Skip to content

Commit 8d08a84

Browse files
sdesmalen-armsdesmalen-arm
committed
[LV] Remove a change that was added in D106164.
This change wasn't strictly necessary for D106164 and could be removed. This patch addresses the post-commit comments from @fhahn on D106164, and also changes sve-widen-gep.ll to use the same IR test as shown in pointer-induction.ll. Reviewed By: fhahn Differential Revision: https://reviews.llvm.org/D106878
1 parent 21bf817 commit 8d08a84

File tree

2 files changed

+97
-53
lines changed

2 files changed

+97
-53
lines changed

‎llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

Lines changed: 5 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -5137,12 +5137,11 @@ void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) {
51375137
if (Worklist.count(I))
51385138
return;
51395139

5140-
// If all users of the pointer will be memory accesses and scalar, place the
5141-
// pointer in ScalarPtrs. Otherwise, place the pointer in
5142-
// PossibleNonScalarPtrs.
5143-
if (llvm::all_of(I->users(), [&](User *U) {
5144-
return (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
5145-
isScalarUse(cast<Instruction>(U), Ptr);
5140+
// If the use of the pointer will be a scalar use, and all users of the
5141+
// pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise,
5142+
// place the pointer in PossibleNonScalarPtrs.
5143+
if (isScalarUse(MemAccess, Ptr) && llvm::all_of(I->users(), [&](User *U) {
5144+
return isa<LoadInst>(U) || isa<StoreInst>(U);
51465145
}))
51475146
ScalarPtrs.insert(I);
51485147
else

‎llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-gep.ll

Lines changed: 92 additions & 47 deletions
Original file line numberDiff line numberDiff line change
@@ -1,68 +1,113 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
12
; REQUIRES: asserts
23
; RUN: opt -loop-vectorize -scalable-vectorization=on -S -mtriple=aarch64 -mattr=+sve -debug-only=loop-vectorize < %s 2>&1 | FileCheck %s
34

45
target triple = "aarch64-unknown-linux-gnu"
56

6-
; In the test below the PHI instruction:
7-
; %0 = phi i8* [ %incdec.ptr190, %loop.body ], [ %src, %entry ]
8-
; has multiple uses, i.e.
9-
; 1. As a uniform address for the load, and
10-
; 2. Non-uniform use by the getelementptr + store, which leads to replication.
11-
12-
; CHECK-LABEL: LV: Checking a loop in "phi_multiple_use"
13-
; CHECK-NOT: LV: Found new scalar instruction: %incdec.ptr190 = getelementptr inbounds i8, i8* %0, i64 1
7+
; CHECK-LABEL: LV: Checking a loop in "pointer_induction_used_as_vector"
8+
; CHECK-NOT: LV: Found {{.*}} scalar instruction: %ptr.iv.2.next = getelementptr inbounds i8, i8* %ptr.iv.2, i64 1
149
;
1510
; CHECK: VPlan 'Initial VPlan for VF={vscale x 2},UF>=1' {
1611
; CHECK-NEXT: loop.body:
17-
; CHECK-NEXT: WIDEN-INDUCTION %index = phi 0, %index.next
18-
; CHECK-NEXT: WIDEN-PHI %curchar = phi %curchar.next, %curptr
19-
; CHECK-NEXT: WIDEN-PHI %0 = phi %incdec.ptr190, %src
20-
; CHECK-NEXT: WIDEN-GEP Var[Inv] ir<%incdec.ptr190> = getelementptr ir<%0>, ir<1>
21-
; CHECK-NEXT: WIDEN store ir<%curchar>, ir<%incdec.ptr190>
22-
; CHECK-NEXT: WIDEN ir<%1> = load ir<%0>
23-
; CHECK-NEXT: WIDEN ir<%2> = add ir<%1>, ir<1>
24-
; CHECK-NEXT: WIDEN store ir<%0>, ir<%2>
12+
; CHECK-NEXT: WIDEN-INDUCTION %iv = phi 0, %iv.next
13+
; CHECK-NEXT: WIDEN-PHI %ptr.iv.1 = phi %start.1, %ptr.iv.1.next
14+
; CHECK-NEXT: WIDEN-PHI %ptr.iv.2 = phi %start.2, %ptr.iv.2.next
15+
; CHECK-NEXT: WIDEN-GEP Var[Inv] ir<%ptr.iv.2.next> = getelementptr ir<%ptr.iv.2>, ir<1>
16+
; CHECK-NEXT: WIDEN store ir<%ptr.iv.1>, ir<%ptr.iv.2.next>
17+
; CHECK-NEXT: WIDEN ir<%lv> = load ir<%ptr.iv.2>
18+
; CHECK-NEXT: WIDEN ir<%add> = add ir<%lv>, ir<1>
19+
; CHECK-NEXT: WIDEN store ir<%ptr.iv.2>, ir<%add>
2520
; CHECK-NEXT: No successors
2621
; CHECK-NEXT: }
2722

