From 330925d9118120863f5e76e01fdfa1a1f44be2b1 Mon Sep 17 00:00:00 2001
From: Luca Barbieri <luca@luca-barbieri.com>
Date: Tue, 23 Feb 2010 16:48:22 +0100
Subject: [PATCH] nvfx: rework validation logic to use code and avoid stateobjs

This makes the code faster due to the lack of indirect calls and also
makes it much easier to understand what is actually going on.
---
 src/gallium/drivers/nvfx/nvfx_context.c       |   6 -
 src/gallium/drivers/nvfx/nvfx_context.h       |  64 ++-----
 src/gallium/drivers/nvfx/nvfx_draw.c          |  14 +-
 src/gallium/drivers/nvfx/nvfx_fragprog.c      |  13 +-
 src/gallium/drivers/nvfx/nvfx_fragtex.c       |  13 +-
 src/gallium/drivers/nvfx/nvfx_screen.c        |   6 -
 src/gallium/drivers/nvfx/nvfx_screen.h        |   3 -
 src/gallium/drivers/nvfx/nvfx_state.c         |  17 ++
 src/gallium/drivers/nvfx/nvfx_state_blend.c   |  20 +--
 src/gallium/drivers/nvfx/nvfx_state_emit.c    | 156 +++++++++---------
 src/gallium/drivers/nvfx/nvfx_state_fb.c      |  10 +-
 .../drivers/nvfx/nvfx_state_rasterizer.c      |   9 +-
 src/gallium/drivers/nvfx/nvfx_state_scissor.c |  12 +-
 src/gallium/drivers/nvfx/nvfx_state_stipple.c |  13 +-
 .../drivers/nvfx/nvfx_state_viewport.c        |  15 +-
 src/gallium/drivers/nvfx/nvfx_state_zsa.c     |  20 +--
 src/gallium/drivers/nvfx/nvfx_vbo.c           |  10 +-
 src/gallium/drivers/nvfx/nvfx_vertprog.c      |  11 +-
 18 files changed, 127 insertions(+), 285 deletions(-)

diff --git a/src/gallium/drivers/nvfx/nvfx_context.c b/src/gallium/drivers/nvfx/nvfx_context.c
index e7b1486d496..1faa0af31fb 100644
--- a/src/gallium/drivers/nvfx/nvfx_context.c
+++ b/src/gallium/drivers/nvfx/nvfx_context.c
@@ -30,12 +30,6 @@ static void
 nvfx_destroy(struct pipe_context *pipe)
 {
 	struct nvfx_context *nvfx = nvfx_context(pipe);
-	unsigned i;
-
-	for (i = 0; i < NVFX_STATE_MAX; i++) {
-		if (nvfx->state.hw[i])
-			so_ref(NULL, &nvfx->state.hw[i]);
-	}
 
 	if (nvfx->draw)
 		draw_destroy(nvfx->draw);
diff --git a/src/gallium/drivers/nvfx/nvfx_context.h b/src/gallium/drivers/nvfx/nvfx_context.h
index cb9049a448e..9d4489da53d 100644
--- a/src/gallium/drivers/nvfx/nvfx_context.h
+++ b/src/gallium/drivers/nvfx/nvfx_context.h
@@ -25,45 +25,6 @@
 #define NOUVEAU_MSG(fmt, args...) \
 	fprintf(stderr, "nouveau: "fmt, ##args);
 
