https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81719
Bug ID: 81719
Summary: Range-based for loop on short fixed size array
generates long unrolled loop
Product: gcc
Version: 7.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: jzwinck at gmail dot com
Target Milestone: ---
C++11 range-based for loops over arrays of size known at compile time result in
bloated, branchy, and unreachable code with -O3 optimization. For example:
typedef int Items[2];
struct ItemArray
{
Items items;
int sum_x2() const;
};
int ItemArray::sum_x2() const
{
int total = 0;
for (int item : items)
{
total += item;
}
return total;
}
Clang compiles the above to [mov, add, ret]. GCC with -O2 compiles it to a few
more than that, and with -O3, a whopping 81 instructions. Add -march=haswell
and behold about 130 instructions to add two ints.
GCC (all versions, 4 to 7) generates code to handle a variable-sized array up
to about 6 to 14 elements, depending on -march. The number of elements is
known at compile time to be 2 (other small values also elicit the bug). GCC
should generate three instructions in both -O2 and -O3. It actually does, if
sum_x2() is a free function instead of a member function. The problem also
goes away if you use a C-style loop.
There are lots of permutations of this, including using a range-based for loop
to assign a common value to every element of an array whose size is known at
compile time (120 instructions to assign a single int:
https://godbolt.org/g/BGYggD).
Discussion on Stack Overflow:
https://stackoverflow.com/questions/45496987/gcc-optimizes-fixed-range-based-for-loop-as-if-it-had-longer-variable-length
