[DebugInfo] Line 0 record not inserted before prologue moved via shrinkwrap optimisation

[Heath123]

In the following code:
https://godbolt.org/z/hh4aYGbr7
in this snippet:

        .loc    1 14 1
        ret
.LBB0_1:
        push    r15
        .cfi_def_cfa_offset 16
        push    r14
        .cfi_def_cfa_offset 24
        push    rbx
        .cfi_def_cfa_offset 32
        .cfi_offset rbx, -32
        .cfi_offset r14, -24
        .cfi_offset r15, -16
        xor     ebx, ebx
        lea     r14, [rip + .Lstr.2]
        lea     r15, [rip + .Lstr]
.Ltmp3:
        .loc    1 0 1 is_stmt 0

there is a location directive inserted into the assembly that points to the end of the function, used for the ret instruction. However, no line 0 directive is inserted before the generated prologue code, meaning the last directive of line 14 continues to apply. This means that the prologue is incorrectly assigned to line 14.

[llvmbot]

@llvm/issue-subscribers-debuginfo

Author: circuit10 (Heath123)

In the following code: https://godbolt.org/z/hh4aYGbr7 in this snippet: ``` .loc 1 14 1 ret .LBB0_1: push r15 .cfi_def_cfa_offset 16 push r14 .cfi_def_cfa_offset 24 push rbx .cfi_def_cfa_offset 32 .cfi_offset rbx, -32 .cfi_offset r14, -24 .cfi_offset r15, -16 xor ebx, ebx lea r14, [rip + .Lstr.2] lea r15, [rip + .Lstr] .Ltmp3: .loc 1 0 1 is_stmt 0 ``` there is a location directive inserted into the assembly that points to the end of the function, used for the `ret` instruction. However, no line 0 directive is inserted before the generated prologue code, meaning the last directive of line 14 continues to apply. This means that the prologue is incorrectly assigned to line 14.

[jmorse]

NB, I took a closer look at this: the push instructions have a frame-setup flag on them, thus they get ignored by DwarfDebug::beginInstruction. Which also means that the line-zero insertion code doesn't run when the block-boundary is encountered, so the earlier line 14 flows on into the setup block.

Annoyingly this isn't a trivial fix because there's five or six different concerns happening in beginInstruction, including things like basic-block-sections, which all need to be considered.

[OCHyams]

With PR #161845 applied we get:

 # %bb.4:                                # %for.end
        .loc    0 14 1                          # test.c:14:1
        retq
 .LBB0_1:                                # %for.body.preheader
+       .loc    0 0 1 is_stmt 0                 # test.c:0:1
        pushq   %r15
        .cfi_def_cfa_offset 16
        pushq   %r14
        .cfi_def_cfa_offset 24
        pushq   %rbx
        .cfi_def_cfa_offset 32
        .cfi_offset %rbx, -32
        .cfi_offset %r14, -24
        .cfi_offset %r15, -16
        xorl    %ebx, %ebx
        leaq    .Lstr.2(%rip), %r14
        leaq    .Lstr(%rip), %r15
-       .loc    0 0 1 is_stmt 0                 # :0:1
-.Ltmp1:
        .p2align        4
 .LBB0_2:                                # %for.body

(the line-zero inserted for the loop alignment padding is omitted as the new line-zero covers it)