radeonsi: (User) SGPR related cleanups.

Use the same user SGPRs for the same purpose in vertex and pixel shaders.

Better calculation of the number of SGPRs to reserve.

src/gallium/drivers/radeonsi/evergreen_state.c

@@ -1927,6 +1927,7 @@ void si_pipe_shader_ps(struct pipe_context *ctx, struct si_pipe_shader *shader)
 	struct r600_pipe_state *rstate = &shader->rstate;
 	struct r600_shader *rshader = &shader->shader;
 	unsigned i, exports_ps, num_cout, spi_ps_in_control, db_shader_control;
+	unsigned num_sgprs, num_user_sgprs;
 	int pos_index = -1, face_index = -1;
 	int ninterp = 0;
 	boolean have_linear = FALSE, have_centroid = FALSE, have_perspective = FALSE;
@@ -2028,16 +2029,22 @@ void si_pipe_shader_ps(struct pipe_context *ctx, struct si_pipe_shader *shader)
 				va >> 40,
 				shader->bo, RADEON_USAGE_READ);
 
+	num_user_sgprs = 6;
+	num_sgprs = shader->num_sgprs;
+	if (num_user_sgprs > num_sgprs)
+		num_sgprs = num_user_sgprs;
 	/* Last 2 reserved SGPRs are used for VCC */
-	/* XXX: Hard-coding 2 SGPRs for constant buffer */
+	num_sgprs += 2;
+	assert(num_sgprs <= 104);
+
 	r600_pipe_state_add_reg(rstate,
 				R_00B028_SPI_SHADER_PGM_RSRC1_PS,
-				S_00B028_VGPRS(shader->num_vgprs / 4) |
-				S_00B028_SGPRS((shader->num_sgprs + 2 + 2 + 1) / 8),
+				S_00B028_VGPRS((shader->num_vgprs - 1) / 4) |
+				S_00B028_SGPRS((num_sgprs - 1) / 8),
 				NULL, 0);
 	r600_pipe_state_add_reg(rstate,
 				R_00B02C_SPI_SHADER_PGM_RSRC2_PS,
-				S_00B02C_USER_SGPR(6),
+				S_00B02C_USER_SGPR(num_user_sgprs),
 				NULL, 0);
 
 	r600_pipe_state_add_reg(rstate, R_02880C_DB_SHADER_CONTROL,
@@ -2052,6 +2059,7 @@ void si_pipe_shader_vs(struct pipe_context *ctx, struct si_pipe_shader *shader)
 	struct r600_context *rctx = (struct r600_context *)ctx;
 	struct r600_pipe_state *rstate = &shader->rstate;
 	struct r600_shader *rshader = &shader->shader;
+	unsigned num_sgprs, num_user_sgprs;
 	unsigned nparams, i;
 	uint64_t va;
 
@@ -2095,16 +2103,22 @@ void si_pipe_shader_vs(struct pipe_context *ctx, struct si_pipe_shader *shader)
 				va >> 40,
 				shader->bo, RADEON_USAGE_READ);
 
+	num_user_sgprs = 8;
+	num_sgprs = shader->num_sgprs;
+	if (num_user_sgprs > num_sgprs)
+		num_sgprs = num_user_sgprs;
 	/* Last 2 reserved SGPRs are used for VCC */
-	/* XXX: Hard-coding 2 SGPRs for constant buffer */
+	num_sgprs += 2;
+	assert(num_sgprs <= 104);
+
 	r600_pipe_state_add_reg(rstate,
 				R_00B128_SPI_SHADER_PGM_RSRC1_VS,
-				S_00B128_VGPRS(shader->num_vgprs / 4) |
-				S_00B128_SGPRS((shader->num_sgprs + 2 + 2 + 2) / 8),
+				S_00B128_VGPRS((shader->num_vgprs - 1) / 4) |
+				S_00B128_SGPRS((num_sgprs - 1) / 8),
 				NULL, 0);
 	r600_pipe_state_add_reg(rstate,
 				R_00B12C_SPI_SHADER_PGM_RSRC2_VS,
-				S_00B12C_USER_SGPR(2 + 2),
+				S_00B12C_USER_SGPR(num_user_sgprs),
 				NULL, 0);
 }
 

src/gallium/drivers/radeonsi/r600_state_common.c

@@ -452,10 +452,10 @@ void r600_set_constant_buffer(struct pipe_context *ctx, uint shader, uint index,
 		rstate = &rctx->vs_const_buffer;
 		rstate->nregs = 0;
 		r600_pipe_state_add_reg(rstate,
-					R_00B138_SPI_SHADER_USER_DATA_VS_2,
+					R_00B130_SPI_SHADER_USER_DATA_VS_0,
 					va_offset, rbuffer, RADEON_USAGE_READ);
 		r600_pipe_state_add_reg(rstate,
-					R_00B13C_SPI_SHADER_USER_DATA_VS_3,
+					R_00B134_SPI_SHADER_USER_DATA_VS_1,
 					va_offset >> 32, NULL, 0);
 		break;
 	case PIPE_SHADER_FRAGMENT:
@@ -638,10 +638,10 @@ static void r600_vertex_buffer_update(struct r600_context *rctx)
 
 	va = r600_resource_va(ctx->screen, (void*)t_list_buffer);
 	r600_pipe_state_add_reg(rstate,
-				R_00B130_SPI_SHADER_USER_DATA_VS_0,
+				R_00B148_SPI_SHADER_USER_DATA_VS_6,
 				va, t_list_buffer, RADEON_USAGE_READ);
 	r600_pipe_state_add_reg(rstate,
-				R_00B134_SPI_SHADER_USER_DATA_VS_1,
+				R_00B14C_SPI_SHADER_USER_DATA_VS_7,
 				va >> 32,
 				NULL, 0);
 

src/gallium/drivers/radeonsi/radeonsi_shader.c

@@ -138,9 +138,9 @@ static void declare_input_vs(
 	unsigned chan;
 
 	/* XXX: Communicate with the rest of the driver about which SGPR the T#
- 	 * list pointer is going to be stored in.  Hard code to SGPR[0-1] for
+	 * list pointer is going to be stored in.  Hard code to SGPR[6:7] for
  	 * now */
-	t_list_ptr = use_sgpr(base->gallivm, SGPR_I64, 0);
+	t_list_ptr = use_sgpr(base->gallivm, SGPR_I64, 3);
 
 	t_offset = lp_build_const_int32(base->gallivm,
 					4 * velem->vertex_buffer_index);
@@ -188,6 +188,9 @@ static void declare_input_fs(
 	 * [32:16] ParamOffset
 	 *
 	 */
+	/* XXX: This register number must be identical to the S_00B02C_USER_SGPR
+	 * register field value
+	 */
 	LLVMValueRef params = use_sgpr(base->gallivm, SGPR_I32, 6);
 
 
@@ -256,8 +259,8 @@ static LLVMValueRef fetch_constant(
 	LLVMValueRef offset;
 
 	/* XXX: Assume the pointer to the constant buffer is being stored in
-	 * SGPR[2:3] */
-	const_ptr = use_sgpr(base->gallivm, SGPR_I64, 1);
+	 * SGPR[0:1] */
+	const_ptr = use_sgpr(base->gallivm, SGPR_I64, 0);
 
 	/* XXX: This assumes that the constant buffer is not packed, so
 	 * CONST[0].x will have an offset of 0 and CONST[1].x will have an