[Bug translator/16615] don't require access to dwarf_query in has_single_line_record()

["jlebon at redhat dot com" <sourceware-bugzilla at sourceware dot org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=16615

--- Comment #3 from Jonathan Lebon <jlebon at redhat dot com> ---
I've just updated the jlebon/pr16615 branch containing many new goodies for
statement probes.

1. SPEED

We no longer rely on dwarf_getsrc_file(), which is slow, imprecise, and has
funny behaviours. Instead, we get our Dwarf_Lines straight from
dwarf_getsrclines(), which elfutils already caches, and then we cache our own
processed version of it. The results are much faster speeds:

$ # OLD STAP
$ time sh -c 'stap -l '\''process.statement("*@*:*")'\'' -c stap | wc -l'
13071

real    25m49.115s
user    25m47.171s
sys     0m0.268s

$ # NEW STAP
$ time sh -c 'stap -l '\''process.statement("*@*:*")'\'' -c stap | wc -l'
54221

real    1m49.447s
user    1m44.032s
sys     0m5.304s

More realistically, even simple kernel statement probes are faster to resolve:

$ # OLD STAP
$ time sh -c 'stap -l '\''kernel.statement("bio_init@bio.c:*")'\'' | wc -l'
1

real    0m8.338s
user    0m8.259s
sys     0m0.039s
$ 
$ # NEW STAP
$ time sh -c 'stap -l '\''kernel.statement("bio_init@bio.c:*")'\'' | wc -l'
4

real    0m0.395s
user    0m0.360s
sys     0m0.032s
$

2. INLINED FUNCTIONS

We previously just skipped linenos with multiple addresses. We're now smarter
about choosing them by tracking which ones are in which DIEs. This greatly
helps inlined functions:

$ nl -b a inlines.c
     1
     2  __attribute__((always_inline))
     3  static void foo(int i)
     4  {
     5     printf("%d\n", i);
     6  }
     7
     8  int main(int argc, char** argv)
     9  {
    10     // This is a comment
    11     foo(argc);
    12     foo(argc*2);
    13     return 0;
    14  }
$ 
$ gcc -g -w inlines.c -o inlines
$ 
$ # OLD STAP
$ stap -e 'probe process.statement("foo@inlines.c:5") { printf("in foo with
i=%d\n", $i) }' -c ./inlines

semantic error: multiple addresses for
/home/yyz/jlebon/codebase/systemtap/systemtap/inlines.c:5 [man error::dwarf]
(try /home/yyz/jlebon/codebase/systemtap/systemtap/inlines.c:4)
semantic error: while resolving probe point: identifier 'process' at
<input>:1:7
        source: probe process.statement("foo@inlines.c:5") { printf("in foo
with i=%d\n", $i) }
                      ^

semantic error: no match
Pass 2: analysis failed.  [man error::pass2]
$ 
$ # NEW STAP
$ stap -e 'probe process.statement("foo@inlines.c:5") { printf("in foo with
i=%d\n", $i) }' -c ./inlines
1
2
in foo with i=1
in foo with i=2
$

3. ACCURACY

Cross-checking against DIEs also allows us to confidently pick the right addr,
even when there are multiple Dwarf_Lines at the same lineno. This gives us
greater coverage in many optimized functions:

$ # OLD STAP
$ stap -l 'kernel.statement("bio_init@bio.c:*")'
kernel.statement("bio_init@fs/bio.c:277")
$ stap -l 'kernel.statement("bio_reset@bio.c:*")'
kernel.statement("bio_reset@fs/bio.c:297")
kernel.statement("bio_reset@fs/bio.c:298")

$ # NEW STAP
$ stap -l 'kernel.statement("bio_init@bio.c:*")'
kernel.statement("bio_init@fs/bio.c:273")
kernel.statement("bio_init@fs/bio.c:274")
kernel.statement("bio_init@fs/bio.c:275")
kernel.statement("bio_init@fs/bio.c:277")
$ stap -l 'kernel.statement("bio_reset@bio.c:*")'
kernel.statement("bio_reset@fs/bio.c:291")
kernel.statement("bio_reset@fs/bio.c:292")
kernel.statement("bio_reset@fs/bio.c:296")
kernel.statement("bio_reset@fs/bio.c:297")
kernel.statement("bio_reset@fs/bio.c:298")

Hopefully, the new code is now more consistent with users' expectations and
easier to understand.

-- 
You are receiving this mail because:
You are the assignee for the bug.