Skip to content

Commit 3a8b488

Browse files
fhahnfhahn
committed
[LoopPeel] Add tests for peeling last iteration with loop guards.
Add additional test coverage for peeling the last iteration where information from loop guards is needed.
1 parent 8e50e88 commit 3a8b488

File tree

2 files changed

+194
-3
lines changed

2 files changed

+194
-3
lines changed

‎llvm/test/Transforms/LoopUnroll/peel-last-iteration-with-guards.ll

Lines changed: 191 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,191 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt -p loop-unroll -unroll-full-max-count=0 -S %s | FileCheck %s
3+
4+
declare void @foo(i32)
5+
6+
define void @peel_with_guard_known_nonnegative_1(i32 %n) {
7+
; CHECK-LABEL: define void @peel_with_guard_known_nonnegative_1(
8+
; CHECK-SAME: i32 [[N:%.*]]) {
9+
; CHECK-NEXT: [[ENTRY:.*:]]
10+
; CHECK-NEXT: [[PRECOND:%.*]] = icmp slt i32 [[N]], 0
11+
; CHECK-NEXT: br i1 [[PRECOND]], label %[[EXIT:.*]], label %[[PH:.*]]
12+
; CHECK: [[PH]]:
13+
; CHECK-NEXT: [[N_EXT:%.*]] = zext i32 [[N]] to i64
14+
; CHECK-NEXT: [[N_1:%.*]] = add i32 [[N]], 1
15+
; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N_1]] to i64
16+
; CHECK-NEXT: br label %[[LOOP:.*]]
17+
; CHECK: [[LOOP]]:
18+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
19+
; CHECK-NEXT: [[C:%.*]] = icmp eq i64 [[IV]], [[N_EXT]]
20+
; CHECK-NEXT: [[SPEC_SELECT:%.*]] = select i1 [[C]], i32 10, i32 20
21+
; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
22+
; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[WIDE_TRIP_COUNT]]
23+
; CHECK-NEXT: br i1 [[EC]], label %[[EXIT_LOOPEXIT:.*]], label %[[LOOP]]
24+
; CHECK: [[EXIT_LOOPEXIT]]:
25+
; CHECK-NEXT: br label %[[EXIT]]
26+
; CHECK: [[EXIT]]:
27+
; CHECK-NEXT: ret void
28+
;
29+
entry:
30+
%precond = icmp slt i32 %n, 0
31+
br i1 %precond, label %exit, label %ph
32+
33+
ph:
34+
%n.ext = zext i32 %n to i64
35+
%n.1 = add i32 %n, 1
36+
%wide.trip.count = zext i32 %n.1 to i64
37+
br label %loop
38+
39+
loop:
40+
%iv = phi i64 [ 0, %ph ], [ %iv.next, %loop ]
41+
%c = icmp eq i64 %iv, %n.ext
42+
%spec.select = select i1 %c, i32 10, i32 20
43+
%iv.next = add i64 %iv, 1
44+
%ec = icmp eq i64 %iv.next, %wide.trip.count
45+
br i1 %ec, label %exit, label %loop
46+
47+
exit:
48+
ret void
49+
}
50+
51+
define i32 @peel_with_guard_known_nonnegative_2(ptr %x, i32 %w) {
52+
; CHECK-LABEL: define i32 @peel_with_guard_known_nonnegative_2(
53+
; CHECK-SAME: ptr [[X:%.*]], i32 [[W:%.*]]) {
54+
; CHECK-NEXT: [[ENTRY:.*]]:
55+
; CHECK-NEXT: [[PRECOND:%.*]] = icmp sgt i32 [[W]], 0
56+
; CHECK-NEXT: br i1 [[PRECOND]], label %[[PH:.*]], label %[[EXIT:.*]]
57+
; CHECK: [[PH]]:
58+
; CHECK-NEXT: [[SUB:%.*]] = add nsw i32 [[W]], -1
59+
; CHECK-NEXT: [[TMP0:%.*]] = zext nneg i32 [[SUB]] to i64
60+
; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext nneg i32 [[W]] to i64
61+
; CHECK-NEXT: br label %[[LOOP_HEADER:.*]]
62+
; CHECK: [[LOOP_HEADER]]:
63+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
64+
; CHECK-NEXT: [[RED:%.*]] = phi i32 [ 0, %[[PH]] ], [ [[ADD:%.*]], %[[LOOP_LATCH]] ]
65+
