bug#16492: Emacs won't bootstrap with clang 3.4 on Cygwin

bug#16492: Emacs won't bootstrap with clang 3.4 on Cygwin

[Paul Eggert]

[This bug was originally reported by mirek.kaim in 
http://lists.gnu.org/archive/html/emacs-devel/2014-01/msg01702.html and 
I'm taking the liberty of copying its contents below.]

editfns.c gets miscompiled at -O > 0, causing emacs to crash.

backtrace:

(gdb) run
Starting program: /y/build/emacs/src/temacs --batch --load loadup bootstrap
[New Thread 3728.0x10dc]
[New Thread 3728.0x10c0]
Loading loadup.el (source)...
Using load-path (/y/emacs/lisp /y/emacs/lisp/emacs-lisp
/y/emacs/lisp/language /y/emacs/lisp/international
/y/emacs/lisp/textmodes /y/emacs/lisp/vc)
[New Thread 3728.0xdd4]
Loading emacs-lisp/byte-run (source)...
Loading emacs-lisp/backquote (source)...
Loading subr (source)...
Loading version (source)...

Program received signal SIGSEGV, Segmentation fault.
strftime_case_ (upcase=false, s=0x228f00 "\001\212\210",
     maxsize=<optimized out>, format=0x0, tp=0xa, ut=<optimized out>,
     ns=<optimized out>) at ../../../emacs/lib/strftime.c:507
507       for (f = format; *f != '\0'; ++f)
(gdb) backtrace
#0  strftime_case_ (upcase=false, s=0x228f00 "\001\212\210",
     maxsize=<optimized out>, format=0x0, tp=0xa, ut=<optimized out>,
     ns=<optimized out>) at ../../../emacs/lib/strftime.c:507
#1  0x005a80c2 in nstrftime (s=0x228f00 "\001\212\210", maxsize=0, 
format=0x0,
     tp=0x61256e80 <tm>, ut=1, ns=484375000)
     at ../../../emacs/lib/strftime.c:1449
#2  0x00513ef0 in format_time_string (format=0x0, formatlen=0, t=...,
     ut=<optimized out>, tmp=<optimized out>)
     at ../../../emacs/src/editfns.c:1672
#3  0x00513d44 in Fformat_time_string (format_string=<optimized out>,
     timeval=2268888, universal=<optimized out>)
     at ../../../emacs/src/editfns.c:1756
#4  0x0051d61c in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2182
#5  0x0051d429 in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2140
#6  0x0051e6e1 in Flet (args=<optimized out>) at 
../../../emacs/src/eval.c:937
#7  0x0051d397 in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2124
#8  0x0052175b in funcall_lambda (fun=<optimized out>, nargs=<optimized 
out>,
     arg_vector=<optimized out>) at ../../../emacs/src/eval.c:459
#9  0x00520346 in apply_lambda (fun=9104334, args=<optimized out>)
     at ../../../emacs/src/eval.c:2915
---Type <return> to continue, or q <return> to quit---
#10 0x0051d53d in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2221
#11 0x0052030b in apply_lambda (fun=9361006, args=<optimized out>)
     at ../../../emacs/src/eval.c:2906
#12 0x0051d53d in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2221
#13 0x0051d429 in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2140
#14 0x0051d4d9 in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2161
#15 0x0053e568 in readevalloop (readcharfun=8496874,
     stream=0x864080 <bss_sbrk_buffer+671560>, sourcename=<optimized out>,
     printflag=<optimized out>, unibyte=<optimized out>, readfun=8130560,
     start=<optimized out>, end=<optimized out>)
     at ../../../emacs/src/lread.c:1934
#16 0x0053d3f4 in Fload (file=<optimized out>, noerror=<optimized out>,
     nomessage=<optimized out>, nosuffix=<optimized out>,
     must_suffix=<optimized out>) at ../../../emacs/src/lread.c:1370
#17 0x0051d675 in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2190
#18 0x0053e568 in readevalloop (readcharfun=8496874,
     stream=0x864010 <bss_sbrk_buffer+671448>, sourcename=<optimized out>,
     printflag=<optimized out>, unibyte=<optimized out>, readfun=8130560,
---Type <return> to continue, or q <return> to quit---
     start=<optimized out>, end=<optimized out>)
     at ../../../emacs/src/lread.c:1934
#19 0x0053d3f4 in Fload (file=<optimized out>, noerror=<optimized out>,
     nomessage=<optimized out>, nosuffix=<optimized out>,
     must_suffix=<optimized out>) at ../../../emacs/src/lread.c:1370
#20 0x0051d675 in eval_sub (form=<optimized out>)
     at ../../../emacs/src/eval.c:2190
#21 0x005200d7 in Feval (form=8430502, lexical=<optimized out>)
     at ../../../emacs/src/eval.c:1994
#22 0x004bd1bd in top_level_2 () at ../../../emacs/src/keyboard.c:1179
#23 0x0051f4af in internal_condition_case (bfun=0x77e170 <pbm_format+676>,
     handlers=<optimized out>, hfun=<optimized out>)
     at ../../../emacs/src/eval.c:1345
#24 0x004bd183 in top_level_1 (ignore=8450074)
     at ../../../emacs/src/keyboard.c:1187
#25 0x0051ee24 in internal_catch (tag=<optimized out>, func=0x0, 
arg=7856496)
     at ../../../emacs/src/eval.c:1109
#26 0x004ac41a in recursive_edit_1 () at ../../../emacs/src/keyboard.c:1148
#27 0x004ac5d5 in Frecursive_edit () at ../../../emacs/src/keyboard.c:841
#28 0x004ab10a in main (
     argc=<error reading variable: Cannot access memory at address 0x0>,
     argv=<optimized out>) at ../../../emacs/src/emacs.c:1637
(gdb)

