Re: ICE on ARM

[<pl...@pi...>]

On 04/05/10 07:24, Ethan Merritt wrote:
> On Monday 03 May 2010, pl...@pi... wrote:
>> Hi,
>>
>> I'm trying to cross-compile gnuplot with maverick crunch fpu on ARM.
>>
>> I have it building without the fpu code generation but it falls at the
>> first hurdle if I use the FPU.
>>
>> CFLAGS="-O2 -mcpu=ep9312 -mfpu=maverick -mfloat-abi=softfp"
>> CXXFLAGS="-O2 -mcpu=ep9312 " ./configure --host=arm --prefix=/rar/root
>> --without-x --without-pdf --without-cairo --without-documentation
>> --disable-wxwidgets --without-lua --without-tutorial
>> --disable-volatile-data  --disable-objects
>>
>> make
>
> [...]
>> arm-maverick-linux-gnueabi-gcc -DHAVE_CONFIG_H -I. -I..  -I../term
>> -I../term -DBINDIR=\"/rar/root/bin\"
>> -DX11_DRIVER_DIR=\"/rar/root/libexec/gnuplot/4.3\"
>                                                ^^^^^
> So your source must be about a year old?  Without a precise date on the source
> it's hard to comment on what pitfalls may lie in that routine.  On the other
> hand, your compiler failure seems to come from the last line of the routine
> so I'm not sure that more precise line numbers would help.
>
>> -DGNUPLOT_PS_DIR=\"/rar/root/share/gnuplot/4.3/PostScript\"
>> -DGNUPLOT_JS_DIR=\"/rar/root/share/gnuplot/4.3/js\"
>> -DCONTACT=\"gnu...@li...\"
>> -DHELPFILE=\"/rar/root/share/gnuplot/4.3/gnuplot.gih\"
>> -DGNUPLOT_X11=\"`echo gnuplot_x11 | sed 's,x,x,'`\"   -O2 -mcpu=ep9312
>> -mfpu=maverick -mfloat-abi=softfp -MT axis.o -MD -MP -MF .deps/axis.Tpo
>> -c -o axis.o axis.c
>> axis.c: In function 'gen_tics':
>> axis.c:1163: internal compiler error: Segmentation fault
>> Please submit a full bug report,
>>
>>
>> gen_tics seems to make heavy use of floats , is there any areas that may
>> be a bit flakey on an unusual arch?
>
> Have you tried with -O0 ?
>
>> Clearly this is a major problem with the compiler itself but I'm
>> wondering if there's anything likely to trip it up.
>
> The code tests for loss-of-precision errors that "shouldn't happen".
> I can imagine that this might confuse a software fp library.
> Somewhere around line 1068 you may see a code section that begins
>     if (1) /* (some-test-for-range-and-or-step-size) */
> You could try setting that to  if (0) instead, thus disabling the
> tests for  (x + fabs(delta)<= x).
>
>

Thanks very much Ethan.

-O0 does compile! That at least  means I can get eabi crunch working.
Os O1 and O2 all fail in exactly the same way.

Those ideas enabled me to track down where it's happening. The for-loop 
even with the guts ripped out causes the segfault.

		/* {{{  process minitics */
		double mplace, mtic;
mtic=0;
/*
		for (mplace = ministart; mplace < miniend; mplace += ministep) {
mtic=0;
		}
*/
		/* }}} */

The above compiles , with the for-loop it segs.

It looks pretty innocuous on the face of it but the compiler spits the 
dummy.

There are more if I continue the build. graphics:606
     if (lkey) {
   TBOOLEAN key_panic = FALSE;



BTW this is gcc arm eabi with maverick-crunch FPU support enabled, not 
software fp emulation.  -mfloat-abi=softfp does not mean what one may 
expect ;)

I'll test it built with -O0 and FPU, hopefully it will still be faster 
than software fp.

best regards.