https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103254
Aldy Hernandez <aldyh at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amacleod at redhat dot com
--- Comment #2 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
GORI is taking forever to query the outgoing range of _14 on the 3->4 edge.
The cunroll pass has expanded the block to this:
<bb 3> [local count: 59700049]:
# j_32 = PHI <j_20(8), 0(5)>
# i_lsm.10_31 = PHI <_3(8), i_lsm.10_6(5)>
_45 = (short int) j_32;
_44 = i_lsm.10_31 & _45;
_43 = (int) _44;
j_42 = _43 & j_32;
_19 = (short int) j_42;
_14 = _19 & _44;
_13 = (int) _14;
j_12 = _13 & j_42;
_29 = (short int) j_12;
_26 = _29 & _14;
_15 = (int) _26;
j_27 = _15 & j_12;
_56 = (short int) j_27;
_57 = _56 & _26;
_58 = (int) _57;
j_59 = _58 & j_27;
_65 = (short int) j_59;
_66 = _65 & _57;
_67 = (int) _66;
j_68 = _67 & j_59;
_74 = (short int) j_68;
_75 = _74 & _66;
_76 = (int) _75;
j_77 = _76 & j_68;
_83 = (short int) j_77;
_84 = _83 & _75;
_85 = (int) _84;
j_86 = _85 & j_77;
_92 = (short int) j_86;
_93 = _92 & _84;
_94 = (int) _93;
j_95 = _94 & j_86;
_101 = (short int) j_95;
_102 = _101 & _93;
_103 = (int) _102;
j_104 = _103 & j_95;
_110 = (short int) j_104;
_111 = _110 & _102;
_112 = (int) _111;
j_113 = _112 & j_104;
_119 = (short int) j_113;
_120 = _119 & _111;
_121 = (int) _120;
j_122 = _121 & j_113;
_128 = (short int) j_122;
_129 = _128 & _120;
_130 = (int) _129;
j_131 = _130 & j_122;
_137 = (short int) j_131;
_138 = _137 & _129;
_139 = (int) _138;
j_140 = _139 & j_131;
_146 = (short int) j_140;
_147 = _146 & _138;
_148 = (int) _147;
j_149 = _148 & j_140;
_155 = (short int) j_149;
_156 = _155 & _147;
_157 = (int) _156;
j_158 = _157 & j_149;
_164 = (short int) j_158;
_165 = _164 & _156;
_166 = (int) _165;
j_167 = _166 & j_158;
_1 = (short int) j_167;
_3 = _1 & _165;
_5 = (int) _3;
j_22 = _5 & j_167;
j_20 = j_22 + 4;
if (j_20 <= 4)
goto <bb 8>; [89.00%]
else
goto <bb 4>; [11.00%]
Andrew, didn't we talk about clamping gori if the series of dependencies was
greater than some number, because a long chain of events was unlikely to yield
meaningful ranges?
Interestingly vrp2 doesn't hang even with -fno-thread-jumps, but
-fno-thread-jumps -fdump-tree-vrp-details
does make gori hang, since it's now asking for all the outgoing ranges, _14 and
friends included.