-enum nvfx_state_index {
-	NVFX_STATE_FB = 0,
-	NVFX_STATE_VIEWPORT = 1,
-	NVFX_STATE_BLEND = 2,
-	NVFX_STATE_RAST = 3,
-	NVFX_STATE_ZSA = 4,
-	NVFX_STATE_BCOL = 5,
-	NVFX_STATE_CLIP = 6,
-	NVFX_STATE_SCISSOR = 7,
-	NVFX_STATE_STIPPLE = 8,
-	NVFX_STATE_FRAGPROG = 9,
-	NVFX_STATE_VERTPROG = 10,
-	NVFX_STATE_FRAGTEX0 = 11,
-	NVFX_STATE_FRAGTEX1 = 12,
-	NVFX_STATE_FRAGTEX2 = 13,
-	NVFX_STATE_FRAGTEX3 = 14,
-	NVFX_STATE_FRAGTEX4 = 15,
-	NVFX_STATE_FRAGTEX5 = 16,
-	NVFX_STATE_FRAGTEX6 = 17,
-	NVFX_STATE_FRAGTEX7 = 18,
-	NVFX_STATE_FRAGTEX8 = 19,
-	NVFX_STATE_FRAGTEX9 = 20,
-	NVFX_STATE_FRAGTEX10 = 21,
-	NVFX_STATE_FRAGTEX11 = 22,
-	NVFX_STATE_FRAGTEX12 = 23,
-	NVFX_STATE_FRAGTEX13 = 24,
-	NVFX_STATE_FRAGTEX14 = 25,
-	NVFX_STATE_FRAGTEX15 = 26,
-	NVFX_STATE_VERTTEX0 = 27,
-	NVFX_STATE_VERTTEX1 = 28,
-	NVFX_STATE_VERTTEX2 = 29,
-	NVFX_STATE_VERTTEX3 = 30,
-	NVFX_STATE_VTXBUF = 31,
-	NVFX_STATE_VTXFMT = 32,
-	NVFX_STATE_VTXATTR = 33,
-	NVFX_STATE_SR = 34,
-	NVFX_STATE_MAX = 35
-};
-
 #include "nvfx_screen.h"
 
 #define NVFX_NEW_BLEND		(1 <<  0)
@@ -106,9 +67,6 @@ struct nvfx_state {
 	unsigned scissor_enabled;
 	unsigned stipple_enabled;
 	unsigned fp_samplers;
-
-	uint64_t dirty;
-	struct nouveau_stateobj *hw[NVFX_STATE_MAX];
 };
 
 struct nvfx_vtxelt_state {
@@ -147,7 +105,6 @@ struct nvfx_context {
 		HW, SWTNL, SWRAST
 	} render_mode;
 	unsigned fallback_swtnl;
-	unsigned fallback_swrast;
 
 	/* Context state */
 	unsigned dirty, draw_dirty;
@@ -190,14 +147,6 @@ nvfx_context(struct pipe_context *pipe)
 	return (struct nvfx_context *)pipe;
 }
 
-struct nvfx_state_entry {
-	boolean (*validate)(struct nvfx_context *nvfx);
-	struct {
-		unsigned pipe;
-		unsigned hw;
-	} dirty;
-};
-
 extern struct nvfx_state_entry nvfx_state_blend;
 extern struct nvfx_state_entry nvfx_state_blend_colour;
 extern struct nvfx_state_entry nvfx_state_fragprog;
@@ -230,18 +179,22 @@ extern void nvfx_draw_elements_swtnl(struct pipe_context *pipe,
 					struct pipe_resource *idxbuf,
 					unsigned ib_size, unsigned mode,
 					unsigned start, unsigned count);
+extern void nvfx_vtxfmt_validate(struct nvfx_context *nvfx);
 
 /* nvfx_fb.c */
+extern void nvfx_state_framebuffer_validate(struct nvfx_context *nvfx);
 void
 nvfx_framebuffer_relocate(struct nvfx_context *nvfx);
 
 /* nvfx_fragprog.c */
 extern void nvfx_fragprog_destroy(struct nvfx_context *,
 				    struct nvfx_fragment_program *);
+extern void nvfx_fragprog_validate(struct nvfx_context *nvfx);
 extern void
 nvfx_fragprog_relocate(struct nvfx_context *nvfx);
 
 /* nvfx_fragtex.c */
+extern void nvfx_fragtex_validate(struct nvfx_context *nvfx);
 extern void
 nvfx_fragtex_relocate(struct nvfx_context *nvfx);
 
@@ -261,6 +214,14 @@ extern void nv40_fragtex_set(struct nvfx_context *nvfx, int unit);
 
 /* nvfx_state.c */
 extern void nvfx_init_state_functions(struct nvfx_context *nvfx);
+extern void nvfx_state_scissor_validate(struct nvfx_context *nvfx);
+extern void nvfx_state_stipple_validate(struct nvfx_context *nvfx);
+extern void nvfx_state_blend_validate(struct nvfx_context *nvfx);
+extern void nvfx_state_blend_colour_validate(struct nvfx_context *nvfx);
+extern void nvfx_state_viewport_validate(struct nvfx_context *nvfx);
+extern void nvfx_state_rasterizer_validate(struct nvfx_context *nvfx);
+extern void nvfx_state_sr_validate(struct nvfx_context *nvfx);
+extern void nvfx_state_zsa_validate(struct nvfx_context *nvfx);
 
