https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107895
Bug ID: 107895
Summary: mt19937 bad performance on LP64
Product: gcc
Version: 12.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: jengelh at inai dot de
Target Milestone: ---
Input
=====
#include <random>
#include <unistd.h>
static std::mt19937 rng;
int main() {
for (size_t j = 0; j < 0x8000; ++j) {
uint32_t numbers[65536];
for (size_t i = 0; i < std::size(numbers); ++i)
numbers[i] = rng();
// ensure number generation is not all optimized away
write(STDOUT_FILENO, numbers, sizeof(numbers));
}
}
Observed
========
Target: x86_64-suse-linux
gcc version 12.2.1 20221020 [revision 0aaef83351473e8f4eb774f8f999bbe87a4866d7]
(SUSE Linux)
$ g++ x.cpp -O2 && time ./a.out >/dev/zero
-m32 -m64
=============== ===== ======
std::mt19937 3.9s 11.5s
std::mt19937_64 14.0s 11.6s
=============== ===== ======
error ±0.1s
With -ftree-loop-if-convert [Bug #80520], but still not at -m32 levels:
+-ftree- -m32 -m64
=============== ===== ======
std::mt19937 3.9s 5.2s
std::mt19937_64 14.0s 5.4s
=============== ===== ======
error ±0.1s
Expected
========
Expected to see <= 4.7s on -m64 at all times. (3.9 + ~20% margin for wider
transfers CPU<->caches/RAM)
The -m64 versions should have somewhat equal runtime or faster runtime (because
more registers, more opportunities); concerns like https://gmplib.org/32vs64
apply to old CPUs, but I do not think it's indicative of how contemporary
x86_64 systems perform.
Additional information
======================
CPUs:
"11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz"
[fam 6 model 140 stepping 1 microcode 0xa4] and
"AMD Ryzen 7 3700X 8-Core Processor"
[fam 23 model 113 stepping 0 microcode 0x8701013]
(about 3.0 and 10.2 seconds runtime, respectively)
