Skip to content

Commit 68210c7

Browse files
committed
[VPlan] Only generate first lane for VPPredInstPHI if no others used.
IF only the first lane of the result is used, only generate the first lane. Fixes #111042.
1 parent 67c0846 commit 68210c7

File tree

4 files changed

+121
-6
lines changed

4 files changed

+121
-6
lines changed

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2185,6 +2185,9 @@ void VPPredInstPHIRecipe::execute(VPTransformState &State) {
21852185
// predicated iteration inserts its generated value in the correct vector.
21862186
State.reset(getOperand(0), VPhi);
21872187
} else {
2188+
if (vputils::onlyFirstLaneUsed(this) && !State.Lane->isFirstLane())
2189+
return;
2190+
21882191
Type *PredInstType = getOperand(0)->getUnderlyingValue()->getType();
21892192
PHINode *Phi = State.Builder.CreatePHI(PredInstType, 2);
21902193
Phi->addIncoming(PoisonValue::get(ScalarPredInst->getType()),

llvm/test/Transforms/LoopVectorize/X86/divs-with-tail-folding.ll

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -133,21 +133,18 @@ define void @sdiv_feeding_gep_predicated(ptr %dst, i32 %x, i64 %M, i64 %conv6, i
133133
; CHECK-NEXT: [[TMP13:%.*]] = sdiv i64 [[M]], [[CONV6]]
134134
; CHECK-NEXT: br label %[[PRED_SDIV_CONTINUE4]]
135135
; CHECK: [[PRED_SDIV_CONTINUE4]]:
136-
; CHECK-NEXT: [[TMP14:%.*]] = phi i64 [ poison, %[[PRED_SDIV_CONTINUE]] ], [ [[TMP13]], %[[PRED_SDIV_IF3]] ]
137136
; CHECK-NEXT: [[TMP15:%.*]] = extractelement <4 x i1> [[TMP8]], i32 2
138137
; CHECK-NEXT: br i1 [[TMP15]], label %[[PRED_SDIV_IF5:.*]], label %[[PRED_SDIV_CONTINUE6:.*]]
139138
; CHECK: [[PRED_SDIV_IF5]]:
140139
; CHECK-NEXT: [[TMP16:%.*]] = sdiv i64 [[M]], [[CONV6]]
141140
; CHECK-NEXT: br label %[[PRED_SDIV_CONTINUE6]]
142141
; CHECK: [[PRED_SDIV_CONTINUE6]]:
143-
; CHECK-NEXT: [[TMP17:%.*]] = phi i64 [ poison, %[[PRED_SDIV_CONTINUE4]] ], [ [[TMP16]], %[[PRED_SDIV_IF5]] ]
144142
; CHECK-NEXT: [[TMP18:%.*]] = extractelement <4 x i1> [[TMP8]], i32 3
145143
; CHECK-NEXT: br i1 [[TMP18]], label %[[PRED_SDIV_IF7:.*]], label %[[PRED_SDIV_CONTINUE8]]
146144
; CHECK: [[PRED_SDIV_IF7]]:
147145
; CHECK-NEXT: [[TMP19:%.*]] = sdiv i64 [[M]], [[CONV6]]
148146
; CHECK-NEXT: br label %[[PRED_SDIV_CONTINUE8]]
149147
; CHECK: [[PRED_SDIV_CONTINUE8]]:
150-
; CHECK-NEXT: [[TMP20:%.*]] = phi i64 [ poison, %[[PRED_SDIV_CONTINUE6]] ], [ [[TMP19]], %[[PRED_SDIV_IF7]] ]
151148
; CHECK-NEXT: [[TMP21:%.*]] = trunc i64 [[TMP11]] to i32
152149
; CHECK-NEXT: [[TMP22:%.*]] = mul i64 [[TMP11]], [[CONV61]]
153150
; CHECK-NEXT: [[TMP23:%.*]] = sub i64 [[TMP5]], [[TMP22]]

