[X86] bt regression with bmi

[sftlbcn]

looks worse with bmi?
godbolt: https://godbolt.org/z/aj3so3noj

int test (int x)
{
  x &= 0xf;
  return (0x6996 >> x) & 1;
}

trunk:

test(int): # @test(int)
  and dil, 15
  movzx ecx, dil
  mov edx, 27030
  xor eax, eax
  bt edx, ecx
  setb al
  ret

clang 14:

test(int): # @test(int)
  mov ecx, edi
  and cl, 15
  mov eax, 27030
  shr eax, cl
  and eax, 1
  ret

clang 14 with bmi2:

test(int): # @test(int)
  and dil, 15
  mov eax, 27030
  shrx eax, eax, edi
  and eax, 1
  ret

[phoebewang]

Didn't verify, but I think it's changed by https://reviews.llvm.org/D122891
It is an optimization, since shr with reg is slow.

[phoebewang]

My mistake, bmi is better https://godbolt.org/z/rWj8bn8j6

[llvmbot]

@llvm/issue-subscribers-backend-x86