r600g: setup basic loop consts on r600 + evergreen.

this sets up a single loop constant like r600c does.

src/gallium/drivers/r600/evergreen_state.c

@@ -1593,6 +1593,10 @@ void evergreen_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shader
 					S_02880C_KILL_ENABLE(1),
 					S_02880C_KILL_ENABLE(1), NULL);
 	}
+
+	r600_pipe_state_add_reg(rstate,
+				R_03A200_SQ_LOOP_CONST_0, 0x01000FFF,
+				0xFFFFFFFF, NULL);
 }
 
 void evergreen_pipe_shader_vs(struct pipe_context *ctx, struct r600_pipe_shader *shader)
@@ -1640,6 +1644,10 @@ void evergreen_pipe_shader_vs(struct pipe_context *ctx, struct r600_pipe_shader
 	r600_pipe_state_add_reg(rstate,
 			R_0288A4_SQ_PGM_START_FS,
 			0x00000000, 0xFFFFFFFF, shader->bo);
+
+	r600_pipe_state_add_reg(rstate,
+				R_03A200_SQ_LOOP_CONST_0 + (32 * 4), 0x01000FFF,
+				0xFFFFFFFF, NULL);
 }
 
 void *evergreen_create_db_flush_dsa(struct r600_pipe_context *rctx)

src/gallium/drivers/r600/evergreend.h

@@ -1910,4 +1910,5 @@
 #define R_03CFF0_SQ_VTX_BASE_VTX_LOC                    0x03CFF0
 #define R_03CFF4_SQ_VTX_START_INST_LOC                  0x03CFF4
 
+#define R_03A200_SQ_LOOP_CONST_0                     0x3A200
 #endif

src/gallium/drivers/r600/r600_shader.c

@@ -81,6 +81,11 @@ static void r600_pipe_shader_vs(struct pipe_context *ctx, struct r600_pipe_shade
 	r600_pipe_state_add_reg(rstate,
 			R_028894_SQ_PGM_START_FS,
 			0x00000000, 0xFFFFFFFF, shader->bo);
+
+	r600_pipe_state_add_reg(rstate,
+				R_03E200_SQ_LOOP_CONST_0 + (32 * 4), 0x01000FFF,
+				0xFFFFFFFF, NULL);
+
 }
 
 int r600_find_vs_semantic_index(struct r600_shader *vs,
@@ -182,6 +187,9 @@ static void r600_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shade
 					S_02880C_KILL_ENABLE(1),
 					S_02880C_KILL_ENABLE(1), NULL);
 	}
+	r600_pipe_state_add_reg(rstate,
+				R_03E200_SQ_LOOP_CONST_0, 0x01000FFF,
+				0xFFFFFFFF, NULL);
 }
 
 static int r600_pipe_shader(struct pipe_context *ctx, struct r600_pipe_shader *shader)

src/gallium/drivers/r600/r600d.h

@@ -3497,6 +3497,8 @@
 #define R_03CFF0_SQ_VTX_BASE_VTX_LOC                 0x03CFF0
 #define R_03CFF4_SQ_VTX_START_INST_LOC               0x03CFF4
 
+#define R_03E200_SQ_LOOP_CONST_0                     0x3E200
+
 #define SQ_TEX_INST_LD 0x03
 #define SQ_TEX_INST_GET_GRADIENTS_H 0x7
 #define SQ_TEX_INST_GET_GRADIENTS_V 0x8

src/gallium/winsys/r600/drm/evergreen_hw_context.c

@@ -493,6 +493,22 @@ static int evergreen_state_sampler_border_init(struct r600_context *ctx, u32 off
 	return 0;
 }
 
+static int evergreen_loop_const_init(struct r600_context *ctx, u32 offset)
+{
+	unsigned nreg = 32;
+	struct r600_reg r600_loop_consts[32];
+	int i;
+
+	for (i = 0; i < nreg; i++) {
+		r600_loop_consts[i].opcode = PKT3_SET_LOOP_CONST;
+		r600_loop_consts[i].offset_base = EVERGREEN_LOOP_CONST_OFFSET;
+		r600_loop_consts[i].offset = EVERGREEN_LOOP_CONST_OFFSET + ((offset + i) * 4);
+		r600_loop_consts[i].need_bo = 0;
+		r600_loop_consts[i].flush_flags = 0;
+	}
+	return r600_context_add_block(ctx, r600_loop_consts, nreg);
+}
+
 int evergreen_context_init(struct r600_context *ctx, struct radeon *radeon)
 {
 	int r;
@@ -566,6 +582,11 @@ int evergreen_context_init(struct r600_context *ctx, struct radeon *radeon)
 			goto out_err;
 	}
 
+	/* PS loop const */
+	evergreen_loop_const_init(ctx, 0);
+	/* VS loop const */
+	evergreen_loop_const_init(ctx, 32);
+
 	/* setup block table */
 	ctx->blocks = calloc(ctx->nblocks, sizeof(void*));
 	for (int i = 0, c = 0; i < 256; i++) {

src/gallium/winsys/r600/drm/r600_hw_context.c

@@ -542,6 +542,22 @@ static int r600_state_sampler_border_init(struct r600_context *ctx, u32 offset)
 	return r600_context_add_block(ctx, r600_shader_sampler_border, nreg);
 }
 
+static int r600_loop_const_init(struct r600_context *ctx, u32 offset)
+{
+	unsigned nreg = 32;
+	struct r600_reg r600_loop_consts[32];
+	int i;
+
+	for (i = 0; i < nreg; i++) {
+		r600_loop_consts[i].opcode = PKT3_SET_LOOP_CONST;
+		r600_loop_consts[i].offset_base = R600_LOOP_CONST_OFFSET;
+		r600_loop_consts[i].offset = R600_LOOP_CONST_OFFSET + ((offset + i) * 4);
+		r600_loop_consts[i].need_bo = 0;
+		r600_loop_consts[i].flush_flags = 0;
+	}
+	return r600_context_add_block(ctx, r600_loop_consts, nreg);
+}
+
 /* initialize */
 void r600_context_fini(struct r600_context *ctx)
 {
@@ -639,6 +655,11 @@ int r600_context_init(struct r600_context *ctx, struct radeon *radeon)
 			goto out_err;
 	}
 
+	/* PS loop const */
+	r600_loop_const_init(ctx, 0);
+	/* VS loop const */
+	r600_loop_const_init(ctx, 32);
+
 	/* setup block table */
 	ctx->blocks = calloc(ctx->nblocks, sizeof(void*));
 	for (int i = 0, c = 0; i < 256; i++) {