Skip to content

Commit 381e742

Browse files
david-armsivan-shani
david-arm
authored and
sivan-shani
committed
[AArch64] Improve getPartialReductionCost for fixed-width VFs (llvm#126538)
NEON does not have a version of udot/sdot that accumulates into 64-bit integer values, so we should return Invalid from getPartialReductionCost for 64-bit types and fixed-width VFs. In theory, if the 64-bit versions of SVE udot/sdot are available we could use those, but we don't currently have lowering support for that.
1 parent de46745 commit 381e742

File tree

4 files changed

+194
-135
lines changed

4 files changed

+194
-135
lines changed

‎llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4705,7 +4705,8 @@ InstructionCost AArch64TTIImpl::getPartialReductionCost(
47054705
if (VFMinValue == Scale)
47064706
return Invalid;
47074707
}
4708-
if (VF.isFixed() && (!ST->isNeonAvailable() || !ST->hasDotProd()))
4708+
if (VF.isFixed() &&
4709+
(!ST->isNeonAvailable() || !ST->hasDotProd() || AccumEVT == MVT::i64))
47094710
return Invalid;
47104711

47114712
if (InputEVT == MVT::i8) {

‎llvm/test/Transforms/LoopVectorize/AArch64/fully-unrolled-cost.ll‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -82,11 +82,11 @@ define i64 @test_two_ivs(ptr %a, ptr %b, i64 %start) #0 {
8282
; CHECK-NEXT: Cost of 0 for VF 8: induction instruction %j.iv = phi i64 [ %start, %entry ], [ %j.iv.next, %for.body ]
8383
; CHECK-NEXT: Cost of 1 for VF 8: exit condition instruction %exitcond.not = icmp eq i64 %i.iv.next, 16
8484
; CHECK-NEXT: Cost of 0 for VF 8: EMIT vp<{{.+}}> = CANONICAL-INDUCTION ir<0>, vp<%index.next>
85-
; CHECK: Cost for VF 8: 24
85+
; CHECK: Cost for VF 8: 27
8686
; CHECK-NEXT: Cost of 0 for VF 16: induction instruction %i.iv = phi i64 [ 0, %entry ], [ %i.iv.next, %for.body ]
8787
; CHECK-NEXT: Cost of 0 for VF 16: induction instruction %j.iv = phi i64 [ %start, %entry ], [ %j.iv.next, %for.body ]
8888
; CHECK-NEXT: Cost of 0 for VF 16: EMIT vp<{{.+}}> = CANONICAL-INDUCTION ir<0>, vp<%index.next>
89-
; CHECK: Cost for VF 16: 42
89+
; CHECK: Cost for VF 16: 48
9090
; CHECK: LV: Selecting VF: 16
9191
entry:
9292
br label %for.body

