Skip to content

Commit b822a32

Browse files
fhahnfhahn
committed
[VPlan] Fix crash when trying to narrow interleave group storing const.
Use dyn_cast_null to handle the case where an interleave groups stores a constant in any of its lanes.
1 parent 347a5f4 commit b822a32

File tree

2 files changed

+138
-1
lines changed

2 files changed

+138
-1
lines changed

‎llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -3251,7 +3251,7 @@ void VPlanTransforms::narrowInterleaveGroups(VPlan &Plan, ElementCount VF,
32513251
if (!WideMember0)
32523252
return;
32533253
for (const auto &[I, V] : enumerate(InterleaveR->getStoredValues())) {
3254-
auto *R = dyn_cast<VPWidenRecipe>(V->getDefiningRecipe());
3254+
auto *R = dyn_cast_or_null<VPWidenRecipe>(V->getDefiningRecipe());
32553255
if (!R || R->getOpcode() != WideMember0->getOpcode() ||
32563256
R->getNumOperands() > 2)
32573257
return;

‎llvm/test/Transforms/LoopVectorize/AArch64/transform-narrow-interleave-to-widen-memory-cost.ll‎

Lines changed: 137 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -141,3 +141,140 @@ loop:
141141
exit:
142142
ret void
143143
}
144+
145+
define void @test_interleave_store_one_constant(ptr noalias %src, ptr noalias %dst, i64 %n) #0 {
146+
; CHECK-LABEL: define void @test_interleave_store_one_constant(
147+
; CHECK-SAME: ptr noalias [[SRC:%.*]], ptr noalias [[DST:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
148+
; CHECK-NEXT: [[ITER_CHECK:.*]]:
149+
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[N]], 1
150+
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 2
151+
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[VEC_EPILOG_SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]]
152+
; CHECK: [[VECTOR_SCEVCHECK]]:
153+
; CHECK-NEXT: [[MUL1:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 16, i64 [[N]])
154+
; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i64, i1 } [[MUL1]], 0
155+
; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i64, i1 } [[MUL1]], 1
156+
; CHECK-NEXT: [[TMP1:%.*]] = sub i64 0, [[MUL_RESULT]]
157+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, ptr [[DST]], i64 [[MUL_RESULT]]
158+
; CHECK-NEXT: [[TMP3:%.*]] = icmp ult ptr [[TMP2]], [[DST]]
159+
; CHECK-NEXT: [[TMP4:%.*]] = or i1 [[TMP3]], [[MUL_OVERFLOW]]
160+
; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[DST]], i64 8
161+
; CHECK-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 16, i64 [[N]])
162+
; CHECK-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0
163+
; CHECK-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1
164+
; CHECK-NEXT: [[TMP5:%.*]] = sub i64 0, [[MUL_RESULT3]]
165+
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i8, ptr [[SCEVGEP]], i64 [[MUL_RESULT3]]
166+
; CHECK-NEXT: [[TMP7:%.*]] = icmp ult ptr [[TMP6]], [[SCEVGEP]]
167+
; CHECK-NEXT: [[TMP8:%.*]] = or i1 [[TMP7]], [[MUL_OVERFLOW4]]
168+
; CHECK-NEXT: [[TMP9:%.*]] = or i1 [[TMP4]], [[TMP8]]
169+
; CHECK-NEXT: br i1 [[TMP9]], label %[[VEC_EPILOG_SCALAR_PH]], label %[[VECTOR_MAIN_LOOP_ITER_CHECK:.*]]
170+
; CHECK: [[VECTOR_MAIN_LOOP_ITER_CHECK]]:
171+
; CHECK-NEXT: [[MIN_ITERS_CHECK5:%.*]] = icmp ult i64 [[TMP0]], 8
172+
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK5]], label %[[VEC_EPILOG_PH:.*]], label %[[VECTOR_PH:.*]]
173+
; CHECK: [[VECTOR_PH]]:
174+
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 8
175+
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
176+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
177+
; CHECK: [[VECTOR_BODY]]:
178+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
179+
; CHECK-NEXT: [[TMP10:%.*]] = add i64 [[INDEX]], 2
180+
; CHECK-NEXT: [[TMP11:%.*]] = add i64 [[INDEX]], 4
181+
; CHECK-NEXT: [[TMP12:%.*]] = add i64 [[INDEX]], 6
182+
