# RUN: llc %s -mattr=+ptx72 -start-after=nvptx-asm-printer -o - | FileCheck %s
#
# Test simple two-level inlining at MIR level - verifies that inlined_at information
# is correctly emitted when foo() calls bar() and kernel() calls foo().
#
#
# Original source:
# __device__ int gg;
#
# __device__ void foo() {
#   if (gg > 7)
#     bar();
# }
#
# __device__ void bar() {
#   ++gg;
# }
#
# __global__ void kernel() {
#   foo();
# }
#
# The inlining chain is:
#   kernel() at line 18 calls foo()
#   foo() at line 10 calls bar()
#   bar() increments gg at line 14
#
# Expected debug locations:
#   !6 = line 9, inlinedAt line 18  (foo inlined into kernel)
#   !11 = line 14, inlinedAt line 10, inlinedAt line 18 (bar inlined into foo, inlined into kernel)
#   !16 = line 10, inlinedAt line 18 (foo's call to bar, inlined into kernel)
#
# CHECK: .loc [[FILENUM:[1-9]]] 18
# CHECK: .loc [[FILENUM]] 9 {{[0-9]*}}, function_name [[FOONAME:\$L__info_string[0-9]+]], inlined_at [[FILENUM]] 18
# CHECK: .loc [[FILENUM]] 10 {{[0-9]*}}, function_name [[FOONAME]], inlined_at [[FILENUM]] 18
# CHECK: .loc [[FILENUM]] 14 {{[0-9]*}}, function_name [[BARNAME:\$L__info_string[0-9]+]], inlined_at [[FILENUM]] 10
# CHECK: .section .debug_str
# CHECK: {
# CHECK: [[FOONAME]]:
# CHECK-NEXT: // {{.*}} _Z3foov
# CHECK: [[BARNAME]]:
# CHECK-NEXT: // {{.*}} _Z3barv
# CHECK: }

--- |
  source_filename = "<unnamed>"
  target datalayout = "e-p6:32:32-i64:64-i128:128-i256:256-v16:16-v32:32-n16:32:64"
  target triple = "nvptx64-nvidia-cuda"

  @gg = internal addrspace(1) global i32 0, align 4
  @llvm.used = appending global [2 x ptr] [ptr @_Z6kernelv, ptr addrspacecast (ptr addrspace(1) @gg to ptr)], section "llvm.metadata"

  define ptx_kernel void @_Z6kernelv() !dbg !3 {
  entry:
    %tmp.i = load i32, ptr addrspace(1) @gg, align 4, !dbg !6
    %cmp.i = icmp sgt i32 %tmp.i, 7, !dbg !6
    %inc.i.i = add nuw nsw i32 %tmp.i, 1
    br i1 %cmp.i, label %if.then.i, label %_Z3foov.exit, !dbg !6

  if.then.i:                                        ; preds = %entry
    store i32 %inc.i.i, ptr addrspace(1) @gg, align 4, !dbg !11
    br label %_Z3foov.exit, !dbg !16

  _Z3foov.exit:                                     ; preds = %if.then.i, %entry
    ret void, !dbg !17
  }

  !llvm.dbg.cu = !{!0}
  !nvvm.annotations = !{}
  !llvm.module.flags = !{!2}

  ; Debug metadata with inlinedAt chains:
  ; foo() is inlined at kernel() line 18
  ; bar() is inlined at foo() line 10, which is inlined at kernel() line 18
  !0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !1, producer: "clang", isOptimized: true, runtimeVersion: 0, emissionKind: DebugDirectivesOnly)
  !1 = !DIFile(filename: "t2.cu", directory: "")
  !2 = !{i32 1, !"Debug Info Version", i32 3}
  !3 = distinct !DISubprogram(name: "kernel", linkageName: "_Z6kernelv", scope: !1, file: !1, line: 17, type: !4, scopeLine: 17, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0)
  !4 = !DISubroutineType(types: !5)
  !5 = !{}
  ; !6: foo's code at line 9, inlined at kernel line 18
  !6 = !DILocation(line: 9, column: 3, scope: !7, inlinedAt: !9)
  !7 = distinct !DILexicalBlock(scope: !8, file: !1, line: 8, column: 29)
  !8 = distinct !DISubprogram(name: "foo", linkageName: "_Z3foov", scope: !1, file: !1, line: 8, type: !4, scopeLine: 8, spFlags: DISPFlagLocalToUnit | DISPFlagDefinition | DISPFlagOptimized, unit: !0)
  !9 = distinct !DILocation(line: 18, column: 3, scope: !10)
  !10 = distinct !DILexicalBlock(scope: !3, file: !1, line: 17, column: 29)
  ; !11: bar's code at line 14, inlined at foo line 10, which is inlined at kernel line 18
  !11 = !DILocation(line: 14, column: 3, scope: !12, inlinedAt: !14)
  !12 = distinct !DILexicalBlock(scope: !13, file: !1, line: 13, column: 29)
  !13 = distinct !DISubprogram(name: "bar", linkageName: "_Z3barv", scope: !1, file: !1, line: 13, type: !4, scopeLine: 13, spFlags: DISPFlagLocalToUnit | DISPFlagDefinition | DISPFlagOptimized, unit: !0)
  ; !14: foo's call to bar at line 10, inlined at kernel line 18
  !14 = distinct !DILocation(line: 10, column: 5, scope: !15, inlinedAt: !9)
  !15 = distinct !DILexicalBlock(scope: !7, file: !1, line: 9, column: 3)
  !16 = !DILocation(line: 10, column: 5, scope: !15, inlinedAt: !9)
  !17 = !DILocation(line: 19, column: 1, scope: !10)
...
---
name:            _Z6kernelv
alignment:       1
tracksRegLiveness: true
noPhis:          true
isSSA:           false
noVRegs:         false
hasFakeUses:     false
registers:
  - { id: 0, class: b32 }
  - { id: 1, class: b32 }
  - { id: 2, class: b1 }
frameInfo:
  maxAlignment:    1
machineFunctionInfo: {}
body:             |
  bb.0.entry:
    successors: %bb.1(0x40000000), %bb.2(0x40000000)

    %1:b32 = LD_i32 0, 0, 1, 3, 32, -1, @gg, 0, debug-location !6 :: (dereferenceable load (s32) from @gg, addrspace 1); t2.cu:9:3 @[ t2.cu:18:3 ]
    %2:b1 = SETP_i32ri %1, 8, 2, debug-location !6; t2.cu:9:3 @[ t2.cu:18:3 ]
    CBranch %2, %bb.2, 0, debug-location !6; t2.cu:9:3 @[ t2.cu:18:3 ]

  bb.1.if.then.i:
  ; predecessors: %bb.0
    successors: %bb.2(0x80000000)

    %0:b32 = nuw nsw ADD32ri %1, 1
    ST_i32 %0, 0, 0, 1, 32, @gg, 0, debug-location !11 :: (store (s32) into @gg, addrspace 1); t2.cu:14:3 @[ t2.cu:10:5 @[ t2.cu:18:3 ] ]


  bb.2._Z3foov.exit:
  ; predecessors: %bb.0, %bb.1
    Return debug-location !17; t2.cu:19:1
...