 /* nvfx_state_emit.c */
 extern void nvfx_state_relocate(struct nvfx_context *nvfx);
@@ -283,6 +244,7 @@ extern void nvfx_draw_elements(struct pipe_context *pipe,
 				  unsigned count);
 
 /* nvfx_vertprog.c */
+extern boolean nvfx_vertprog_validate(struct nvfx_context *nvfx);
 extern void nvfx_vertprog_destroy(struct nvfx_context *,
 				  struct nvfx_vertex_program *);
 
diff --git a/src/gallium/drivers/nvfx/nvfx_draw.c b/src/gallium/drivers/nvfx/nvfx_draw.c
index 06f5dc39063..2003be10204 100644
--- a/src/gallium/drivers/nvfx/nvfx_draw.c
+++ b/src/gallium/drivers/nvfx/nvfx_draw.c
@@ -303,8 +303,8 @@ emit_attrib(struct nvfx_context *nvfx, unsigned hw, unsigned emit,
 	nvfx->swtnl.draw[a] = draw_out;
 }
 
-static boolean
-nvfx_state_vtxfmt_validate(struct nvfx_context *nvfx)
+void
+nvfx_vtxfmt_validate(struct nvfx_context *nvfx)
 {
 	struct nvfx_fragment_program *fp = nvfx->fragprog;
 	unsigned colour = 0, texcoords = 0, fog = 0, i;
@@ -348,14 +348,4 @@ nvfx_state_vtxfmt_validate(struct nvfx_context *nvfx)
 	}
 
 	emit_attrib(nvfx, 0, 0xff, TGSI_SEMANTIC_POSITION, 0);
-
-	return FALSE;
 }
-
-struct nvfx_state_entry nvfx_state_vtxfmt = {
-	.validate = nvfx_state_vtxfmt_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_ARRAYS | NVFX_NEW_FRAGPROG,
-		.hw = 0
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_fragprog.c b/src/gallium/drivers/nvfx/nvfx_fragprog.c
index eb815302485..301ad82c08b 100644
--- a/src/gallium/drivers/nvfx/nvfx_fragprog.c
+++ b/src/gallium/drivers/nvfx/nvfx_fragprog.c
@@ -864,7 +864,7 @@ nvfx_fragprog_upload(struct nvfx_context *nvfx,
 	}
 }
 
-boolean
+void
 nvfx_fragprog_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
@@ -878,7 +878,6 @@ nvfx_fragprog_validate(struct nvfx_context *nvfx)
 	if (fp->translated)
 		goto update_constants;
 
-	nvfx->fallback_swrast &= ~NVFX_NEW_FRAGPROG;
 	nvfx_fragprog_translate(nvfx, fp);
 	if (!fp->translated) {
 		static unsigned dummy[8] = {1, 0, 0, 0, 1, 0, 0, 0};
@@ -894,7 +893,7 @@ nvfx_fragprog_validate(struct nvfx_context *nvfx)
 		fp->insn = malloc(sizeof(dummy));
 		memcpy(fp->insn, dummy, sizeof(dummy));
 		fp->insn_len = sizeof(dummy) / sizeof(dummy[0]);
-		return FALSE;
+		return;
 	}
 
 	fp->buffer = pipe_buffer_create(pscreen,
@@ -948,7 +947,6 @@ update_constants:
 		OUT_RING(chan, RING_3D(NV34TCL_TX_UNITS_ENABLE, 1));
 		OUT_RING(chan, fp->samplers);
 	}
-	return TRUE;
 }
 
 void
@@ -977,10 +975,3 @@ nvfx_fragprog_destroy(struct nvfx_context *nvfx,
 		FREE(fp->insn);
 }
 
-struct nvfx_state_entry nvfx_state_fragprog = {
-	.validate = nvfx_fragprog_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_FRAGPROG | NVFX_NEW_FRAGCONST,
-		.hw = 0
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_fragtex.c b/src/gallium/drivers/nvfx/nvfx_fragtex.c
index e239235c3f5..f5f6b0c0cbd 100644
--- a/src/gallium/drivers/nvfx/nvfx_fragtex.c
+++ b/src/gallium/drivers/nvfx/nvfx_fragtex.c
@@ -1,7 +1,7 @@
 #include "nvfx_context.h"
 #include "nvfx_resource.h"
 