; CHECK-NEXT: [[TMP13:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDEX]]
183+
; CHECK-NEXT: [[TMP14:%.*]] = getelementptr double, ptr [[TMP13]], i32 0
184+
; CHECK-NEXT: [[TMP15:%.*]] = getelementptr double, ptr [[TMP13]], i32 2
185+
; CHECK-NEXT: [[TMP16:%.*]] = getelementptr double, ptr [[TMP13]], i32 4
186+
; CHECK-NEXT: [[TMP17:%.*]] = getelementptr double, ptr [[TMP13]], i32 6
187+
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x double>, ptr [[TMP14]], align 8
188+
; CHECK-NEXT: [[WIDE_LOAD6:%.*]] = load <2 x double>, ptr [[TMP15]], align 8
189+
; CHECK-NEXT: [[WIDE_LOAD7:%.*]] = load <2 x double>, ptr [[TMP16]], align 8
190+
; CHECK-NEXT: [[WIDE_LOAD8:%.*]] = load <2 x double>, ptr [[TMP17]], align 8
191+
; CHECK-NEXT: [[TMP18:%.*]] = fmul <2 x double> [[WIDE_LOAD]], splat (double 5.000000e+00)
192+
; CHECK-NEXT: [[TMP19:%.*]] = fmul <2 x double> [[WIDE_LOAD6]], splat (double 5.000000e+00)
193+
; CHECK-NEXT: [[TMP20:%.*]] = fmul <2 x double> [[WIDE_LOAD7]], splat (double 5.000000e+00)
194+
; CHECK-NEXT: [[TMP21:%.*]] = fmul <2 x double> [[WIDE_LOAD8]], splat (double 5.000000e+00)
195+
; CHECK-NEXT: [[TMP22:%.*]] = getelementptr [2 x double], ptr [[DST]], i64 [[INDEX]]
196+
; CHECK-NEXT: [[TMP23:%.*]] = getelementptr [2 x double], ptr [[DST]], i64 [[TMP10]]
197+
; CHECK-NEXT: [[TMP24:%.*]] = getelementptr [2 x double], ptr [[DST]], i64 [[TMP11]]
198+
; CHECK-NEXT: [[TMP25:%.*]] = getelementptr [2 x double], ptr [[DST]], i64 [[TMP12]]
199+
; CHECK-NEXT: [[TMP26:%.*]] = shufflevector <2 x double> [[TMP18]], <2 x double> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
200+
; CHECK-NEXT: [[INTERLEAVED_VEC:%.*]] = shufflevector <4 x double> [[TMP26]], <4 x double> poison, <4 x i32> <i32 0, i32 2, i32 1, i32 3>
201+
; CHECK-NEXT: store <4 x double> [[INTERLEAVED_VEC]], ptr [[TMP22]], align 8
202+
; CHECK-NEXT: [[TMP27:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
203+
; CHECK-NEXT: [[INTERLEAVED_VEC9:%.*]] = shufflevector <4 x double> [[TMP27]], <4 x double> poison, <4 x i32> <i32 0, i32 2, i32 1, i32 3>
204+
; CHECK-NEXT: store <4 x double> [[INTERLEAVED_VEC9]], ptr [[TMP23]], align 8
205+
; CHECK-NEXT: [[TMP28:%.*]] = shufflevector <2 x double> [[TMP20]], <2 x double> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
206+
; CHECK-NEXT: [[INTERLEAVED_VEC10:%.*]] = shufflevector <4 x double> [[TMP28]], <4 x double> poison, <4 x i32> <i32 0, i32 2, i32 1, i32 3>
207+
; CHECK-NEXT: store <4 x double> [[INTERLEAVED_VEC10]], ptr [[TMP24]], align 8
208+
; CHECK-NEXT: [[TMP29:%.*]] = shufflevector <2 x double> [[TMP21]], <2 x double> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
209+
; CHECK-NEXT: [[INTERLEAVED_VEC11:%.*]] = shufflevector <4 x double> [[TMP29]], <4 x double> poison, <4 x i32> <i32 0, i32 2, i32 1, i32 3>
210+
; CHECK-NEXT: store <4 x double> [[INTERLEAVED_VEC11]], ptr [[TMP25]], align 8
211+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
212+
; CHECK-NEXT: [[TMP30:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
213+
; CHECK-NEXT: br i1 [[TMP30]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
214+
; CHECK: [[MIDDLE_BLOCK]]:
215+
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]
216+
; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[VEC_EPILOG_ITER_CHECK:.*]]
217+
; CHECK: [[VEC_EPILOG_ITER_CHECK]]:
218+
; CHECK-NEXT: [[N_VEC_REMAINING:%.*]] = sub i64 [[TMP0]], [[N_VEC]]
219+
; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_VEC_REMAINING]], 2
220+
; CHECK-NEXT: br i1 [[MIN_EPILOG_ITERS_CHECK]], label %[[VEC_EPILOG_SCALAR_PH]], label %[[VEC_EPILOG_PH]]
221+
; CHECK: [[VEC_EPILOG_PH]]:
222+
; CHECK-NEXT: [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[VEC_EPILOG_ITER_CHECK]] ], [ 0, %[[VECTOR_MAIN_LOOP_ITER_CHECK]] ]
