Re: [Bug-gsl] SV decomp failure

[Steve Brosnan]

Hi Rhys,

Well, I took a stab at it. But the results are confusing. 

First, some info I've learned since my last message: 
(1) The compiler symbolically linked by 'gcc' on the Mac is 'llvm-gcc-4.2', not 
gnu gcc. The llvm version is supposed to be interoperable, but that may not be 
fully true, particularly where optimization is concerned.
(2) On Lion (OS X 10.7.4) the default (preferred?) compiler in Xcode is the 
'Apple LLVM Compiler 4.0'. I think this is, or is related to, the Clang 
compiler, and is the direction that Apple is headed in.
(3) Using the LLVM Compiler 4.0, with -O2, the SV-decomp tests succeed.
(4) As reported before, the tests also succeed with the llvm-gcc-4.2 compiler 
with -Os optimization selected. HOWEVER, my prior statement that the SV-decomp 
tests fail if I 'export CFLAGS="-Os -g"' before the gsl configure/make/make 
check sequence using the gsl build mechanism is WRONG. -Os SUCCEEDS. (I 
discovered my error, probably a dumb typo in the prior test, after doing the 
tests described in the following paragraph.)

That said, I tried to do as you requested and tabulate the SV-decomp test 
results as the optimization switches are added within Xcode. The compiler was 
llvm-gcc-4.2 in all cases, with -O1 selected. Under the 'Other C flags' heading 
in Xcode, I added the following switches one by one (the gcc man page says that 
they are what separate -O1 from -O2):
        -fthread-jumps -fcrossjumping -foptimize-sibling-calls 
-fcse-follow-jumps  -fcse-skip-blocks -fgcse  -fgcse-lm 
        -fexpensive-optimizations -frerun-cse-after-loop -fcaller-saves 
-fpeephole2 -fschedule-insns  -fschedule-insns2 
        -fsched-interblock  -fsched-spec -fregmove -fstrict-aliasing 
-fstrict-overflow
        -fdelete-null-pointer-checks -freorder-blocks  -freorder-functions 
-falign-functions  -falign-jumps -falign-loops 
        -falign-labels -ftree-vrp -ftree-pre
However, the tests never failed. -O1 succeeded. -O2 failed. But adding all the 
switches that separate -O1 and -O2 also  succeeded! I'm baffled. If anyone can 
educate me, please do.

As far as options of how to proceed to a solution, here are my, as a gsl 
noobie, suggestions:
(1) Accept -O1 as the baseline and modify the makefiles accordingly.
(2) Modify 'configure' to use the -Os optimization for Macs running OS X 10.7.4 
and later. Is this hard?

But in the meantime, Mac people can use the 'export CFLAGS="-Os -g"' trick.

Hope this helps. If you have other suggestions, I will do the best I can.

Regards,

Steve

On Sep 23, 2012, at 12:49 PM, Rhys Ulerich wrote:

> Hi Steve,
> 
> Any chance you'd be willing to dig into the gcc optimization delta
> between -O1 and -O2 to determine exactly which optimization causes the
> SVD failure on recent OS X systems?  You'd have to dig up a
> version-specific set of GCC documentation, but the details should look
> something like http://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html.
> 
> It would be nice to either a) report the problem upstream and/or b)
> disable this specific optimization at the smallest possible scope when
> detecting OS X along with issuing a compiler warning.
> 
> - Rhys