I am writing software that does some rather complex static analysis and dynamic tracing of other programs. This program uses a lot of static DWARF information to assist in the tracing, including line/column info from the .debug_line DWARF section. In order for this program to have the accuracy that we need, it must have fine-grained and accurate row and column number information to be populated in the DWARF debugging info. Using clang I can force row and column info to be populated using the -g -Xclang -dwarf-column-info options together.
However, there are some cases where clang does not produce fine-grained enough column information. One particular instance is for for loops. Take the following example program which I will refer to as source01.c:
1
2 int main ()
3 {
4 int number1 = 10, number2 = 20;
5 for (int i=0; i < 10; ++i) {
6 number1++;
7 number2++;
8 }
9 return 0;
10 }
I can compile it like so:
clang -g -Xclang -dwarf-column-info source01.c
Which produces the executable a.out. I then use dwarfdump to inspect how the row/column info has been populated:
dwarfdump a.out > dwarf_info
Taking a look at the .debug_line section, I see all of the row/col pairs that are contained in the debug info this executable:
.debug_line: line number info for a single cu
Source lines (from CU-DIE at .debug_info offset 0x0000000b):
<pc> [row,col] NS BB ET PE EB IS= DI= uri: "filepath"
NS new statement, BB new basic block, ET end of text sequence
PE prologue end, EB epilogue begin
IA=val ISA number, DI=val discriminator value
0x004004f0 [ 3, 0] NS uri: "/xxx/loop_01/source01.c"
0x004004fb [ 4, 5] NS PE
0x00400509 [ 5,10] NS
0x0040051d [ 6, 9] NS
0x00400528 [ 7, 9] NS
0x00400533 [ 5,27] NS
0x00400548 [ 9, 5] NS
0x0040054a [ 9, 5] NS ET
As you can see, there is the pair (5,10), which corresponds to int i=0;, and the pair (5,27), which corresponds to ++i. However, I would expect (and need) there to also be the pair (5,19), which would correspond to i < 10, but it isn't there. I have inspected the executable's instructions with objdump, and have confirmed that there are indeed instructions which correspond to the comparison i < 10 (Thus, it has not simply been "optimized away").
Do you have any intuition as to why clang would not populate this info? Or is there a way to force clang to produce more fine-grained column info? It seems like clang should have this capability, because the ASTs that clang generates have extremely fine-grained mappings between itself and the source code row and columns.
Thank you.
