llvm dev - Oct 2020 - Optimizing assembly generated for tail call

Hello,

I recently found that LLVM generates sub-optimal assembly for a tail call
optimization case. Below is an example (https://godbolt.org/z/ao15xE):
> void g1();
> void g2();
> void f(bool v) {
>    if (v) {
>     g1();
>    } else {
>      g2();
>    }
> }
>
The assembly generated is as follow:
> f(bool): # @f(bool)
> testb %dil, %dil
> je .LBB0_2
> jmp g1() # TAILCALL
> .LBB0_2:
> jmp g2() # TAILCALL
>
However, in this specific case (where no function epilogue is needed), one
can actually change 'je .LBB0_2' to 'je g2()' directly, thus
saving a jump.
Is there any way I could instruct LLVM to do this? For my use case, it is
acceptable to do this at any level (C++ level /IR level /MachineInst level
is all fine).

Thanks!

Best,
Haoran
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20201006/400bcc5a/attachment.html>

On Tue, Oct 06, 2020 at 02:51:46AM -0700, Haoran Xu via llvm-dev
wrote:
> Hello,
> 
> I recently found that LLVM generates sub-optimal assembly for a tail call
> optimization case. Below is an example (https://godbolt.org/z/ao15xE):
> 
> > void g1();
> > void g2();
> > void f(bool v) {
> >    if (v) {
> >     g1();
> >    } else {
> >      g2();
> >    }
> > }
> >
> The assembly generated is as follow:
> 
> > f(bool): # @f(bool)
> > testb %dil, %dil
> > je .LBB0_2
> > jmp g1() # TAILCALL
> > .LBB0_2:
> > jmp g2() # TAILCALL
> >
> However, in this specific case (where no function epilogue is needed), one
> can actually change 'je .LBB0_2' to 'je g2()' directly,
thus saving a jump.
> Is there any way I could instruct LLVM to do this? For my use case, it is
> acceptable to do this at any level (C++ level /IR level /MachineInst level
> is all fine).
That translation doesn't normally work as conditional jumps have a much
more restricted range.

Joerg

I checked Intel instruction guide and it seems like conditional jumps can
take rip-relative 32-bit operands, same as unconditional jumps. Did I miss
something?


Joerg Sonnenberger via llvm-dev <llvm-dev at lists.llvm.org> 于2020年10月6日周二
下午4:14写道：
> On Tue, Oct 06, 2020 at 02:51:46AM -0700, Haoran Xu via llvm-dev wrote:
> > Hello,
> >
> > I recently found that LLVM generates sub-optimal assembly for a tail
call
> > optimization case. Below is an example (https://godbolt.org/z/ao15xE):
> >
> > > void g1();
> > > void g2();
> > > void f(bool v) {
> > >    if (v) {
> > >     g1();
> > >    } else {
> > >      g2();
> > >    }
> > > }
> > >
> > The assembly generated is as follow:
> >
> > > f(bool): # @f(bool)
> > > testb %dil, %dil
> > > je .LBB0_2
> > > jmp g1() # TAILCALL
> > > .LBB0_2:
> > > jmp g2() # TAILCALL
> > >
> > However, in this specific case (where no function epilogue is needed),
> one
> > can actually change 'je .LBB0_2' to 'je g2()'
directly, thus saving a
> jump.
> > Is there any way I could instruct LLVM to do this? For my use case, it
is
> > acceptable to do this at any level (C++ level /IR level /MachineInst
> level
> > is all fine).
>
> That translation doesn't normally work as conditional jumps have a much
> more restricted range.
>
> Joerg
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20201006/ecc189c1/attachment-0001.html>