On Mon, Nov 12, 2018 at 3:07 PM Eric Anholt <[email protected]> wrote: > > Matt Turner <[email protected]> writes: > > > Prior to this patch sizeof(linear_header) was 20 bytes in a > > non-debug build on 32-bit platforms. We do some pointer arithmetic to > > calculate the next available location with > > > > ptr = (linear_size_chunk *)((char *)&latest[1] + latest->offset); > > > > in linear_alloc_child(). The &latest[1] adds 20 bytes, so an allocation > > would only be 4-byte aligned. > > > > On 32-bit SPARC a 'sttw' instruction (which stores a consecutive pair of > > 4-byte registers to memory) requires an 8-byte aligned address. Such an > > instruction is used to store to an 8-byte integer type, like intmax_t > > which is used in glcpp's expression_value_t struct. > > > > As a result of the 4-byte alignment returned by linear_alloc_child() we > > would generate a SIGBUS (unaligned exception) on SPARC. > > > > According to the GNU libc manual malloc() always returns memory that has > > at least an alignment of 8-bytes [1]. I think our allocator should do > > the same. > > > > So, simple fix with two parts: > > > > (1) Increase SUBALLOC_ALIGNMENT to 8 unconditionally. > > (2) Mark linear_header with an aligned attribute, which will cause > > its sizeof to be rounded up to that alignment. (We already do > > this for ralloc_header) > > > > With this done, all Mesa's unit tests now pass on SPARC. > > > > [1] > > https://www.gnu.org/software/libc/manual/html_node/Aligned-Memory-Blocks.html > > > > Fixes: 47e17586924f ("glcpp: use the linear allocator for most objects") > > Bug: https://bugs.gentoo.org/636326 > > --- > > src/util/ralloc.c | 14 ++++++++++++-- > > 1 file changed, 12 insertions(+), 2 deletions(-) > > > > diff --git a/src/util/ralloc.c b/src/util/ralloc.c > > index 745b4cf1226..fc35661996d 100644 > > --- a/src/util/ralloc.c > > +++ b/src/util/ralloc.c > > @@ -552,10 +552,18 @@ ralloc_vasprintf_rewrite_tail(char **str, size_t > > *start, const char *fmt, > > */ > > > > #define MIN_LINEAR_BUFSIZE 2048 > > -#define SUBALLOC_ALIGNMENT sizeof(uintptr_t) > > +#define SUBALLOC_ALIGNMENT 8 > > #define LMAGIC 0x87b9c7d3 > > > > -struct linear_header { > > +struct > > +#ifdef _MSC_VER > > + __declspec(align(8)) > > +#elif defined(__LP64__) > > + __attribute__((aligned(16))) > > +#else > > + __attribute__((aligned(8))) > > +#endif > > + linear_header { > > #ifndef NDEBUG > > unsigned magic; /* for debugging */ > > #endif > > @@ -647,6 +655,8 @@ linear_alloc_child(void *parent, unsigned size) > > ptr = (linear_size_chunk *)((char*)&latest[1] + latest->offset); > > ptr->size = size; > > latest->offset += full_size; > > + > > + assert((uintptr_t)&ptr[1] % SUBALLOC_ALIGNMENT == 0); > > return &ptr[1]; > > } > > These patches are: > > Reviewed-by: Eric Anholt <[email protected]>
Thanks a bunch! I hope this is useful for you on arm as well. > However, shouldn't we also bump SUBALLOC_ALIGNMENT to 16 on LP64, too, > if that's what glibc is doing for malloc? 16-byte alignment is necessary for SSE aligned vector load/store instructions. I suppose we're not getting any vectorized SSE load/store instructions to memory allocated by linear_alloc_* and that's why we haven't seen problems? Seems reasonable to bump it to 16-bytes. _______________________________________________ mesa-dev mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/mesa-dev