-static boolean
+void
 nvfx_fragtex_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
@@ -9,7 +9,7 @@ nvfx_fragtex_validate(struct nvfx_context *nvfx)
 
 	samplers = nvfx->dirty_samplers;
 	if(!samplers)
-		return FALSE;
+		return;
 
 	while (samplers) {
 		unit = ffs(samplers) - 1;
@@ -29,7 +29,6 @@ nvfx_fragtex_validate(struct nvfx_context *nvfx)
 		}
 	}
 	nvfx->dirty_samplers = 0;
-	return FALSE;
 }
 
 void
@@ -54,11 +53,3 @@ nvfx_fragtex_relocate(struct nvfx_context *nvfx)
 				NV34TCL_TX_FORMAT_DMA0, NV34TCL_TX_FORMAT_DMA1);
 	}
 }
-
-struct nvfx_state_entry nvfx_state_fragtex = {
-	.validate = nvfx_fragtex_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_SAMPLER | NVFX_NEW_FRAGPROG,
-		.hw = 0
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_screen.c b/src/gallium/drivers/nvfx/nvfx_screen.c
index 5bf921db99c..893086cae85 100644
--- a/src/gallium/drivers/nvfx/nvfx_screen.c
+++ b/src/gallium/drivers/nvfx/nvfx_screen.c
@@ -179,12 +179,6 @@ static void
 nvfx_screen_destroy(struct pipe_screen *pscreen)
 {
 	struct nvfx_screen *screen = nvfx_screen(pscreen);
-	unsigned i;
-
-	for (i = 0; i < NVFX_STATE_MAX; i++) {
-		if (screen->state[i])
-			so_ref(NULL, &screen->state[i]);
-	}
 
 	nouveau_resource_destroy(&screen->vp_exec_heap);
 	nouveau_resource_destroy(&screen->vp_data_heap);
diff --git a/src/gallium/drivers/nvfx/nvfx_screen.h b/src/gallium/drivers/nvfx/nvfx_screen.h
index 3302e1aa586..14bd6339206 100644
--- a/src/gallium/drivers/nvfx/nvfx_screen.h
+++ b/src/gallium/drivers/nvfx/nvfx_screen.h
@@ -27,9 +27,6 @@ struct nvfx_screen {
 	/* Vtxprog resources */
 	struct nouveau_resource *vp_exec_heap;
 	struct nouveau_resource *vp_data_heap;
-
-	/* Current 3D state of channel */
-	struct nouveau_stateobj *state[NVFX_STATE_MAX];
 };
 
 static INLINE struct nvfx_screen *
diff --git a/src/gallium/drivers/nvfx/nvfx_state.c b/src/gallium/drivers/nvfx/nvfx_state.c
index 4053f28a651..0f03c2eb3d5 100644
--- a/src/gallium/drivers/nvfx/nvfx_state.c
+++ b/src/gallium/drivers/nvfx/nvfx_state.c
@@ -303,6 +303,23 @@ nvfx_rasterizer_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nvfx_context *nvfx = nvfx_context(pipe);
 
+	if(nvfx->rasterizer && hwcso)
+	{
+		if(!nvfx->rasterizer || ((struct nvfx_rasterizer_state*)hwcso)->pipe.scissor
+					!= nvfx->rasterizer->pipe.scissor)
+		{
+			nvfx->dirty |= NVFX_NEW_SCISSOR;
+			nvfx->draw_dirty |= NVFX_NEW_SCISSOR;
+		}
+
+		if(((struct nvfx_rasterizer_state*)hwcso)->pipe.poly_stipple_enable
+					!= nvfx->rasterizer->pipe.poly_stipple_enable)
+		{
+			nvfx->dirty |= NVFX_NEW_STIPPLE;
+			nvfx->draw_dirty |= NVFX_NEW_STIPPLE;
+		}
+	}
+
 	nvfx->rasterizer = hwcso;
 	nvfx->dirty |= NVFX_NEW_RAST;
 	nvfx->draw_dirty |= NVFX_NEW_RAST;
diff --git a/src/gallium/drivers/nvfx/nvfx_state_blend.c b/src/gallium/drivers/nvfx/nvfx_state_blend.c
index f54eabd6428..fe34e98364c 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_blend.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_blend.c
@@ -1,21 +1,13 @@
 #include "nvfx_context.h"
 
