r600g: avoid dynamic allocation of states

Make state statically allocated, this kills a bunch of code
and avoid intensive use of malloc/free. There is still a lot
of useless duplicate function wrapping that can be kill. This
doesn't improve yet performance, needs to avoid memcpy states
in radeon_ctx_set_draw and to avoid rebuilding vs_resources,
dsa, scissor, cb_cntl, ... states at each draw command.

Signed-off-by: Jerome Glisse <[email protected]>

6 files changed, 122 insertions, 308 deletions

diff --git a/src/gallium/winsys/r600/drm/r600_state.c b/src/gallium/winsys/r600/drm/r600_state.c
index f6a428e884d..71d65f0feab 100644
--- a/src/gallium/winsys/r600/drm/r600_state.c
+++ b/src/gallium/winsys/r600/drm/r600_state.c
@@ -367,7 +367,6 @@ static int r600_state_pm4_vgt(struct radeon_state *state)
 
 static int r600_state_pm4_draw(struct radeon_state *state)
 {
-	unsigned i;
 	int r;
 
 	if (state->nbo) {
@@ -485,11 +484,10 @@ static void r600_build_types_array(struct radeon *radeon)
 		} else {
 			for (j = 0; j < R600_SHADER_MAX; j++) {
 				if (r600_stypes[i].reginfo[j].shader_type)
-					id += r600_stypes[i].num;					
+					id += r600_stypes[i].num;
 			}
 		}
 	}
-	radeon->nstate = id;
 	radeon->stype = r600_stypes;
 	radeon->nstype = STYPES_SIZE;
 
diff --git a/src/gallium/winsys/r600/drm/radeon.c b/src/gallium/winsys/r600/drm/radeon.c
index 2b16e3ce884..e2d813ebac7 100644
--- a/src/gallium/winsys/r600/drm/radeon.c
+++ b/src/gallium/winsys/r600/drm/radeon.c
@@ -42,24 +42,13 @@ static int radeon_get_device(struct radeon *radeon)
 	return r;
 }
 
-/* symbol missing drove me crazy hack to get symbol exported */
-static void fake(void)
-{
-	struct radeon_ctx *ctx;
-	struct radeon_draw *draw;
-
-	ctx = radeon_ctx(NULL);
-	draw = radeon_draw(NULL);
-}
-
 struct radeon *radeon_new(int fd, unsigned device)
 {
 	struct radeon *radeon;
-	int r;
+	int r, i, id;
 
 	radeon = calloc(1, sizeof(*radeon));
 	if (radeon == NULL) {
-		fake();
 		return NULL;
 	}
 	radeon->fd = fd;
@@ -131,6 +120,19 @@ struct radeon *radeon_new(int fd, unsigned device)
 			__func__, radeon->device);
 		break;
 	}
+	radeon->state_type_id = calloc(radeon->nstype, sizeof(unsigned));
+	if (radeon->state_type_id == NULL) {
+		return radeon_decref(radeon);
+	}
+	for (i = 0, id = 0; i < radeon->nstype; i++) {
+		radeon->state_type_id[i] = id;
+		for (int j = 0; j < radeon->nstype; j++) {
+			if (radeon->stype[j].stype != i)
+				continue;
+			id += radeon->stype[j].num;
+		}
+	}
+	radeon->nstate_per_shader = id;
 	return radeon;
 }
 
diff --git a/src/gallium/winsys/r600/drm/radeon_ctx.c b/src/gallium/winsys/r600/drm/radeon_ctx.c
index 1a12c1023e1..50c5e4741e1 100644
--- a/src/gallium/winsys/r600/drm/radeon_ctx.c
+++ b/src/gallium/winsys/r600/drm/radeon_ctx.c
@@ -30,21 +30,16 @@
 #include "radeon_drm.h"
 #include "bof.h"
 
-int radeon_ctx_set_bo_new(struct radeon_ctx *ctx, struct radeon_bo *bo)
+static int radeon_ctx_set_bo_new(struct radeon_ctx *ctx, struct radeon_bo *bo)
 {
-	void *ptr;
-
-	ptr = realloc(ctx->bo, sizeof(struct radeon_bo) * (ctx->nbo + 1));
-	if (ptr == NULL) {
-		return -ENOMEM;
-	}
-	ctx->bo = ptr;
+	if (ctx->nbo >= RADEON_CTX_MAX_PM4)
+		return -EBUSY;
 	ctx->bo[ctx->nbo] = bo;
 	ctx->nbo++;
 	return 0;
 }
 
