[Bug tree-optimization/96187] GCC at -O2 generates branch for code that should be branch-free

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96187

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
               Host|x86_64                      |
             Status|UNCONFIRMED                 |NEW
             Target|x86_64                      |x86_64-*-*
           Keywords|                            |missed-optimization
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2020-07-14

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  The issue seems to be PRE removing the RTL if-conversion
opportunity by eliding the final bitwise OR on the path with a literal zero:

   _2 = MEM[(const uint8_t *)b_10(D) + 1B];
   if (_2 != 0)
-    goto <bb 6>; [50.00%]
-  else
     goto <bb 5>; [50.00%]
+  else
+    goto <bb 6>; [50.00%]

-  <bb 5> [local count: 536870913]:
+  <bb 5> [local count: 536870912]:
+  _19 = iftmp.3_7 | 4;

   <bb 6> [local count: 1073741824]:
-  # iftmp.4_8 = PHI <4(4), 0(5)>
-  _3 = iftmp.3_7 | iftmp.4_8;
-  _4 = _3 * 4;
+  # iftmp.4_8 = PHI <4(5), 0(4)>
+  # prephitmp_20 = PHI <_19(5), iftmp.3_7(4)>
+  _4 = prephitmp_20 * 4;
   _5 = 4837447714461450240 >> _4;
   _6 = (int) _5;
   len_11 = _6 & 15;


in the get_length4 case this pattern doesn't occur since the branch has
been removed by early folding and/or gimplification already.