On Mon, Jul 25, 2011 at 11:10 AM, Ulrich Weigand <uweig...@de.ibm.com> wrote: > Richard Guenther wrote: >> On Sun, Jul 24, 2011 at 2:02 PM, Ira Rosen <ira.ro...@linaro.org> wrote: >> > On 21 July 2011 15:19, Ira Rosen <ira.ro...@linaro.org> wrote: >> >> I reproduced the failure. It occurs without Richard's >> >> (http://gcc.gnu.org/ml/gcc-patches/2011-07/msg01022.html) and this >> >> patches too. Obviously the vectorized loop is executed, but at the >> >> moment I don't understand why. I'll have a better look on Sunday. >> > >> > Actually it doesn't choose the vectorized code. But the scalar version >> > gets optimized in a harmful way for SPU, AFAIU. >> > Here is the scalar loop after vrp2 >> > >> > <bb 8>: >> > # ivtmp.42_50 = PHI <ivtmp.42_59(3), ivtmp.42_45(10)> >> > D.4593_42 = (void *) ivtmp.53_32; >> > D.4520_33 = MEM[base: D.4593_42, offset: 0B]; >> > D.4521_34 = D.4520_33 + 1; >> > MEM[symbol: a, index: ivtmp.42_50, offset: 0B] = D.4521_34; >> > ivtmp.42_45 = ivtmp.42_50 + 4; >> > if (ivtmp.42_45 != 16) >> > goto <bb 10>; >> > else >> > goto <bb 5>; >> > >> > and the load is changed by dom2 to: >> > >> > <bb 4>: >> > ... >> > D.4520_33 = MEM[base: vect_pa.9_19, offset: 0B]; >> > ... >> > >> > where vector(4) int * vect_pa.9; >> > >> > And the scalar loop has no rotate for that load: >> >> Hum. This smells like we are hiding sth from the tree optimizers? > > Well, the back-end assumes a pointer to vector type is always > naturally aligned, and therefore the data it points to can be > accessed via a simple load, with no extra rotate needed.
I can't see any use of VECTOR_TYPE in config/spu/, and assuming anything about alignment just because of the kind of the pointer is bogus - the scalar code does a scalar read using that pointer. So the backend better should look at the memory operation, not at the pointer type. That said, I can't find any code that looks suspicious in the spu backend. > It seems what happened here is that somehow, a pointer to int > gets replaced by a pointer to vector, even though their alignment > properties are different. No, they are not. They get replaced if they are value-equivalent in which case they are also alignment-equivalent. But well, the dump snippet wasn't complete and I don't feel like building a SPU cross to verify myself. > This vector pointer must originate somehow in the vectorizer, > however, since the original C source does not contain any > vector types at all ... That's for sure true, it must be the initial pointer we then increment in the vectorized loop. Richard. > Bye, > Ulrich > > -- > Dr. Ulrich Weigand > GNU Toolchain for Linux on System z and Cell BE > ulrich.weig...@de.ibm.com >
