[Bug target/69274] [6 Regression] 435.gromacs performance regression after r231814 on x86 Haswell and bdver2

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69274

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Samples: 2M of event 'cycles', Event count (approx.): 1928893785632             
 36.40%  gromacs_base.am  gromacs_base.amd64-m64-gcc42-nn  [.] inl1130_        
 28.60%  gromacs_peak.am  gromacs_peak.amd64-m64-gcc42-nn  [.] inl1130_        
  7.51%  gromacs_base.am  gromacs_base.amd64-m64-gcc42-nn  [.] search_neighbour
  7.38%  gromacs_peak.am  gromacs_peak.amd64-m64-gcc42-nn  [.] search_neighbour
  2.00%  gromacs_base.am  gromacs_base.amd64-m64-gcc42-nn  [.] inl1100_        
  2.00%  gromacs_peak.am  gromacs_peak.amd64-m64-gcc42-nn  [.] inl1100_        

so that's innerf.f


Ok, so I spot one non-scheduling/RA difference:

-       vmovss  52(%rsp), %xmm6
-       vsubss  -4(%r13,%rdi,4), %xmm6, %xmm6
-.LVL250:
+       vsubss  -4(%r13,%rdi,4), %xmm6, %xmm4
+.LVL253:
        leaq    (%r15,%rsi,4), %r12
        .loc 1 662 0
-       vmulss  %xmm4, %xmm4, %xmm2
-       vmovss  %xmm4, 24(%rsp)
-       vmovss  %xmm5, 20(%rsp)
-       vmovss  %xmm6, 16(%rsp)
-       vfmadd231ss     %xmm5, %xmm5, %xmm2
-       vfmadd231ss     %xmm6, %xmm6, %xmm2
-.LVL251:
+       vmulss  %xmm2, %xmm2, %xmm1
+       vmovss  %xmm2, 24(%rsp)
+       vmovss  %xmm3, 20(%rsp)
+       vmovss  %xmm4, 16(%rsp)
+       vfmadd231ss     %xmm3, %xmm3, %xmm1
+       vmovaps %xmm1, %xmm7
+       vfmadd231ss     %xmm4, %xmm4, %xmm7
+.LVL254:

thus there seems to be some more spilling.