[PATCH] D39331: Switch -mcount and -finstrument-functions to emit EnterExitInstrumenter attributes

[Reid Kleckner via Phabricator via cfe-commits]

rnk added inline comments.


================
Comment at: include/clang/Driver/Options.td:1032
+def fno_instrument_functions_inline : Flag<["-"], "fno-instrument-functions-inline">, Group<f_Group>, Flags<[CC1Option]>,
+  HelpText<"When using -finstrument-functions, insert the calls after inlining">;
 
----------------
hans wrote:
> rnk wrote:
> > This actually describes GCC's behavior for `-finstrument-functions`. I suspect the current behavior is pretty useless for most people. I'd rather make the new behavior the default and add a flag to request the old behavior.
> At least with the GCC I tried, they do insert the calls before inlining, and inline them:
> 
> ```
> $ echo 'static void f() {} void g() { f(); }' | gcc -x c -O3 -finstrument-functions -S -o - - | grep -A50 g:
> g:
> .LFB1:
>         .cfi_startproc
>         subq    $8, %rsp
>         .cfi_def_cfa_offset 16
>         movl    $g, %edi
>         movq    8(%rsp), %rsi
>         call    __cyg_profile_func_enter
>         movq    8(%rsp), %rsi
>         movl    $f, %edi
>         call    __cyg_profile_func_enter
>         movq    8(%rsp), %rsi
>         movl    $f, %edi
>         call    __cyg_profile_func_exit
>         movq    8(%rsp), %rsi
>         movl    $g, %edi
>         addq    $8, %rsp
>         .cfi_def_cfa_offset 8
>         jmp     __cyg_profile_func_exit
>         .cfi_endproc
> ```
You're right, I misread what I did locally.

I feel like there are too many negatives here, though. -finstrument-functions-after-inlining would probably be more readable, especially considering we don't have a finstrument-inline-functions. -finstrument-functions-after-inlining could imply -finstrument-functions, and users would have one less flag to set.


https://reviews.llvm.org/D39331