[llvm-dev] Where's the optimiser gone? Or: who reviews the code you ship?

[Jordan Frank via llvm-dev]

In my experience, all humans strive for self-improvement. However,
such endeavors can be difficult without good feedback. Thus, I present
to you, my friend Stefan, a version of your contribution to our
mailing list in which the author does not come across as a
condescending <explicative deleted>, and whose contents holds the
possibility of being taken in good faith and acted upon.

---

Subject: Potential opportunity for generating more efficient code.

Hi folks,

Thank you for all the hard work you do to produce LLVM. I am finding
it helpful in my work.

I believe I may have identified instances of suboptimal code being
generated by the compiler. While digging in and understanding the
cause of this issue or proposing a fix is well outside both my
interests and skillset, I hope that in highlighting the issue I might
spur someone's interest and potentially lead to improvements to this
awesome project.

I have observed the following sequence of instructions in 4 places in
the builtins library:

mov eax, 0
cmp ecx, edx
jb @f
cmp edx, ecx
mov eax, 1
adc eax, 0
@@
ret

which I believe could be more compactly expressed as:

xor eax, eax
cmp edx, ecx
ja @f
adc eax, 1
@@
ret

I believe this code saves 10 bytes and 2 instructions. I could also be
making incorrect assumptions about why the former code is produced (my
assumption being incompetent developers), in which case I would love
the opportunity to learn and grow through an explanation. I could also
have tried to produce sample code or provide an explanation of how I
identified this, but, to be honest, my intention here is not to help
but to attempt to embarrass the developers on this project as I felt I
was treated unfairly when previously trolling this mailing list.

Best,
Stefan Kanthak
73rd-most prolific Windows bug report in 2016.

---

Best wishes,
Jordan Frank,
Whom does not enjoy being subjected to the repeated trolling of this list by OP.


On 9/6/20, 10:07 AM, "llvm-dev on behalf of Stefan Kanthak via
llvm-dev" <llvm-dev-bounces at lists.llvm.org on behalf of
llvm-dev at lists.llvm.org> wrote:

    Hi @ll,

    who can tell me (or the world) why an optimising compiler emits the
    code sequence shown in the left column (found 4x in the "builtins"
    library)?
    In the middle column the expected code sequence.

        mov    eax, 0         #     xor   eax, eax         # -3 bytes
        cmp    ecx, edx       #     cmp   edx, ecx
        jb     @f             #     ja    @f
        cmp    edx, ecx       #                            # -2 bytes
        mov    eax, 1         #                            # -5 bytes
        adc    eax, 0         #     adc   eax, 1           # -2 insns
    @@:                       # @@:
        ret                   #     ret

    Stefan
    --