[Mesa-dev] [PATCH 15/47] i965/fs: Retype 16-bit/stride2 movs to UD on nir_op_vecX
Alejandro PiƱeiro
apinheiro at igalia.com
Thu Aug 24 13:54:30 UTC 2017
Due to 32-bit alignment, we are fully using those registers even if
the type is 16-bit.
This change allows copy propagation to work on those registers without
making the optimization pass too complicated.
---
src/intel/compiler/brw_fs_nir.cpp | 19 +++++++++++++++++++
1 file changed, 19 insertions(+)
diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index c469baf42a1..eadd49ff13a 100644
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -635,6 +635,20 @@ fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr)
}
}
+ /* For 16-bit data, we are using the full 32-bit register, due alignment
+ * requirements. So it is safe to retype a 16-bit to 16-bit raw mov to
+ * UD, and that would make these movs easier to be optimized during copy
+ * propagation later.
+ */
+ bool retype_to_ud = (nir_dest_bit_size(instr->dest.dest) == 16 &&
+ nir_src_bit_size(instr->src[0].src) == 16 &&
+ temp.stride == 2 &&
+ op[0].stride == 2);
+ if (retype_to_ud) {
+ temp.type = BRW_REGISTER_TYPE_UD;
+ temp.stride = 1;
+ }
+
for (unsigned i = 0; i < 4; i++) {
if (!(instr->dest.write_mask & (1 << i)))
continue;
@@ -643,6 +657,11 @@ fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr)
inst = bld.MOV(offset(temp, bld, i),
offset(op[0], bld, instr->src[0].swizzle[i]));
} else {
+ if (retype_to_ud) {
+ op[i].type = BRW_REGISTER_TYPE_UD;
+ op[i].stride = 1;
+ }
+
inst = bld.MOV(offset(temp, bld, i),
offset(op[i], bld, instr->src[i].swizzle[0]));
}
--
2.11.0
More information about the mesa-dev
mailing list