28-
define void @phi_multiple_use(i8** noalias %curptr, i8* noalias %src, i64 %N) #0 {
29-
; CHECK-LABEL: @phi_multiple_use(
23+
; In the test below the pointer phi %ptr.iv.2 is used as
24+
; 1. As a uniform address for the load, and
25+
; 2. Non-uniform use by the getelementptr which is stored. This requires the
26+
; vector value.
27+
define void @pointer_induction_used_as_vector(i8** noalias %start.1, i8* noalias %start.2, i64 %N) {
28+
; CHECK-LABEL: @pointer_induction_used_as_vector(
29+
; CHECK-NEXT: entry:
30+
; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
31+
; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 2
32+
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]]
33+
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
34+
; CHECK: vector.ph:
35+
; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
36+
; CHECK-NEXT: [[TMP3:%.*]] = mul i64 [[TMP2]], 2
37+
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]
38+
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
39+
; CHECK-NEXT: [[IND_END:%.*]] = getelementptr i8*, i8** [[START_1:%.*]], i64 [[N_VEC]]
40+
; CHECK-NEXT: [[IND_END3:%.*]] = getelementptr i8, i8* [[START_2:%.*]], i64 [[N_VEC]]
41+
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
3042
; CHECK: vector.body:
31-
; CHECK-NEXT: [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ {{.*}}, %vector.body ]
32-
; CHECK-NEXT: {{.*}} = add i64 [[INDEX1]], 0
33-
; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[INDEX1]], 0
34-
; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8*, i8** %curptr, i64 [[TMP1]]
35-
; CHECK-NEXT: [[TMP2:%.*]] = call <vscale x 2 x i64> @llvm.experimental.stepvector.nxv2i64()
36-
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[INDEX1]], i32 0
43+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
44+
; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[INDEX]], 0
45+
; CHECK-NEXT: [[TMP5:%.*]] = add i64 [[INDEX]], 0
46+
; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8*, i8** [[START_1]], i64 [[TMP5]]
47+
; CHECK-NEXT: [[TMP6:%.*]] = call <vscale x 2 x i64> @llvm.experimental.stepvector.nxv2i64()
48+
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[INDEX]], i32 0
3749
; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[DOTSPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
38-
; CHECK-NEXT: [[TMP3:%.*]] = add <vscale x 2 x i64> shufflevector (<vscale x 2 x i64> insertelement (<vscale x 2 x i64> poison, i64 0, i32 0), <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer), [[TMP2]]
39-
; CHECK-NEXT: [[TMP4:%.*]] = add <vscale x 2 x i64> [[DOTSPLAT]], [[TMP3]]
40-
; CHECK-NEXT: [[NEXT_GEP6:%.*]] = getelementptr i8, i8* %src, <vscale x 2 x i64> [[TMP4]]
41-
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds i8, <vscale x 2 x i8*> [[NEXT_GEP6]], i64 1
42-
; CHECK: store <vscale x 2 x i8*> [[TMP5]], <vscale x 2 x i8*>*
43-
; CHECK-NEXT: [[TMP6:%.*]] = extractelement <vscale x 2 x i8*> [[NEXT_GEP6]], i32 0
44-
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr i8, i8* [[TMP6]], i32 0
45-
; CHECK-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to <vscale x 2 x i8>*
46-
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 2 x i8>, <vscale x 2 x i8>* [[TMP8]]
47-
; CHECK-NEXT: [[TMP9:%.*]] = add <vscale x 2 x i8> [[WIDE_LOAD]],
48-
; CHECK: store <vscale x 2 x i8> [[TMP9]], <vscale x 2 x i8>*
50+
