Segfault when 0.12.2 built with gcc 5.1.0 on illumos

[alhazred]

# ./node
Segmentation Fault (core dumped)
# mdb core
Loading modules: [ libumem.so.1 libc.so.1 ld.so.1 ]
> $G
C++ symbol demangling enabled
> $c
v8::internal::String::WriteToFlat<unsigned short>+0x7fc(9e025971, 9e026d36, 1f, 3c, 804787c, 6a7)
_ZN2v88internalL39StringReplaceGlobalAtomRegExpWithStringINS0_16SeqTwoByteStringEEEPNS0_6ObjectEPNS0_7IsolateENS0_6HandleINS0_6StringEEENS7_
INS0_8JSRegExpEEES9_NS7_INS0_7JSArrayEEE.constprop.415+0x15f(9044008, 80479a8, 80479a4, 80479a0, 804799c, 8)
v8::internal::Runtime_StringReplaceGlobalRegExpWithString+0x50f(4, 80479a8, 9044008, bcb0a361, 8047980, 80479d8)
0xbcb0a3d6(98a69809, 9ff234f1, 9e026bf9, 9e025971, 98a08091, 98a08091)
0xbcb5bab3(9ff234f1, 9e026bf9, 9e025971, 98a2f271, 9ff360d5, 98a3978d)
0xbcb591c8(98a410b1, 98a3978d, 98a08091, 98a08091, 98a08091, 98a08091)
0xbcb4da13(98a364f1, 98a410b1, 98a41115, 98a40ff5, 8047a6c, bcb46f15)
0xbcb4d0ec(98a41139, 98a364f1, 98a41115, bcb46e81, 10, 0)
0xbcb46f15(0, 0, bcb25601, 0, 0, 0)
0xbcb256aa(bcb4cae0, 9e041d65, 98a364f1, 1, 8047c0c, 90b9cb8)
v8::internal::Execution::Call+0x129(8047b3c, 9044008, 908d014, 908d018, 1, 8047c0c)
v8::Function::Call+0x99(8047bb0, 908d014, 908d018, 1, 8047c0c, 908d050)
node::LoadEnvironment+0x278(909cac8, 0, 9043fd8, 833535f)
node::Start+0x14d(1, 8047d20, 8047cf8, 8324a2a, feffb0a4, 8047d00)
main+0x1b(8047cdc, fe8586e8, 8047d14, 8239df3, 1, 8047d20)
_start+0x83(1, 8047de0, 0, 8047de7, 8047df8, 8047e03)

[misterdjules]

It seems that it might be related to #9227 and https://code.google.com/p/chromium/issues/detail?id=152407#c11.

Could you please show us the complete log of the build process. I'm especially interested in seeing the optimization flags used by g++ to build V8.

Are you able to reproduce the problem with gcc 4.x?

[alhazred]

On 13 May 2015 at 02:44:25, Julien Gilli (notifications@github.com) wrote:
It seems that it might be related to #9227 and https://code.google.com/p/chromium/issues/detail?id=152407#c11.

Could you please show us the complete log of the build process. I'm especially interested in seeing the optimization flags used by g++ to build V8.

Are you able to reproduce the problem with gcc 4.x?

You can find build log on Dropbox: https://www.dropbox.com/s/fsyqp9g9t2gqbyv/log5?dl=0
No problem with using gcc 4.8.1. Looks like there are no different flags when building with 4.8.1 or 5.1.0.

But problem is gone when I use -gO2 instead of -gO3.

Alexander

—
Reply to this email directly or view it on GitHub.

[misterdjules]

Thank you. Can you try building io.js, which comes with a newer V8, with GCC 5.1 and -O3, and let us know if you get the same problem?

[alhazred]

No, io.js doesn't have such problem ;)

[misterdjules]

This issue in Chromium seems like it could be related too: https://code.google.com/p/chromium/issues/detail?id=412967.

@joyent/node-tsc I don't have the time to investigate potential fixes right now, but if we can validate that the fix mentioned above fixes this specific issue (using GCC 5.1.0 on SmartOS 32 bits), then we should also investigate what other platforms are affected and the extent of the problem.

[misterdjules]

@joyent/node-tsc Set P-1 because I suspect we would get a lot more reporters of this problem if it affects every recent version of GCC on 32 bits systems. If it turns out this issue is more specific to a single platform or a restricted set of GCC versions, then we can lower the priority.

[misterdjules]

@orangemocha That's potentially related to the issues you're seeing on the build platform.

[orangemocha]

The stack trace is very similar to the issue described https://code.google.com/p/chromium/issues/detail?id=412967. It's almost certainly the same issue.

I am going to investigate a couple of options:

1. Backporting the patches listed in the Chromium issue
2. build deps/v8 with -fvect-cost-model=cheap. @joaocgreis verified this solves the issue. But it might have performance implications.

[misterdjules]

@alhazred Do you have some time to try the fix in #25556?

[alhazred]

I've built janeasystems/node with 5.1, unfortunately same core.

[misterdjules]

@alhazred Did you build the joaocgreis-pick-v8-smartos branch from janeasystems/node?

[alhazred]

Ah.. End of the day :) give me a min

[misterdjules]

@alhazred No worries, your help is very much appreciated 👍

[alhazred]

Bingo! Thank you very much for this :)

[misterdjules]

That's all thanks to @joaocgreis' work :)

@alhazred Thank you very much to you for helping us solve this! 👍

[joaocgreis]

Fixed by 48b0ca2 and 13ea50e