Nice find! We should always build yasm in opt to address the 3x debug slowdown (precedent: https://cs.chromium.org/chromium/src/third_party/ffmpeg/BUILD.gn?q=file:ffmpeg.*build.gn+package:%5Echromium$&dr&l=207).

In the case of yasm adding the optimize_max config is not sufficient to resolve the problem. That only gives about a 10-15% speedup.

The rest of the performance difference comes from the debug CRT and from having _DEBUG defined. It turns out that the static_crt/dynamic_crt configs are the most important ones for improving the performance of yasm. The /MTd flag causes about a 2.3x slowdown because it pulls in the debug CRT (with lots of extra locks and allocation overhead) and because it defines _DEBUG (which presumably triggers various asserts and other debug checks).

I'll have a CL for review soon. The test results are:

      # Build with full optimizations even on debug configurations, because
      # some yasm build steps (highbd_sad4d_sse2.asm) can take ~33 seconds
      # or more in debug component builds. Enabling compiler optimizations
      # saves ~5 seconds.
      configs -= [ "//build/config/compiler:default_optimization" ]
      configs += [ "//build/config/compiler:optimize_max" ]
      # This avoids explicitly defining _DEBUG - this doesn't make a
      # noticeable difference to the run time.
      configs -= [ "//build/config:debug" ]
      configs += [ "//build/config:release" ]
      # This switches to using the debug CRT. On debug-component builds of
      # highbd_sad4d_sse2.asm this saves about 15 s.
      configs -= [ "//build/config/win:default_crt" ]
      configs += [ "//build/config/win:release_crt" ]

For simplicity not all of these changes will be included.

Manually adding link to CL that fixed this issue:

commit 9c90eb122f3327c8c03bbf2171a97cb157f89459
Author: brucedawson <brucedawson@chromium.org>
Date:   Thu May 18 20:45:28 2017 -0700

    Optimize yasm even in debug builds

    Running yasm on highbd_sad4d_sse2.asm takes about ~33 seconds on debug
    component builds on Windows - closer to 42 seconds when the system is
    under load. This is partially because compiler optimizations are
    disabled in debug builds (costs ~5 seconds) and partially because we
    link with the debug CRT (costs ~15 seconds). This change makes it so
    that we always enable optimizations and always link with the release
    CRT, thus reducing the time to run yasm on highbd_sad4d_sse2.asm from
    ~33 seconds to ~13 seconds.

    Further improvements could be obtained by only running yasm once on the
    .asm files, but such a change is left for later.

    The total CPU-time savings is tiny compared to the cost of a full build,
    but on some goma builds this step ends up being the long pole which
    serializes the build and costs an estimated 5% of elapsed build time.

    BUG= 722617 

    Review-Url: https://codereview.chromium.org/2885213002
    Cr-Commit-Position: refs/heads/master@{#473066}