https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86423
Bug ID: 86423
Summary: Omnetpp is slower on PowerPC using -ffast-math than
not using -ffast-math
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: meissner at gcc dot gnu.org
Target Milestone: ---
I did some Spec 2006 runs, and one of the runs I was comparing normal speed
with -ffast-math. As expected a lot of SpecFP benchmarks were faster using the
-ffast-math option. However, 3 of the benchmarks were slower, including
omnetpp.
In particular, cMessageHeap::shiftup(int) [clone .part.16], which is the hotest
function took 2,442,614 cycles when compiled with -ffast-math, and 2,378,917
cycles when compiled without -ffast-math.
Looking at the profile within the function, we see:
The C++ name is cMessageHeap::shiftup(int) [clone .part.16].
Percent Percent | Samples Samples | Line # Filename
slowmath fastmath | slowmath fastmath | Line # Filename
-------- -------- | -------- -------- | ------ -----------
39.3521% 40.3151% | 936,158 984,740 | 198 cmsgheap.cc
17.4123% 10.5189% | 414,223 256,938 | 200 cmsgheap.cc
4.6070% 16.3806% | 109,595 400,114 | 43 cmsgheap.cc
14.2110% 14.2290% | 338,071 347,561 | 199 cmsgheap.cc
9.0102% 7.1333% | 214,345 174,237 | 45 cmsgheap.cc
7.4063% 1.8450% | 176,191 45,066 | 44 cmsgheap.cc
3.4912% 4.9701% | 83,054 121,401 | 196 cmsgheap.cc
2.3383% 2.3092% | 55,627 56,405 | 204 cmsgheap.cc
2.1165% 2.2862% | 50,347 55,841 | 46 cmsgheap.cc