exact steps to reproduce:
http://aurora.irc.arloria.net/~unic0rn/emacs_clang_optimization_bug.html

my workaround is kinda silly, but at least it works. i've tried to
modify editfns.c to trick llvm's optimizers into generating something
that will work, to no avail. it seems like the first param doesn't end
up on the stack somehow. since llvm/clang 3.4 is a current release
version, i guess it'll land on osx and freebsd sooner than later, and
if they won't fix that until then - well, it may cause some problems.
unfortunately, i can't test it with higher -march settings, because my
cpu doesn't even support SSE2.

i've filled a bug for llvm, but i'm not sure how far it'll go:
http://llvm.org/bugs/show_bug.cgi?id=18537

bug#16492: can't reproduce it with clang 3.4 on Fedora x86-64

[Paul Eggert]

For what it's worth, I cannot reproduce the problem with clang 3.4 
(which I downloaded and built myself) on Fedora 20 x86-64.  'make 
bootstrap' works just fine, with trunk bzr 116067.  Most likely the 
problem is specific to x86, and perhaps it needs the optimization 
parameters '-march=pentium3 -mtune=pentium3 -O3' that were given in the 
exact steps to reproduce.

bug#16492: Emacs won't bootstrap with clang 3.4 on Cygwin

[Angelo Graziosi]

Paul Eggert wrote:
> This bug was originally reported by mirek.kaim

For the sake of completeness, Emacs (GTK on Cygwin 32) builds just fine 
with the official Cygwin CLANG package, i.e. version 3.1. I do this 
weekly and it is more than an year... The last build is from yesterday 
trunk... But, perhaps, the issue regards only CLANG 3.4..


Ciao,
Angelo.

bug#16492: can't reproduce it with clang 3.4 on Fedora x86-64

[unic0rn]

looks like the bug exists in clang 3.3 as well, and that's quite old
actually (3.1 packaged for cygwin is ancient). but you may be right
that this is x86-specific problem. if that's the case, it may take
some time until llvm devs will try to fix it.

http://llvm.org/bugs/show_bug.cgi?id=18171

On Sun, Jan 19, 2014 at 9:07 AM, Paul Eggert <eggert@cs.ucla.edu> wrote:
> For what it's worth, I cannot reproduce the problem with clang 3.4 (which I
> downloaded and built myself) on Fedora 20 x86-64.  'make bootstrap' works
> just fine, with trunk bzr 116067.  Most likely the problem is specific to
> x86, and perhaps it needs the optimization parameters '-march=pentium3
> -mtune=pentium3 -O3' that were given in the exact steps to reproduce.

bug#16492: can't reproduce it with clang 3.4 on Fedora x86-64

[npostavs]

close 16492
quit

unic0rn <mirek.kaim@gmail.com> writes:

> looks like the bug exists in clang 3.3 as well, and that's quite old
> actually (3.1 packaged for cygwin is ancient). but you may be right
> that this is x86-specific problem. if that's the case, it may take
> some time until llvm devs will try to fix it.
>
> http://llvm.org/bugs/show_bug.cgi?id=18171

According to this link, it was fixed in clang.