223+
; CHECK-NEXT: [[N_MOD_VF12:%.*]] = urem i64 [[TMP0]], 2
224+
; CHECK-NEXT: [[N_VEC13:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF12]]
225+
; CHECK-NEXT: br label %[[VEC_EPILOG_VECTOR_BODY:.*]]
226+
; CHECK: [[VEC_EPILOG_VECTOR_BODY]]:
227+
; CHECK-NEXT: [[INDEX14:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], %[[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT17:%.*]], %[[VEC_EPILOG_VECTOR_BODY]] ]
228+
; CHECK-NEXT: [[TMP31:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDEX14]]
229+
; CHECK-NEXT: [[TMP32:%.*]] = getelementptr double, ptr [[TMP31]], i32 0
230+
; CHECK-NEXT: [[WIDE_LOAD15:%.*]] = load <2 x double>, ptr [[TMP32]], align 8
231+
; CHECK-NEXT: [[TMP33:%.*]] = fmul <2 x double> [[WIDE_LOAD15]], splat (double 5.000000e+00)
232+
; CHECK-NEXT: [[TMP34:%.*]] = getelementptr [2 x double], ptr [[DST]], i64 [[INDEX14]]
233+
; CHECK-NEXT: [[TMP35:%.*]] = shufflevector <2 x double> [[TMP33]], <2 x double> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
234+
; CHECK-NEXT: [[INTERLEAVED_VEC16:%.*]] = shufflevector <4 x double> [[TMP35]], <4 x double> poison, <4 x i32> <i32 0, i32 2, i32 1, i32 3>
235+
; CHECK-NEXT: store <4 x double> [[INTERLEAVED_VEC16]], ptr [[TMP34]], align 8
236+
; CHECK-NEXT: [[INDEX_NEXT17]] = add nuw i64 [[INDEX14]], 2
237+
; CHECK-NEXT: [[TMP36:%.*]] = icmp eq i64 [[INDEX_NEXT17]], [[N_VEC13]]
238+
; CHECK-NEXT: br i1 [[TMP36]], label %[[VEC_EPILOG_MIDDLE_BLOCK:.*]], label %[[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP7:![0-9]+]]
239+
; CHECK: [[VEC_EPILOG_MIDDLE_BLOCK]]:
240+
; CHECK-NEXT: [[CMP_N18:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC13]]
241+
; CHECK-NEXT: br i1 [[CMP_N18]], label %[[EXIT]], label %[[VEC_EPILOG_SCALAR_PH]]
242+
; CHECK: [[VEC_EPILOG_SCALAR_PH]]:
243+
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC13]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[N_VEC]], %[[VEC_EPILOG_ITER_CHECK]] ], [ 0, %[[VECTOR_SCEVCHECK]] ], [ 0, %[[ITER_CHECK]] ]
244+
; CHECK-NEXT: br label %[[LOOP:.*]]
245+
; CHECK: [[LOOP]]:
246+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[VEC_EPILOG_SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
247+
; CHECK-NEXT: [[GEP_SRC:%.*]] = getelementptr double, ptr [[SRC]], i64 [[IV]]
248+
; CHECK-NEXT: [[L:%.*]] = load double, ptr [[GEP_SRC]], align 8
249+
; CHECK-NEXT: [[MUL:%.*]] = fmul double [[L]], 5.000000e+00
250+
; CHECK-NEXT: [[GEP_DST:%.*]] = getelementptr [2 x double], ptr [[DST]], i64 [[IV]]
251+
; CHECK-NEXT: store double [[MUL]], ptr [[GEP_DST]], align 8
252+
; CHECK-NEXT: [[GEP_DST_1:%.*]] = getelementptr [2 x double], ptr [[DST]], i64 [[IV]], i32 1
253+
; CHECK-NEXT: store double 0.000000e+00, ptr [[GEP_DST_1]], align 8
254+
; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
255+
; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV]], [[N]]
256+
; CHECK-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP8:![0-9]+]]
257+
; CHECK: [[EXIT]]:
258+
; CHECK-NEXT: ret void
259+
;
260+
entry:
261+
br label %loop
262+
263+
loop: ; preds = %loop, %entry
264+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
265+
%gep.src = getelementptr double, ptr %src, i64 %iv
266+
%l = load double, ptr %gep.src, align 8
267+
%mul = fmul double %l, 5.000000e+00
268+
%gep.dst = getelementptr [2 x double], ptr %dst, i64 %iv
269+
store double %mul, ptr %gep.dst, align 8
270+
%gep.dst.1 = getelementptr [2 x double], ptr %dst, i64 %iv, i32 1
271+
store double 0.000000e+00, ptr %gep.dst.1, align 8
272+
%iv.next = add i64 %iv, 1
273+
%ec = icmp eq i64 %iv, %n
274+
br i1 %ec, label %exit, label %loop
275+
276+
exit:
277+
ret void
278+
}
279+
280+
attributes #0 = { "target-cpu"="neoverse-v2" }

0 commit comments

Comments
 (0)