Hi, 2011/11/11 Måns Rullgård <[email protected]>: > "Ronald S. Bultje" <[email protected]> writes: >> 2011/11/10 Måns Rullgård <[email protected]>: >>> "Ronald S. Bultje" <[email protected]> writes: >>>> 2011/11/10 Måns Rullgård <[email protected]>: >>>>> Sean McGovern <[email protected]> writes: >>>>>> valgrind is not fond of the pointer math in hScale_altivec_real(), >>>>> >>>>> This is a weird description. Valgrind never complains about pointer >>>>> maths as such, only about actual memory accesses. I'm guessing it has >>>>> to do with loading from an unaligned array, which will overread the end >>>>> up to a 16-byte aligned position. >>>> >>>> The array is aligned. Doesn't it just read 32 pixels at once, i.e. the >>>> FFALIGN(width, 16) should be changed to 32? >>> >>> The function is doing unaligned loads. Maybe it doesn't need to. >> >> I see. OK, patch is fine then. Maybe add a comment that the +16 is to >> account for buggy altivec overreads. > > They are not buggy. It's the only way to read unaligned data in > altivec, so unless you consider altivec itself buggy, the code is > correct provided it needs to handle unaligned buffers.
But it doesn't need to; convertData is guaranteed to be aligned. It was allocated using av_malloc(). Ronald _______________________________________________ libav-devel mailing list [email protected] https://lists.libav.org/mailman/listinfo/libav-devel
