Bug 149621

r190288 is the last good revision - https://build.webkit.org/builders/Apple%20Yosemite%20Release%20WK2%20%28Perf%29/builds/3031

I am able to reproduce this in both release and debug builds although the symptoms are a little different. In a debug build, we are OSR exiting from a function due to BadType, but the real reason is that the tag registers (r14 & r15 on X86-64) do not have the tag constants. it appears that they were not properly restored after a call (likely to an FTL function). More debugging.

Should this be duped to bug 149647 now (the rollout bug)?

(In reply to comment #3) > Should this be duped to bug 149647 now (the rollout bug)? The rollout bug is closed. I'll still work on the issue with this defect. I'll add this defect and https://bugs.webkit.org/show_bug.cgi?id=149619 to bug 149647.

A little more detail. We are calling a function that makes a varargs tail call. When the FTL makes a tail call and it needs to take the slow path, it doesn't properly restore the callee save registers. Moved the restoration of callee saves to common path before deciding to take the fast or slow path.

Created attachment 262192 [details] Patch

Landed as part of change set r190370: <http://trac.webkit.org/changeset/190370>.

Committed r190372: <http://trac.webkit.org/changeset/190372>

(In reply to comment #8) > Committed r190372: <http://trac.webkit.org/changeset/190372> This was really a change for an unrelated build fix.

(In reply to comment #7) > Landed as part of change set r190370: > <http://trac.webkit.org/changeset/190370>. Reopen, because r190370 didn't fix the orignal issue, Speedometer/Full.html still fails everywhere. Additionally it can't be an FTL issue, because FTL isn't enabled on Linux platforms at all. - https://build.webkit.org/builders/Apple%20Mavericks%20Release%20WK2%20%28Perf%29/builds/5900 - https://build.webkit.org/builders/EFL%20Linux%2064-bit%20Release%20WK2%20%28Perf%29/builds/6975 To make sure if r190370 is the culprit, I started force builds on perf bots: - r190368 - https://build.webkit.org/builders/Apple%20El%20Capitan%20Release%20WK2%20%28Perf%29/builds/12 - r190370 - https://build.webkit.org/builders/Apple%20Mavericks%20Release%20WK2%20%28Perf%29/builds/5904 (r190369 is an unrelated change)

(In reply to comment #10) > (In reply to comment #7) > > Landed as part of change set r190370: > > <http://trac.webkit.org/changeset/190370>. > > Reopen, because r190370 didn't fix the orignal issue, Speedometer/Full.html > still fails everywhere. Additionally it can't be an FTL issue, because > FTL isn't enabled on Linux platforms at all. > > - > https://build.webkit.org/builders/ > Apple%20Mavericks%20Release%20WK2%20%28Perf%29/builds/5900 > - > https://build.webkit.org/builders/EFL%20Linux%2064- > bit%20Release%20WK2%20%28Perf%29/builds/6975 > > To make sure if r190370 is the culprit, I started force builds on perf bots: > - r190368 - > https://build.webkit.org/builders/ > Apple%20El%20Capitan%20Release%20WK2%20%28Perf%29/builds/12 > - r190370 - > https://build.webkit.org/builders/ > Apple%20Mavericks%20Release%20WK2%20%28Perf%29/builds/5904 > > (r190369 is an unrelated change) The patch that was landed solves the problem encountered on Mac builds. Is it possible for you to run Speedometer/Full.html from a debug build in a debugger and post stack trace?

I didn't notice crashes (in release mode), but I get this console message locally (EFL Linux x86_64) and you can see it on EFL and Apple Mac performance bots too: ERROR: CONSOLE MESSAGE: line 208: Miss the info bar? Run TodoMVC from a server to avoid a cross-origin error. After grepping a little bit, base.js:140 in SpeedoMeter is responsible for printing this message to console.

(In reply to comment #12) > I didn't notice crashes (in release mode), but I get this console > message locally (EFL Linux x86_64) and you can see it on EFL and > Apple Mac performance bots too: > > ERROR: CONSOLE MESSAGE: line 208: Miss the info bar? Run TodoMVC from a > server to avoid a cross-origin error. > > After grepping a little bit, base.js:140 in SpeedoMeter > is responsible for printing this message to console. Looks like: https://bugs.webkit.org/show_bug.cgi?id=149720

*** Bug 149720 has been marked as a duplicate of this bug. ***

*** Bug 149720 has been marked as a duplicate of this bug. ***

<rdar://problem/22995488> In order to reproduce this failure, revert http://trac.webkit.org/changeset/190535

The bug is that console.info called in line 140 of base.js (there are a few copies of it inside PerformanceTests/Speedometer/) is reported erroneously at line 208. Given: 138: function getFile(file, callback) { 139: if (!location.host) { 140: return console.info('Miss the info bar? Run TodoMVC from a server to avoid a cross-origin error.'); 141: } 207: redirect(); 208: getFile('learn.json', Learn); 209: })(); I'd imagine that the console.info in "getFile" is tail-call optimized into the caller, and erroneously reporting the line number.