On 11/11/17 19:14 +0100, Marc Glisse wrote:
Hello,
operator new can clobber memory, it is hard to teach the compiler
otherwise since it is replaceable. Here I cache a couple values before
the call to the allocator. I checked the result on this simple
example:
#include <vector>
void f(std::vector<int>&v){ v.push_back(0); }
The patch does not affect the fast path where no reallocation is
needed. Compiling with -O3 (everything gets inlined into f), I see a
nice decrease in code size on x86_64
$ size old.o new.o
text data bss dec hex filename
462 0 0 462 1ce old.o
376 0 0 376 178 new.o
Even at -O2 where _M_realloc_insert is not inlined, I get a slight
decrease in code size (490 -> 470). On x86, that's 531 -> 519 (465 ->
387 at -O3).
I'm not going to modify every function like that, I just happened to
be looking at this example for other reasons, and the size gain is
larger than I expected, so I am posting the patch.
That's a nice improvement, OK for trunk. Thanks.
