Re: [RFC PATCH v3 0/3] Introduce BPF map tracing capability

[Yonghong Song]

On 11/3/21 10:49 AM, Alexei Starovoitov wrote:
> On Wed, Nov 3, 2021 at 10:45 AM Joe Burton <jevburton.kernel@xxxxxxxxx> wrote:
> > Sort of - I hit issues when defining the function in the same
> > compilation unit as the call site. For example:
> >
> >    static noinline int bpf_array_map_trace_update(struct bpf_map *map,
> >                  void *key, void *value, u64 map_flags)
>
> Not quite :)
> You've had this issue because of 'static noinline'.
> Just 'noinline' would not have such issues even in the same file.

This seems not true. With latest trunk clang,

[$ ~/tmp2] cat t.c
int __attribute__((noinline)) foo() { return 1; }
int bar() { return foo() + foo(); }
[$ ~/tmp2] clang -O2 -c t.c
[$ ~/tmp2] llvm-objdump -d t.o

t.o:    file format elf64-x86-64

Disassembly of section .text:

0000000000000000 <foo>:
       0: b8 01 00 00 00                movl    $1, %eax
       5: c3                            retq
       6: 66 2e 0f 1f 84 00 00 00 00 00 nopw    %cs:(%rax,%rax)

0000000000000010 <bar>:
      10: b8 02 00 00 00                movl    $2, %eax
      15: c3                            retq
[$ ~/tmp2]

The compiler did the optimization and the original noinline function 
still in the binary.

With a single foo() in bar() has the same effect.

asm("") indeed helped preserve the call.

[$ ~/tmp2] cat t.c
int __attribute__((noinline)) foo() { asm(""); return 1; }
int bar() { return foo() + foo(); }
[$ ~/tmp2] clang -O2 -c t.c
[$ ~/tmp2] llvm-objdump -d t.o

t.o:    file format elf64-x86-64

Disassembly of section .text:

0000000000000000 <foo>:
       0: b8 01 00 00 00                movl    $1, %eax
       5: c3                            retq
       6: 66 2e 0f 1f 84 00 00 00 00 00 nopw    %cs:(%rax,%rax)

0000000000000010 <bar>:
      10: 50                            pushq   %rax
      11: e8 00 00 00 00                callq   0x16 <bar+0x6>
      16: e8 00 00 00 00                callq   0x1b <bar+0xb>
      1b: b8 02 00 00 00                movl    $2, %eax
      20: 59                            popq    %rcx
      21: c3                            retq
[$ ~/tmp2]

Note with asm(""), foo() is called twice, but the compiler optimization
knows foo()'s return value is 1 so it did calculation at compiler time,
assign the 2 to %eax and returns.

Having a single foo() in bar() has the same effect.

[$ ~/tmp2] cat t.c
int __attribute__((noinline)) foo() { return 1; }
int bar() { return foo(); }
[$ ~/tmp2] clang -O2 -c t.c
[$ ~/tmp2] llvm-objdump -d t.o

t.o:    file format elf64-x86-64

Disassembly of section .text:

0000000000000000 <foo>:
       0: b8 01 00 00 00                movl    $1, %eax
       5: c3                            retq
       6: 66 2e 0f 1f 84 00 00 00 00 00 nopw    %cs:(%rax,%rax)

0000000000000010 <bar>:
      10: b8 01 00 00 00                movl    $1, %eax
      15: c3                            retq
[$ ~/tmp2]

I checked with a few llvm compiler engineers in Facebook.
They mentioned there is nothing preventing compiler from doing
optimization like poking inside the noinline function and doing
some optimization based on that knowledge.

> Reminder: please don't top post and trim your replies.