This makes sure that unspills restore the exact contents of the
variable in scratch space into the GRF without applying channel
masking, which is incorrect under control flow for things like message
headers or vectors of heterogeneous types that don't properly respect
channel boundaries.
---
src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp | 12 ++++++++++--
1 file changed, 10 insertions(+), 2 deletions(-)
diff --git a/src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp
b/src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp
index 87ade2d..09f7876 100644
--- a/src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp
@@ -944,8 +944,16 @@ fs_visitor::spill_reg(int spill_reg)
const unsigned width =
dispatch_width == 16 && regs_read % 2 == 0 ? 16 : 8;
- emit_unspill(ibld.group(width, 0), unspill_dst,
- subset_spill_offset, regs_read);
+
+ /* Set exec_all() on unspill messages under the (rather
+ * pessimistic) assumption that there is no one-to-one
+ * correspondence between channels of the spilled variable in
+ * scratch space and the scratch read message, which operates on
+ * 32 bit channels. It shouldn't hurt in any case because the
+ * unspill destination is a block-local temporary.
+ */
+ emit_unspill(ibld.exec_all().group(width, 0),
+ unspill_dst, subset_spill_offset, regs_read);
}
}
--
2.7.3
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
