xref: /llvm-project/llvm/test/ThinLTO/X86/memprof-inlined2.ll (revision 176889868024d98db032842bc47b416997d9e349)
1;; Test callsite context graph generation for call graph with two memprof
2;; contexts and multiple levels of inlining, requiring generation of new
3;; fused nodes to represent the inlined sequence while matching callsite
4;; nodes onto the graph. In particular this tests the case where a function
5;; has inlined a callee containing an inlined callee.
6;;
7;; Original code looks like:
8;;
9;; char *bar() __attribute__((noinline)) {
10;;   return new char[10];
11;; }
12;;
13;; char *baz() {
14;;   return bar();
15;; }
16;;
17;; char *foo() {
18;;   return baz();
19;; }
20;;
21;; int main(int argc, char **argv) {
22;;   char *x = foo();
23;;   char *y = foo();
24;;   memset(x, 0, 10);
25;;   memset(y, 0, 10);
26;;   delete[] x;
27;;   sleep(10);
28;;   delete[] y;
29;;   return 0;
30;; }
31;;
32;; Code compiled with -mllvm -memprof-ave-lifetime-cold-threshold=5 so that the
33;; memory freed after sleep(10) results in cold lifetimes.
34;;
35;; Both foo and baz are inlined into main, at both foo callsites.
36;; We should update the graph for new fused nodes for both of those inlined
37;; callsites to bar.
38;;
39;; Note that baz and bar are both dead due to the inlining, but have been left
40;; in the input IR to ensure that the MIB call chain is matched to the longer
41;; inline sequences from main.
42;;
43;; The IR was then reduced using llvm-reduce with the expected FileCheck input.
44
45; RUN: opt -thinlto-bc %s >%t.o
46; RUN: llvm-lto2 run %t.o -enable-memprof-context-disambiguation \
47; RUN:  -supports-hot-cold-new \
48; RUN:  -r=%t.o,main,plx \
49; RUN:  -r=%t.o,_Z3barv,plx \
50; RUN:  -r=%t.o,_Z3bazv,plx \
51; RUN:  -r=%t.o,_Z3foov,plx \
52; RUN:  -r=%t.o,_ZdaPv, \
53; RUN:  -r=%t.o,sleep, \
54; RUN:  -r=%t.o,_Znam, \
55; RUN:  -memprof-verify-ccg -memprof-verify-nodes -memprof-dump-ccg \
56; RUN:  -o %t.out 2>&1 | FileCheck %s --check-prefix=DUMP
57
58
59target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
60target triple = "x86_64-unknown-linux-gnu"
61
62define ptr @_Z3barv() {
63entry:
64  %call = call ptr @_Znam(i64 0), !memprof !0, !callsite !5
65  ret ptr null
66}
67
68declare ptr @_Znam(i64)
69
70declare ptr @_Z3bazv()
71
72declare ptr @_Z3foov()
73
74define i32 @main() {
75delete.end5:
76  %call.i.i = call ptr @_Z3barv(), !callsite !6
77  %call.i.i8 = call ptr @_Z3barv(), !callsite !7
78  ret i32 0
79}
80
81declare void @_ZdaPv()
82
83declare i32 @sleep()
84
85!0 = !{!1, !3}
86!1 = !{!2, !"notcold"}
87!2 = !{i64 9086428284934609951, i64 -5964873800580613432, i64 2732490490862098848, i64 8632435727821051414}
88!3 = !{!4, !"cold"}
89!4 = !{i64 9086428284934609951, i64 -5964873800580613432, i64 2732490490862098848, i64 -3421689549917153178}
90!5 = !{i64 9086428284934609951}
91!6 = !{i64 -5964873800580613432, i64 2732490490862098848, i64 8632435727821051414}
92!7 = !{i64 -5964873800580613432, i64 2732490490862098848, i64 -3421689549917153178}
93
94
95; DUMP: CCG before cloning:
96; DUMP: Callsite Context Graph:
97; DUMP: Node [[BAR:0x[a-z0-9]+]]
98; DUMP: 	Versions: 1 MIB:
99; DUMP: 		AllocType 1 StackIds: 0, 1, 2
100; DUMP: 		AllocType 2 StackIds: 0, 1, 3
101; DUMP: 	(clone 0)
102; DUMP: 	AllocTypes: NotColdCold
103; DUMP: 	ContextIds: 1 2
104; DUMP: 	CalleeEdges:
105; DUMP: 	CallerEdges:
106; DUMP: 		Edge from Callee [[BAR]] to Caller: [[MAIN1:0x[a-z0-9]+]] AllocTypes: NotCold ContextIds: 1
107; DUMP: 		Edge from Callee [[BAR]] to Caller: [[MAIN2:0x[a-z0-9]+]] AllocTypes: Cold ContextIds: 2
108
109;; This is the node synthesized for the first inlined call chain of main->foo->baz
110; DUMP: Node [[MAIN1]]
111; DUMP: 	Callee: 17377440600225628772 (_Z3barv) Clones: 0 StackIds: 0, 1, 2	(clone 0)
112; DUMP: 	AllocTypes: NotCold
113; DUMP: 	ContextIds: 1
114; DUMP: 	CalleeEdges:
115; DUMP: 		Edge from Callee [[BAR]] to Caller: [[MAIN1]] AllocTypes: NotCold ContextIds: 1
116; DUMP: 	CallerEdges:
117
118;; This is the node synthesized for the second inlined call chain of main->foo->baz
119; DUMP: Node [[MAIN2]]
120; DUMP: 	Callee: 17377440600225628772 (_Z3barv) Clones: 0 StackIds: 0, 1, 3	(clone 0)
121; DUMP: 	AllocTypes: Cold
122; DUMP: 	ContextIds: 2
123; DUMP: 	CalleeEdges:
124; DUMP: 		Edge from Callee [[BAR]] to Caller: [[MAIN2]] AllocTypes: Cold ContextIds: 2
125; DUMP: 	CallerEdges:
126