-static boolean
+void
 nvfx_state_blend_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
 	sb_emit(chan, nvfx->blend->sb, nvfx->blend->sb_len);
-	return TRUE;
 }
 
-struct nvfx_state_entry nvfx_state_blend = {
-	.validate = nvfx_state_blend_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_BLEND,
-	}
-};
-
-static boolean
+void
 nvfx_state_blend_colour_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
@@ -27,12 +19,4 @@ nvfx_state_blend_colour_validate(struct nvfx_context *nvfx)
 		       (float_to_ubyte(bcol->color[0]) << 16) |
 		       (float_to_ubyte(bcol->color[1]) <<  8) |
 		       (float_to_ubyte(bcol->color[2]) <<  0)));
-	return TRUE;
 }
-
-struct nvfx_state_entry nvfx_state_blend_colour = {
-	.validate = nvfx_state_blend_colour_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_BCOL,
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_state_emit.c b/src/gallium/drivers/nvfx/nvfx_state_emit.c
index 2bc7927a697..4137849bf0b 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_emit.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_emit.c
@@ -2,86 +2,92 @@
 #include "nvfx_state.h"
 #include "draw/draw_context.h"
 
-#define RENDER_STATES(name, vbo) \
-static struct nvfx_state_entry *name##render_states[] = { \
-	&nvfx_state_framebuffer, \
-	&nvfx_state_rasterizer, \
-	&nvfx_state_scissor, \
-	&nvfx_state_stipple, \
-	&nvfx_state_fragprog, \
-	&nvfx_state_fragtex, \
-	&nvfx_state_vertprog, \
-	&nvfx_state_blend, \
-	&nvfx_state_blend_colour, \
-	&nvfx_state_zsa, \
-	&nvfx_state_sr, \
-	&nvfx_state_viewport, \
-	&nvfx_state_##vbo, \
-	NULL \
-}
-
-RENDER_STATES(, vbo);
-RENDER_STATES(swtnl_, vtxfmt);
-
-static void
-nvfx_state_do_validate(struct nvfx_context *nvfx,
-		       struct nvfx_state_entry **states)
+static boolean
+nvfx_state_validate_common(struct nvfx_context *nvfx)
 {
-	while (*states) {
-		struct nvfx_state_entry *e = *states;
-
-		if (nvfx->dirty & e->dirty.pipe) {
-			if (e->validate(nvfx))
-				nvfx->state.dirty |= (1ULL << e->dirty.hw);
+	struct nouveau_channel* chan = nvfx->screen->base.channel;
+	unsigned dirty = nvfx->dirty;
+
+	if(nvfx != nvfx->screen->cur_ctx)
+		dirty = ~0;
+
+	if(nvfx->render_mode == HW)
+	{
+		if(dirty & (NVFX_NEW_VERTPROG | NVFX_NEW_VERTCONST | NVFX_NEW_UCP))
+		{
+			if(!nvfx_vertprog_validate(nvfx))
+				return FALSE;
 		}
 
-		states++;
+		if(dirty & (NVFX_NEW_ARRAYS))
+		{
+			if(!nvfx_vbo_validate(nvfx))
+				return FALSE;
+		}
+	}
+	else
+	{
+		/* TODO: this looks a bit misdesigned */
+		if(dirty & (NVFX_NEW_VERTPROG | NVFX_NEW_UCP))
+			nvfx_vertprog_validate(nvfx);
+
+		if(dirty & (NVFX_NEW_ARRAYS | NVFX_NEW_FRAGPROG))
+			nvfx_vtxfmt_validate(nvfx);
 	}
-	nvfx->dirty = 0;
-}
 
-void
-nvfx_state_emit(struct nvfx_context *nvfx)
-{
-	struct nvfx_state *state = &nvfx->state;
-	struct nvfx_screen *screen = nvfx->screen;
-	struct nouveau_channel *chan = screen->base.channel;
-	struct nouveau_grobj *eng3d = screen->eng3d;
-	unsigned i;
-	uint64_t states;
-
-	/* XXX: race conditions
-	 */
-	if (nvfx != screen->cur_ctx) {
-		for (i = 0; i < NVFX_STATE_MAX; i++) {
-			if (state->hw[i] && screen->state[i] != state->hw[i])
-				state->dirty |= (1ULL << i);
-		}
+	if(dirty & NVFX_NEW_FB)
+		nvfx_state_framebuffer_validate(nvfx);
 