-struct radeon_bo *radeon_ctx_get_bo(struct radeon_ctx *ctx, unsigned reloc)
+static struct radeon_bo *radeon_ctx_get_bo(struct radeon_ctx *ctx, unsigned reloc)
 {
 	struct radeon_cs_reloc *greloc;
 	unsigned i;
@@ -59,7 +54,7 @@ struct radeon_bo *radeon_ctx_get_bo(struct radeon_ctx *ctx, unsigned reloc)
 	return NULL;
 }
 
-void radeon_ctx_get_placement(struct radeon_ctx *ctx, unsigned reloc, u32 *placement)
+static void radeon_ctx_get_placement(struct radeon_ctx *ctx, unsigned reloc, u32 *placement)
 {
 	struct radeon_cs_reloc *greloc;
 	unsigned i;
@@ -76,50 +71,58 @@ void radeon_ctx_get_placement(struct radeon_ctx *ctx, unsigned reloc, u32 *place
 	}
 }
 
-struct radeon_ctx *radeon_ctx(struct radeon *radeon)
+void radeon_ctx_clear(struct radeon_ctx *ctx)
 {
-	struct radeon_ctx *ctx;
-
-	if (radeon == NULL)
-		return NULL;
-	ctx = calloc(1, sizeof(*ctx));
-	if (ctx == NULL)
-		return NULL;
-	ctx->radeon = radeon_incref(radeon);
-	return ctx;
+	for (int i = 0; i < ctx->nbo; i++) {
+		ctx->bo[i] = radeon_bo_decref(ctx->radeon, ctx->bo[i]);
+	}
+	ctx->ndwords = RADEON_CTX_MAX_PM4;
+	ctx->cdwords = 0;
+	ctx->nreloc = 0;
+	ctx->nbo = 0;
 }
 
-struct radeon_ctx *radeon_ctx_incref(struct radeon_ctx *ctx)
+int radeon_ctx_init(struct radeon_ctx *ctx, struct radeon *radeon)
 {
-	ctx->refcount++;
-	return ctx;
+	if (radeon == NULL)
+		return -EINVAL;
+	memset(ctx, 0, sizeof(struct radeon_ctx));
+	ctx->radeon = radeon_incref(radeon);
+	radeon_ctx_clear(ctx);
+	ctx->pm4 = malloc(RADEON_CTX_MAX_PM4 * 4);
+	if (ctx->pm4 == NULL) {
+		radeon_ctx_fini(ctx);
+		return -ENOMEM;
+	}
+	ctx->reloc = malloc(sizeof(struct radeon_cs_reloc) * RADEON_CTX_MAX_PM4);
+	if (ctx->reloc == NULL) {
+		radeon_ctx_fini(ctx);
+		return -ENOMEM;
+	}
+	ctx->bo = malloc(sizeof(void *) * RADEON_CTX_MAX_PM4);
+	if (ctx->bo == NULL) {
+		radeon_ctx_fini(ctx);
+		return -ENOMEM;
+	}
+	return 0;
 }
 
-struct radeon_ctx *radeon_ctx_decref(struct radeon_ctx *ctx)
+void radeon_ctx_fini(struct radeon_ctx *ctx)
 {
 	unsigned i;
 
 	if (ctx == NULL)
-		return NULL;
-	if (--ctx->refcount > 0) {
-		return NULL;
-	}
+		return;
 
-	for (i = 0; i < ctx->ndraw; i++) {
-		ctx->draw[i] = radeon_draw_decref(ctx->draw[i]);
-	}
 	for (i = 0; i < ctx->nbo; i++) {
 		ctx->bo[i] = radeon_bo_decref(ctx->radeon, ctx->bo[i]);
 	}
 	ctx->radeon = radeon_decref(ctx->radeon);
-	free(ctx->state);
-	free(ctx->draw);
 	free(ctx->bo);
 	free(ctx->pm4);
 	free(ctx->reloc);
 	memset(ctx, 0, sizeof(*ctx));
 	free(ctx);
-	return NULL;
 }
 
 static int radeon_ctx_state_bo(struct radeon_ctx *ctx, struct radeon_state *state)
@@ -152,17 +155,17 @@ int radeon_ctx_submit(struct radeon_ctx *ctx)
 	uint64_t chunk_array[2];
 	int r = 0;
 
