https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117801
Bug ID: 117801
Summary: aarch64: 20% regression in TSVC s278 since
r15-3509-gd34cda72098867
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: dhruvc at nvidia dot com
Target Milestone: ---
The s278 kernel has the following code:
===
real_t s278(struct args_t * func_args)
{
// control flow
// if/goto to block if-then-else
initialise_arrays(__func__);
gettimeofday(&func_args->t1, NULL);
for (int nl = 0; nl < iterations; nl++) {
for (int i = 0; i < LEN_1D; i++) {
if (a[i] > (real_t)0.) {
goto L20;
}
b[i] = -b[i] + d[i] * e[i];
goto L30;
L20:
c[i] = -c[i] + d[i] * e[i];
L30:
a[i] = b[i] + c[i] * d[i];
}
dummy(a, b, c, d, e, aa, bb, cc, 0.);
}
gettimeofday(&func_args->t2, NULL);
return calc_checksum(__func__);
}
===
Since r15-3509-gd34cda720988674bcf8a24267c9e1ec61335d6de, an extra mov is being
generated in the inner loop. This (likely along with instruction ordering
differences) is causing a 20% slowdown in the execution of the kernel.
Assembly differences between GCC 15 and GCC
14.2:https://godbolt.org/z/Kvaj5W6ve
Assembly differences on LLVM-MCA: https://godbolt.org/z/9jWKf4c6n
