[PATCH] reduce size penalty for including C++11 <algorithm> on x86 systems

[Jonathan Wakely]

On 13/10/15 21:44 -0400, Nathan Froyd wrote:
>Including <algorithm> in C++11 mode (typically done for
>std::{min,max,swap}) includes <random>, for
>std::uniform_int_distribution.  On x86 platforms, <random> manages to
>drag in <x86intrin.h> through x86's opt_random.h header, and
><x86intrin.h> has gotten rather large recently with the addition of AVX
>intrinsics.  The comparison between C++03 mode and C++11 mode is not
>quite exact, but it gives an idea of the penalty we're talking about
>here:
>
>froydnj@thor:~/src$ echo '#include <algorithm>' | g++ -x c++ - -o - -E -std=c++11 | wc
>  53460  127553 1401268
>froydnj@thor:~/src$ echo '#include <algorithm>' | g++ -x c++ - -o - -E -std=c++03 | wc
>   9202   18933  218189
>
>That's approximately a 7x penalty in C++11 mode (granted, C++11 includes
>more than just <x86intrin.h>) with GCC 4.9.2 on a Debian system; current
>mainline is somewhat worse:
>
>froydnj@thor: gcc-build$ echo '#include <algorithm>' | xgcc [...] -std=c++11 | wc
>  84851  210475 2369616
>froydnj@thor: gcc-build$ echo '#include <algorithm>' | xgcc [...] -std=c++03 | wc
>   9383   19402  239676
>
><x86intrin.h> itself clocks in at 1.3MB+ of preprocessed text.

Yep, that's been bothering me for a while.

>This patch aims to reduce that size penalty by recognizing that both of
>the places that #include <x86intrin.h> do not need the full set of x86
>intrinsics, but can get by with a smaller, more focused header in each
>case.  <ext/random> needs only <emmintrin.h> to declare __m128i, while
>x86's opt_random.h must include <pmmintrin.h> for declarations of
>various intrinsic functions.
>
>The net result is that the size of mainline's <algorithm> is significantly reduced:
>
>froydnj@thor: gcc-build$ echo '#include <algorithm>' | xgcc [...] -std=c++11 | wc
>  39174   88538 1015281
>
>which seems like a win.

Indeed!

>Bootstrapped on x86_64-pc-linux-gnu with --enable-languages=c,c++,
>tested with check-target-libstdc++-v3, no regressions.  Also verified
>that <algorithm> and <ext/random> pass -fsyntax-check with
>-march=native (on a recent Haswell chip); if an -march=native bootstrap
>is necessary, I am happy to do that if somebody instructs me in getting
>everything properly set up.
>
>OK?

OK, thanks.