1; RUN: opt < %s -inline -inline-threshold=20 -S | FileCheck %s
2
3define internal i32 @callee1(i32 %A, i32 %B) {
4  %C = sdiv i32 %A, %B
5  ret i32 %C
6}
7
8define i32 @caller1() {
9; CHECK: define i32 @caller1
10; CHECK-NEXT: ret i32 3
11
12  %X = call i32 @callee1( i32 10, i32 3 )
13  ret i32 %X
14}
15
16define i32 @caller2() {
17; Check that we can constant-prop through instructions after inlining callee21
18; to get constants in the inlined callsite to callee22.
19; FIXME: Currently, the threshold is fixed at 20 because we don't perform
20; *recursive* cost analysis to realize that the nested call site will definitely
21; inline and be cheap. We should eventually do that and lower the threshold here
22; to 1.
23;
24; CHECK: @caller2
25; CHECK-NOT: call void @callee2
26; CHECK: ret
27
28  %x = call i32 @callee21(i32 42, i32 48)
29  ret i32 %x
30}
31
32define i32 @callee21(i32 %x, i32 %y) {
33  %sub = sub i32 %y, %x
34  %result = call i32 @callee22(i32 %sub)
35  ret i32 %result
36}
37
38declare i8* @getptr()
39
40define i32 @callee22(i32 %x) {
41  %icmp = icmp ugt i32 %x, 42
42  br i1 %icmp, label %bb.true, label %bb.false
43bb.true:
44  ; This block musn't be counted in the inline cost.
45  %x1 = add i32 %x, 1
46  %x2 = add i32 %x1, 1
47  %x3 = add i32 %x2, 1
48  %x4 = add i32 %x3, 1
49  %x5 = add i32 %x4, 1
50  %x6 = add i32 %x5, 1
51  %x7 = add i32 %x6, 1
52  %x8 = add i32 %x7, 1
53
54  ret i32 %x8
55bb.false:
56  ret i32 %x
57}
58
59define i32 @caller3() {
60; Check that even if the expensive path is hidden behind several basic blocks,
61; it doesn't count toward the inline cost when constant-prop proves those paths
62; dead.
63;
64; CHECK: @caller3
65; CHECK-NOT: call
66; CHECK: ret i32 6
67
68entry:
69  %x = call i32 @callee3(i32 42, i32 48)
70  ret i32 %x
71}
72
73define i32 @callee3(i32 %x, i32 %y) {
74  %sub = sub i32 %y, %x
75  %icmp = icmp ugt i32 %sub, 42
76  br i1 %icmp, label %bb.true, label %bb.false
77
78bb.true:
79  %icmp2 = icmp ult i32 %sub, 64
80  br i1 %icmp2, label %bb.true.true, label %bb.true.false
81
82bb.true.true:
83  ; This block musn't be counted in the inline cost.
84  %x1 = add i32 %x, 1
85  %x2 = add i32 %x1, 1
86  %x3 = add i32 %x2, 1
87  %x4 = add i32 %x3, 1
88  %x5 = add i32 %x4, 1
89  %x6 = add i32 %x5, 1
90  %x7 = add i32 %x6, 1
91  %x8 = add i32 %x7, 1
92  br label %bb.merge
93
94bb.true.false:
95  ; This block musn't be counted in the inline cost.
96  %y1 = add i32 %y, 1
97  %y2 = add i32 %y1, 1
98  %y3 = add i32 %y2, 1
99  %y4 = add i32 %y3, 1
100  %y5 = add i32 %y4, 1
101  %y6 = add i32 %y5, 1
102  %y7 = add i32 %y6, 1
103  %y8 = add i32 %y7, 1
104  br label %bb.merge
105
106bb.merge:
107  %result = phi i32 [ %x8, %bb.true.true ], [ %y8, %bb.true.false ]
108  ret i32 %result
109
110bb.false:
111  ret i32 %sub
112}
113
114
115define i32 @PR13412.main() {
116; This is a somewhat complicated three layer subprogram that was reported to
117; compute the wrong value for a branch due to assuming that an argument
118; mid-inline couldn't be equal to another pointer.
119;
120; After inlining, the branch should point directly to the exit block, not to
121; the intermediate block.
122; CHECK: @PR13412.main
123; CHECK: br i1 true, label %[[TRUE_DEST:.*]], label %[[FALSE_DEST:.*]]
124; CHECK: [[FALSE_DEST]]:
125; CHECK-NEXT: call void @PR13412.fail()
126; CHECK: [[TRUE_DEST]]:
127; CHECK-NEXT: ret i32 0
128
129entry:
130  %i1 = alloca i64
131  store i64 0, i64* %i1
132  %arraydecay = bitcast i64* %i1 to i32*
133  %call = call i1 @PR13412.first(i32* %arraydecay, i32* %arraydecay)
134  br i1 %call, label %cond.end, label %cond.false
135
136cond.false:
137  call void @PR13412.fail()
138  br label %cond.end
139
140cond.end:
141  ret i32 0
142}
143
144define internal i1 @PR13412.first(i32* %a, i32* %b) {
145entry:
146  %call = call i32* @PR13412.second(i32* %a, i32* %b)
147  %cmp = icmp eq i32* %call, %b
148  ret i1 %cmp
149}
150
151declare void @PR13412.fail()
152
153define internal i32* @PR13412.second(i32* %a, i32* %b) {
154entry:
155  %sub.ptr.lhs.cast = ptrtoint i32* %b to i64
156  %sub.ptr.rhs.cast = ptrtoint i32* %a to i64
157  %sub.ptr.sub = sub i64 %sub.ptr.lhs.cast, %sub.ptr.rhs.cast
158  %sub.ptr.div = ashr exact i64 %sub.ptr.sub, 2
159  %cmp = icmp ugt i64 %sub.ptr.div, 1
160  br i1 %cmp, label %if.then, label %if.end3
161
162if.then:
163  %0 = load i32* %a
164  %1 = load i32* %b
165  %cmp1 = icmp eq i32 %0, %1
166  br i1 %cmp1, label %return, label %if.end3
167
168if.end3:
169  br label %return
170
171return:
172  %retval.0 = phi i32* [ %b, %if.end3 ], [ %a, %if.then ]
173  ret i32* %retval.0
174}
175