-	if (!ctx->cpm4)
+	if (!ctx->cdwords)
 		return 0;
 #if 0
-	for (r = 0; r < ctx->cpm4; r++) {
+	for (r = 0; r < ctx->cdwords; r++) {
 		fprintf(stderr, "0x%08X\n", ctx->pm4[r]);
 	}
 #endif
 	drmib.num_chunks = 2;
 	drmib.chunks = (uint64_t)(uintptr_t)chunk_array;
 	chunks[0].chunk_id = RADEON_CHUNK_ID_IB;
-	chunks[0].length_dw = ctx->cpm4;
+	chunks[0].length_dw = ctx->cdwords;
 	chunks[0].chunk_data = (uint64_t)(uintptr_t)ctx->pm4;
 	chunks[1].chunk_id = RADEON_CHUNK_ID_RELOCS;
 	chunks[1].length_dw = ctx->nreloc * sizeof(struct radeon_cs_reloc) / 4;
@@ -180,7 +183,6 @@ static int radeon_ctx_reloc(struct radeon_ctx *ctx, struct radeon_bo *bo,
 			unsigned id, unsigned *placement)
 {
 	unsigned i;
-	struct radeon_cs_reloc *ptr;
 
 	for (i = 0; i < ctx->nreloc; i++) {
 		if (ctx->reloc[i].handle == bo->handle) {
@@ -188,14 +190,13 @@ static int radeon_ctx_reloc(struct radeon_ctx *ctx, struct radeon_bo *bo,
 			return 0;
 		}
 	}
-	ptr = realloc(ctx->reloc, sizeof(struct radeon_cs_reloc) * (ctx->nreloc + 1));
-	if (ptr == NULL)
-		return -ENOMEM;
-	ctx->reloc = ptr;
-	ptr[ctx->nreloc].handle = bo->handle;
-	ptr[ctx->nreloc].read_domain = placement[0] | placement [1];
-	ptr[ctx->nreloc].write_domain = placement[0] | placement [1];
-	ptr[ctx->nreloc].flags = 0;
+	if (ctx->nreloc >= RADEON_CTX_MAX_PM4) {
+		return -EBUSY;
+	}
+	ctx->reloc[ctx->nreloc].handle = bo->handle;
+	ctx->reloc[ctx->nreloc].read_domain = placement[0] | placement [1];
+	ctx->reloc[ctx->nreloc].write_domain = placement[0] | placement [1];
+	ctx->reloc[ctx->nreloc].flags = 0;
 	ctx->pm4[id] = ctx->nreloc * sizeof(struct radeon_cs_reloc) / 4;
 	ctx->nreloc++;
 	return 0;
@@ -208,11 +209,14 @@ static int radeon_ctx_state_schedule(struct radeon_ctx *ctx, struct radeon_state
 
 	if (state == NULL)
 		return 0;
-	memcpy(&ctx->pm4[ctx->id], state->pm4, state->cpm4 * 4);
+	if (state->cpm4 > ctx->ndwords) {
+		return -EBUSY;
+	}
+	memcpy(&ctx->pm4[ctx->cdwords], state->pm4, state->cpm4 * 4);
 	for (i = 0; i < state->nreloc; i++) {
 		rid = state->reloc_pm4_id[i];
 		bid = state->reloc_bo_id[i];
-		cid = ctx->id + rid;
+		cid = ctx->cdwords + rid;
 		r = radeon_ctx_reloc(ctx, state->bo[bid], cid,
 					&state->placement[bid * 2]);
 		if (r) {
@@ -220,120 +224,65 @@ static int radeon_ctx_state_schedule(struct radeon_ctx *ctx, struct radeon_state
 			return r;
 		}
 	}
-	ctx->id += state->cpm4;
+	ctx->cdwords += state->cpm4;
+	ctx->ndwords -= state->cpm4;
 	return 0;
 }
 
 int radeon_ctx_set_query_state(struct radeon_ctx *ctx, struct radeon_state *state)
 {
-	void *tmp;
 	int r = 0;
 
