https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78288
Bug ID: 78288
Summary: Compile time hog (with var-tracking) in
libsanitizer/asan/asan_interceptors.cc
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: debug
Assignee: unassigned at gcc dot gnu.org
Reporter: marxin at gcc dot gnu.org
Target Milestone: ---
I've noticed that compilation of sanitizer interceptors take quite significant
time.
With -fno-checking it takes 0m18.664s, while adding -gtoggle time shrinks to
0m12.680s.
As var-tracking eats 1/3 of compile time, I'm wondering whether there's a space
for improvement?
perf report:
5.72% cc1plus cc1plus [.] canonicalize_values_star
3.64% cc1plus cc1plus [.] set_slot_part
1.50% cc1plus cc1plus [.] bitmap_set_bit
1.39% cc1plus libc-2.24.so [.] _int_malloc
1.35% cc1plus cc1plus [.] ggc_set_mark
1.23% cc1plus cc1plus [.]
hash_table<variable_hasher, xcallocator>::find_slot_with_hash
1.21% cc1plus cc1plus [.] loc_cmp
perf annotate of the 2 hottest functions:
│ if (GET_CODE (node->loc) == VALUE)
0.13 │ 88: mov 0x8(%rax),%rdx
0.99 │ cmpw $0x1,(%rdx)
1.86 │ ↑ jne 80
│ {
│ has_value = true;
│ if (VALUE_RECURSED_INTO (node->loc))
│ has_marks = true;
5.04 │ testb $0x20,0x3(%rdx)
0.96 │ mov $0x1,%edi
│ cmovne %edi,%ecx
│ canon_value_cmp():
│ || CSELIB_VAL_PTR (tval)->uid < CSELIB_VAL_PTR (cval)->uid;
5.74 │ test %rbx,%rbx
│ ↓ je 455
│ mov 0x8(%rbx),%r8
17.32 │ mov 0x8(%rdx),%rdi
│ _Z24canonicalize_values_starPP8variableP12dataflow_set():
│ for (node = var->var_part[0].loc_chain; node; node = node->next)
0.04 │ mov (%rax),%rax
│ canon_value_cmp():
│ || CSELIB_VAL_PTR (tval)->uid < CSELIB_VAL_PTR (cval)->uid;
0.38 │ mov 0x4(%rdi),%edi
3.81 │ cmp %edi,0x4(%r8)
31.32 │ mov $0x1,%edi
0.02 │ cmovg %rdx,%rbx
│ _Z24canonicalize_values_starPP8variableP12dataflow_set():
│ for (node = var->var_part[0].loc_chain; node; node = node->next)
10.26 │ test %rax,%rax
│ ↑ jne 88
│ if (canon_value_cmp (node->loc, cval))
│ cval = node->loc;
│ }
│
│ if (!has_value)
│ c7: test %dil,%dil
│ ↓ je 310
│ return 1;
│
│ if (cval == val)
0.02 │ cmp %rbx,%r12
│ ↓ je 4ad
│ else if (REG_P (node->loc) || MEM_P (node->loc))
0.14 │650: cmp $0x2f,%cx
│ ↓ je 660
│ cmp $0x29,%cx
│ ↑ jne 2c0
│ nextp = &node->next)
14.58 │660: mov %rbx,%rbp
│ for (nextp = &var->var_part[0].loc_chain; (node =
*nextp);
0.04 │ mov (%rbx),%rbx
│ c++;
2.01 │ mov %eax,%r14d
│ add $0x1,%eax
│ for (nextp = &var->var_part[0].loc_chain; (node =
*nextp);
7.20 │ test %rbx,%rbx
│ ↑ je 2c0
│ if (GET_CODE (node->loc) == VALUE)
0.14 │675: mov 0x8(%rbx),%rsi
4.02 │ lea -0x1(%rax),%r14d
0.04 │ movzwl (%rsi),%ecx
12.35 │ cmp $0x1,%cx
│ ↑ jne 650
│ if (node->loc == loc)
1.27 │ cmp %r13,%rsi
│ ↑ je 503
│ canon_value_cmp():
│ || CSELIB_VAL_PTR (tval)->uid < CSELIB_VAL_PTR (cval)->uid;
1.52 │ mov 0x8(%r13),%rdi
0.07 │ mov 0x8(%rsi),%rcx
9.25 │ mov 0x4(%rcx),%edx
42.46 │ cmp %edx,0x4(%rdi)
Giving me an advise, I can help with that.
Thanks