-		screen->cur_ctx = nvfx;
-	}
+	if(dirty & NVFX_NEW_RAST)
+		sb_emit(chan, nvfx->rasterizer->sb, nvfx->rasterizer->sb_len);
 
-	for (i = 0, states = state->dirty; states; i++) {
-		if (!(states & (1ULL << i)))
-			continue;
-		so_ref (state->hw[i], &nvfx->screen->state[i]);
-		if (state->hw[i])
-			so_emit(chan, nvfx->screen->state[i]);
-		states &= ~(1ULL << i);
-	}
+	if(dirty & NVFX_NEW_SCISSOR)
+		nvfx_state_scissor_validate(nvfx);
+
+	if(dirty & NVFX_NEW_STIPPLE)
+		nvfx_state_stipple_validate(nvfx);
+
+	if(dirty & (NVFX_NEW_FRAGPROG | NVFX_NEW_FRAGCONST))
+		nvfx_fragprog_validate(nvfx);
+
+	if(dirty & NVFX_NEW_SAMPLER)
+		nvfx_fragtex_validate(nvfx);
+
+	if(dirty & NVFX_NEW_BLEND)
+		sb_emit(chan, nvfx->blend->sb, nvfx->blend->sb_len);
+
+	if(dirty & NVFX_NEW_BCOL)
+		nvfx_state_blend_colour_validate(nvfx);
+
+	if(dirty & NVFX_NEW_ZSA)
+		sb_emit(chan, nvfx->zsa->sb, nvfx->zsa->sb_len);
+
+	if(dirty & NVFX_NEW_SR)
+		nvfx_state_sr_validate(nvfx);
+
+/* Having this depend on FB looks wrong, but it seems
+   necessary to make this work on nv3x
+   TODO: find the right fix
+*/
+	if(dirty & (NVFX_NEW_VIEWPORT | NVFX_NEW_FB))
+		nvfx_state_viewport_validate(nvfx);
 
 	/* TODO: could nv30 need this or something similar too? */
-	if(nvfx->is_nv4x) {
-		if (state->dirty & ((1ULL << NVFX_STATE_FRAGPROG) |
-				    (1ULL << NVFX_STATE_FRAGTEX0))) {
-			BEGIN_RING(chan, eng3d, NV40TCL_TEX_CACHE_CTL, 1);
-			OUT_RING  (chan, 2);
-			BEGIN_RING(chan, eng3d, NV40TCL_TEX_CACHE_CTL, 1);
-			OUT_RING  (chan, 1);
-		}
+	if((dirty & (NVFX_NEW_FRAGPROG | NVFX_NEW_SAMPLER)) && nvfx->is_nv4x) {
+		WAIT_RING(chan, 4);
+		OUT_RING(chan, RING_3D(NV40TCL_TEX_CACHE_CTL, 1));
+		OUT_RING(chan, 2);
+		OUT_RING(chan, RING_3D(NV40TCL_TEX_CACHE_CTL, 1));
+		OUT_RING(chan, 1);
 	}
-	state->dirty = 0;
+	nvfx->dirty = 0;
+	return TRUE;
+}
 
+void
+nvfx_state_emit(struct nvfx_context *nvfx)
+{
+	struct nouveau_channel* chan = nvfx->screen->base.channel;
 	/* we need to ensure there is enough space to output relocations in one go */
 	unsigned max_relocs = 0
 	      + 16 /* vertex buffers, incl. dma flag */
@@ -119,16 +125,13 @@ nvfx_state_validate(struct nvfx_context *nvfx)
 			return FALSE;
 
 		/* Attempt to go to hwtnl again */
-		nvfx->pipe.flush(&nvfx->pipe, 0, NULL);
 		nvfx->dirty |= (NVFX_NEW_VIEWPORT |
 				NVFX_NEW_VERTPROG |
 				NVFX_NEW_ARRAYS);
 		nvfx->render_mode = HW;
 	}
 
-	nvfx_state_do_validate(nvfx, render_states);
-
-	if (nvfx->fallback_swtnl || nvfx->fallback_swrast)
+	if(!nvfx_state_validate_common(nvfx))
 		return FALSE;
 
 	if (was_sw)
@@ -169,12 +172,7 @@ nvfx_state_validate_swtnl(struct nvfx_context *nvfx)
 		draw_set_vertex_elements(draw, nvfx->vtxelt->num_elements, nvfx->vtxelt->pipe);
 	}
 
