gcc 3.2 no longer works

[flub]

Compiling greenlet with gcc 3.2 (RHEL3) no longer works. The assembler generated by for calling the functions by pointers is different then from the 0.3.1 release but I can't figure out why. Also to my untrained eye both generated assembler files seem like they should work.

[schmir]

what's the error message? which prior version does work on RHEL 3?

[flub]

It coredumps, 0.3.1 worked fine. Newer version of gcc work fine too, which is why I have been concentrating on the noinline stuff and comparing the assembler for the functions called by pointer between 0.3.3 and 0.3.1.

I wanted to attach the generated assembler but failed to figure out how to do that on github...

[schmir]

I can't reproduce the issue with gcc 3.2.3 on a 32 bit linux. The testsuite runs fine.

[schmir]

which machine architecture are you using?

[flub]

Odd, I was also using gcc 3.2.3 on 32-bit linux. Using python 2.7.2 but I didn't think that would matter.

I'm starting to think that my investigations into noinline where a red-herring. I've put the assembler code I get generated at http://paste.pocoo.org/show/543585/, might be interesting to compare to the assembler generated by your test platform.

[snaury]

Hi, you're hitting a gcc bug. :( Right after calling slp_save_state it destroys local variables, then calls slp_restore_state, which overwrites those variables, then it uses the frame pointer to restore registers saved on the stack and of course it results in their corruption.

Please compile with -Os which seemed to help in a similar case. I don't think there's anything wrong with switch code, the asm input constraints specifically tell gcc that those variables are used at the end of the function, but it does "addl $16, %esp" anyway. :(

[schmir]

rm -rf build; CFLAGS=-Os python setup.py build install does that or
rm -rf build; CFLAGS=-Os python run-tests.py to run the tests.

I can confirm that the test suite succeeds if compiled that way.

[schmir]

I've added a workaround in 199563f.
I've tested it on CentOS 3.9. Can you confirm that it works on RHEL 3?

[flub]

Passing -Os does indeed fix the problem. However the fix you committed does not quite work on my host since platform.platform() returns "Linux-2.4.21-63.EL-i686-athlon-with-redhat-3.9-Final". I changed it to the following if statement:

if ((sys.platform == "openbsd4" and os.uname()[-1] == "i386")
    or ("-with-redhat-3." in platform.platform()
        and platform.machine() == 'i686')):
os.environ["CFLAGS"] = ("%s %s" % (os.environ.get("CFLAGS", ""), "-Os")).lstrip()

Thanks for so quickly finding out what the issue is!

[schmir]

I've committed that in 01c9e8a

thanks flub for testing and snaury for confirming that this is a compiler bug!

[muggenhor]

I'm seeing exactly the same issue on Solaris 10 with a747442.

$ uname -a
SunOS ws200390 5.10 Generic_147440-12 sun4v sparc sun4v

Python 2.7 (r27:82500, Feb  8 2011, 11:39:07) 
[GCC 3.4.3] on sunos5
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys, os, platform
>>> import sys, os, platform
>>> sys.platform, os.uname, platform.platform(), platform.machine()
('sunos5', <built-in function uname>, 'SunOS-5.10-sun4v-sparc-32bit-ELF', 'sun4v')

Compiling with anything but -O3 (e.g. -Os, -O0, -O1, -O2) solves this.

A GDB backtrace shows the problem appears to reside in the invocation of PyThreadState_GET():

g_switchstack (target=0xffbfc584, args=<value optimized out>, kwargs=<value optimized out>) at greenlet.c:487
487                     PyThreadState* tstate = PyThreadState_GET();
(gdb) bt
#0  g_switchstack (target=0xffbfc584, args=<value optimized out>, kwargs=<value optimized out>) at greenlet.c:487
#1  g_switch (target=0xffbfc584, args=<value optimized out>, kwargs=<value optimized out>) at greenlet.c:570
#2  0xfeae3810 in g_switch (target=0x290da0, args=<value optimized out>, kwargs=<value optimized out>) at greenlet.c:576
#3  0xfeae474c in green_switch (self=0x290da0, args=0xfeaf6058, kwargs=0x0) at greenlet.c:1092

Using this patch prevents my problems:

diff --git a/setup.py b/setup.py
index bc03c14..e503574 100755
--- a/setup.py
+++ b/setup.py
@@ -6,7 +6,9 @@ import sys, os, glob, platform
 # https://bitbucket.org/ambroff/greenlet/issue/11/segfault-on-openbsd-i386
 # https://github.com/python-greenlet/greenlet/issues/4
 if ((sys.platform == "openbsd4" and os.uname()[-1] == "i386")
-    or ("-with-redhat-3." in platform.platform() and platform.machine() == 'i686')):
+    or ("-with-redhat-3." in platform.platform() and platform.machine() == 'i686')
+    or (sys.platform == "sunos5" and os.uname()[-1] == "sun4v")
+    or ("SunOS" in platform.platform() and platform.machine() == "sun4v")):
     os.environ["CFLAGS"] = ("%s %s" % (os.environ.get("CFLAGS", ""), "-Os")).lstrip()

 try:

[schmir]

I've applied that in 229624e

Thanks and please test that again!