Missed optimization in math expression: tan(a) * cos(a) == sin(a)

[zamazan4ik]

Bugzilla Link	35602
Version	trunk
OS	All
Blocks	#34959
CC	@hfinkel,@rotateright

Extended Description

clang(trunk) with '--std=c++17 -O3 -march=native -ffast-math' flags for this code:

#include <cmath>

double test(double a)
{
    return tan(a) * cos(a);
}

generates this assembly:

test(double): # @test(double)
  sub rsp, 24
  vmovsd qword ptr [rsp + 8], xmm0 # 8-byte Spill
  call tan
  vmovsd qword ptr [rsp + 16], xmm0 # 8-byte Spill
  vmovsd xmm0, qword ptr [rsp + 8] # 8-byte Reload
  call cos
  vmulsd xmm0, xmm0, qword ptr [rsp + 16] # 8-byte Folded Reload
  add rsp, 24
  ret

gcc(trunk) with '--std=c++17 -O3 -march=native -ffast-math' flags generates this:

test(double):
        jmp     sin

[hfinkel]

Could be in InstCombine.

[zamazan4ik]

See a lot of optimizations here: https://github.com/gcc-mirror/gcc/blob/07b69d3f1cd3dd8ebb0af1fbff95914daee477d2/gcc/match.pd

[llvmbot]

See https://reviews.llvm.org/D41283

[llvmbot]

Hi,
Is this still open?
I have seen that it is assigned for Adil. But since then no activity here.
If it is open, I will take it.
Thanks

[rotateright]

Hi,
Is this still open?
I have seen that it is assigned for Adil. But since then no activity here.
If it is open, I will take it.
Thanks

Yes, this is still open. I'm not sure if Dmitry is able to respond, but the last comment in D41283 suggested that we would add an llvm.tan* intrinsic corresponding to sin/cos. I still think that's the right 1st step to move this forward.

[llvmbot]

Sorry for the delay. I'm quit busy at the moment doing my primary job, but still working on this. In a couple of weeks I'm returning to this patch.

[RKSimon]

mentioned in issue #34959

[elhewaty]

Hi @zamazan4ik , I am new to LLVM. Can I work on this issue? please mention any resources that can help.

[llvmbot]

@llvm/issue-subscribers-good-first-issue