-	nvfx_state_do_validate(nvfx, swtnl_render_states);
-
-	if (nvfx->fallback_swrast) {
-		NOUVEAU_ERR("swtnl->swrast 0x%08x\n", nvfx->fallback_swrast);
-		return FALSE;
-	}
+	nvfx_state_validate_common(nvfx);
 
 	nvfx->draw_dirty = 0;
 	return TRUE;
diff --git a/src/gallium/drivers/nvfx/nvfx_state_fb.c b/src/gallium/drivers/nvfx/nvfx_state_fb.c
index 93bc0a3feef..8c215980e23 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_fb.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_fb.c
@@ -4,7 +4,7 @@
 
 
 
-static void
+void
 nvfx_state_framebuffer_validate(struct nvfx_context *nvfx)
 {
 	struct pipe_framebuffer_state *fb = &nvfx->framebuffer;
@@ -248,11 +248,3 @@ nvfx_framebuffer_relocate(struct nvfx_context *nvfx)
 
 	DO_(nvfx->hw_zeta, NV34, ZETA);
 }
-
-struct nvfx_state_entry nvfx_state_framebuffer = {
-	.validate = nvfx_state_framebuffer_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_FB,
-		.hw = NVFX_STATE_FB
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_state_rasterizer.c b/src/gallium/drivers/nvfx/nvfx_state_rasterizer.c
index a5d21e88b23..7f14ae85d5a 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_rasterizer.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_rasterizer.c
@@ -1,16 +1,9 @@
 #include "nvfx_context.h"
 
-static boolean
+void
 nvfx_state_rasterizer_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
 	sb_emit(chan, nvfx->rasterizer->sb, nvfx->rasterizer->sb_len);
-	return TRUE;
 }
 
-struct nvfx_state_entry nvfx_state_rasterizer = {
-	.validate = nvfx_state_rasterizer_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_RAST,
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_state_scissor.c b/src/gallium/drivers/nvfx/nvfx_state_scissor.c
index dd6ae757df7..9077266120a 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_scissor.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_scissor.c
@@ -1,6 +1,6 @@
 #include "nvfx_context.h"
 
-static boolean
+void
 nvfx_state_scissor_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel *chan = nvfx->screen->base.channel;
@@ -8,7 +8,7 @@ nvfx_state_scissor_validate(struct nvfx_context *nvfx)
 	struct pipe_scissor_state *s = &nvfx->scissor;
 
 	if ((rast->scissor == 0 && nvfx->state.scissor_enabled == 0))
-		return FALSE;
+		return;
 	nvfx->state.scissor_enabled = rast->scissor;
 
 	WAIT_RING(chan, 3);
@@ -20,12 +20,4 @@ nvfx_state_scissor_validate(struct nvfx_context *nvfx)
 		OUT_RING(chan, 4096 << 16);
 		OUT_RING(chan, 4096 << 16);
 	}
-	return TRUE;
 }
-
-struct nvfx_state_entry nvfx_state_scissor = {
-	.validate = nvfx_state_scissor_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_SCISSOR | NVFX_NEW_RAST,
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_state_stipple.c b/src/gallium/drivers/nvfx/nvfx_state_stipple.c
index 828113c4dfe..4da968f093f 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_stipple.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_stipple.c
@@ -1,13 +1,13 @@
 #include "nvfx_context.h"
 
-static boolean
+void
 nvfx_state_stipple_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel *chan = nvfx->screen->base.channel;
 	struct pipe_rasterizer_state *rast = &nvfx->rasterizer->pipe;
 
 	if ((rast->poly_stipple_enable == 0 && nvfx->state.stipple_enabled == 0))
-		return FALSE;
+		return;
 
 	if (rast->poly_stipple_enable) {
 		unsigned i;
@@ -23,13 +23,4 @@ nvfx_state_stipple_validate(struct nvfx_context *nvfx)
 		OUT_RING(chan, RING_3D(NV34TCL_POLYGON_STIPPLE_ENABLE, 1));
 		OUT_RING(chan, 0);
 	}
-
-	return TRUE;
 }
-
-struct nvfx_state_entry nvfx_state_stipple = {
-	.validate = nvfx_state_stipple_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_STIPPLE | NVFX_NEW_RAST,
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_state_viewport.c b/src/gallium/drivers/nvfx/nvfx_state_viewport.c
index 73db0984cdb..e983b16f321 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_viewport.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_viewport.c
@@ -1,11 +1,6 @@
 #include "nvfx_context.h"
 
