[ACCUM_NEXT_LCSSA]]
7777
;
7878
entry:
@@ -114,7 +114,7 @@ define i32 @test_access_size_multiple_of_align_but_offset_by_1(i64 %len, ptr %te
114114
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
115115
; CHECK: vector.body:
116116
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_LOAD_CONTINUE2:%.*]] ]
117-
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP16:%.*]], [[PRED_LOAD_CONTINUE2]] ]
117+
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[PRED_LOAD_CONTINUE2]] ]
118118
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[INDEX]], 0
119119
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[TEST_BASE:%.*]], i64 [[TMP0]]
120120
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 0
@@ -140,16 +140,16 @@ define i32 @test_access_size_multiple_of_align_but_offset_by_1(i64 %len, ptr %te
140140
; CHECK: pred.load.continue2:
141141
; CHECK-NEXT: [[TMP14:%.*]] = phi <2 x i32> [ [[TMP8]], [[PRED_LOAD_CONTINUE]] ], [ [[TMP13]], [[PRED_LOAD_IF1]] ]
142142
; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP3]], <2 x i32> [[TMP14]], <2 x i32> zeroinitializer
143-
; CHECK-NEXT: [[TMP16]] = add <2 x i32> [[VEC_PHI]], [[PREDPHI]]
143+
; CHECK-NEXT: [[TMP15]] = add <2 x i32> [[VEC_PHI]], [[PREDPHI]]
144144
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
145-
; CHECK-NEXT: [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
146-
; CHECK-NEXT: br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
145+
; CHECK-NEXT: [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
146+
; CHECK-NEXT: br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
147147
; CHECK: middle.block:
148-
; CHECK-NEXT: [[TMP18:%.*]] = call i32 @llvm.vector.reduce.add.v2i32(<2 x i32> [[TMP16]])
148+
; CHECK-NEXT: [[TMP17:%.*]] = call i32 @llvm.vector.reduce.add.v2i32(<2 x i32> [[TMP15]])
149149
; CHECK-NEXT: br i1 true, label [[LOOP_EXIT:%.*]], label [[SCALAR_PH]]
150150
; CHECK: scalar.ph:
151151
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
152-
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP18]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
152+
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP17]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
153153
; CHECK-NEXT: br label [[LOOP:%.*]]
154154
; CHECK: loop:
155155
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LATCH:%.*]] ]
@@ -169,7 +169,7 @@ define i32 @test_access_size_multiple_of_align_but_offset_by_1(i64 %len, ptr %te
169169
; CHECK-NEXT: [[EXIT:%.*]] = icmp eq i64 [[IV]], 4095
170170
; CHECK-NEXT: br i1 [[EXIT]], label [[LOOP_EXIT]], label [[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
171171
; CHECK: loop_exit:
172-
; CHECK-NEXT: [[ACCUM_NEXT_LCSSA:%.*]] = phi i32 [ [[ACCUM_NEXT]], [[LATCH]] ], [ [[TMP18]], [[MIDDLE_BLOCK]] ]
172+
; CHECK-NEXT: [[ACCUM_NEXT_LCSSA:%.*]] = phi i32 [ [[ACCUM_NEXT]], [[LATCH]] ], [ [[TMP17]], [[MIDDLE_BLOCK]] ]
173173
; CHECK-NEXT: ret i32 [[ACCUM_NEXT_LCSSA]]
174174
;
175175
entry:
@@ -198,3 +198,101 @@ latch:
198198
loop_exit:
199199
ret i32 %accum.next
200200
}
201+
202+
203+
; Test where offset relative to alloca is negative and we shouldn't
204+
; treat predicated loads as being always dereferenceable.