 	/* !!! ONLY ACCEPT QUERY STATE HERE !!! */
-	if (state->stype->stype != R600_STATE_QUERY_BEGIN && state->stype->stype != R600_STATE_QUERY_END) {
-		return -EINVAL;
-	}
 	r = radeon_state_pm4(state);
 	if (r)
 		return r;
-	if ((ctx->draw_cpm4 + state->cpm4) > RADEON_CTX_MAX_PM4) {
-		/* need to flush */
-		return -EBUSY;
-	}
-	if (state->cpm4 >= RADEON_CTX_MAX_PM4) {
-		fprintf(stderr, "%s single state too big %d, max %d\n",
-			__func__, state->cpm4, RADEON_CTX_MAX_PM4);
-		return -EINVAL;
-	}
-	tmp = realloc(ctx->state, (ctx->nstate + 1) * sizeof(void*));
-	if (tmp == NULL)
-		return -ENOMEM;
-	ctx->state = tmp;
-	ctx->state[ctx->nstate++] = radeon_state_incref(state);
 	/* BEGIN/END query are balanced in the same cs so account for END
 	 * END query when scheduling BEGIN query
 	 */
-	if (state->stype->stype == R600_STATE_QUERY_BEGIN) {
-		ctx->draw_cpm4 += state->cpm4 * 2;
+	switch (state->stype->stype) {
+	case R600_STATE_QUERY_BEGIN:
+		/* is there enough place for begin & end */
+		if ((state->cpm4 * 2) > ctx->ndwords)
+			return -EBUSY;
+		ctx->ndwords -= state->cpm4;
+		break;
+	case R600_STATE_QUERY_END:
+		ctx->ndwords += state->cpm4;
+		break;
+	default:
+		return -EINVAL;
 	}
-	return 0;
+	return radeon_ctx_state_schedule(ctx, state);
 }
 
-int radeon_ctx_set_draw_new(struct radeon_ctx *ctx, struct radeon_draw *draw)
+int radeon_ctx_set_draw(struct radeon_ctx *ctx, struct radeon_draw *draw)
 {
-	struct radeon_draw *pdraw = NULL;
-	struct radeon_draw **ndraw;
-	struct radeon_state *nstate, *ostate;
-	unsigned cpm4, i, cstate;
-	void *tmp;
+	unsigned previous_cdwords;
 	int r = 0;
 
-	ndraw = realloc(ctx->draw, sizeof(void*) * (ctx->ndraw + 1));
-	if (ndraw == NULL)
-		return -ENOMEM;
-	ctx->draw = ndraw;
-	for (i = 0; i < draw->nstate; i++) {
+	for (int i = 0; i < (ctx->radeon->nstate_per_shader * R600_SHADER_MAX); i++) {
 		r = radeon_ctx_state_bo(ctx, draw->state[i]);
 		if (r)
 			return r;
 	}
-	r = radeon_draw_check(draw);
-	if (r)
-		return r;
-	if (draw->cpm4 >= RADEON_CTX_MAX_PM4) {
-		fprintf(stderr, "%s single draw too big %d, max %d\n",
-			__func__, draw->cpm4, RADEON_CTX_MAX_PM4);
-		return -EINVAL;
-	}
-	tmp = realloc(ctx->state, (ctx->nstate + draw->nstate) * sizeof(void*));
-	if (tmp == NULL)
-		return -ENOMEM;
-	ctx->state = tmp;
-	pdraw = ctx->cdraw;
-	for (i = 0, cpm4 = 0, cstate = ctx->nstate; i < draw->nstate - 1; i++) {
-		nstate = draw->state[i];
-		if (nstate) {
-			if (pdraw && pdraw->state[i]) {
-				ostate = pdraw->state[i];
-				if (ostate->pm4_crc != nstate->pm4_crc) {
-					ctx->state[cstate++] = nstate;
-					cpm4 += nstate->cpm4;
+	previous_cdwords = ctx->cdwords;
+	for (int i = 0, id = 0; i < ctx->radeon->nstate_per_shader; i++) {
+		for (int j = 0; j < R600_SHADER_MAX; j++) {
+			id = j * ctx->radeon->nstate_per_shader + i;
+			if (draw->state[id]) {
+				r = radeon_ctx_state_schedule(ctx, draw->state[id]);
+				if (r) {
+					ctx->cdwords = previous_cdwords;
+					return r;
 				}
-			} else {
-				ctx->state[cstate++] = nstate;
-				cpm4 += nstate->cpm4;
 			}
 		}
 	}
-	/* The last state is the draw state always add it */
-	if (draw->state[i] == NULL) {
-		fprintf(stderr, "%s no draw command\n", __func__);
-		return -EINVAL;
-	}
-	ctx->state[cstate++] = draw->state[i];
-	cpm4 += draw->state[i]->cpm4;
-	if ((ctx->draw_cpm4 + cpm4) > RADEON_CTX_MAX_PM4) {
-		/* need to flush */
-		return -EBUSY;
-	}
-	ctx->draw_cpm4 += cpm4;
-	ctx->nstate = cstate;
-	ctx->draw[ctx->ndraw++] = draw;
-	ctx->cdraw = draw;
 	return 0;
 }
 