‎llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-epilogue.ll‎

Lines changed: 38 additions & 38 deletions
Original file line numberDiff line numberDiff line change
@@ -122,7 +122,7 @@ define void @dotp_small_epilogue_vf(i64 %idx.neg, i8 %a) #1 {
122122
; CHECK-SAME: i64 [[IDX_NEG:%.*]], i8 [[A:%.*]]) #[[ATTR1:[0-9]+]] {
123123
; CHECK-NEXT: iter.check:
124124
; CHECK-NEXT: [[TMP0:%.*]] = sub i64 1, [[IDX_NEG]]
125-
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 8
125+
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 4
126126
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[ENTRY:%.*]]
127127
; CHECK: vector.main.loop.iter.check:
128128
; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ult i64 [[TMP0]], 16
@@ -132,78 +132,78 @@ define void @dotp_small_epilogue_vf(i64 %idx.neg, i8 %a) #1 {
132132
; CHECK-NEXT: [[IV_NEXT:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
133133
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <16 x i8> poison, i8 [[A]], i64 0
134134
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <16 x i8> [[BROADCAST_SPLATINSERT]], <16 x i8> poison, <16 x i32> zeroinitializer
135-
; CHECK-NEXT: [[TMP1:%.*]] = sext <16 x i8> [[BROADCAST_SPLAT]] to <16 x i64>
135+
; CHECK-NEXT: [[TMP1:%.*]] = sext <16 x i8> [[BROADCAST_SPLAT]] to <16 x i32>
136136
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
137137
; CHECK: vector.body:
138138
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
139-
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], [[VECTOR_BODY]] ]
139+
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], [[VECTOR_BODY]] ]
140140
; CHECK-NEXT: [[TMP2:%.*]] = load i8, ptr null, align 1
141141
; CHECK-NEXT: [[BROADCAST_SPLATINSERT2:%.*]] = insertelement <16 x i8> poison, i8 [[TMP2]], i64 0
142142
; CHECK-NEXT: [[BROADCAST_SPLAT3:%.*]] = shufflevector <16 x i8> [[BROADCAST_SPLATINSERT2]], <16 x i8> poison, <16 x i32> zeroinitializer
143-
; CHECK-NEXT: [[TMP3:%.*]] = sext <16 x i8> [[BROADCAST_SPLAT3]] to <16 x i64>
144-
; CHECK-NEXT: [[TMP4:%.*]] = mul <16 x i64> [[TMP3]], [[TMP1]]
145-
; CHECK-NEXT: [[PARTIAL_REDUCE]] = call <2 x i64> @llvm.experimental.vector.partial.reduce.add.v2i64.v16i64(<2 x i64> [[VEC_PHI]], <16 x i64> [[TMP4]])
143+
; CHECK-NEXT: [[TMP3:%.*]] = sext <16 x i8> [[BROADCAST_SPLAT3]] to <16 x i32>
144+
; CHECK-NEXT: [[TMP4:%.*]] = mul <16 x i32> [[TMP3]], [[TMP1]]
145+
; CHECK-NEXT: [[PARTIAL_REDUCE]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI]], <16 x i32> [[TMP4]])
146146
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 16
147147
; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[IV_NEXT]]
148148
; CHECK-NEXT: br i1 [[TMP5]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
149149
; CHECK: middle.block:
150-
; CHECK-NEXT: [[ADD:%.*]] = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> [[PARTIAL_REDUCE]])
150+
; CHECK-NEXT: [[TMP6:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE]])
151151
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[IV_NEXT]]
152152
; CHECK-NEXT: br i1 [[CMP_N]], label [[WHILE_END_LOOPEXIT:%.*]], label [[WHILE_BODY:%.*]]
153153
; CHECK: vec.epilog.iter.check:
154154
; CHECK-NEXT: [[IND_END6:%.*]] = add i64 [[IDX_NEG]], [[IV_NEXT]]
155155
; CHECK-NEXT: [[N_VEC_REMAINING:%.*]] = sub i64 [[TMP0]], [[IV_NEXT]]
156-
; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_VEC_REMAINING]], 8
156+
; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_VEC_REMAINING]], 4
157157
; CHECK-NEXT: br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]]
158158
; CHECK: vec.epilog.ph:
159159
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT]], [[WHILE_BODY]] ], [ 0, [[ENTRY]] ]
160-
; CHECK-NEXT: [[ACCUM:%.*]] = phi i64 [ [[ADD]], [[WHILE_BODY]] ], [ 0, [[ENTRY]] ]
161-
; CHECK-NEXT: [[N_MOD_VF4:%.*]] = urem i64 [[TMP0]], 8
