[Pixman] Test suite failures on 32-bit x86?

[Matt Turner]

Building the 0.32.2 release and from git with

CC="gcc -m32" ./autogen.sh && make check

PASS: prng-test
PASS: a1-trap-test
PASS: region-translate-test
PASS: pdf-op-test
PASS: region-test
PASS: fetch-test
../test-driver: line 95:  3312 Segmentation fault      "$@" > $log_file 2>&1
FAIL: rotate-test
PASS: oob-test
PASS: infinite-loop
PASS: combiner-test
PASS: pixel-test
PASS: trap-crasher
PASS: alpha-loop
PASS: thread-test
PASS: scaling-helpers-test
PASS: scaling-crash-test
../test-driver: line 95:  3571 Segmentation fault      "$@" > $log_file 2>&1
FAIL: matrix-test
PASS: gradient-crash-test
../test-driver: line 95:  3637 Segmentation fault      "$@" > $log_file 2>&1
FAIL: blitters-test
../test-driver: line 95:  3659 Segmentation fault      "$@" > $log_file 2>&1
FAIL: glyph-test
../test-driver: line 95:  3681 Segmentation fault      "$@" > $log_file 2>&1
FAIL: scaling-test
../test-driver: line 95:  3703 Segmentation fault      "$@" > $log_file 2>&1
FAIL: affine-test
PASS: alphamap
PASS: composite-traps-test
PASS: region-contains-test
PASS: stress-test
PASS: composite

Manually running the tests shows that they all crash in
prng_rand_128_r (utils-prng.h:138):

> uint32x4 e = x->a - ((x->b << 27) + (x->b >> (32 - 27)));

which is code inside an #ifdef GCC_VECTOR_EXTENSIONS_SUPPORTED block.

I realize this may be a gcc bug, so I tested with 4.8.1 and 4.7.2 and
got the same results. Testing with 4.6.3 leads to only a single
failure, in matrix-test (with a different backtrace, so probably
different).

Do we need some kind of configure check to make sure that our use of
gcc's vector extensions is actually going to work?