205+
define i8 @test_negative_off(i16 %len, ptr %test_base) {
206+
; CHECK-LABEL: @test_negative_off(
207+
; CHECK-NEXT: entry:
208+
; CHECK-NEXT: [[ALLOCA:%.*]] = alloca [64638 x i8], align 1
209+
; CHECK-NEXT: call void @init(ptr [[ALLOCA]])
210+
; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
211+
; CHECK: vector.ph:
212+
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
213+
; CHECK: vector.body:
214+
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_LOAD_CONTINUE2:%.*]] ]
215+
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i8> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP18:%.*]], [[PRED_LOAD_CONTINUE2]] ]
216+
; CHECK-NEXT: [[DOTCAST:%.*]] = trunc i32 [[INDEX]] to i16
217+
; CHECK-NEXT: [[OFFSET_IDX:%.*]] = add i16 -1000, [[DOTCAST]]
218+
; CHECK-NEXT: [[TMP0:%.*]] = add i16 [[OFFSET_IDX]], 0
219+
; CHECK-NEXT: [[TMP1:%.*]] = add i16 [[OFFSET_IDX]], 1
220+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i1, ptr [[TEST_BASE:%.*]], i16 [[TMP0]]
221+
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i1, ptr [[TEST_BASE]], i16 [[TMP1]]
222+
; CHECK-NEXT: [[TMP4:%.*]] = load i1, ptr [[TMP2]], align 1
223+
; CHECK-NEXT: [[TMP5:%.*]] = load i1, ptr [[TMP3]], align 1
224+
; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x i1> poison, i1 [[TMP4]], i32 0
225+
; CHECK-NEXT: [[TMP7:%.*]] = insertelement <2 x i1> [[TMP6]], i1 [[TMP5]], i32 1
226+
; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x i1> [[TMP7]], i32 0
227+
; CHECK-NEXT: br i1 [[TMP8]], label [[PRED_LOAD_IF:%.*]], label [[PRED_LOAD_CONTINUE:%.*]]
228+
; CHECK: pred.load.if:
229+
; CHECK-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[ALLOCA]], i16 [[TMP0]]
230+
; CHECK-NEXT: [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 1
231+
; CHECK-NEXT: [[TMP11:%.*]] = insertelement <2 x i8> poison, i8 [[TMP10]], i32 0
232+
; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE]]
233+
; CHECK: pred.load.continue:
234+
; CHECK-NEXT: [[TMP12:%.*]] = phi <2 x i8> [ poison, [[VECTOR_BODY]] ], [ [[TMP11]], [[PRED_LOAD_IF]] ]
235+
; CHECK-NEXT: [[TMP13:%.*]] = extractelement <2 x i1> [[TMP7]], i32 1
236+
; CHECK-NEXT: br i1 [[TMP13]], label [[PRED_LOAD_IF1:%.*]], label [[PRED_LOAD_CONTINUE2]]
237+
; CHECK: pred.load.if1:
238+
; CHECK-NEXT: [[TMP14:%.*]] = getelementptr i8, ptr [[ALLOCA]], i16 [[TMP1]]
239+
; CHECK-NEXT: [[TMP15:%.*]] = load i8, ptr [[TMP14]], align 1
240+
; CHECK-NEXT: [[TMP16:%.*]] = insertelement <2 x i8> [[TMP12]], i8 [[TMP15]], i32 1
241+
; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE2]]
242+
; CHECK: pred.load.continue2:
243+
; CHECK-NEXT: [[TMP17:%.*]] = phi <2 x i8> [ [[TMP12]], [[PRED_LOAD_CONTINUE]] ], [ [[TMP16]], [[PRED_LOAD_IF1]] ]
244+
; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP7]], <2 x i8> [[TMP17]], <2 x i8> zeroinitializer
245+
