https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77689
--- Comment #17 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
I posted the patch. With it we split the loop, but we don't get really big
improvements from that
h@ryzen3:~/gcc/build3/gcc> ./xgcc -B ./ -Ofast c.ii -S -fopt-info 2>&1 | grep
split ; perf stat ./a.out
c.C:15:9: optimized: loop split
Performance counter stats for './a.out':
862.32 msec task-clock:u # 0.978 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
53,443 page-faults:u # 61.976 K/sec
3,295,805,448 cycles:u # 3.822 GHz
(83.44%)
81,606,129 stalled-cycles-frontend:u # 2.48% frontend
cycles idle (83.33%)
8,205,437 stalled-cycles-backend:u # 0.25% backend
cycles idle (83.32%)
7,420,801,599 instructions:u # 2.25 insn per
cycle
# 0.01 stalled cycles per
insn (82.88%)
903,367,479 branches:u # 1.048 G/sec
(83.50%)
54,872 branch-misses:u # 0.01% of all
branches (83.53%)
0.881716607 seconds time elapsed
0.782798000 seconds user
0.079877000 seconds sys
jh@ryzen3:~/gcc/build3/gcc> ~/trunk-install/bin/g++ -Ofast c.ii -S -fopt-info
2>&1 | grep split ; perf stat ./a.out
Performance counter stats for './a.out':
905.76 msec task-clock:u # 0.998 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
51,910 page-faults:u # 57.311 K/sec
3,459,244,533 cycles:u # 3.819 GHz
(83.24%)
83,603,137 stalled-cycles-frontend:u # 2.42% frontend
cycles idle (83.24%)
13,908,621 stalled-cycles-backend:u # 0.40% backend
cycles idle (83.25%)
7,422,922,864 instructions:u # 2.15 insn per
cycle
# 0.01 stalled cycles per
insn (83.30%)
899,226,266 branches:u # 992.791 M/sec
(83.67%)
52,719 branch-misses:u # 0.01% of all
branches (83.31%)
0.907459830 seconds time elapsed
0.810481000 seconds user
0.095820000 seconds sys
optimized dump is:
<bb 6> [local count: 679982665]:
# ivtmp.62_101 = PHI <1(5), ivtmp.62_16(6)>
_122 = MEM[(value_type &)_44 + ivtmp.62_101 * 4];
_120 = MEM[(value_type &)_41 + ivtmp.62_101 * 4];
_119 = _120 + _122;
_114 = MEM[(value_type &)_41 + 18446744073709551612 + ivtmp.62_101 * 4];
_113 = _114 * _119;
_112 = (double) _113;
_111 = (signed int) ivtmp.62_101;
_110 = (double) _111;
_109 = __builtin_log (_110);
_108 = _109 * _112;
_107 = (float) _108;
MEM[(value_type &)_35 + ivtmp.62_101 * 4] = _107;
ivtmp.62_16 = ivtmp.62_101 + 1;
if (ivtmp.62_16 != 100000000)
goto <bb 6>; [98.42%]
else
goto <bb 7>; [1.58%]
which looks reasonable. Vectorizer says
c.ii.174t.vect:c.C:18:26: missed: versioning for alias required: can't
determine dependence between *_123 and *_106
c.ii.174t.vect:c.C:18:26: missed: versioning for alias required: can't
determine dependence between *_121 and *_106
c.ii.174t.vect:c.C:18:34: missed: versioning for alias required: can't
determine dependence between *_115 and *_106
c.ii.174t.vect:c.C:13:27: missed: Unknown misalignment, naturally aligned
c.ii.174t.vect:c.C:13:27: missed: Unknown misalignment, naturally aligned
c.ii.174t.vect:c.C:13:27: missed: Unknown misalignment, naturally aligned
c.ii.174t.vect:c.C:13:27: missed: Unknown misalignment, naturally aligned
c.ii.174t.vect:c.C:13:27: missed: conversion not supported by target.
c.ii.174t.vect:c.C:13:27: missed: no optab.
c.ii.174t.vect:c.C:13:27: missed: function is not vectorizable.
c.ii.174t.vect:/home/jh/trunk-install/include/c++/14.0.0/cmath:353:27: missed:
not vectorized: relevant stmt not supported: _109 = __builtin_log (_110);
c.ii.174t.vect:c.C:13:27: missed: bad operation or unsupported loop bound.
c.ii.174t.vect:c.C:18:26: missed: versioning for alias required: can't
determine dependence between *_123 and *_106
c.ii.174t.vect:c.C:18:26: missed: versioning for alias required: can't
determine dependence between *_121 and *_106
c.ii.174t.vect:c.C:18:34: missed: versioning for alias required: can't
determine dependence between *_115 and *_106
c.ii.174t.vect:c.C:13:27: missed: Unknown misalignment, naturally aligned
c.ii.174t.vect:c.C:13:27: missed: Unknown misalignment, naturally aligned
c.ii.174t.vect:c.C:13:27: missed: Unknown misalignment, naturally aligned
c.ii.174t.vect:c.C:13:27: missed: Unknown misalignment, naturally aligned
c.ii.174t.vect:c.C:13:27: missed: conversion not supported by target.
c.ii.174t.vect:c.C:13:27: missed: no optab.
c.ii.174t.vect:c.C:13:27: missed: function is not vectorizable.
c.ii.174t.vect:/home/jh/trunk-install/include/c++/14.0.0/cmath:353:27: missed:
not vectorized: relevant stmt not supported: _109 = __builtin_log (_110);
c.ii.174t.vect:c.C:13:27: missed: bad operation or unsupported loop bound.
c.ii.174t.vect:c.C:18:26: missed: versioning for alias required: can't
determine dependence between *_123 and *_106
c.ii.174t.vect:c.C:18:26: missed: versioning for alias required: can't
determine dependence between *_121 and *_106
c.ii.174t.vect:c.C:18:34: missed: versioning for alias required: can't
determine dependence between *_115 and *_106
c.ii.174t.vect:c.C:18:34: missed: not vectorized: unsupported data-type
double
c.ii.174t.vect:c.C:13:27: missed: can't determine vectorization factor.
c.ii.174t.vect:c.C:13:27: missed: not vectorized: no vectype for stmt: _122 =
*_123;
c.ii.174t.vect:c.C:18:26: missed: not vectorized: no vectype for stmt: _122 =
*_123;
c.ii.174t.vect:c.C:13:27: missed: bad data references.
c.ii.174t.vect:c.C:13:27: missed: couldn't vectorize loop
At loop vectorization time the counter goes backwards:
<bb 36> [local count: 679982665]:
# i_127 = PHI <1(7), i_17(40)>
# ivtmp_31 = PHI <99999999(7), ivtmp_28(40)>
_125 = (long unsigned int) i_127;
_124 = _125 * 4;
_123 = _44 + _124;
_122 = *_123;
_121 = _41 + _124;
_120 = *_121;
_119 = _122 + _120;
_118 = i_127 + 4294967295;
_117 = (long unsigned int) _118;
_116 = _117 * 4;
_115 = _41 + _116;
_114 = *_115;
_113 = _119 * _114;
_112 = (double) _113;
_111 = (signed int) i_127;
_110 = (double) _111;
_109 = __builtin_log (_110);
_108 = _112 * _109;
_107 = (float) _108;
_106 = _35 + _124;
*_106 = _107;
i_17 = i_127 + 1;
ivtmp_28 = ivtmp_31 - 1;
if (ivtmp_28 != 0)
goto <bb 40>; [98.42%]
else
goto <bb 12>; [1.58%]
<bb 40> [local count: 669250617]:
goto <bb 36>; [100.00%]