-int radeon_ctx_set_draw(struct radeon_ctx *ctx, struct radeon_draw *draw)
-{
-	int r;
-
-	radeon_draw_incref(draw);
-	r = radeon_ctx_set_draw_new(ctx, draw);
-	if (r)
-		radeon_draw_decref(draw);
-	return r;
-}
-
+#if 0
 int radeon_ctx_pm4(struct radeon_ctx *ctx)
 {
 	unsigned i;
@@ -345,9 +294,6 @@ int radeon_ctx_pm4(struct radeon_ctx *ctx)
 	if (ctx->pm4 == NULL)
 		return -EINVAL;
 	for (i = 0, ctx->id = 0; i < ctx->nstate; i++) {
-		r = radeon_ctx_state_schedule(ctx, ctx->state[i]);
-		if (r)
-			return r;
 	}
 	if (ctx->id != ctx->draw_cpm4) {
 		fprintf(stderr, "%s miss predicted pm4 size %d for %d\n",
@@ -357,6 +303,7 @@ int radeon_ctx_pm4(struct radeon_ctx *ctx)
 	ctx->cpm4 = ctx->draw_cpm4;
 	return 0;
 }
+#endif
 
 void radeon_ctx_dump_bof(struct radeon_ctx *ctx, const char *file)
 {
@@ -384,8 +331,8 @@ printf("%d relocs\n", ctx->nreloc);
 	bof_decref(blob);
 	blob = NULL;
 	/* dump cs */
-printf("%d pm4\n", ctx->cpm4);
-	blob = bof_blob(ctx->cpm4 * 4, ctx->pm4);
+printf("%d pm4\n", ctx->cdwords);
+	blob = bof_blob(ctx->cdwords * 4, ctx->pm4);
 	if (blob == NULL)
 		goto out_err;
 	if (bof_object_set(root, "pm4", blob))
diff --git a/src/gallium/winsys/r600/drm/radeon_draw.c b/src/gallium/winsys/r600/drm/radeon_draw.c
index 0eacdc74a03..b992c4a55dc 100644
--- a/src/gallium/winsys/r600/drm/radeon_draw.c
+++ b/src/gallium/winsys/r600/drm/radeon_draw.c
@@ -31,116 +31,27 @@
 /*
  * draw functions
  */
-struct radeon_draw *radeon_draw(struct radeon *radeon)
+int radeon_draw_init(struct radeon_draw *draw, struct radeon *radeon)
 {
-	struct radeon_draw *draw;
-
-	draw = calloc(1, sizeof(*draw));
-	if (draw == NULL)
-		return NULL;
-	draw->nstate = radeon->nstate;
 	draw->radeon = radeon;
-	draw->refcount = 1;
-	draw->state = calloc(1, sizeof(void*) * draw->nstate);
-	if (draw->state == NULL) {
-		free(draw);
-		return NULL;
-	}
-	return draw;
-}
-
-struct radeon_draw *radeon_draw_incref(struct radeon_draw *draw)
-{
-	draw->refcount++;
-	return draw;
-}
-
-struct radeon_draw *radeon_draw_decref(struct radeon_draw *draw)
-{
-	unsigned i;
-
-	if (draw == NULL)
-		return NULL;
-	if (--draw->refcount > 0)
-		return NULL;
-	for (i = 0; i < draw->nstate; i++) {
-		draw->state[i] = radeon_state_decref(draw->state[i]);
-	}
-	free(draw->state);
-	memset(draw, 0, sizeof(*draw));
-	free(draw);
-	return NULL;
-}
-
-int radeon_draw_set_new(struct radeon_draw *draw, struct radeon_state *state)
-{
-	int id;
-	if (state == NULL)
-		return 0;
-
-	id = state->stype->base_id + (state->id + (state->stype->num * state->shader_index));
-	if (id > draw->radeon->nstate)
-	{
-		return -EINVAL;
-	}
-	draw->state[id] = radeon_state_decref(draw->state[id]);
-	draw->state[id] = state;
+	draw->state = calloc(radeon->nstate_per_shader * R600_SHADER_MAX, sizeof(void*));
+	if (draw->state == NULL)
+		return -ENOMEM;
 	return 0;
 }
 
