https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119943
Bug ID: 119943
Summary: -O3 forgets trivial code shift. causing significant
slowdown
Product: gcc
Version: 12.4.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: aleks at physik dot tu-berlin.de
Target Milestone: ---
the test below caused significant slowdown (3 vs 11 us) using "-O3"
optimization on x64 arch (Intel 12 gen).
the critical line of code is a struct initialization of 4x uint8.
this indicates 2 issues:
1. why moving the intialization around changes runtime in O3 ?
gcc should detect this easy trick (low hanging fruit).
2. moving up another line "uint8 aa,bb=0,cc=3,dd=4;" is not harmful.
struct filling is more critical, even if only {0,0,0,0} = 0u (32bit)
// the code reduced to the idea
typedef unsigned int uint32;
typedef unsigned char uint8;
// time measure via (average over 10 runs)
struct timespec ts0 = {0, 0};
clock_gettime(CLOCK_REALTIME, &ts0);
typedef struct {
uint8 a, b, c, d;
} cfg_t;
// test (2 vs 10 us)
uint32 test1(elem_t array[], const uint32 n) { // n=60
// if critical line is here, then slowdown !!
if (n <= 60u) {
if (n > 1u) quickSort(array, 0, n - 1);
return 0u;
}
// we never go here, as (n<=60)
uint8 aa, bb = 0, cc = 3, dd = 4; // the other not harmful line
cfg_t cfg = {0, 0, 0, 0}; // critical line !!!!!!!!!!!
// something ABC, same in test2()
// we use cfg struct here many times
return n + cfg.a;
}
