[Bug tree-optimization/126239] [17 Regression] 5% slowdown of milc on Zen4 since r17-2213-g09021961ea0271

[tnfchris at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126239

--- Comment #4 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
imp_gauge_force seems to have changed but that's doing what I had expected

the pointer base calculation is now lifted outside the loop and the addressing
modes
inside the loop becomes simpler:

  @@
    lea    0x0(%r13,%r13,2),%rsi
    xor    %edi,%edi
    shl    $0x7,%rsi
  + add    $0x429ea0,%rsi
    test   %r8d,%r8d
    jg     <imp_gauge_force+0x23a>
    inc    %edi
    add    $0x18,%rsi

  @@
  - data16 cs nopw 0x0(%rax,%rax,1)
  - nopl   (%rax)
  + nopl   0x0(%rax)

  @@
  - vmovdqu 0x429ea0(%rsi),%xmm6
  + vmovdqu (%rsi),%xmm6

  @@
  - vmovdqu32 0x429ea0(%rsi,%rax,4),%xmm6{%k1}{z}
  + vmovdqu32 (%rsi,%rax,4),%xmm6{%k1}{z}

which has a knock on effect on the alignment of the loop (and of many other
functions in the file).
So I wonder if you're not seeing the effect of alignment changes, or is the
more complicated addressing modes faster somehow?

But at the GIMPLE level the transformation looks correct.