[llvm-commits] [llvm] r37843 - /llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

[Evan Cheng]

Hi Dan,

I've committed this:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of- 
Mon-20070702/051079.html

as a workaround. Please figure out a proper fix. Thanks!

Evan

On Jul 5, 2007, at 5:52 PM, Evan Cheng wrote:

> Hi Dan,
>
> Ok, this is the problem:
>
>          li r4, 124
>          li r30, 0
>          mr r2, r30
>          mr r29, r30
> LBB1_1: ;bb
>          li r3, lo16(LCPI1_0)
>          rlwinm r5, r4, 3, 29, 29
>          addis r3, r3, ha16(LCPI1_0)
>          stw r4, 64(r1)
>          stw r30, 68(r1)
>
> The last two instructions are storing the wrong values. They should be
>          stw r30, 64(r1)
>          stw r4, 68(r1)
>
> Please take a look at this asap. Thanks.
>
> Evan
>
> On Jul 5, 2007, at 4:57 PM, Evan Cheng wrote:
>
>> Here is a much simpler test case:
>>
>> SingleSource/UnitTests/2005-05-12-Int64ToFP.c:
>>
>> I am still trying to figure out what exactly is wrong.
>>
>> Evan
>>
>> On Jul 5, 2007, at 4:33 PM, Bill Wendling wrote:
>>
>>> On 7/5/07, Evan Cheng <evan.cheng at apple.com> wrote:
>>>>
>>>> On Jul 5, 2007, at 7:49 AM, Dan Gohman wrote:
>>>>
>>>>> Hi Evan,
>>>>>
>>>>> I'm currently testing the attached patch, which seems to fix the
>>>>> problems for PPC. It extends getCopyToParts and getCopyFromParts
>>>>> to have an extra parameter to indicate when endian-swapping is
>>>>> needed. I don't have access to a PPC test environment; I've
>>>>> included
>>>>> a few trivial regression tests in this patch; I'll do some more
>>>>> soon.
>>>>
>>>> Thanks. I'll do some testing.
>>>>
>>>>>
>>>>> Copying to virtual registers is always done in little-endian  
>>>>> order,
>>>>> while copying to physical registers, arguments, or return values,
>>>>> is done in target-endian order. I wonder if it would make sense to
>>>>> change to using target-endian order for virtual registers as well?
>>>>
>>>> I don't think it matters, does it? If it makes the code cleaner  
>>>> then
>>>> please do so.
>>>>
>>> FWIW, the PPC bootstrap is still failing with this patch. It's doing
>>> so at an earlier stage now. Here's part of the log file that shows
>>> the
>>> failure:
>>>
>>> # When building multilibbed target libraries, all the required
>>> # When building multilibbed target libraries, all the required
>>> # libraries are expected to exist in the multilib directory.
>>> MLIBS=` ./xgcc -B./
>>> -B/usr/local/llvm-gcc-install/powerpc-apple-darwin9/bin/ -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/include -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/sys-include
>>> -L/Volumes/SandBox/Clean/llvm-1040-01.roots/llvm-1040-01~obj/obj-
>>> powerpc-powerpc/gcc/../ld
>>> --print-multi-lib \
>>> 		| sed -e 's/;.*$//' -e '/^\.$/d'` ; \
>>> 	for mlib in $MLIBS ; do \
>>> 	  rm -f ${mlib}/libgcc_s.10.4.dylib || exit 1 ; \
>>> 	  ln -s ../libgcc_s.10.4.dylib ${mlib}/libgcc_s.10.4.dylib || exit
>>> 1 ; \
>>> 	done
>>> # libraries are expected to exist in the multilib directory.
>>> MLIBS=` ./xgcc -B./
