All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=750923

(For debugging:) Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?sid=f1ba06349eb79b27e5f729fb332e6d9eb01141fbb8c84306fec896d9b0043a9f


Bot(s) for this bug's original alert(s):

chromium-rel-win8-dual

Started bisect job https://chromeperf.appspot.com/buildbucket_job_status/8972511192980415776

=== Auto-CCing suspected CL author hans@chromium.org ===

Hi hans@chromium.org, the bisect results pointed to your CL, please take a look at the
results.


=== BISECT JOB RESULTS ===
Perf regression found with culprit

Suspected Commit
  Author : Hans Wennborg
  Commit : d2c91228a51bdf37ae3b2e501fb53c0528f1629c
  Date   : Fri Jul 28 20:11:05 2017
  Subject: win: Set is_clang=true by default

Bisect Details
  Configuration: win_8_perf_bisect
  Benchmark    : media_perftests
  Metric       : audio_converter/convert_fifo_only
  Change       : 53.55% | 1369035.03985 -> 635869.919234

Revision             Result                  N
chromium@490447      1369035 +- 18625.6      6      good
chromium@490481      1372256 +- 12191.3      6      good
chromium@490490      1361809 +- 27941.8      6      good
chromium@490492      1371743 +- 38374.6      6      good
chromium@490493      1367999 +- 11075.7      6      good
chromium@490494      636643 +- 7339.42       6      bad       <--
chromium@490498      637718 +- 6185.32       6      bad
chromium@490515      635870 +- 5850.33       6      bad

To Run This Test
  .\src\out\Release\media_perftests.exe --single-process-tests

More information on addressing performance regressions:
  http://g.co/ChromePerformanceRegressions

Debug information about this bisect:
  https://chromeperf.appspot.com/buildbucket_job_status/8972511192980415776


For feedback, file a bug with component Speed>Bisection

 Issue 750672  has been merged into this issue.

Issue 750703 has been merged into this issue.

Started looking.

Issue 750703 has been merged into this issue.

Issue 750703 has been merged into this issue.

The perf drop is due to FMAC_FUNC in media/base/vector_math.cc chosing the non-intrinsic version with win/clang. Clang is supposed to be able to autovectorize that to get better performance, but for some reason that's not happening here.

(Kind of related: https://chromium-review.googlesource.com/557059)

(The audio converter code is valling it as vector_math::FMAC)

*calling it

In my build, FMAC_C looks like this:

?FMAC_C@vector_math@media@@YAXQEBMMHQEAM@Z (void __cdecl media::vector_math::FMAC_C(float const * const,float,int,float * const)):
  0000000000000000: 45 85 C0           test        r8d,r8d
  0000000000000003: 7E 22              jle         0000000000000027
  0000000000000005: 44 89 C0           mov         eax,r8d
  0000000000000008: F3 0F 10 01        movss       xmm0,dword ptr [rcx]
  000000000000000C: F3 0F 59 C1        mulss       xmm0,xmm1
  0000000000000010: F3 41 0F 58 01     addss       xmm0,dword ptr [r9]
  0000000000000015: F3 41 0F 11 01     movss       dword ptr [r9],xmm0
  000000000000001A: 48 83 C1 04        add         rcx,4
  000000000000001E: 49 83 C1 04        add         r9,4
  0000000000000022: 48 FF C8           dec         rax
  0000000000000025: 75 E1              jne         0000000000000008
  0000000000000027: C3                 ret


Clang is getting invoked with /O1 for this code, which we interpret as "optimize for size" (it becomes -Os in the cc1 invocation).

If I pass -O2 instead, we get fancier code:

