llvm-project

1
; RUN: opt %loadNPMPolly -passes=polly-codegen -polly-codegen-perf-monitoring \
2
; RUN:   -S < %s | FileCheck %s
3

4
; void f(long A[], long N) {
5
;   long i;
6
;   if (true)
7
;     for (i = 0; i < N; ++i)
8
;       A[i] = i;
9
; }
10

11
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
12
target triple = "x86_64-unknown-linux-gnu"
13

14
define void @f(ptr %A, i64 %N) nounwind {
15
entry:
16
  fence seq_cst
17
  br label %next
18

19
next:
20
  br i1 true, label %for.i, label %return
21

22
for.i:
23
  %indvar = phi i64 [ 0, %next], [ %indvar.next, %for.i ]
24
  %scevgep = getelementptr i64, ptr %A, i64 %indvar
25
  store i64 %indvar, ptr %scevgep
26
  %indvar.next = add nsw i64 %indvar, 1
27
  %exitcond = icmp eq i64 %indvar.next, %N
28
  br i1 %exitcond, label %return, label %for.i
29

30
return:
31
  fence seq_cst
32
  ret void
33
}
34

35
; CHECK:      @__polly_perf_cycles_total_start = weak thread_local(initialexec) constant i64 0
36
; CHECK-NEXT: @__polly_perf_initialized = weak thread_local(initialexec) constant i1 false
37
; CHECK-NEXT: @__polly_perf_cycles_in_scops = weak thread_local(initialexec) constant i64 0
38
; CHECK-NEXT: @__polly_perf_cycles_in_scop_start = weak thread_local(initialexec) constant i64 0
39

40
; CHECK:      polly.split_new_and_old:                          ; preds = %entry
41
; CHECK-NEXT:   %0 = call { i64, i32 } @llvm.x86.rdtscp()
42
; CHECK-NEXT:   %1 = extractvalue { i64, i32 } %0, 0
43
; CHECK-NEXT:   store volatile i64 %1, ptr @__polly_perf_cycles_in_scop_start
44

45
; CHECK:      polly.merge_new_and_old:                          ; preds = %polly.exiting, %return.region_exiting
46
; CHECK-NEXT:   %6 = load volatile i64, ptr @__polly_perf_cycles_in_scop_start
47
; CHECK-NEXT:   %7 = call { i64, i32 } @llvm.x86.rdtscp()
48
; CHECK-NEXT:   %8 = extractvalue { i64, i32 } %7, 0
49
; CHECK-NEXT:   %9 = sub i64 %8, %6
50
; CHECK-NEXT:   %10 = load volatile i64, ptr @__polly_perf_cycles_in_scops
51
; CHECK-NEXT:   %11 = add i64 %10, %9
52
; CHECK-NEXT:   store volatile i64 %11, ptr @__polly_perf_cycles_in_scops
53

54

55
; CHECK:      define weak_odr void @__polly_perf_final() {
56
; CHECK-NEXT: start:
57
; CHECK-NEXT:   %0 = call { i64, i32 } @llvm.x86.rdtscp()
58
; CHECK-NEXT:   %1 = extractvalue { i64, i32 } %0, 0
59
; CHECK-NEXT:   %2 = load volatile i64, ptr @__polly_perf_cycles_total_start
60
; CHECK-NEXT:   %3 = sub i64 %1, %2
61
; CHECK-NEXT:   %4 = load volatile i64, ptr @__polly_perf_cycles_in_scops
62
; CHECK-NEXT:   %5 = call i32 (...) @printf(ptr @1, ptr addrspace(4) @0)
63
; CHECK-NEXT:   %6 = call i32 @fflush(ptr null)
64
; CHECK-NEXT:   %7 = call i32 (...) @printf(ptr @3, ptr addrspace(4) @2)
65
; CHECK-NEXT:   %8 = call i32 @fflush(ptr null)
66
; CHECK-NEXT:   %9 = call i32 (...) @printf(ptr @6, ptr addrspace(4) @4, i64 %3, ptr addrspace(4) @5)
67
; CHECK-NEXT:   %10 = call i32 @fflush(ptr null)
68
; CHECK-NEXT:   %11 = call i32 (...) @printf(ptr @9, ptr addrspace(4) @7, i64 %4, ptr addrspace(4) @8)
69
; CHECK-NEXT:   %12 = call i32 @fflush(ptr null)
70

71

72
; CHECK:      define weak_odr void @__polly_perf_init() {
73
; CHECK-NEXT: start:
74
; CHECK-NEXT:   %0 = load i1, ptr @__polly_perf_initialized
75
; CHECK-NEXT:   br i1 %0, label %earlyreturn, label %initbb
76

77
; CHECK:      earlyreturn:                                      ; preds = %start
78
; CHECK-NEXT:   ret void
79

80
; CHECK:      initbb:                                           ; preds = %start
81
; CHECK-NEXT:   store i1 true, ptr @__polly_perf_initialized
82
; CHECK-NEXT:   %1 = call i32 @atexit(ptr @__polly_perf_final)
83
; CHECK-NEXT:   %2 = call { i64, i32 } @llvm.x86.rdtscp()
84
; CHECK-NEXT:   %3 = extractvalue { i64, i32 } %2, 0
85
; CHECK-NEXT:   store volatile i64 %3, ptr @__polly_perf_cycles_total_start
86
; CHECK-NEXT:   ret void
87
; CHECK-NEXT: }
88