r600: Add shader precompile and shader-db support.

This should reduce draw-time jank, and was useful for me in evaluating
NIR-to-TGSI's impact for r600.

Acked-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14427>

src/gallium/drivers/r600/evergreen_compute.c

@@ -444,6 +444,13 @@ static void *evergreen_create_compute_state(struct pipe_context *ctx,
 	if (shader->ir_type == PIPE_SHADER_IR_TGSI ||
 	    shader->ir_type == PIPE_SHADER_IR_NIR) {
 		shader->sel = r600_create_shader_state_tokens(ctx, cso->prog, cso->ir_type, PIPE_SHADER_COMPUTE);
+
+		/* Precompile the shader with the expected shader key, to reduce jank at
+		 * draw time. Also produces output for shader-db.
+		 */
+		bool dirty;
+		r600_shader_select(ctx, shader->sel, &dirty, true);
+
 		return shader;
 	}
 #ifdef HAVE_OPENCL
@@ -506,7 +513,7 @@ static void evergreen_bind_compute_state(struct pipe_context *ctx, void *state)
 	    cstate->ir_type == PIPE_SHADER_IR_NIR) {
 		bool compute_dirty;
 		cstate->sel->ir_type = cstate->ir_type;
-		if (r600_shader_select(ctx, cstate->sel, &compute_dirty))
+		if (r600_shader_select(ctx, cstate->sel, &compute_dirty, false))
 			R600_ERR("Failed to select compute shader\n");
 	}
 	
@@ -736,7 +743,7 @@ static void compute_emit_cs(struct r600_context *rctx,
 
 	if (rctx->cs_shader_state.shader->ir_type == PIPE_SHADER_IR_TGSI||
 	    rctx->cs_shader_state.shader->ir_type == PIPE_SHADER_IR_NIR) {
-		if (r600_shader_select(&rctx->b.b, rctx->cs_shader_state.shader->sel, &compute_dirty)) {
+		if (r600_shader_select(&rctx->b.b, rctx->cs_shader_state.shader->sel, &compute_dirty, false)) {
 			R600_ERR("Failed to select compute shader\n");
 			return;
 		}

src/gallium/drivers/r600/r600_pipe.h

@@ -1065,7 +1065,7 @@ struct r600_pipe_shader_selector *r600_create_shader_state_tokens(struct pipe_co
 								  unsigned pipe_shader_type);
 int r600_shader_select(struct pipe_context *ctx,
 		       struct r600_pipe_shader_selector* sel,
-		       bool *dirty);
+		       bool *dirty, bool precompile);
 
 void r600_delete_shader_selector(struct pipe_context *ctx,
 				 struct r600_pipe_shader_selector *sel);

src/gallium/drivers/r600/r600_shader.c

@@ -372,6 +372,15 @@ int r600_pipe_shader_create(struct pipe_context *ctx,
 		r = -EINVAL;
 		goto error;
 	}
+
+	pipe_debug_message(&rctx->b.debug, SHADER_INFO, "%s shader: %d dw, %d gprs, %d loops, %d cf, %d stack",
+		           _mesa_shader_stage_to_abbrev(tgsi_processor_to_shader_stage(processor)),
+	                   shader->shader.bc.ndw,
+	                   shader->shader.bc.ngpr,
+			   shader->shader.num_loops,
+			   shader->shader.bc.ncf,
+			   shader->shader.bc.nstack);
+
 	return 0;
 
 error:
@@ -3451,6 +3460,8 @@ static int r600_shader_from_tgsi(struct r600_context *rctx,
 	shader->uses_helper_invocation = false;
 	shader->uses_doubles = ctx.info.uses_doubles;
 	shader->uses_atomics = ctx.info.file_mask[TGSI_FILE_HW_ATOMIC];
+	shader->num_loops = ctx.info.opcode_count[TGSI_OPCODE_BGNLOOP];
+
 	shader->nsys_inputs = 0;
 
 	shader->uses_images = ctx.info.file_count[TGSI_FILE_IMAGE] > 0 ||

src/gallium/drivers/r600/r600_shader.h

@@ -114,6 +114,8 @@ struct r600_shader {
 	unsigned                tes_as_es;
 	unsigned                tcs_prim_mode;
 	unsigned                ps_prim_id_input;
+	unsigned                num_loops;
+
 	struct r600_shader_array * arrays;
 
 	boolean			uses_doubles;

src/gallium/drivers/r600/r600_state_common.c

@@ -867,17 +867,63 @@ static inline void r600_shader_selector_key(const struct pipe_context *ctx,
 	}
 }
 
+static void
+r600_shader_precompile_key(const struct pipe_context *ctx,
+			   const struct r600_pipe_shader_selector *sel,
+			   union r600_shader_key *key)
+{
+	memset(key, 0, sizeof(*key));
+
+	switch (sel->type) {
+	case PIPE_SHADER_VERTEX:
+	case PIPE_SHADER_TESS_EVAL:
+		/* Assume no tess or GS for setting .as_es.  In order to
+		 * precompile with es, we'd need the other shaders we're linked
+		 * with (see the link_shader screen method)
+		 */
+		break;
+
+	case PIPE_SHADER_GEOMETRY:
+		break;
+
+	case PIPE_SHADER_FRAGMENT:
+		key->ps.image_size_const_offset = sel->info.file_max[TGSI_FILE_IMAGE];
+
+		/* This is used for gl_FragColor output expansion to the number
+		 * of color buffers bound, but also with sb it'll drop outputs
+		 * to unused cbufs.
+		 */
+		key->ps.nr_cbufs = sel->info.file_max[TGSI_FILE_OUTPUT] + 1;
+		break;
+
+	case PIPE_SHADER_TESS_CTRL:
+		/* Prim mode comes from the TES, but we need some valid value. */
+		key->tcs.prim_mode = PIPE_PRIM_TRIANGLES;
+		break;
+
+	case PIPE_SHADER_COMPUTE:
+		break;
+
+	default:
+		unreachable("bad shader stage");
+		break;
+	}
+}
+
 /* Select the hw shader variant depending on the current state.
  * (*dirty) is set to 1 if current variant was changed */
 int r600_shader_select(struct pipe_context *ctx,
         struct r600_pipe_shader_selector* sel,
-        bool *dirty)
+        bool *dirty, bool precompile)
 {
 	union r600_shader_key key;
 	struct r600_pipe_shader * shader = NULL;
 	int r;
 
-	r600_shader_selector_key(ctx, sel, &key);
+	if (precompile)
+		r600_shader_precompile_key(ctx, sel, &key);
+	else
+		r600_shader_selector_key(ctx, sel, &key);
 
 	/* Check if we don't need to change anything.
 	 * This path is also used for most shaders that don't need multiple
@@ -997,6 +1043,12 @@ static void *r600_create_shader_state(struct pipe_context *ctx,
 		break;
 	}
 
+	/* Precompile the shader with the expected shader key, to reduce jank at
+	 * draw time. Also produces output for shader-db.
+	 */
+	bool dirty;
+	r600_shader_select(ctx, sel, &dirty, true);
+
 	return sel;
 }
 
@@ -1771,7 +1823,7 @@ void r600_setup_scratch_buffers(struct r600_context *rctx) {
 }
 
 #define SELECT_SHADER_OR_FAIL(x) do {					\
-		r600_shader_select(ctx, rctx->x##_shader, &x##_dirty);	\
+		r600_shader_select(ctx, rctx->x##_shader, &x##_dirty, false);	\
 		if (unlikely(!rctx->x##_shader->current))		\
 			return false;					\
 	} while(0)