[Bug target/124042] [16 regression] Miscompare of 482.sphinx3 on aarch64

[pheeck at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124042

--- Comment #4 from Filip Kastl <pheeck at gcc dot gnu.org> ---
-mearly-ra=none helps.  The miscompare doesn't reproduce with it.


The relevant lines from an4.log (-Ofast -march=native PGO -mearly-ra=none, test
workload):

   750   126   162     -537204  -148846    0 T    
   947   163   185     -540794   -74100   -1 <sil>    
  1281   186   200     -229685  -148846    0 N    
  1935   201   284    -1552966   -74100   -1 <sil>    
  2336   285   341    -1028411   -74100   -1 <sil>    
  2700   342   365     -307116  -148846    0 U    
  3054   366   397     -458916   -74100   -1 <sil>    
  3056   398   398           0  -148846    0 </s>

INFO: utt.c(337):  398 frm;  1703 sen,  3704 gau/fr, Sen 0.00 CPU 0.00 Clk
[Ovrhd 0.00 CPU 0.00 Clk];    841 hmm,   8 wd/fr, 0.00 CPU 0.00 Clk
(an406-fcaw-b)    

INFO: lm.c(826):    29545 bg(),    29533 bo;     1 fills,        1 in mem
(50.0%)