-int radeon_draw_set(struct radeon_draw *draw, struct radeon_state *state)
+void radeon_draw_bind(struct radeon_draw *draw, struct radeon_state *state)
 {
 	if (state == NULL)
-		return 0;
-	radeon_state_incref(state);
-	return radeon_draw_set_new(draw, state);
+		return;
+	draw->state[state->state_id] = state;
 }
 
-int radeon_draw_check(struct radeon_draw *draw)
+void radeon_draw_unbind(struct radeon_draw *draw, struct radeon_state *state)
 {
-	unsigned i;
-	int r;
-
-	r = radeon_draw_pm4(draw);
-	if (r)
-		return r;
-	for (i = 0, draw->cpm4 = 0; i < draw->nstate; i++) {
-		if (draw->state[i]) {
-			draw->cpm4 += draw->state[i]->cpm4;
-		}
-	}
-	return 0;
-}
-
-struct radeon_draw *radeon_draw_duplicate(struct radeon_draw *draw)
-{
-	struct radeon_draw *ndraw;
-	unsigned i;
-
-	if (draw == NULL)
-		return NULL;
-	ndraw = radeon_draw(draw->radeon);
-	if (ndraw == NULL) {
-		return NULL;
-	}
-	for (i = 0; i < draw->nstate; i++) {
-		if (radeon_draw_set(ndraw, draw->state[i])) {
-			radeon_draw_decref(ndraw);
-			return NULL;
-		}
-	}
-	return ndraw;
-}
-
-int radeon_draw_pm4(struct radeon_draw *draw)
-{
-	unsigned i;
-	int r;
-
-	for (i = 0; i < draw->nstate; i++) {
-		r = radeon_state_pm4(draw->state[i]);
-		if (r)
-			return r;
+	if (state == NULL)
+		return;
+	if (draw->state[state->state_id] == state) {
+		draw->state[state->state_id] = NULL;
 	}
-	return 0;
 }
diff --git a/src/gallium/winsys/r600/drm/radeon_priv.h b/src/gallium/winsys/r600/drm/radeon_priv.h
index af5319efd1f..84e552ba4d3 100644
--- a/src/gallium/winsys/r600/drm/radeon_priv.h
+++ b/src/gallium/winsys/r600/drm/radeon_priv.h
@@ -58,9 +58,10 @@ struct radeon {
 	int				refcount;
 	unsigned			device;
 	unsigned			family;
-	unsigned			nstate;
-	unsigned                        nstype;
-	struct radeon_stype_info *stype;
+	unsigned			nstype;
+	unsigned			nstate_per_shader;
+	unsigned			*state_type_id;
+	struct radeon_stype_info	*stype;
 };
 
 extern struct radeon *radeon_new(int fd, unsigned device);
@@ -69,12 +70,6 @@ extern struct radeon *radeon_decref(struct radeon *radeon);
 extern unsigned radeon_family_from_device(unsigned device);
 extern int radeon_is_family_compatible(unsigned family1, unsigned family2);
 