llvm/test/Transforms/LoopVectorize/X86/drop-poison-generating-flags.ll

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -509,21 +509,18 @@ define void @pr70590_recipe_without_underlying_instr(i64 %n, ptr noalias %dst) {
509509
; CHECK-NEXT: [[TMP7:%.*]] = srem i64 3, 0
510510
; CHECK-NEXT: br label [[PRED_SREM_CONTINUE2]]
511511
; CHECK: pred.srem.continue2:
512-
; CHECK-NEXT: [[TMP8:%.*]] = phi i64 [ poison, [[PRED_SREM_CONTINUE]] ], [ [[TMP7]], [[PRED_SREM_IF1]] ]
513512
; CHECK-NEXT: [[TMP9:%.*]] = extractelement <4 x i1> [[TMP2]], i32 2
514513
; CHECK-NEXT: br i1 [[TMP9]], label [[PRED_SREM_IF3:%.*]], label [[PRED_SREM_CONTINUE4:%.*]]
515514
; CHECK: pred.srem.if3:
516515
; CHECK-NEXT: [[TMP10:%.*]] = srem i64 3, 0
517516
; CHECK-NEXT: br label [[PRED_SREM_CONTINUE4]]
518517
; CHECK: pred.srem.continue4:
519-
; CHECK-NEXT: [[TMP11:%.*]] = phi i64 [ poison, [[PRED_SREM_CONTINUE2]] ], [ [[TMP10]], [[PRED_SREM_IF3]] ]
520518
; CHECK-NEXT: [[TMP12:%.*]] = extractelement <4 x i1> [[TMP2]], i32 3
521519
; CHECK-NEXT: br i1 [[TMP12]], label [[PRED_SREM_IF5:%.*]], label [[PRED_SREM_CONTINUE6]]
522520
; CHECK: pred.srem.if5:
523521
; CHECK-NEXT: [[TMP13:%.*]] = srem i64 3, 0
524522
; CHECK-NEXT: br label [[PRED_SREM_CONTINUE6]]
525523
; CHECK: pred.srem.continue6:
526-
; CHECK-NEXT: [[TMP14:%.*]] = phi i64 [ poison, [[PRED_SREM_CONTINUE4]] ], [ [[TMP13]], [[PRED_SREM_IF5]] ]
527524
; CHECK-NEXT: [[TMP15:%.*]] = add i64 [[TMP5]], -3
528525
; CHECK-NEXT: [[TMP16:%.*]] = add i64 [[TMP0]], [[TMP15]]
529526
; CHECK-NEXT: [[TMP17:%.*]] = getelementptr [5 x i8], ptr @c, i64 0, i64 [[TMP16]]
Lines changed: 118 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,118 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt -p loop-vectorize -force-vector-width=4 -force-vector-interleave=2 -S %s | FileCheck %s
3+
4+
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
5+
target triple = "x86_64-unknown-linux-gnu"
6+
7+
; Test case for https://github.com/llvm/llvm-project/issues/111042.
8+
define void @replicate_udiv_with_only_first_lane_used(i32 %x, ptr %dst, i64 %d) {
9+
; CHECK-LABEL: define void @replicate_udiv_with_only_first_lane_used(
10+
; CHECK-SAME: i32 [[X:%.*]], ptr [[DST:%.*]], i64 [[D:%.*]]) {
11+
; CHECK-NEXT: [[ENTRY:.*]]:
12+
; CHECK-NEXT: [[C:%.*]] = icmp eq i32 [[X]], 10
13+
; CHECK-NEXT: br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
14+
; CHECK: [[VECTOR_PH]]:
15+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
16+
; CHECK: [[VECTOR_BODY]]:
17+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_UDIV_CONTINUE14:.*]] ]
18+
; CHECK-NEXT: br i1 false, label %[[PRED_UDIV_IF:.*]], label %[[PRED_UDIV_CONTINUE:.*]]
19+
; CHECK: [[PRED_UDIV_IF]]:
20+
; CHECK-NEXT: [[TMP0:%.*]] = udiv i64 99, [[D]]
21+
; CHECK-NEXT: br label %[[PRED_UDIV_CONTINUE]]
22+
; CHECK: [[PRED_UDIV_CONTINUE]]:
23+
; CHECK-NEXT: [[TMP1:%.*]] = phi i64 [ poison, %[[VECTOR_BODY]] ], [ [[TMP0]], %[[PRED_UDIV_IF]] ]
24+
; CHECK-NEXT: br i1 false, label %[[PRED_UDIV_IF1:.*]], label %[[PRED_UDIV_CONTINUE2:.*]]
25+
; CHECK: [[PRED_UDIV_IF1]]:
26+
; CHECK-NEXT: [[TMP2:%.*]] = udiv i64 99, [[D]]
27+
; CHECK-NEXT: br label %[[PRED_UDIV_CONTINUE2]]
28+
; CHECK: [[PRED_UDIV_CONTINUE2]]:
29+
; CHECK-NEXT: br i1 false, label %[[PRED_UDIV_IF3:.*]], label %[[PRED_UDIV_CONTINUE4:.*]]
30+
; CHECK: [[PRED_UDIV_IF3]]:
31+
; CHECK-NEXT: [[TMP3:%.*]] = udiv i64 99, [[D]]
32+
; CHECK-NEXT: br label %[[PRED_UDIV_CONTINUE4]]
33+
; CHECK: [[PRED_UDIV_CONTINUE4]]:
34+
; CHECK-NEXT: br i1 false, label %[[PRED_UDIV_IF5:.*]], label %[[PRED_UDIV_CONTINUE6:.*]]
35+
; CHECK: [[PRED_UDIV_IF5]]:
36+
; CHECK-NEXT: [[TMP4:%.*]] = udiv i64 99, [[D]]
37+
; CHECK-NEXT: br label %[[PRED_UDIV_CONTINUE6]]
38+
; CHECK: [[PRED_UDIV_CONTINUE6]]:
39+
; CHECK-NEXT: br i1 false, label %[[PRED_UDIV_IF7:.*]], label %[[PRED_UDIV_CONTINUE8:.*]]
