New Linux Issue? - Page 3

Adrian · 08-11-2016, 07:56 PM

Well, we finally got to the real problem!

You never had a 64-bit version working in the past. We released the 64-bit version in Steam some weeks ago. You were running the 32-bit version and everything was fine until we made available the 64-bit one.

Even if we set the flag to avoid lahf/sahf, LuaJIT needs it and, of course, adds it.

The possible solutions would be:

- Downloading the 32-bit installer from our website and play outside Steam.
- Find a way (I don't know if it's possible) to force Steam to use the 32-bit architecture for CoR.

So, there is no way for us to fix it. Sorry!

_Enio_ · 08-11-2016, 08:47 PM

I was curious, and turned out - there is another option.

However first a quick note - even when supplying "-mno-sahf" to gcc when compiling LuaJIT - it will not remove it - because it is INLINE asm, this will be translated verbatim.

So what are we using.

Code:

[enio@thinkpad ~]$ cd regnum/live; strings game | grep LuaJ
LuaJIT 2.0.2

Have to say i checked 2.0.4 first, thats why i tested 2.0.3 too if it already works:

Code:

[enio@thinkpad LuaJIT-2.0.2]$ objdump -d src/libluajit.a | grep -A3 -B2 fprem
    2924:    dd 42 08                 fldl   0x8(%rdx)
    2927:    dd 02                    fldl   (%rdx)
    2929:    d9 f8                    fprem  
    292b:    df e0                    fnstsw %ax
    292d:    9e                       sahf   
    292e:    7a f9                    jp     2929 <lj_ff_math_fmod+0x28>

[enio@thinkpad LuaJIT-2.0.3]$ objdump -d src/libluajit.a | grep -A3 -B2 fprem
    2921:    dd 42 08                 fldl   0x8(%rdx)
    2924:    dd 02                    fldl   (%rdx)
    2926:    d9 f8                    fprem  
    2928:    df e0                    fnstsw %ax
    292a:    9e                       sahf   
    292b:    7a f9                    jp     2926 <lj_ff_math_fmod+0x28>

[enio@thinkpad LuaJIT-2.0.4]$ objdump -d src/libluajit.a | grep -A3 -B2 fprem
    291f:    dd 42 08                 fldl   0x8(%rdx)
    2922:    dd 02                    fldl   (%rdx)
    2924:    d9 f8                    fprem  
    2926:    df e0                    fnstsw %ax
    2928:    66 25 00 04              and    $0x400,%ax
    292c:    75 f6                    jne    2924 <lj_ff_math_fmod+0x28>

Isnt that funny? Thats the only occurrence of remainder calculation in LuaJIT, and they changed it in 2.0.4 (current stable release) to use and instead of sahf (which does the same as test in the other thread, maybe its faster).

Here is the github commit in question, sadly no reasoning stated, however its labelled "x86/x64: Avoid use of SAHF instruction." (https://github.com/LuaJIT/LuaJIT/com...412c20c3eb0128)

tl;dr: SHAF is not used anymore in LuaJit since version 2.0.4, so get it updated! :P

Adrian · 08-11-2016, 09:02 PM

Quote:

Originally Posted by _Enio_

tl;dr: SHAF is not used anymore in LuaJit since version 2.0.4, so get it updated! :P

OMG, I feel ashamed

Tomorrow we will test updating it and see if everything still works, ha!

_Enio_ · 08-11-2016, 09:44 PM

Quote:

Originally Posted by Adrian

OMG, I feel ashamed

Tomorrow we will test updating it and see if everything still works, ha!

Lol, don't be! It was a lucky find driven by curiosity on the workings of the "-no-sahf" switch.

In case there are incompatibilities regarding other changes in 2.0.3/2.0.4 - it should be very easy to cherry pick this change into 2.0.2 too.

Good luck!

Adrian · 08-13-2016, 07:28 PM

Quote:

Originally Posted by _Enio_

Lol, don't be! It was a lucky find driven by curiosity on the workings of the "-no-sahf" switch.

In case there are incompatibilities regarding other changes in 2.0.3/2.0.4 - it should be very easy to cherry pick this change into 2.0.2 too.

Good luck!

Good news. It works and sahf is not there anymore. We'll be uploading this on Monday and I'll let you know here so you can test.

_Enio_ · 08-13-2016, 08:32 PM

Quote:

Originally Posted by Adrian

Good news. It works and sahf is not there anymore. We'll be uploading this on Monday and I'll let you know here so you can test.

Awesome, thanks for the update.

Adrian · 08-15-2016, 11:14 AM

We uploaded the 64-bit version with the changes to our experimental repository. Lebeau, please tick in your launcher advanced options "Use experimental client" and let us know if the problem is gone. Thanks!

Lebeau · 08-15-2016, 01:30 PM

Quote:

Originally Posted by Adrian

We uploaded the 64-bit version with the changes to our experimental repository. Lebeau, please tick in your launcher advanced options "Use experimental client" and let us know if the problem is gone. Thanks!