-/* Having this depend on FB and RAST looks wrong, but it seems
-   necessary to make this work on nv3x
-   TODO: find the right fix
-*/
-
-static boolean
+void
 nvfx_state_viewport_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel *chan = nvfx->screen->base.channel;
@@ -37,12 +32,4 @@ nvfx_state_viewport_validate(struct nvfx_context *nvfx)
 		OUT_RING(chan, RING_3D(0x1d78, 1));
 		OUT_RING(chan, nvfx->is_nv4x ? 0x110 : 1);
 	}
-	return FALSE;
 }
-
-struct nvfx_state_entry nvfx_state_viewport = {
-	.validate = nvfx_state_viewport_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_VIEWPORT | NVFX_NEW_FB | NVFX_NEW_RAST,
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_state_zsa.c b/src/gallium/drivers/nvfx/nvfx_state_zsa.c
index ba2261c4228..608605d32bd 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_zsa.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_zsa.c
@@ -1,21 +1,13 @@
 #include "nvfx_context.h"
 
-static boolean
+void
 nvfx_state_zsa_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
 	sb_emit(chan, nvfx->zsa->sb, nvfx->zsa->sb_len);
-	return TRUE;
 }
 
-struct nvfx_state_entry nvfx_state_zsa = {
-	.validate = nvfx_state_zsa_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_ZSA,
-	}
-};
-
-static boolean
+void
 nvfx_state_sr_validate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
@@ -26,12 +18,4 @@ nvfx_state_sr_validate(struct nvfx_context *nvfx)
 	OUT_RING(chan, sr->ref_value[0]);
 	OUT_RING(chan, RING_3D(NV34TCL_STENCIL_BACK_FUNC_REF, 1));
 	OUT_RING(chan, sr->ref_value[1]);
-	return TRUE;
 }
-
-struct nvfx_state_entry nvfx_state_sr = {
-	.validate = nvfx_state_sr_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_SR,
-	}
-};
diff --git a/src/gallium/drivers/nvfx/nvfx_vbo.c b/src/gallium/drivers/nvfx/nvfx_vbo.c
index 49f0cba9688..74de3a799e1 100644
--- a/src/gallium/drivers/nvfx/nvfx_vbo.c
+++ b/src/gallium/drivers/nvfx/nvfx_vbo.c
@@ -575,17 +575,9 @@ nvfx_vbo_validate(struct nvfx_context *nvfx)
 	}
 
 	nvfx->hw_vtxelt_nr = nvfx->vtxelt->num_elements;
-	return FALSE;
+	return TRUE;
 }
 
-struct nvfx_state_entry nvfx_state_vbo = {
-	.validate = nvfx_vbo_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_ARRAYS,
-		.hw = 0,
-	}
-};
-
 void
 nvfx_vbo_relocate(struct nvfx_context *nvfx)
 {
diff --git a/src/gallium/drivers/nvfx/nvfx_vertprog.c b/src/gallium/drivers/nvfx/nvfx_vertprog.c
index 4637212b18e..6b847f3cbfd 100644
--- a/src/gallium/drivers/nvfx/nvfx_vertprog.c
+++ b/src/gallium/drivers/nvfx/nvfx_vertprog.c
@@ -830,7 +830,7 @@ out_err:
 	FREE(vpc);
 }
 
-static boolean
+boolean
 nvfx_vertprog_validate(struct nvfx_context *nvfx)
 {
 	struct pipe_context *pipe = &nvfx->pipe;
@@ -1007,7 +1007,7 @@ nvfx_vertprog_validate(struct nvfx_context *nvfx)
 		OUT_RING(chan, vp->clip_ctrl);
 	}
 
-	return FALSE;
+	return TRUE;
 }
 
 void
@@ -1035,10 +1035,3 @@ nvfx_vertprog_destroy(struct nvfx_context *nvfx, struct nvfx_vertex_program *vp)
 
 	vp->ir = vp->or = vp->clip_ctrl = 0;
 }
-
-struct nvfx_state_entry nvfx_state_vertprog = {
-	.validate = nvfx_vertprog_validate,
-	.dirty = {
-		.pipe = NVFX_NEW_VERTPROG | NVFX_NEW_VERTCONST | NVFX_NEW_UCP,
-	}
-};
-- 
2.30.2