40+
; CHECK: [[PRED_UDIV_IF7]]:
41+
; CHECK-NEXT: [[TMP5:%.*]] = udiv i64 99, [[D]]
42+
; CHECK-NEXT: br label %[[PRED_UDIV_CONTINUE8]]
43+
; CHECK: [[PRED_UDIV_CONTINUE8]]:
44+
; CHECK-NEXT: [[TMP6:%.*]] = phi i64 [ poison, %[[PRED_UDIV_CONTINUE6]] ], [ [[TMP5]], %[[PRED_UDIV_IF7]] ]
45+
; CHECK-NEXT: br i1 false, label %[[PRED_UDIV_IF9:.*]], label %[[PRED_UDIV_CONTINUE10:.*]]
46+
; CHECK: [[PRED_UDIV_IF9]]:
47+
; CHECK-NEXT: [[TMP7:%.*]] = udiv i64 99, [[D]]
48+
; CHECK-NEXT: br label %[[PRED_UDIV_CONTINUE10]]
49+
; CHECK: [[PRED_UDIV_CONTINUE10]]:
50+
; CHECK-NEXT: br i1 false, label %[[PRED_UDIV_IF11:.*]], label %[[PRED_UDIV_CONTINUE12:.*]]
51+
; CHECK: [[PRED_UDIV_IF11]]:
52+
; CHECK-NEXT: [[TMP8:%.*]] = udiv i64 99, [[D]]
53+
; CHECK-NEXT: br label %[[PRED_UDIV_CONTINUE12]]
54+
; CHECK: [[PRED_UDIV_CONTINUE12]]:
55+
; CHECK-NEXT: br i1 false, label %[[PRED_UDIV_IF13:.*]], label %[[PRED_UDIV_CONTINUE14]]
56+
; CHECK: [[PRED_UDIV_IF13]]:
57+
; CHECK-NEXT: [[TMP9:%.*]] = udiv i64 99, [[D]]
58+
; CHECK-NEXT: br label %[[PRED_UDIV_CONTINUE14]]
59+
; CHECK: [[PRED_UDIV_CONTINUE14]]:
60+
; CHECK-NEXT: [[PREDPHI:%.*]] = select i1 true, i64 0, i64 [[TMP1]]
61+
; CHECK-NEXT: [[PREDPHI15:%.*]] = select i1 true, i64 0, i64 [[TMP6]]
62+
; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i16, ptr [[DST]], i64 [[PREDPHI]]
63+
; CHECK-NEXT: [[TMP11:%.*]] = getelementptr i16, ptr [[DST]], i64 [[PREDPHI15]]
64+
; CHECK-NEXT: store i16 0, ptr [[TMP10]], align 2
65+
; CHECK-NEXT: store i16 0, ptr [[TMP11]], align 2
66+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
67+
; CHECK-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 96
68+
; CHECK-NEXT: br i1 [[TMP12]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
69+
; CHECK: [[MIDDLE_BLOCK]]:
70+
; CHECK-NEXT: br i1 false, label %[[EXIT:.*]], label %[[SCALAR_PH]]
71+
; CHECK: [[SCALAR_PH]]:
72+
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 96, %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
73+
; CHECK-NEXT: br label %[[LOOP_HEADER:.*]]
74+
; CHECK: [[LOOP_HEADER]]:
75+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
76+
; CHECK-NEXT: br i1 true, label %[[LOOP_LATCH]], label %[[ELSE:.*]]
77+
; CHECK: [[ELSE]]:
78+
; CHECK-NEXT: [[DIV_I:%.*]] = udiv i64 99, [[D]]
79+
; CHECK-NEXT: br label %[[LOOP_LATCH]]
80+
; CHECK: [[LOOP_LATCH]]:
81+
; CHECK-NEXT: [[RETVAL_0_I:%.*]] = phi i64 [ [[DIV_I]], %[[ELSE]] ], [ 0, %[[LOOP_HEADER]] ]
82+
; CHECK-NEXT: [[GEP:%.*]] = getelementptr i16, ptr [[DST]], i64 [[RETVAL_0_I]]
83+
; CHECK-NEXT: store i16 0, ptr [[GEP]], align 2
84+
; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
85+
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i64 [[IV_NEXT]], 101
86+
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP_HEADER]], label %[[EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
87+
; CHECK: [[EXIT]]:
88+
; CHECK-NEXT: ret void
89+
;
90+
entry:
91+
%c = icmp eq i32 %x, 10
92+
br label %loop.header
93+
94+
loop.header: ; preds = %loop.latch, %entry
95+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]
96+
br i1 true, label %loop.latch, label %else
97+
98+
else:
99+
%div.i = udiv i64 99, %d
100+
br label %loop.latch
101+
102+
loop.latch:
103+
%retval.0.i = phi i64 [ %div.i, %else ], [ 0, %loop.header ]
104+
%gep = getelementptr i16, ptr %dst, i64 %retval.0.i
105+
store i16 0, ptr %gep, align 2
106+
%iv.next = add i64 %iv, 1
107+
%cmp = icmp ult i64 %iv.next, 101
108+
br i1 %cmp, label %loop.header, label %exit
109+
110+
exit:
111+
ret void
112+
}
113+
;.
114+
; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
115+
; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
116+
; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
117+
; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
118+
;.

0 commit comments

Comments
 (0)