Re: [PATCH] KCOV: Introduced tracing unique covered PCs

[Dmitry Vyukov]

On Wed, Mar 17, 2021 at 10:10 PM Alexander Lochmann
<info@xxxxxxxxxxxxxxxxxxxxx> wrote:
> On 15.03.21 09:02, Dmitry Vyukov wrote:
> >>> Does this introduce an additional real of t->kcov_mode?
> >>> If yes, please reuse the value read in check_kcov_mode.
> >> Okay. How do I get that value from check_kcov_mode() to the caller?
> >> Shall I add an additional parameter to check_kcov_mode()?
> >
> > Yes, I would try to add an additional pointer parameter for mode. I
> > think after inlining the compiler should be able to regestrize it.
> First, I'll go for the extra argument. However, the compiler doesn't
> seem to inline check_kcov_mode(). Can I enforce inlining?
> I'm using GCC 9.3 on Debian Testing.

That's very strange and wrong. Maybe you use something like
CONFIG_CC_OPTIMIZE_FOR_SIZE=y?

With gcc-10 I am getting:

ffffffff81529ba0 <__sanitizer_cov_trace_pc>:
ffffffff81529ba0:       65 8b 05 59 53 af 7e    mov
%gs:0x7eaf5359(%rip),%eax        # 1ef00 <__preempt_count>
ffffffff81529ba7:       89 c1                   mov    %eax,%ecx
ffffffff81529ba9:       48 8b 34 24             mov    (%rsp),%rsi
ffffffff81529bad:       81 e1 00 01 00 00       and    $0x100,%ecx
ffffffff81529bb3:       65 48 8b 14 25 40 ef    mov    %gs:0x1ef40,%rdx
ffffffff81529bba:       01 00
ffffffff81529bbc:       a9 00 01 ff 00          test   $0xff0100,%eax
ffffffff81529bc1:       74 0e                   je
ffffffff81529bd1 <__sanitizer_cov_trace_pc+0x31>
ffffffff81529bc3:       85 c9                   test   %ecx,%ecx
ffffffff81529bc5:       74 35                   je
ffffffff81529bfc <__sanitizer_cov_trace_pc+0x5c>
ffffffff81529bc7:       8b 82 d4 14 00 00       mov    0x14d4(%rdx),%eax
ffffffff81529bcd:       85 c0                   test   %eax,%eax
ffffffff81529bcf:       74 2b                   je
ffffffff81529bfc <__sanitizer_cov_trace_pc+0x5c>
ffffffff81529bd1:       8b 82 b0 14 00 00       mov    0x14b0(%rdx),%eax
ffffffff81529bd7:       83 f8 02                cmp    $0x2,%eax
ffffffff81529bda:       75 20                   jne
ffffffff81529bfc <__sanitizer_cov_trace_pc+0x5c>
ffffffff81529bdc:       48 8b 8a b8 14 00 00    mov    0x14b8(%rdx),%rcx
ffffffff81529be3:       8b 92 b4 14 00 00       mov    0x14b4(%rdx),%edx
ffffffff81529be9:       48 8b 01                mov    (%rcx),%rax
ffffffff81529bec:       48 83 c0 01             add    $0x1,%rax
ffffffff81529bf0:       48 39 c2                cmp    %rax,%rdx
ffffffff81529bf3:       76 07                   jbe
ffffffff81529bfc <__sanitizer_cov_trace_pc+0x5c>
ffffffff81529bf5:       48 89 34 c1             mov    %rsi,(%rcx,%rax,8)
ffffffff81529bf9:       48 89 01                mov    %rax,(%rcx)
ffffffff81529bfc:       c3                      retq

Oh, wait gcc-9 indeed does not inline:

0000000000000070 <__sanitizer_cov_trace_pc>:
      70:       65 48 8b 0c 25 00 00    mov    %gs:0x0,%rcx
      77:       00 00
      79:       bf 02 00 00 00          mov    $0x2,%edi
      7e:       48 89 ce                mov    %rcx,%rsi
      81:       4c 8b 04 24             mov    (%rsp),%r8
      85:       e8 76 ff ff ff          callq  0 <check_kcov_mode>
      8a:       84 c0                   test   %al,%al
      8c:       74 20                   je     ae
<__sanitizer_cov_trace_pc+0x3e>
      8e:       48 8b 91 b8 14 00 00    mov    0x14b8(%rcx),%rdx
      95:       8b 89 b4 14 00 00       mov    0x14b4(%rcx),%ecx
      9b:       48 8b 02                mov    (%rdx),%rax
      9e:       48 83 c0 01             add    $0x1,%rax
      a2:       48 39 c1                cmp    %rax,%rcx
      a5:       76 07                   jbe    ae
<__sanitizer_cov_trace_pc+0x3e>
      a7:       4c 89 04 c2             mov    %r8,(%rdx,%rax,8)
      ab:       48 89 02                mov    %rax,(%rdx)
      ae:       c3                      retq

This looks like a bug in gcc-8/9. gcc-6 inlines again as well as
clang-11/12 inline.

Please add __always_inline for check_kcov_mode.