; CHECK-NEXT: [[CMP1:%.*]] = icmp eq i64 [[IV]], [[TMP0]]
66+
; CHECK-NEXT: br i1 [[CMP1]], label %[[IF_THEN:.*]], label %[[LOOP_LATCH]]
67+
; CHECK: [[IF_THEN]]:
68+
; CHECK-NEXT: tail call void @foo()
69+
; CHECK-NEXT: br label %[[LOOP_LATCH]]
70+
; CHECK: [[LOOP_LATCH]]:
71+
; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds nuw i32, ptr [[X]], i64 [[IV]]
72+
; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
73+
; CHECK-NEXT: [[ADD]] = add nsw i32 [[TMP1]], [[RED]]
74+
; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
75+
; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[WIDE_TRIP_COUNT]]
76+
; CHECK-NEXT: br i1 [[EC]], label %[[EXIT_LOOPEXIT:.*]], label %[[LOOP_HEADER]]
77+
; CHECK: [[EXIT_LOOPEXIT]]:
78+
; CHECK-NEXT: [[ADD_LCSSA:%.*]] = phi i32 [ [[ADD]], %[[LOOP_LATCH]] ]
79+
; CHECK-NEXT: br label %[[EXIT]]
80+
; CHECK: [[EXIT]]:
81+
; CHECK-NEXT: [[RED_LCSSA:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[ADD_LCSSA]], %[[EXIT_LOOPEXIT]] ]
82+
; CHECK-NEXT: ret i32 [[RED_LCSSA]]
83+
;
84+
entry:
85+
%precond = icmp sgt i32 %w, 0
86+
br i1 %precond, label %ph, label %exit
87+
88+
ph:
89+
%sub = add nsw i32 %w, -1
90+
%0 = zext nneg i32 %sub to i64
91+
%wide.trip.count = zext nneg i32 %w to i64
92+
br label %loop.header
93+
94+
loop.header:
95+
%iv = phi i64 [ 0, %ph ], [ %iv.next, %loop.latch ]
96+
%red = phi i32 [ 0, %ph ], [ %add, %loop.latch ]
97+
%cmp1 = icmp eq i64 %iv, %0
98+
br i1 %cmp1, label %if.then, label %loop.latch
99+
100+
if.then:
101+
tail call void @foo()
102+
br label %loop.latch
103+
104+
loop.latch:
105+
%arrayidx = getelementptr inbounds nuw i32, ptr %x, i64 %iv
106+
%1 = load i32, ptr %arrayidx, align 4
107+
%add = add nsw i32 %1, %red
108+
%iv.next = add nuw nsw i64 %iv, 1
109+
%ec = icmp eq i64 %iv.next, %wide.trip.count
110+
br i1 %ec, label %exit, label %loop.header
111+
112+
exit:
113+
%red.lcssa = phi i32 [ 0, %entry ], [ %add, %loop.latch ]
114+
ret i32 %red.lcssa
115+
}
116+
117+
define void @peel_with_guard2(i32 %n) {
118+
; CHECK-LABEL: define void @peel_with_guard2(
119+
; CHECK-SAME: i32 [[N:%.*]]) {
120+
; CHECK-NEXT: [[ENTRY:.*:]]
121+
; CHECK-NEXT: [[SUB:%.*]] = add i32 [[N]], -1
122+
; CHECK-NEXT: [[PRECOND:%.*]] = icmp eq i32 [[SUB]], 0
123+
; CHECK-NEXT: br i1 [[PRECOND]], label %[[EXIT:.*]], label %[[LOOP_HEADER_PREHEADER:.*]]
124+
; CHECK: [[LOOP_HEADER_PREHEADER]]:
125+
; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -2
126+
; CHECK-NEXT: [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
127+
; CHECK-NEXT: br i1 [[TMP1]], label %[[LOOP_HEADER_PREHEADER_SPLIT:.*]], label %[[EXIT_LOOPEXIT_PEEL_BEGIN:.*]]
128+
; CHECK: [[LOOP_HEADER_PREHEADER_SPLIT]]:
129+
; CHECK-NEXT: br label %[[LOOP_HEADER:.*]]
130+
; CHECK: [[LOOP_HEADER]]:
131+
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ], [ 1, %[[LOOP_HEADER_PREHEADER_SPLIT]] ]
132+
; CHECK-NEXT: br i1 false, label %[[THEN:.*]], label %[[LOOP_LATCH]]
133+
; CHECK: [[THEN]]:
134+
; CHECK-NEXT: call void @foo()
135+
; CHECK-NEXT: br label %[[LOOP_LATCH]]
136+
; CHECK: [[LOOP_LATCH]]:
137+
; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 1
138+
