freedreno/ir3: split ubo analysis/lowering passes

Since binning pass variants share the same const_state with their
draw-pass counterpart, we should re-use the draw-pass variant's ubo
range analysis.  So split the two functions of the existing pass
into two parts.

Signed-off-by: Rob Clark <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5526>

4 files changed, 31 insertions, 12 deletions

diff --git a/src/freedreno/ir3/ir3_nir.c b/src/freedreno/ir3/ir3_nir.c
index 93b8012e1d5..b1f5fb9669c 100644
--- a/src/freedreno/ir3/ir3_nir.c
+++ b/src/freedreno/ir3/ir3_nir.c
@@ -344,7 +344,10 @@ ir3_nir_lower_variant(struct ir3_shader_variant *so, nir_shader *s)
 		progress |= OPT(s, nir_lower_tex, &tex_options);
 	}
 
-	progress |= OPT(s, ir3_nir_analyze_ubo_ranges, so);
+	if (!so->binning_pass)
+		OPT_V(s, ir3_nir_analyze_ubo_ranges, so);
+
+	progress |= OPT(s, ir3_nir_lower_ubo_loads, so);
 
 	/* UBO offset lowering has to come after we've decided what will
 	 * be left as load_ubo
diff --git a/src/freedreno/ir3/ir3_nir.h b/src/freedreno/ir3/ir3_nir.h
index 915f1638419..b84525bd69f 100644
--- a/src/freedreno/ir3/ir3_nir.h
+++ b/src/freedreno/ir3/ir3_nir.h
@@ -57,7 +57,8 @@ void ir3_nir_lower_variant(struct ir3_shader_variant *so, nir_shader *s);
 
 void ir3_setup_const_state(nir_shader *nir, struct ir3_shader_variant *v,
 		struct ir3_const_state *const_state);
-bool ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader_variant *v);
+void ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader_variant *v);
+bool ir3_nir_lower_ubo_loads(nir_shader *nir, struct ir3_shader_variant *v);
 
 nir_ssa_def *
 ir3_nir_try_propagate_bit_shift(nir_builder *b, nir_ssa_def *offset, int32_t shift);
diff --git a/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c b/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
index 3feb60ed711..2cccee24dc8 100644
--- a/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
+++ b/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
@@ -229,9 +229,10 @@ track_ubo_use(nir_intrinsic_instr *instr, nir_builder *b, int *num_ubos)
 	}
 }
 
-static void
+static bool
 lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
-		struct ir3_ubo_analysis_state *state, int *num_ubos, uint32_t alignment)
+		const struct ir3_ubo_analysis_state *state,
+		int *num_ubos, uint32_t alignment)
 {
 	b->cursor = nir_before_instr(&instr->instr);
 
@@ -242,14 +243,14 @@ lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
 	const struct ir3_ubo_range *range = get_existing_range(instr, state);
 	if (!range) {
 		track_ubo_use(instr, b, num_ubos);
-		return;
+		return false;
 	}
 
 	/* We don't have a good way of determining the range of the dynamic
 	 * access in general, so for now just fall back to pulling.
 	 */
 	if (!nir_src_is_const(instr->src[1]) && !ubo_is_gl_uniforms(&range->ubo))
-		return;
+		return false;
 
 	/* After gathering the UBO access ranges, we limit the total
 	 * upload. Don't lower if this load is outside the range.
@@ -258,7 +259,7 @@ lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
 			instr, alignment);
 	if (!(range->start <= r.start && r.end <= range->end)) {
 		track_ubo_use(instr, b, num_ubos);
-		return;
+		return false;
 	}
 
 	nir_ssa_def *ubo_offset = nir_ssa_for_src(b, instr->src[1], 1);
@@ -313,7 +314,7 @@ lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
 
 	nir_instr_remove(&instr->instr);
 
-	state->lower_count++;
+	return true;
 }
 
 static bool
@@ -330,7 +331,7 @@ instr_is_load_ubo(nir_instr *instr)
 	return op == nir_intrinsic_load_ubo;
 }
 
-bool
+void
 ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader_variant *v)
 {
 	struct ir3_const_state *const_state = ir3_const_state(v);
@@ -394,15 +395,30 @@ ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader_variant *v)
 
 	}
 	state->size = offset;
+}
+
+bool
+ir3_nir_lower_ubo_loads(nir_shader *nir, struct ir3_shader_variant *v)
+{
+	struct ir3_compiler *compiler = v->shader->compiler;
+	/* For the binning pass variant, we re-use the corresponding draw-pass
+	 * variants const_state and ubo state.  To make these clear, in this
+	 * pass it is const (read-only)
+	 */
+	const struct ir3_const_state *const_state = ir3_const_state(v);
+	const struct ir3_ubo_analysis_state *state = &const_state->ubo_state;
 
 	int num_ubos = 0;
+	bool progress = false;
 	nir_foreach_function (function, nir) {
 		if (function->impl) {
 			nir_builder builder;
 			nir_builder_init(&builder, function->impl);
 			nir_foreach_block (block, function->impl) {
 				nir_foreach_instr_safe (instr, block) {
-					if (instr_is_load_ubo(instr))
+					if (!instr_is_load_ubo(instr))
+						continue;
+					progress |=
 						lower_ubo_load_to_uniform(nir_instr_as_intrinsic(instr),
 								&builder, state, &num_ubos,
 								compiler->const_upload_unit);
@@ -420,5 +436,5 @@ ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader_variant *v)
 	if (nir->info.first_ubo_is_default_ubo)
 	    nir->info.num_ubos = num_ubos;
 
-	return state->lower_count > 0;
+	return progress;
 }
diff --git a/src/freedreno/ir3/ir3_shader.h b/src/freedreno/ir3/ir3_shader.h
index e7e6f381293..5ccbd8cb658 100644
--- a/src/freedreno/ir3/ir3_shader.h
+++ b/src/freedreno/ir3/ir3_shader.h
@@ -97,7 +97,6 @@ struct ir3_ubo_analysis_state {
 	struct ir3_ubo_range range[IR3_MAX_UBO_PUSH_RANGES];
 	uint32_t num_enabled;
 	uint32_t size;
-	uint32_t lower_count;
 	uint32_t cmdstream_size; /* for per-gen backend to stash required cmdstream size */
 };