---
src/glsl/nir/nir_lower_alu_to_scalar.c | 56 ++++++++++++++++++++++------------
1 file changed, 37 insertions(+), 19 deletions(-)
diff --git a/src/glsl/nir/nir_lower_alu_to_scalar.c
b/src/glsl/nir/nir_lower_alu_to_scalar.c
index 9313fc0..e4d247f 100644
--- a/src/glsl/nir/nir_lower_alu_to_scalar.c
+++ b/src/glsl/nir/nir_lower_alu_to_scalar.c
@@ -30,6 +30,11 @@
* arguments with individual per-channel operations.
*/
+struct alu_to_scalar_state {
+ nir_builder builder;
+ bool progress;
+};
+
static void
nir_alu_ssa_dest_init(nir_alu_instr *instr, unsigned num_components)
{
@@ -70,7 +75,7 @@ lower_reduction(nir_alu_instr *instr, nir_op chan_op, nir_op
merge_op,
nir_instr_remove(&instr->instr);
}
-static void
+static bool
lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
{
unsigned num_src = nir_op_infos[instr->op].num_inputs;
@@ -86,7 +91,7 @@ lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
case name##3: \
case name##4: \
lower_reduction(instr, chan, merge, b); \
- return;
+ return true;
switch (instr->op) {
case nir_op_vec4:
@@ -95,7 +100,7 @@ lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
/* We don't need to scalarize these ops, they're the ones generated to
* group up outputs into a value that can be SSAed.
*/
- return;
+ return false;
case nir_op_unpack_unorm_4x8:
case nir_op_unpack_snorm_4x8:
@@ -104,13 +109,13 @@ lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder
*b)
/* There is no scalar version of these ops, unless we were to break it
* down to bitshifts and math (which is definitely not intended).
*/
- return;
+ return false;
case nir_op_unpack_half_2x16:
/* We could split this into unpack_half_2x16_split_[xy], but should
* we?
*/
- return;
+ return false;
case nir_op_fdph: {
nir_ssa_def *sum[4];
@@ -127,7 +132,7 @@ lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
nir_instr_remove(&instr->instr);
- return;
+ return true;
}
LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd);
@@ -147,7 +152,7 @@ lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
}
if (instr->dest.dest.ssa.num_components == 1)
- return;
+ return false;
unsigned num_components = instr->dest.dest.ssa.num_components;
nir_ssa_def *comps[] = { NULL, NULL, NULL, NULL };
@@ -182,33 +187,46 @@ lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder
*b)
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec));
nir_instr_remove(&instr->instr);
+
+ return true;
}
static bool
-lower_alu_to_scalar_block(nir_block *block, void *builder)
+lower_alu_to_scalar_block(nir_block *block, void *void_state)
{
+ struct alu_to_scalar_state *state = void_state;
+
nir_foreach_instr_safe(block, instr) {
- if (instr->type == nir_instr_type_alu)
- lower_alu_instr_scalar(nir_instr_as_alu(instr), builder);
+ if (instr->type == nir_instr_type_alu) {
+ if (lower_alu_instr_scalar(nir_instr_as_alu(instr), &state->builder))
+ state->progress = true;
+ }
}
return true;
}
-static void
-nir_lower_alu_to_scalar_impl(nir_function_impl *impl)
+static bool
+nir_lower_alu_to_scalar_impl(nir_function_impl *impl, void *unused)
{
- nir_builder builder;
- nir_builder_init(&builder, impl);
+ struct alu_to_scalar_state state;
+
+ nir_builder_init(&state.builder, impl);
+ state.progress = false;
- nir_foreach_block(impl, lower_alu_to_scalar_block, &builder);
+ nir_foreach_block(impl, lower_alu_to_scalar_block, &state);
+
+ return state.progress;
}
+static const nir_pass nir_lower_alu_to_scalar_pass = {
+ NULL, /* shader_pass_func */
+ nir_lower_alu_to_scalar_impl,
+ nir_metadata_block_index | nir_metadata_dominance,
+};
+
void
nir_lower_alu_to_scalar(nir_shader *shader)
{
- nir_foreach_overload(shader, overload) {
- if (overload->impl)
- nir_lower_alu_to_scalar_impl(overload->impl);
- }
+ nir_shader_run_pass(shader, &nir_lower_alu_to_scalar_pass);
}
--
2.5.0.400.gff86faf
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