Adrian, Enio, I'm extremely grateful for all of your time & effort that have been expended on my (& perhaps others') behalf in this matter, & I truly thank you both very much indeed. There's currently a hitch of sorts though, entirely on my end. Currently, I'm away from home & desktop, with only a tablet for internet access, & so, cannot test this revamped 64-bit game version. Regrettably, I'll be away for another two weeks, but will install, test it, & then, post my findings here as soon as I get settled back in at home. I duly apologize for being the roadblock in this & for this unavoidable delay, but summer vacation at the lake, cookouts, vino & an exceptionally large & comfy hammock hung in the shade are not at all to be missed if at all possible..../me raises glass in salute....

_Enio_ · 08-15-2016, 05:05 PM

Quote:

Originally Posted by Adrian

We uploaded the 64-bit version with the changes to our experimental repository.

I can confirm the absence of that evil sahf instruction, and i did not notice any side effects on my machine (A recent intel i5 cpu).

Code:

[15/08/2016 19:03:59] [ClientBase][client_base.cpp(343)] Client build: 39674
[enio@thinkpad live]$ strings game | grep LuaJ
LuaJIT 2.0.4
[enio@thinkpad live]$ objdump -d game | grep sahf
[enio@thinkpad live]$

Thank you Adrian!

Adrian · 08-15-2016, 08:43 PM

Quote:

Originally Posted by _Enio_

I can confirm the absence of that evil sahf instruction, and i did not notice any side effects on my machine (A recent intel i5 cpu).

Code:

[15/08/2016 19:03:59] [ClientBase][client_base.cpp(343)] Client build: 39674
[enio@thinkpad live]$ strings game | grep LuaJ
LuaJIT 2.0.4
[enio@thinkpad live]$ objdump -d game | grep sahf
[enio@thinkpad live]$

Thank you Adrian!

Thank you for your stoic help!

08-11-2016, 07:56 PM	#21
Adrian NGD·Studios Join Date: May 2007 Location: Wien, Österreich Posts: 2,859	Well, we finally got to the real problem! You never had a 64-bit version working in the past. We released the 64-bit version in Steam some weeks ago. You were running the 32-bit version and everything was fine until we made available the 64-bit one. Even if we set the flag to avoid lahf/sahf, LuaJIT needs it and, of course, adds it. The possible solutions would be: - Downloading the 32-bit installer from our website and play outside Steam. - Find a way (I don't know if it's possible) to force Steam to use the 32-bit architecture for CoR. So, there is no way for us to fix it. Sorry! __________________

08-11-2016, 08:47 PM	#22
_Enio_ Marquis Join Date: Aug 2007 Location: Germany Posts: 1,843	I was curious, and turned out - there is another option. However first a quick note - even when supplying "-mno-sahf" to gcc when compiling LuaJIT - it will not remove it - because it is INLINE asm, this will be translated verbatim. So what are we using. Code: [enio@thinkpad ~]$ cd regnum/live; strings game \| grep LuaJ LuaJIT 2.0.2 Have to say i checked 2.0.4 first, thats why i tested 2.0.3 too if it already works: Code: [enio@thinkpad LuaJIT-2.0.2]$ objdump -d src/libluajit.a \| grep -A3 -B2 fprem 2924: dd 42 08 fldl 0x8(%rdx) 2927: dd 02 fldl (%rdx) 2929: d9 f8 fprem 292b: df e0 fnstsw %ax 292d: 9e sahf 292e: 7a f9 jp 2929 <lj_ff_math_fmod+0x28> [enio@thinkpad LuaJIT-2.0.3]$ objdump -d src/libluajit.a \| grep -A3 -B2 fprem 2921: dd 42 08 fldl 0x8(%rdx) 2924: dd 02 fldl (%rdx) 2926: d9 f8 fprem 2928: df e0 fnstsw %ax 292a: 9e sahf 292b: 7a f9 jp 2926 <lj_ff_math_fmod+0x28> [enio@thinkpad LuaJIT-2.0.4]$ objdump -d src/libluajit.a \| grep -A3 -B2 fprem 291f: dd 42 08 fldl 0x8(%rdx) 2922: dd 02 fldl (%rdx) 2924: d9 f8 fprem 2926: df e0 fnstsw %ax 2928: 66 25 00 04 and $0x400,%ax 292c: 75 f6 jne 2924 <lj_ff_math_fmod+0x28> Isnt that funny? Thats the only occurrence of remainder calculation in LuaJIT, and they changed it in 2.0.4 (current stable release) *to use and* instead of sahf** (which does the same as test in the other thread, maybe its faster). Here is the github commit in question, sadly no reasoning stated, however its labelled "x86/x64: Avoid use of SAHF instruction." (https://github.com/LuaJIT/LuaJIT/com...412c20c3eb0128) tl;dr: SHAF is not used anymore in LuaJit since version 2.0.4, so get it updated! :P __________________ Fix the Marksman subclass: Suggestion

08-15-2016, 11:14 AM	#27
Adrian NGD·Studios Join Date: May 2007 Location: Wien, Österreich Posts: 2,859	We uploaded the 64-bit version with the changes to our experimental repository. Lebeau, please tick in your launcher advanced options "Use experimental client" and let us know if the problem is gone. Thanks! __________________