; CHECK-NEXT: [[TMP7:%.*]] = add <vscale x 2 x i64> shufflevector (<vscale x 2 x i64> insertelement (<vscale x 2 x i64> poison, i64 0, i32 0), <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer), [[TMP6]]
51+
; CHECK-NEXT: [[TMP8:%.*]] = add <vscale x 2 x i64> [[DOTSPLAT]], [[TMP7]]
52+
; CHECK-NEXT: [[NEXT_GEP4:%.*]] = getelementptr i8, i8* [[START_2]], <vscale x 2 x i64> [[TMP8]]
53+
; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8, <vscale x 2 x i8*> [[NEXT_GEP4]], i64 1
54+
; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i8*, i8** [[NEXT_GEP]], i32 0
55+
; CHECK-NEXT: [[TMP11:%.*]] = bitcast i8** [[TMP10]] to <vscale x 2 x i8*>*
56+
; CHECK-NEXT: store <vscale x 2 x i8*> [[TMP9]], <vscale x 2 x i8*>* [[TMP11]], align 8
57+
; CHECK-NEXT: [[TMP12:%.*]] = extractelement <vscale x 2 x i8*> [[NEXT_GEP4]], i32 0
58+
; CHECK-NEXT: [[TMP13:%.*]] = getelementptr i8, i8* [[TMP12]], i32 0
59+
; CHECK-NEXT: [[TMP14:%.*]] = bitcast i8* [[TMP13]] to <vscale x 2 x i8>*
60+
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 2 x i8>, <vscale x 2 x i8>* [[TMP14]], align 1
61+
; CHECK-NEXT: [[TMP15:%.*]] = add <vscale x 2 x i8> [[WIDE_LOAD]], shufflevector (<vscale x 2 x i8> insertelement (<vscale x 2 x i8> poison, i8 1, i32 0), <vscale x 2 x i8> poison, <vscale x 2 x i32> zeroinitializer)
62+
; CHECK-NEXT: [[TMP16:%.*]] = bitcast i8* [[TMP13]] to <vscale x 2 x i8>*
63+
; CHECK-NEXT: store <vscale x 2 x i8> [[TMP15]], <vscale x 2 x i8>* [[TMP16]], align 1
64+
; CHECK-NEXT: [[TMP17:%.*]] = call i64 @llvm.vscale.i64()
65+
; CHECK-NEXT: [[TMP18:%.*]] = mul i64 [[TMP17]], 2
66+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP18]]
67+
; CHECK-NEXT: [[TMP19:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
68+
; CHECK-NEXT: br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
69+
; CHECK: middle.block:
70+
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
71+
; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
72+
; CHECK: scalar.ph:
73+
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
74+
; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi i8** [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[START_1]], [[ENTRY]] ]
75+
; CHECK-NEXT: [[BC_RESUME_VAL2:%.*]] = phi i8* [ [[IND_END3]], [[MIDDLE_BLOCK]] ], [ [[START_2]], [[ENTRY]] ]
76+
; CHECK-NEXT: br label [[LOOP_BODY:%.*]]
77+
; CHECK: loop.body:
78+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP_BODY]] ]
79+
; CHECK-NEXT: [[PTR_IV_1:%.*]] = phi i8** [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ], [ [[PTR_IV_1_NEXT:%.*]], [[LOOP_BODY]] ]
80+
; CHECK-NEXT: [[PTR_IV_2:%.*]] = phi i8* [ [[BC_RESUME_VAL2]], [[SCALAR_PH]] ], [ [[PTR_IV_2_NEXT:%.*]], [[LOOP_BODY]] ]
81+
; CHECK-NEXT: [[PTR_IV_1_NEXT]] = getelementptr inbounds i8*, i8** [[PTR_IV_1]], i64 1
82+
; CHECK-NEXT: [[PTR_IV_2_NEXT]] = getelementptr inbounds i8, i8* [[PTR_IV_2]], i64 1
83+
; CHECK-NEXT: store i8* [[PTR_IV_2_NEXT]], i8** [[PTR_IV_1]], align 8
84+
; CHECK-NEXT: [[LV:%.*]] = load i8, i8* [[PTR_IV_2]], align 1
85+
; CHECK-NEXT: [[ADD:%.*]] = add i8 [[LV]], 1
86+
; CHECK-NEXT: store i8 [[ADD]], i8* [[PTR_IV_2]], align 1
87+
; CHECK-NEXT: [[IV_NEXT]] = add nuw i64 [[IV]], 1
88+
; CHECK-NEXT: [[C:%.*]] = icmp ne i64 [[IV_NEXT]], [[N]]
89+
; CHECK-NEXT: br i1 [[C]], label [[LOOP_BODY]], label [[EXIT]], !llvm.loop [[LOOP2:![0-9]+]]
90+
; CHECK: exit:
91+
; CHECK-NEXT: ret void
92+
;
93+
4994

