It should be possible to write a library function that accepts vector
arguments in a variable-length argument list, with that function
compiled with -mno-altivec, the caller compiled with -maltivec, and
both using the same Altivec ABI. With mainline this works for -m64 but
not with -m32. The hammer-3.3 branch has the same problem; I haven't
tried it for other branches but assume that 3.4 has does the same thing.
The tarball includes two source files plus a Makefile that will build
them in different combinations. The specific problem described here
is demonstrated by the following, using a biarch powerpc64-linux
compiler from mainline sources. Results for this test are not affected
by patches from Alan Modra on April 23 or from Fariborz Jahanian on
dufur% $GCC -m32 -maltivec -mabi=altivec -c -o bug_passer.o bug_passer.c
dufur% $GCC -m32 -mno-altivec -mabi=altivec -c -o bug_passee.o bug_passee.c
dufur% $GCC -m32 bug_passer.o bug_passee.o
If other failing combinations generated by the Makefile are not supported
then they should be documented as not expected to work.
Created attachment 6154 [details]
test case: two source files plus Makefile
Confirmed. -m32 vector varargs looks badly broken.
prototype patch at http://gcc.gnu.org/ml/gcc/2004-04/msg01316.html
Fixed on mainline and hammer-3_3-branch. Patches listed in