?FMAC_C@vector_math@media@@YAXQEBMMHQEAM@Z (void __cdecl media::vector_math::FMAC_C(float const * const,float,int,float * const)):
  0000000000000000: 56                 push        rsi
  0000000000000001: 45 85 C0           test        r8d,r8d
  0000000000000004: 0F 8E 40 01 00 00  jle         000000000000014A
  000000000000000A: 45 89 C2           mov         r10d,r8d
  000000000000000D: 41 83 F8 07        cmp         r8d,7
  0000000000000011: 76 1A              jbe         000000000000002D
  0000000000000013: 4A 8D 04 91        lea         rax,[rcx+r10*4]
  0000000000000017: 4C 39 C8           cmp         rax,r9
  000000000000001A: 0F 86 C9 00 00 00  jbe         00000000000000E9
  0000000000000020: 4B 8D 04 91        lea         rax,[r9+r10*4]
  0000000000000024: 48 39 C8           cmp         rax,rcx
  0000000000000027: 0F 86 BC 00 00 00  jbe         00000000000000E9
  000000000000002D: 31 D2              xor         edx,edx
  000000000000002F: 44 89 D0           mov         eax,r10d
  0000000000000032: 29 D0              sub         eax,edx
  0000000000000034: 4D 8D 42 FF        lea         r8,[r10-1]
  0000000000000038: 49 29 D0           sub         r8,rdx
  000000000000003B: 48 83 E0 03        and         rax,3
  000000000000003F: 74 2C              je          000000000000006D
  0000000000000041: 48 F7 D8           neg         rax
  0000000000000044: 66 66 66 2E 0F 1F  nop         word ptr cs:[rax+rax+0000000000000000h]
                    84 00 00 00 00 00
  0000000000000050: F3 0F 10 04 91     movss       xmm0,dword ptr [rcx+rdx*4]
  0000000000000055: F3 0F 59 C1        mulss       xmm0,xmm1
  0000000000000059: F3 41 0F 58 04 91  addss       xmm0,dword ptr [r9+rdx*4]
  000000000000005F: F3 41 0F 11 04 91  movss       dword ptr [r9+rdx*4],xmm0
  0000000000000065: 48 FF C2           inc         rdx
  0000000000000068: 48 FF C0           inc         rax
  000000000000006B: 75 E3              jne         0000000000000050
  000000000000006D: 49 83 F8 03        cmp         r8,3
  0000000000000071: 0F 82 D3 00 00 00  jb          000000000000014A
  0000000000000077: 49 29 D2           sub         r10,rdx
  000000000000007A: 49 8D 44 91 0C     lea         rax,[r9+rdx*4+0Ch]
  000000000000007F: 48 8D 4C 91 0C     lea         rcx,[rcx+rdx*4+0Ch]
  0000000000000084: 66 66 66 2E 0F 1F  nop         word ptr cs:[rax+rax+0000000000000000h]
                    84 00 00 00 00 00
  0000000000000090: F3 0F 10 41 F4     movss       xmm0,dword ptr [rcx-0Ch]
  0000000000000095: F3 0F 59 C1        mulss       xmm0,xmm1
  0000000000000099: F3 0F 58 40 F4     addss       xmm0,dword ptr [rax-0Ch]
  000000000000009E: F3 0F 11 40 F4     movss       dword ptr [rax-0Ch],xmm0
  00000000000000A3: F3 0F 10 41 F8     movss       xmm0,dword ptr [rcx-8]
  00000000000000A8: F3 0F 59 C1        mulss       xmm0,xmm1
  00000000000000AC: F3 0F 58 40 F8     addss       xmm0,dword ptr [rax-8]
  00000000000000B1: F3 0F 11 40 F8     movss       dword ptr [rax-8],xmm0
  00000000000000B6: F3 0F 10 41 FC     movss       xmm0,dword ptr [rcx-4]
  00000000000000BB: F3 0F 59 C1        mulss       xmm0,xmm1
  00000000000000BF: F3 0F 58 40 FC     addss       xmm0,dword ptr [rax-4]
  00000000000000C4: F3 0F 11 40 FC     movss       dword ptr [rax-4],xmm0
  00000000000000C9: F3 0F 10 01        movss       xmm0,dword ptr [rcx]
  00000000000000CD: F3 0F 59 C1        mulss       xmm0,xmm1
  00000000000000D1: F3 0F 58 00        addss       xmm0,dword ptr [rax]
  00000000000000D5: F3 0F 11 00        movss       dword ptr [rax],xmm0
  00000000000000D9: 48 83 C0 10        add         rax,10h
  00000000000000DD: 48 83 C1 10        add         rcx,10h
  00000000000000E1: 49 83 C2 FC        add         r10,0FFFFFFFFFFFFFFFCh
  00000000000000E5: 75 A9              jne         0000000000000090
  00000000000000E7: EB 61              jmp         000000000000014A
  00000000000000E9: 41 83 E0 07        and         r8d,7
  00000000000000ED: 4C 89 D2           mov         rdx,r10
  00000000000000F0: 4C 29 C2           sub         rdx,r8
  00000000000000F3: 0F 28 C1           movaps      xmm0,xmm1
  00000000000000F6: 0F C6 C0 00        shufps      xmm0,xmm0,0
  00000000000000FA: 4C 8D 59 10        lea         r11,[rcx+10h]
  00000000000000FE: 49 8D 41 10        lea         rax,[r9+10h]
  0000000000000102: 48 89 D6           mov         rsi,rdx
  0000000000000105: 66 66 2E 0F 1F 84  nop         word ptr cs:[rax+rax+0000000000000000h]
                    00 00 00 00 00
  0000000000000110: 41 0F 10 53 F0     movups      xmm2,xmmword ptr [r11-10h]
  0000000000000115: 41 0F 10 1B        movups      xmm3,xmmword ptr [r11]
  0000000000000119: 0F 59 D0           mulps       xmm2,xmm0
  000000000000011C: 0F 59 D8           mulps       xmm3,xmm0
  000000000000011F: 0F 10 60 F0        movups      xmm4,xmmword ptr [rax-10h]
  0000000000000123: 0F 58 E2           addps       xmm4,xmm2
  0000000000000126: 0F 10 10           movups      xmm2,xmmword ptr [rax]
  0000000000000129: 0F 58 D3           addps       xmm2,xmm3
  000000000000012C: 0F 11 60 F0        movups      xmmword ptr [rax-10h],xmm4
  0000000000000130: 0F 11 10           movups      xmmword ptr [rax],xmm2
  0000000000000133: 49 83 C3 20        add         r11,20h
  0000000000000137: 48 83 C0 20        add         rax,20h
  000000000000013B: 48 83 C6 F8        add         rsi,0FFFFFFFFFFFFFFF8h
  000000000000013F: 75 CF              jne         0000000000000110
  0000000000000141: 45 85 C0           test        r8d,r8d
  0000000000000144: 0F 85 E5 FE FF FF  jne         000000000000002F
  000000000000014A: 5E                 pop         rsi
  000000000000014B: C3                 ret

https://chromium-review.googlesource.com/597017

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/7c7dcc2872d7edd2ca4b9121eadf9ed5a6a664bf

commit 7c7dcc2872d7edd2ca4b9121eadf9ed5a6a664bf
Author: Hans Wennborg <hans@chromium.org>
Date: Wed Aug 02 15:00:08 2017

media: Build vector_math.cc with max optimization

It was currently built at "optimize for size", which means important
function such as FMAC_C weren't getting autovectorized and unrolled.
Since this file is performance sensitive, using max optimization seems
reasonable.

BUG= 750923 

Change-Id: Id38e66365ee1b8b636e391ba07b83a804142eba1
Reviewed-on: https://chromium-review.googlesource.com/597017
Reviewed-by: Dale Curtis <dalecurtis@chromium.org>
Reviewed-by: Nico Weber <thakis@chromium.org>
Commit-Queue: Nico Weber <thakis@chromium.org>
Cr-Commit-Position: refs/heads/master@{#491372}
[modify] https://crrev.com/7c7dcc2872d7edd2ca4b9121eadf9ed5a6a664bf/media/BUILD.gn

 Issue 751634  has been merged into this issue.

The media_perftests issue is fixed. Please don't dupe other issues against this.

 Issue 752057  has been merged into this issue.

 Issue 752089  has been merged into this issue.

> The media_perftests issue is fixed. Please don't dupe other issues against this.

hans@ dupe happens automatically by bisect bots once they figure out the regression is caused by the same CL. In this case, I expect your CL which
enables clang on window to have caused regression on a few different metrics.
So I suggest you just assign each duped issue to yourself when needs a
different fix. I cannot think of any better way to do this.

 

I see, thanks for clarifying.

 Issue 753655  has been merged into this issue.

 Issue 751639  has been merged into this issue.

 Issue 750822  has been merged into this issue.