I can reproduce the `i = 9` behavior using the Fedora 29 system gcc.

The DWARF for "i":

 <2><6b>: Abbrev Number: 6 (DW_TAG_variable)
    <6c>   DW_AT_name        : i
    <6e>   DW_AT_decl_file   : 1
    <6f>   DW_AT_decl_line   : 3
    <70>   DW_AT_decl_column : 7
    <71>   DW_AT_type        : <0xc2>
    <75>   DW_AT_location    : 0x16 (location list)
    <79>   DW_AT_GNU_locviews: 0x0


So, let's look at location list 0x16:

    00000016 v000000000000003 v000000000000000 views at 00000000 for:
             0000000000400420 0000000000400431 (DW_OP_lit0; DW_OP_stack_value)
    0000002a v000000000000000 v000000000000000 views at 00000002 for:
             0000000000400431 000000000040043c (DW_OP_lit1; DW_OP_stack_value)
    0000003e v000000000000000 v000000000000000 views at 00000004 for:
             000000000040043c 0000000000400447 (DW_OP_lit2; DW_OP_stack_value)
    00000052 v000000000000000 v000000000000000 views at 00000006 for:
             0000000000400447 0000000000400452 (DW_OP_lit3; DW_OP_stack_value)
    00000066 v000000000000000 v000000000000000 views at 00000008 for:
             0000000000400452 000000000040045d (DW_OP_lit4; DW_OP_stack_value)
    0000007a v000000000000000 v000000000000000 views at 0000000a for:
             000000000040045d 0000000000400468 (DW_OP_lit5; DW_OP_stack_value)
    0000008e v000000000000000 v000000000000000 views at 0000000c for:
             0000000000400468 0000000000400473 (DW_OP_lit6; DW_OP_stack_value)
    000000a2 v000000000000000 v000000000000000 views at 0000000e for:
             0000000000400473 000000000040047e (DW_OP_lit7; DW_OP_stack_value)
    000000b6 v000000000000000 v000000000000000 views at 00000010 for:
             000000000040047e 0000000000400489 (DW_OP_lit8; DW_OP_stack_value)
    000000ca v000000000000000 v000000000000000 views at 00000012 for:
             0000000000400489 0000000000400495 (DW_OP_lit9; DW_OP_stack_value)
    000000de v000000000000000 v000000000000000 views at 00000014 for:
             0000000000400495 000000000040049c (DW_OP_lit2; DW_OP_stack_value)
    000000f2 <End of list>


Now in gdb, let's find the PC in question:

(gdb) b 11
Breakpoint 1 at 0x400489: file abc.c, line 11.
(gdb) r
Starting program: /tmp/qq 

Breakpoint 1, main () at abc.c:11
11	      optimize_me_not();
(gdb) p $pc
$1 = (void (*)()) 0x400489 <main+105>

Looking this up in the location list:

             0000000000400489 0000000000400495 (DW_OP_lit9; DW_OP_stack_value)

In other words, gcc tells gdb that the value is 9.


Now, what is interesting is that lldb sets a different breakpoint location:

(lldb) b 11
Breakpoint 1: where = qq`main + 4 at abc.c:11, address = 0x0000000000400424

Reproducing this in gdb:

(gdb) b *0x400424
Breakpoint 1 at 0x400424: file abc.c, line 6.
(gdb) r
Starting program: /tmp/qq 
Missing separate debuginfos, use: dnf debuginfo-install glibc-2.28-27.fc29.x86_64

Breakpoint 1, 0x0000000000400424 in main () at abc.c:6
6	    a = i;
(gdb) p i
$1 = 0


So maybe gdb is misinterpreting the line table somehow; but at the same
time there is a gcc bug here IMO.

> So maybe gdb is misinterpreting the line table somehow; but at the same
time there is a gcc bug here IMO.

I think what is happening is that gcc -O3 unrolls the second loop. In which case the value of i doesn't really matter, since it isn't actually used anymore in the code.

Also if you look at the linetable with eu-readelf --debug-dump=decodedline you'll see that line 11 expands to multiple addresses. Two of which have a statement marker, so are equally ok to set a breakpoint on.

I haven't looked at the location views, but maybe those can help determine which view the user expects of the variable?

There are actually 4 addresses that expand to line 11:

 CU [b] abc.c
  line:col SBPE* disc isa op address (Statement Block Prologue Epilogue *End)
  /tmp/abc.c (mtime: 0, length: 0)
     2:12  S        0   0  0 +0x0000000000001040 <main>
     3:3   S        0   0  0 +0x0000000000001040 <main>
     4:3   S        0   0  0 +0x0000000000001040 <main>
     5:3   S        0   0  0 +0x0000000000001040 <main>
     6:5   S        0   0  0 +0x0000000000001040 <main>
     2:12           0   0  0 +0x0000000000001040 <main>
->  11:7            0   0  0 +0x0000000000001044 <main+0x4>
     6:7            0   0  0 +0x0000000000001046 <main+0x6>
     6:5   S        0   0  0 +0x0000000000001051 <main+0x11>
     6:7            0   0  0 +0x0000000000001051 <main+0x11>
     6:5   S        0   0  0 +0x000000000000105c <main+0x1c>
     6:7            0   0  0 +0x000000000000105c <main+0x1c>
     6:5   S        0   0  0 +0x0000000000001067 <main+0x27>
     6:7            0   0  0 +0x0000000000001067 <main+0x27>
     6:5   S        0   0  0 +0x0000000000001072 <main+0x32>
     6:7            0   0  0 +0x0000000000001072 <main+0x32>
     6:5   S        0   0  0 +0x000000000000107d <main+0x3d>
     6:7            0   0  0 +0x000000000000107d <main+0x3d>
     6:5   S        0   0  0 +0x0000000000001088 <main+0x48>
     6:7            0   0  0 +0x0000000000001088 <main+0x48>
     6:5   S        0   0  0 +0x0000000000001093 <main+0x53>
     6:7            0   0  0 +0x0000000000001093 <main+0x53>
     6:5   S        0   0  0 +0x000000000000109e <main+0x5e>
     6:7            0   0  0 +0x000000000000109e <main+0x5e>
->  11:7   S        0   0  0 +0x00000000000010a9 <main+0x69>
->  11:7   S        0   0  0 +0x00000000000010ae <main+0x6e>
->  11:7            0   0  0 +0x00000000000010b5 <main+0x75>
    13:1            0   0  0 +0x00000000000010b5 <main+0x75>
    13:1       *    0   0  0 +0x00000000000010bb <main+0x7b>

So the two places where i = 0 were "merged" by the compiler.