-int radeon_ctx_set_bo_new(struct radeon_ctx *ctx, struct radeon_bo *bo);
-struct radeon_bo *radeon_ctx_get_bo(struct radeon_ctx *ctx, unsigned reloc);
-void radeon_ctx_get_placement(struct radeon_ctx *ctx, unsigned reloc, u32 *placement);
-int radeon_ctx_set_draw_new(struct radeon_ctx *ctx, struct radeon_draw *draw);
-int radeon_ctx_draw(struct radeon_ctx *ctx);
-
 /*
  * r600/r700 context functions
  */
diff --git a/src/gallium/winsys/r600/drm/radeon_state.c b/src/gallium/winsys/r600/drm/radeon_state.c
index d4e622cf7fd..ac60485b280 100644
--- a/src/gallium/winsys/r600/drm/radeon_state.c
+++ b/src/gallium/winsys/r600/drm/radeon_state.c
@@ -32,9 +32,8 @@
 /*
  * state core functions
  */
-struct radeon_state *radeon_state_shader(struct radeon *radeon, u32 stype, u32 id, u32 shader_type)
+int radeon_state_init(struct radeon_state *state, struct radeon *radeon, u32 stype, u32 id, u32 shader_type)
 {
-	struct radeon_state *state;
 	struct radeon_stype_info *found = NULL;
 	int i, j, shader_index = -1;
 
@@ -67,12 +66,11 @@ struct radeon_state *radeon_state_shader(struct radeon *radeon, u32 stype, u32 i
 
 	if (!found) {
 		fprintf(stderr, "%s invalid type %d/id %d/shader class %d\n", __func__, stype, id, shader_type);
-		return NULL;
+		return -EINVAL;
 	}
 
-	state = calloc(1, sizeof(*state));
-	if (state == NULL)
-		return NULL;
+	memset(state, 0, sizeof(struct radeon_state));
+	state->state_id = radeon->nstate_per_shader * shader_index + radeon->state_type_id[stype] + id;
 	state->stype = found;
 	state->radeon = radeon;
 	state->id = id;
@@ -80,7 +78,7 @@ struct radeon_state *radeon_state_shader(struct radeon *radeon, u32 stype, u32 i
 	state->refcount = 1;
 	state->npm4 = found->npm4;
 	state->nstates = found->reginfo[shader_index].nstates;
-	return state;
+	return 0;
 }
 
 int radeon_state_convert(struct radeon_state *state, u32 stype, u32 id, u32 shader_type)
@@ -130,57 +128,20 @@ int radeon_state_convert(struct radeon_state *state, u32 stype, u32 id, u32 shad
 	state->stype = found;
 	state->id = id;
 	state->shader_index = shader_index;
+	state->state_id = state->radeon->nstate_per_shader * shader_index + state->radeon->state_type_id[stype] + id;
 	return radeon_state_pm4(state);
 }
 
-struct radeon_state *radeon_state(struct radeon *radeon, u32 type, u32 id)
-{
-	return radeon_state_shader(radeon, type, id, 0);
-}
-
-struct radeon_state *radeon_state_duplicate(struct radeon_state *state)
-{
-	struct radeon_state *nstate = radeon_state_shader(state->radeon, state->stype->stype, state->id, (1 << state->shader_index));
-	unsigned i;
-
-	if (state == NULL)
-		return NULL;
-	nstate->cpm4 = state->cpm4;
-	nstate->nbo = state->nbo;
-	nstate->nreloc = state->nreloc;
-	memcpy(nstate->states, state->states, state->nstates * 4);
-	memcpy(nstate->pm4, state->pm4, state->npm4 * 4);
-	memcpy(nstate->placement, state->placement, 8 * 4);
-	memcpy(nstate->reloc_pm4_id, state->reloc_pm4_id, 8 * 4);
-	memcpy(nstate->reloc_bo_id, state->reloc_bo_id, 8 * 4);
-	memcpy(nstate->bo_dirty, state->bo_dirty, 4 * 4);
-	for (i = 0; i < state->nbo; i++) {
-		nstate->bo[i] = radeon_bo_incref(state->radeon, state->bo[i]);
-	}
-	return nstate;
-}
-
-struct radeon_state *radeon_state_incref(struct radeon_state *state)
-{
-	state->refcount++;
-	return state;
-}
-
-struct radeon_state *radeon_state_decref(struct radeon_state *state)
+void radeon_state_fini(struct radeon_state *state)
 {
 	unsigned i;
 
 	if (state == NULL)
 		return NULL;
-	if (--state->refcount > 0) {
-		return NULL;
-	}
 	for (i = 0; i < state->nbo; i++) {
 		state->bo[i] = radeon_bo_decref(state->radeon, state->bo[i]);
 	}
-	memset(state, 0, sizeof(*state));
-	free(state);
-	return NULL;
+	memset(state, 0, sizeof(struct radeon_state));
 }
 
 int radeon_state_replace_always(struct radeon_state *ostate,