https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96252
Bug ID: 96252
Summary: mis-optimization where identical functions have very
different codegen since gcc 10
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: wjwray at gmail dot com
Target Milestone: ---
(Stumbled on this odd effect while examining codegen for operator<=>)
The reduced sample compiles c++11 and up (std version likely irrelevant)
with the different codegen occurring since GCC 10.
Two identical functions:
bool cmp_x(cmp l, cmp r) noexcept {
return std::lexicographical_compare(begin(l),end(l)
,begin(r),end(r)); }
bool cmp_y(cmp l, cmp r) noexcept {
return std::lexicographical_compare(begin(l),end(l)
,begin(r),end(r)); }
generate very different code for cmp = array<int,64> at -O2 and -O3.
The first is looping, the second has much longer unrolled codegen.
My benchmarks show 40% difference in runtime, quick-bench shows 30%.
Compiler Explorer https://godbolt.org/z/97box6
Quick-bench 1.3x https://quick-bench.com/q/480qkw1sP4OWOH6JBxsm-J_9uOk
(Adding -fno-inline may provide a clue.)