Fix a segmentation fault in multi-thread code
The segmentation fault happened while using arch=x86-linux-gcc, and
-DSANITIZE=integer. It crashed at
memcpy(&lf_data->planes[i].dst, &pd[i].dst,
sizeof(lf_data->planes[i].dst));
That is in loop_filter_data_reset() function.
It seemed that MOVAPS was used in memcpy(), which required that the
memory operand must be aligned on a 16-byte boundary.
First fix(patch set 1): Aligned lf_sync->lfdata to 16 bytes during
memory allocation.
Second fix(patch set 2): The above fix was not quite robust. Changed
to copy each element instead of calling memcpy. (Note: we tried
"lf_data->planes[i].dst = pd[i].dst;", that didn't work either.)
BUG=aomedia:1856
BUG=aomedia:1859
Change-Id: I2d2d4bae1c1546ed7b7e61bafea2334124079719
diff --git a/av1/common/thread_common.c b/av1/common/thread_common.c
index d3ce1d4..3884954 100644
--- a/av1/common/thread_common.c
+++ b/av1/common/thread_common.c
@@ -124,7 +124,18 @@
lf_data->cm = cm;
lf_data->xd = xd;
for (int i = 0; i < MAX_MB_PLANE; i++) {
- memcpy(&lf_data->planes[i].dst, &pd[i].dst, sizeof(lf_data->planes[i].dst));
+ // TODO(yunqing): Copying dst struct by "lf_data->planes[i].dst =
+ // pd[i].dst;" or "memcpy(&lf_data->planes[i].dst, &pd[i].dst,
+ // sizeof(lf_data->planes[i].dst));" would cause a segmentation fault while
+ // using arch=x86-linux-gcc, and -DSANITIZE=integer. It seems that MOVAPS
+ // is used in memcpy(), which requires that the memory operand to be
+ // aligned on a 16-byte boundary. This copy needs to be written efficiently
+ // once Clang fixes this issue.
+ lf_data->planes[i].dst.buf = pd[i].dst.buf;
+ lf_data->planes[i].dst.buf0 = pd[i].dst.buf0;
+ lf_data->planes[i].dst.width = pd[i].dst.width;
+ lf_data->planes[i].dst.height = pd[i].dst.height;
+ lf_data->planes[i].dst.stride = pd[i].dst.stride;
lf_data->planes[i].subsampling_x = pd[i].subsampling_x;
lf_data->planes[i].subsampling_y = pd[i].subsampling_y;
}
