Use MRF for 1-Oword read messages again to avoid problems when
sending scratch read messages. We cannot reuse the destination as the DF
scratch reads on IVB are splitted in several instructions and that could
end up having invalid data.
Signed-off-by: Samuel Iglesias Gonsálvez <sigles...@igalia.com>
---
src/intel/compiler/brw_eu_emit.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/src/intel/compiler/brw_eu_emit.c b/src/intel/compiler/brw_eu_emit.c
index bd6f46c776..fa6dc0d5ff 100644
--- a/src/intel/compiler/brw_eu_emit.c
+++ b/src/intel/compiler/brw_eu_emit.c
@@ -2267,7 +2267,9 @@ brw_oword_block_read_scratch(struct brw_codegen *p,
if (devinfo->gen >= 6)
offset /= 16;
- if (p->devinfo->gen >= 7) {
+ if (p->devinfo->gen >= 7 &&
+ (p->devinfo->gen > 7 || p->devinfo->is_haswell ||
+ type_sz(dest.type) != 8)) {
/* On gen 7 and above, we no longer have message registers and we can
* send from any register we want. By using the destination register
* for the message, we guarantee that the implied message write won't
--
2.11.0
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