; CHECK-NEXT: [[TMP18]] = add <2 x i8> [[VEC_PHI]], [[PREDPHI]]
246+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
247+
; CHECK-NEXT: [[TMP19:%.*]] = icmp eq i32 [[INDEX_NEXT]], 12
248+
; CHECK-NEXT: br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
249+
; CHECK: middle.block:
250+
; CHECK-NEXT: [[TMP20:%.*]] = call i8 @llvm.vector.reduce.add.v2i8(<2 x i8> [[TMP18]])
251+
; CHECK-NEXT: br i1 true, label [[LOOP_EXIT:%.*]], label [[SCALAR_PH]]
252+
; CHECK: scalar.ph:
253+
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i16 [ -988, [[MIDDLE_BLOCK]] ], [ -1000, [[ENTRY:%.*]] ]
254+
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i8 [ [[TMP20]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
255+
; CHECK-NEXT: br label [[LOOP:%.*]]
256+
; CHECK: loop:
257+
; CHECK-NEXT: [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LATCH:%.*]] ]
258+
; CHECK-NEXT: [[ACCUM:%.*]] = phi i8 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[ACCUM_NEXT:%.*]], [[LATCH]] ]
259+
; CHECK-NEXT: [[IV_NEXT]] = add i16 [[IV]], 1
260+
; CHECK-NEXT: [[TEST_ADDR:%.*]] = getelementptr inbounds i1, ptr [[TEST_BASE]], i16 [[IV]]
261+
; CHECK-NEXT: [[EARLYCND:%.*]] = load i1, ptr [[TEST_ADDR]], align 1
262+
; CHECK-NEXT: br i1 [[EARLYCND]], label [[PRED:%.*]], label [[LATCH]]
263+
; CHECK: pred:
264+
; CHECK-NEXT: [[ADDR:%.*]] = getelementptr i8, ptr [[ALLOCA]], i16 [[IV]]
265+
; CHECK-NEXT: [[VAL:%.*]] = load i8, ptr [[ADDR]], align 1
266+
; CHECK-NEXT: br label [[LATCH]]
267+
; CHECK: latch:
268+
; CHECK-NEXT: [[VAL_PHI:%.*]] = phi i8 [ 0, [[LOOP]] ], [ [[VAL]], [[PRED]] ]
269+
; CHECK-NEXT: [[ACCUM_NEXT]] = add i8 [[ACCUM]], [[VAL_PHI]]
270+
; CHECK-NEXT: [[EXIT:%.*]] = icmp ugt i16 [[IV]], -990
271+
; CHECK-NEXT: br i1 [[EXIT]], label [[LOOP_EXIT]], label [[LOOP]], !llvm.loop [[LOOP7:![0-9]+]]
272+
; CHECK: loop_exit:
273+
; CHECK-NEXT: [[ACCUM_NEXT_LCSSA:%.*]] = phi i8 [ [[ACCUM_NEXT]], [[LATCH]] ], [ [[TMP20]], [[MIDDLE_BLOCK]] ]
274+
; CHECK-NEXT: ret i8 [[ACCUM_NEXT_LCSSA]]
275+
;
276+
entry:
277+
%alloca = alloca [64638 x i8]
278+
call void @init(ptr %alloca)
279+
br label %loop
280+
loop:
281+
%iv = phi i16 [ -1000, %entry ], [ %iv.next, %latch ]
282+
%accum = phi i8 [ 0, %entry ], [ %accum.next, %latch ]
283+
%iv.next = add i16 %iv, 1
284+
%test_addr = getelementptr inbounds i1, ptr %test_base, i16 %iv
285+
%earlycnd = load i1, ptr %test_addr
286+
br i1 %earlycnd, label %pred, label %latch
287+
pred:
288+
%addr = getelementptr i8, ptr %alloca, i16 %iv
289+
%val = load i8, ptr %addr
290+
br label %latch
291+
latch:
292+
%val.phi = phi i8 [ 0, %loop ], [ %val, %pred ]
293+
%accum.next = add i8 %accum, %val.phi
294+
%exit = icmp ugt i16 %iv, -990
295+
br i1 %exit, label %loop_exit, label %loop
296+
loop_exit:
297+
ret i8 %accum.next
298+
}

0 commit comments

Comments
 (0)