[Mesa-dev] [PATCH 67/92] ac/nir: use shader_info pass to determine whether instance_id is used

[Nicolai Hähnle]

From: Nicolai Hähnle <nicolai.haehnle at amd.com>

This improves the separation of ABI and NIR translation.
---
 src/amd/common/ac_nir_to_llvm.c | 7 +++++--
 src/amd/common/ac_shader_info.c | 3 +++
 src/amd/common/ac_shader_info.h | 1 +
 3 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 9d06a19..ecb04e4 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -3914,22 +3914,20 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
 		result = load_sample_pos(ctx->nctx);
 		break;
 	case nir_intrinsic_load_sample_mask_in:
 		result = ctx->nctx->sample_coverage;
 		break;
 	case nir_intrinsic_load_front_face:
 		result = ctx->nctx->front_face;
 		break;
 	case nir_intrinsic_load_instance_id:
 		result = LLVMGetParam(ctx->main_function, ctx->abi->param_instance_id);
-		ctx->nctx->shader_info->vs.vgpr_comp_cnt = MAX2(3,
-		                            ctx->nctx->shader_info->vs.vgpr_comp_cnt);
 		break;
 	case nir_intrinsic_load_num_work_groups:
 		result = ctx->nctx->num_work_groups;
 		break;
 	case nir_intrinsic_load_local_invocation_index:
 		result = visit_load_local_invocation_index(ctx->nctx);
 		break;
 	case nir_intrinsic_load_push_constant:
 		result = visit_load_push_constant(ctx->nctx, instr);
 		break;
@@ -6054,20 +6052,25 @@ LLVMModuleRef ac_translate_nir_to_llvm(LLVMTargetMachineRef tm,
 							  LOCAL_ADDR_SPACE);
 			LLVMSetAlignment(var, 4);
 			ctx.shared_memory = LLVMBuildBitCast(ctx.builder, var, i8p, "");
 		}
 	} else if (nir->stage == MESA_SHADER_GEOMETRY) {
 		ctx.gs_next_vertex = ac_build_alloca(&ctx.ac, ctx.i32, "gs_next_vertex");
 
 		ctx.gs_max_out_vertices = nir->info.gs.vertices_out;
 	} else if (nir->stage == MESA_SHADER_TESS_EVAL) {
 		ctx.tes_primitive_mode = nir->info.tess.primitive_mode;
+	} else if (nir->stage == MESA_SHADER_VERTEX) {
+		if (shader_info->info.vs.needs_instance_id) {
+			ctx.shader_info->vs.vgpr_comp_cnt =
+				MAX2(3, ctx.shader_info->vs.vgpr_comp_cnt);
+		}
 	}
 
 	ac_setup_rings(&ctx);
 
 	ctx.num_output_clips = nir->info.clip_distance_array_size;
 	ctx.num_output_culls = nir->info.cull_distance_array_size;
 
 	nir_foreach_variable(variable, &nir->inputs)
 		handle_shader_input_decl(&ctx, variable);
 
diff --git a/src/amd/common/ac_shader_info.c b/src/amd/common/ac_shader_info.c
index 13d73df..7d34535 100644
--- a/src/amd/common/ac_shader_info.c
+++ b/src/amd/common/ac_shader_info.c
@@ -32,20 +32,23 @@ static void mark_sampler_desc(nir_variable *var, struct ac_shader_info *info)
 static void
 gather_intrinsic_info(nir_intrinsic_instr *instr, struct ac_shader_info *info)
 {
 	switch (instr->intrinsic) {
 	case nir_intrinsic_interp_var_at_sample:
 		info->ps.needs_sample_positions = true;
 		break;
 	case nir_intrinsic_load_draw_id:
 		info->vs.needs_draw_id = true;
 		break;
+	case nir_intrinsic_load_instance_id:
+		info->vs.needs_instance_id = true;
+		break;
 	case nir_intrinsic_load_num_work_groups:
 		info->cs.grid_components_used = instr->num_components;
 		break;
 	case nir_intrinsic_vulkan_resource_index:
 		info->desc_set_used_mask |= (1 << nir_intrinsic_desc_set(instr));
 		break;
 	case nir_intrinsic_image_load:
 	case nir_intrinsic_image_store:
 	case nir_intrinsic_image_atomic_add:
 	case nir_intrinsic_image_atomic_min:
diff --git a/src/amd/common/ac_shader_info.h b/src/amd/common/ac_shader_info.h
index 5f03e79..5bc16cc 100644
--- a/src/amd/common/ac_shader_info.h
+++ b/src/amd/common/ac_shader_info.h
@@ -26,20 +26,21 @@
 
 struct nir_shader;
 struct ac_nir_compiler_options;
 
 struct ac_shader_info {
 	bool needs_push_constants;
 	uint32_t desc_set_used_mask;
 	struct {
 		bool has_vertex_buffers; /* needs vertex buffers and base/start */
 		bool needs_draw_id;
+		bool needs_instance_id;
 	} vs;
 	struct {
 		bool needs_sample_positions;
 	} ps;
 	struct {
 		uint8_t grid_components_used;
 	} cs;
 };
 
 /* A NIR pass to gather all the info needed to optimise the allocation patterns
-- 
2.9.3