; CHECK-NEXT: [[TMP2:%.*]] = sub i32 [[N]], 1
139+
; CHECK-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[TMP2]]
140+
; CHECK-NEXT: br i1 [[EC]], label %[[EXIT_LOOPEXIT_PEEL_BEGIN_LOOPEXIT:.*]], label %[[LOOP_HEADER]], !llvm.loop [[LOOP0:![0-9]+]]
141+
; CHECK: [[EXIT_LOOPEXIT_PEEL_BEGIN_LOOPEXIT]]:
142+
; CHECK-NEXT: [[DOTPH:%.*]] = phi i32 [ [[IV_NEXT]], %[[LOOP_LATCH]] ]
143+
; CHECK-NEXT: br label %[[EXIT_LOOPEXIT_PEEL_BEGIN]]
144+
; CHECK: [[EXIT_LOOPEXIT_PEEL_BEGIN]]:
145+
; CHECK-NEXT: [[TMP3:%.*]] = phi i32 [ 1, %[[LOOP_HEADER_PREHEADER]] ], [ [[DOTPH]], %[[EXIT_LOOPEXIT_PEEL_BEGIN_LOOPEXIT]] ]
146+
; CHECK-NEXT: br label %[[LOOP_HEADER_PEEL:.*]]
147+
; CHECK: [[LOOP_HEADER_PEEL]]:
148+
; CHECK-NEXT: [[C_PEEL:%.*]] = icmp eq i32 [[TMP3]], [[SUB]]
149+
; CHECK-NEXT: br i1 [[C_PEEL]], label %[[THEN_PEEL:.*]], label %[[LOOP_LATCH_PEEL:.*]]
150+
; CHECK: [[THEN_PEEL]]:
151+
; CHECK-NEXT: call void @foo()
152+
; CHECK-NEXT: br label %[[LOOP_LATCH_PEEL]]
153+
; CHECK: [[LOOP_LATCH_PEEL]]:
154+
; CHECK-NEXT: [[IV_NEXT_PEEL:%.*]] = add nuw i32 [[TMP3]], 1
155+
; CHECK-NEXT: [[EC_PEEL:%.*]] = icmp eq i32 [[IV_NEXT_PEEL]], [[N]]
156+
; CHECK-NEXT: br i1 [[EC_PEEL]], label %[[EXIT_LOOPEXIT_PEEL_NEXT:.*]], label %[[EXIT_LOOPEXIT_PEEL_NEXT]]
157+
; CHECK: [[EXIT_LOOPEXIT_PEEL_NEXT]]:
158+
; CHECK-NEXT: br label %[[LOOP_HEADER_PEEL_NEXT:.*]]
159+
; CHECK: [[LOOP_HEADER_PEEL_NEXT]]:
160+
; CHECK-NEXT: br label %[[EXIT_LOOPEXIT:.*]]
161+
; CHECK: [[EXIT_LOOPEXIT]]:
162+
; CHECK-NEXT: br label %[[EXIT]]
163+
; CHECK: [[EXIT]]:
164+
; CHECK-NEXT: ret void
165+
;
166+
entry:
167+
%sub = add i32 %n, -1
168+
%precond = icmp eq i32 %sub, 0
169+
br i1 %precond, label %exit, label %loop.header
170+
171+
loop.header:
172+
%iv = phi i32 [ 1, %entry ], [ %iv.next, %loop.latch ]
173+
%c = icmp eq i32 %iv, %sub
174+
br i1 %c, label %then, label %loop.latch
175+
176+
then:
177+
call void @foo()
178+
br label %loop.latch
179+
180+
loop.latch:
181+
%iv.next = add nuw i32 %iv, 1
182+
%ec = icmp eq i32 %iv.next, %n
183+
br i1 %ec, label %exit, label %loop.header
184+
185+
exit:
186+
ret void
187+
}
188+
;.
189+
; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]]}
190+
; CHECK: [[META1]] = !{!"llvm.loop.peeled.count", i32 1}
191+
;.

‎llvm/test/Transforms/LoopUnroll/peel-last-iteration-with-variable-trip-count.ll

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,6 @@
11
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
22
; RUN: opt -p loop-unroll -unroll-full-max-count=0 -S %s | FileCheck %s
33

4-
54
declare void @foo(i32)
65

76
define i32 @peel_last_with_trip_count_check_lcssa_phi(i32 %n) {
@@ -235,9 +234,10 @@ then:
235234

236235
inner.latch:
237236
%iv.next = add i32 %iv, 1
238-
%exitcond.not = icmp eq i32 %iv.next, %n
239-
br i1 %exitcond.not, label %outer.header, label %inner.header
237+
%ec = icmp eq i32 %iv.next, %n
238+
br i1 %ec, label %outer.header, label %inner.header
240239
}
240+
241241
;.
242242
; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]]}
243243
; CHECK: [[META1]] = !{!"llvm.loop.peeled.count", i32 1}

0 commit comments

Comments
 (0)