summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--src/compiler/nir/nir.h2
-rw-r--r--src/compiler/nir/nir_lower_alu_to_scalar.c42
-rw-r--r--src/gallium/drivers/freedreno/ir3/ir3_nir.c2
-rw-r--r--src/gallium/drivers/vc4/vc4_program.c2
-rw-r--r--src/mesa/drivers/dri/i965/brw_nir.c2
5 files changed, 30 insertions, 20 deletions
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 6f85bdb6bb6..3b3d6ae8393 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -2410,7 +2410,7 @@ bool nir_remove_dead_variables(nir_shader *shader, nir_variable_mode modes);
void nir_move_vec_src_uses_to_dest(nir_shader *shader);
bool nir_lower_vec_to_movs(nir_shader *shader);
-void nir_lower_alu_to_scalar(nir_shader *shader);
+bool nir_lower_alu_to_scalar(nir_shader *shader);
void nir_lower_load_const_to_scalar(nir_shader *shader);
void nir_lower_phis_to_scalar(nir_shader *shader);
diff --git a/src/compiler/nir/nir_lower_alu_to_scalar.c b/src/compiler/nir/nir_lower_alu_to_scalar.c
index a84fbdfd32b..fa18debd850 100644
--- a/src/compiler/nir/nir_lower_alu_to_scalar.c
+++ b/src/compiler/nir/nir_lower_alu_to_scalar.c
@@ -73,7 +73,7 @@ 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;
@@ -90,7 +90,7 @@ 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:
@@ -99,11 +99,11 @@ 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_pack_half_2x16:
if (!b->shader->options->lower_pack_half_2x16)
- return;
+ return false;
nir_ssa_def *val =
nir_pack_half_2x16_split(b, nir_channel(b, instr->src[0].src.ssa,
@@ -113,7 +113,7 @@ 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;
case nir_op_unpack_unorm_4x8:
case nir_op_unpack_snorm_4x8:
@@ -122,11 +122,11 @@ 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: {
if (!b->shader->options->lower_unpack_half_2x16)
- return;
+ return false;
nir_ssa_def *comps[2];
comps[0] = nir_unpack_half_2x16_split_x(b, instr->src[0].src.ssa);
@@ -135,7 +135,7 @@ lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec));
nir_instr_remove(&instr->instr);
- return;
+ return true;
}
case nir_op_pack_uvec2_to_uint: {
@@ -185,11 +185,11 @@ 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;
}
case nir_op_unpack_double_2x32:
- return;
+ return false;
LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd);
LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand);
@@ -204,7 +204,7 @@ 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 };
@@ -240,30 +240,40 @@ 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 void
+static bool
nir_lower_alu_to_scalar_impl(nir_function_impl *impl)
{
nir_builder builder;
nir_builder_init(&builder, impl);
+ bool progress = false;
nir_foreach_block(block, impl) {
nir_foreach_instr_safe(instr, block) {
- if (instr->type == nir_instr_type_alu)
- lower_alu_instr_scalar(nir_instr_as_alu(instr), &builder);
+ if (instr->type == nir_instr_type_alu) {
+ progress = lower_alu_instr_scalar(nir_instr_as_alu(instr),
+ &builder) || progress;
+ }
}
}
nir_metadata_preserve(impl, nir_metadata_block_index |
nir_metadata_dominance);
+
+ return progress;
}
-void
+bool
nir_lower_alu_to_scalar(nir_shader *shader)
{
+ bool progress = false;
+
nir_foreach_function(function, shader) {
if (function->impl)
- nir_lower_alu_to_scalar_impl(function->impl);
+ progress = nir_lower_alu_to_scalar_impl(function->impl) || progress;
}
+
+ return progress;
}
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_nir.c b/src/gallium/drivers/freedreno/ir3/ir3_nir.c
index 023da3be292..25262224f98 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_nir.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_nir.c
@@ -90,7 +90,7 @@ ir3_optimize_loop(nir_shader *s)
progress = false;
OPT_V(s, nir_lower_vars_to_ssa);
- OPT_V(s, nir_lower_alu_to_scalar);
+ progress |= OPT(s, nir_lower_alu_to_scalar);
OPT_V(s, nir_lower_phis_to_scalar);
progress |= OPT(s, nir_copy_prop);
diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c
index 52b938d066e..24e4699b92e 100644
--- a/src/gallium/drivers/vc4/vc4_program.c
+++ b/src/gallium/drivers/vc4/vc4_program.c
@@ -1423,7 +1423,7 @@ vc4_optimize_nir(struct nir_shader *s)
progress = false;
NIR_PASS_V(s, nir_lower_vars_to_ssa);
- NIR_PASS_V(s, nir_lower_alu_to_scalar);
+ NIR_PASS(progress, s, nir_lower_alu_to_scalar);
NIR_PASS_V(s, nir_lower_phis_to_scalar);
NIR_PASS(progress, s, nir_copy_prop);
diff --git a/src/mesa/drivers/dri/i965/brw_nir.c b/src/mesa/drivers/dri/i965/brw_nir.c
index e8dafaebc91..27be201db6f 100644
--- a/src/mesa/drivers/dri/i965/brw_nir.c
+++ b/src/mesa/drivers/dri/i965/brw_nir.c
@@ -375,7 +375,7 @@ nir_optimize(nir_shader *nir, bool is_scalar)
OPT_V(nir_lower_vars_to_ssa);
if (is_scalar) {
- OPT_V(nir_lower_alu_to_scalar);
+ OPT(nir_lower_alu_to_scalar);
}
OPT(nir_copy_prop);