160+
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP6]], [[WHILE_BODY]] ], [ 0, [[ENTRY]] ]
161+
; CHECK-NEXT: [[N_MOD_VF4:%.*]] = urem i64 [[TMP0]], 4
162162
; CHECK-NEXT: [[N_VEC5:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF4]]
163163
; CHECK-NEXT: [[IND_END:%.*]] = add i64 [[IDX_NEG]], [[N_VEC5]]
164-
; CHECK-NEXT: [[BROADCAST_SPLATINSERT7:%.*]] = insertelement <8 x i8> poison, i8 [[A]], i64 0
165-
; CHECK-NEXT: [[BROADCAST_SPLAT8:%.*]] = shufflevector <8 x i8> [[BROADCAST_SPLATINSERT7]], <8 x i8> poison, <8 x i32> zeroinitializer
166-
; CHECK-NEXT: [[TMP7:%.*]] = sext <8 x i8> [[BROADCAST_SPLAT8]] to <8 x i64>
167-
; CHECK-NEXT: [[TMP8:%.*]] = insertelement <1 x i64> zeroinitializer, i64 [[ACCUM]], i32 0
164+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT6:%.*]] = insertelement <4 x i8> poison, i8 [[A]], i64 0
165+
; CHECK-NEXT: [[BROADCAST_SPLAT7:%.*]] = shufflevector <4 x i8> [[BROADCAST_SPLATINSERT6]], <4 x i8> poison, <4 x i32> zeroinitializer
166+
; CHECK-NEXT: [[TMP8:%.*]] = sext <4 x i8> [[BROADCAST_SPLAT7]] to <4 x i32>
167+
; CHECK-NEXT: [[TMP10:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[BC_MERGE_RDX]], i32 0
168168
; CHECK-NEXT: br label [[VEC_EPILOG_VECTOR_BODY:%.*]]
169169
; CHECK: vec.epilog.vector.body:
170170
; CHECK-NEXT: [[INDEX9:%.*]] = phi i64 [ [[IV]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT14:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
171-
; CHECK-NEXT: [[VEC_PHI10:%.*]] = phi <1 x i64> [ [[TMP8]], [[VEC_EPILOG_PH]] ], [ [[PARTIAL_REDUCE13:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
171+
; CHECK-NEXT: [[VEC_PHI9:%.*]] = phi <4 x i32> [ [[TMP10]], [[VEC_EPILOG_PH]] ], [ [[TMP13:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
172172
; CHECK-NEXT: [[TMP9:%.*]] = load i8, ptr null, align 1
173-
; CHECK-NEXT: [[BROADCAST_SPLATINSERT11:%.*]] = insertelement <8 x i8> poison, i8 [[TMP9]], i64 0
174-
; CHECK-NEXT: [[BROADCAST_SPLAT12:%.*]] = shufflevector <8 x i8> [[BROADCAST_SPLATINSERT11]], <8 x i8> poison, <8 x i32> zeroinitializer
175-
; CHECK-NEXT: [[TMP10:%.*]] = sext <8 x i8> [[BROADCAST_SPLAT12]] to <8 x i64>
176-
; CHECK-NEXT: [[TMP11:%.*]] = mul <8 x i64> [[TMP10]], [[TMP7]]
177-
; CHECK-NEXT: [[PARTIAL_REDUCE13]] = call <1 x i64> @llvm.experimental.vector.partial.reduce.add.v1i64.v8i64(<1 x i64> [[VEC_PHI10]], <8 x i64> [[TMP11]])
178-
; CHECK-NEXT: [[INDEX_NEXT14]] = add nuw i64 [[INDEX9]], 8
173+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT10:%.*]] = insertelement <4 x i8> poison, i8 [[TMP9]], i64 0
174+
; CHECK-NEXT: [[BROADCAST_SPLAT11:%.*]] = shufflevector <4 x i8> [[BROADCAST_SPLATINSERT10]], <4 x i8> poison, <4 x i32> zeroinitializer
175+
; CHECK-NEXT: [[TMP11:%.*]] = sext <4 x i8> [[BROADCAST_SPLAT11]] to <4 x i32>
176+
; CHECK-NEXT: [[TMP14:%.*]] = mul <4 x i32> [[TMP11]], [[TMP8]]
177+
; CHECK-NEXT: [[TMP13]] = add <4 x i32> [[TMP14]], [[VEC_PHI9]]
178+
; CHECK-NEXT: [[INDEX_NEXT14]] = add nuw i64 [[INDEX9]], 4
179179
; CHECK-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT14]], [[N_VEC5]]
180180
; CHECK-NEXT: br i1 [[TMP12]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
181181
; CHECK: vec.epilog.middle.block:
182-
; CHECK-NEXT: [[TMP13:%.*]] = call i64 @llvm.vector.reduce.add.v1i64(<1 x i64> [[PARTIAL_REDUCE13]])
182+
; CHECK-NEXT: [[TMP15:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP13]])
183183
; CHECK-NEXT: [[CMP_N15:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC5]]
184184
; CHECK-NEXT: br i1 [[CMP_N15]], label [[WHILE_END_LOOPEXIT]], label [[VEC_EPILOG_SCALAR_PH]]
185185
; CHECK: vec.epilog.scalar.ph:
186186
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IND_END6]], [[WHILE_BODY]] ], [ [[IDX_NEG]], [[ITER_CHECK:%.*]] ]
187-
; CHECK-NEXT: [[BC_RESUME_VAL16:%.*]] = phi i64 [ [[N_VEC5]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IV_NEXT]], [[WHILE_BODY]] ], [ 0, [[ITER_CHECK]] ]
188-
; CHECK-NEXT: [[BC_MERGE_RDX17:%.*]] = phi i64 [ [[TMP13]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[ADD]], [[WHILE_BODY]] ], [ 0, [[ITER_CHECK]] ]
187+
; CHECK-NEXT: [[BC_MERGE_RDX17:%.*]] = phi i64 [ [[N_VEC5]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IV_NEXT]], [[WHILE_BODY]] ], [ 0, [[ITER_CHECK]] ]
188+
; CHECK-NEXT: [[BC_MERGE_RDX15:%.*]] = phi i32 [ [[TMP15]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[TMP6]], [[WHILE_BODY]] ], [ 0, [[ITER_CHECK]] ]
189189
; CHECK-NEXT: br label [[WHILE_BODY1:%.*]]
190190
; CHECK: while.body:
191191
; CHECK-NEXT: [[IV_NEG:%.*]] = phi i64 [ [[IV_NEG_NEXT:%.*]], [[WHILE_BODY1]] ], [ [[BC_RESUME_VAL]], [[VEC_EPILOG_SCALAR_PH]] ]
192-
; CHECK-NEXT: [[IV1:%.*]] = phi i64 [ [[IV_NEXT1:%.*]], [[WHILE_BODY1]] ], [ [[BC_RESUME_VAL16]], [[VEC_EPILOG_SCALAR_PH]] ]
193192
; CHECK-NEXT: [[ACCUM1:%.*]] = phi i64 [ [[ADD1:%.*]], [[WHILE_BODY1]] ], [ [[BC_MERGE_RDX17]], [[VEC_EPILOG_SCALAR_PH]] ]
193+
; CHECK-NEXT: [[ACCUM:%.*]] = phi i32 [ [[ADD:%.*]], [[WHILE_BODY1]] ], [ [[BC_MERGE_RDX15]], [[VEC_EPILOG_SCALAR_PH]] ]
194194
; CHECK-NEXT: [[IV_NEG_NEXT]] = add i64 [[IV_NEG]], 1
195-
; CHECK-NEXT: [[EXT_A:%.*]] = sext i8 [[A]] to i64
196-
; CHECK-NEXT: [[IV_NEXT1]] = add i64 [[IV1]], 1
195+
; CHECK-NEXT: [[EXT_A:%.*]] = sext i8 [[A]] to i32
196+
; CHECK-NEXT: [[ADD1]] = add i64 [[ACCUM1]], 1
197197
; CHECK-NEXT: [[B:%.*]] = load i8, ptr null, align 1
198-
; CHECK-NEXT: [[EXT_B:%.*]] = sext i8 [[B]] to i64
199-
; CHECK-NEXT: [[MUL:%.*]] = mul i64 [[EXT_B]], [[EXT_A]]
200-
; CHECK-NEXT: [[ADD1]] = add i64 [[MUL]], [[ACCUM1]]
198+
; CHECK-NEXT: [[EXT_B:%.*]] = sext i8 [[B]] to i32
199+
; CHECK-NEXT: [[MUL:%.*]] = mul i32 [[EXT_B]], [[EXT_A]]
200+
; CHECK-NEXT: [[ADD]] = add i32 [[MUL]], [[ACCUM]]
201201
; CHECK-NEXT: [[CMP_IV_NEG:%.*]] = icmp ugt i64 [[IV_NEG]], 0
202-
; CHECK-NEXT: [[CMP_IV:%.*]] = icmp ne i64 [[IV1]], -1
202+
; CHECK-NEXT: [[CMP_IV:%.*]] = icmp ne i64 [[ACCUM1]], -1
203203
; CHECK-NEXT: [[EXITCOND:%.*]] = and i1 [[CMP_IV_NEG]], [[CMP_IV]]
204204
; CHECK-NEXT: br i1 [[EXITCOND]], label [[WHILE_BODY1]], label [[WHILE_END_LOOPEXIT]], !llvm.loop [[LOOP7:![0-9]+]]
205205
; CHECK: while.end.loopexit:
206-
; CHECK-NEXT: [[RESULT:%.*]] = phi i64 [ [[ADD1]], [[WHILE_BODY1]] ], [ [[ADD]], [[MIDDLE_BLOCK]] ], [ [[TMP13]], [[VEC_EPILOG_MIDDLE_BLOCK]] ]
206+
; CHECK-NEXT: [[RESULT:%.*]] = phi i32 [ [[ADD]], [[WHILE_BODY1]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ], [ [[TMP15]], [[VEC_EPILOG_MIDDLE_BLOCK]] ]
207207
; CHECK-NEXT: ret void
208208
;
209209
entry:
@@ -212,21 +212,21 @@ entry:
212212
while.body: ; preds = %while.body, %entry
213213
%iv.neg = phi i64 [ %iv.neg.next, %while.body ], [ %idx.neg, %entry ]
214214
%iv = phi i64 [ %iv.next, %while.body ], [ 0, %entry ]
215-
%accum = phi i64 [ %add, %while.body ], [ 0, %entry ]
215+
%accum = phi i32 [ %add, %while.body ], [ 0, %entry ]
216216
%iv.neg.next = add i64 %iv.neg, 1
217-
%ext.a = sext i8 %a to i64
217+
%ext.a = sext i8 %a to i32
218218
%iv.next = add i64 %iv, 1
219219
%b = load i8, ptr null, align 1
220-
%ext.b = sext i8 %b to i64
221-
%mul = mul i64 %ext.b, %ext.a
222-
%add = add i64 %mul, %accum
220+
%ext.b = sext i8 %b to i32
221+
%mul = mul i32 %ext.b, %ext.a
222+
%add = add i32 %mul, %accum
223223
%cmp.iv.neg = icmp ugt i64 %iv.neg, 0
224224
%cmp.iv = icmp ne i64 %iv, -1
225225
%exitcond = and i1 %cmp.iv.neg, %cmp.iv
226226
br i1 %exitcond, label %while.body, label %while.end.loopexit
227227

228228
while.end.loopexit: ; preds = %while.body
229-
%result = phi i64 [ %add, %while.body ]
229+
%result = phi i32 [ %add, %while.body ]
230230
ret void
231231
}
232232

0 commit comments

Comments
 (0)