5095
entry:
5196
br label %loop.body
5297

5398
loop.body: ; preds = %loop.body, %entry
54-
%index = phi i64 [ 0, %entry ], [ %index.next, %loop.body ]
55-
%curchar = phi i8** [ %curchar.next, %loop.body ], [ %curptr, %entry ]
56-
%0 = phi i8* [ %incdec.ptr190, %loop.body ], [ %src, %entry ]
57-
%incdec.ptr190 = getelementptr inbounds i8, i8* %0, i64 1
58-
%curchar.next = getelementptr inbounds i8*, i8** %curchar, i64 1
59-
store i8* %incdec.ptr190, i8** %curchar, align 8
60-
%1 = load i8, i8* %0, align 1
61-
%2 = add i8 %1, 1
62-
store i8 %2, i8* %0, align 1
63-
%index.next = add nuw i64 %index, 1
64-
%3 = icmp ne i64 %index.next, %N
65-
br i1 %3, label %loop.body, label %exit, !llvm.loop !0
99+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.body ]
100+
%ptr.iv.1 = phi i8** [ %start.1, %entry ], [ %ptr.iv.1.next, %loop.body ]
101+
%ptr.iv.2 = phi i8* [ %start.2, %entry ], [ %ptr.iv.2.next, %loop.body ]
102+
%ptr.iv.1.next = getelementptr inbounds i8*, i8** %ptr.iv.1, i64 1
103+
%ptr.iv.2.next = getelementptr inbounds i8, i8* %ptr.iv.2, i64 1
104+
store i8* %ptr.iv.2.next, i8** %ptr.iv.1, align 8
105+
%lv = load i8, i8* %ptr.iv.2, align 1
106+
%add = add i8 %lv, 1
107+
store i8 %add, i8* %ptr.iv.2, align 1
108+
%iv.next = add nuw i64 %iv, 1
109+
%c = icmp ne i64 %iv.next, %N
110+
br i1 %c, label %loop.body, label %exit, !llvm.loop !0
66111

67112
exit: ; preds = %loop.body
68113
ret void

0 commit comments

Comments
 (0)