>>> -B/usr/local/llvm-gcc-install/powerpc-apple-darwin9/bin/ -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/include -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/sys-include
>>> -L/Volumes/SandBox/Clean/llvm-1040-01.roots/llvm-1040-01~obj/obj-
>>> powerpc-powerpc/gcc/../ld
>>> --print-multi-lib \
>>> 		| sed -e 's/;.*$//' -e '/^\.$/d'` ; \
>>> 	for mlib in $MLIBS ; do \
>>> 	  rm -f ${mlib}/libgcc_s.10.5.dylib || exit 1 ; \
>>> 	  ln -s ../libgcc_s.10.5.dylib ${mlib}/libgcc_s.10.5.dylib || exit
>>> 1 ; \
>>> 	done
>>> MLIBS=` ./xgcc -B./
>>> -B/usr/local/llvm-gcc-install/powerpc-apple-darwin9/bin/ -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/include -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/sys-include
>>> -L/Volumes/SandBox/Clean/llvm-1040-01.roots/llvm-1040-01~obj/obj-
>>> powerpc-powerpc/gcc/../ld
>>> --print-multi-lib \
>>> 		| sed -e 's/;.*$//' -e '/^\.$/d'` ; \
>>> 	for mlib in '' $MLIBS ; do \
>>> 	  strip -o libgcc_s.10.5.dylib_T${mlib} \
>>> 	    -s /Volumes/SandBox/Clean/llvm-1040-01.roots/llvm-1040-01~obj/
>>> src/gcc/config/rs6000/darwin-libgcc.10.5.ver
>>> -c -u \
>>> 	    ./${mlib}/libgcc_s.1.dylib.tmp || exit 1 ; \
>>> 	done
>>> MLIBS=` ./xgcc -B./
>>> -B/usr/local/llvm-gcc-install/powerpc-apple-darwin9/bin/ -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/include -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/sys-include
>>> -L/Volumes/SandBox/Clean/llvm-1040-01.roots/llvm-1040-01~obj/obj-
>>> powerpc-powerpc/gcc/../ld
>>> --print-multi-lib \
>>> 		| sed -e 's/;.*$//' -e '/^\.$/d'` ; \
>>> 	for mlib in '' $MLIBS ; do \
>>> 	  strip -o libgcc_s.10.4.dylib_T${mlib} \
>>> 	    -s /Volumes/SandBox/Clean/llvm-1040-01.roots/llvm-1040-01~obj/
>>> src/gcc/config/rs6000/darwin-libgcc.10.4.ver
>>> -c -u \
>>> 	    ./${mlib}/libgcc_s.1.dylib.tmp || exit 1 ; \
>>> 	done
>>> lipo -output libgcc_s.10.4.dylib -create libgcc_s.10.4.dylib_T*
>>> lipo -output libgcc_s.10.5.dylib -create libgcc_s.10.5.dylib_T*
>>> rm libgcc_s.10.5.dylib_T*
>>> rm libgcc_s.10.4.dylib_T*
>>> # When building multilibbed target libraries, all the required
>>> # libraries are expected to exist in the multilib directory.
>>> MLIBS=` ./xgcc -B./
>>> -B/usr/local/llvm-gcc-install/powerpc-apple-darwin9/bin/ -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/include -isystem
>>> /usr/local/llvm-gcc-install/powerpc-apple-darwin9/sys-include
>>> -L/Volumes/SandBox/Clean/llvm-1040-01.roots/llvm-1040-01~obj/obj-
>>> powerpc-powerpc/gcc/../ld
>>> --print-multi-lib \
>>> 		| sed -e 's/;.*$//' -e '/^\.$/d'` ; \
>>> 	for mlib in '' $MLIBS ; do \
>>> 	  cp ./${mlib}/libgcc_s.1.dylib.tmp \
>>> 	    ./libgcc_s.1.dylib_T_${mlib} || exit 1 ; \
>>> 	done
>>> lipo -output libgcc_s.1.dylib \
>>> 	  -create libgcc_s.1.dylib_T*
>>> rm libgcc_s.1.dylib_T*
>>> make[3]: *** [libgcc_s.1.dylib] Segmentation fault
>>> make[3]: *** Deleting file `libgcc_s.1.dylib'
>>> rm cpp.pod gfdl.pod fsf-funding.pod gcov.pod gcc.pod gpl.pod
>>> make[2]: *** [stage1_build] Error 2
>>> make[1]: *** [bootstrap] Error 2
>>> + exit 1
>>> make: *** [install] Error 1
>>>
>>> -bw
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits