diff --git a/src/gallium/drivers/nv40/nv40_context.h b/src/gallium/drivers/nv40/nv40_context.h
index 432204b8252..ce0933fea4a 100644
--- a/src/gallium/drivers/nv40/nv40_context.h
+++ b/src/gallium/drivers/nv40/nv40_context.h
@@ -37,6 +37,7 @@
 #define NV40_NEW_UCP		(1 << 12)
 
 #define NV40_FALLBACK_TNL (1 << 0)
+#define NV40_FALLBACK_RAST (1 << 1)
 
 struct nv40_channel_context {
 	struct nouveau_winsys *nvws;
@@ -72,6 +73,8 @@ struct nv40_state {
 		unsigned enabled;
 		struct nouveau_stateobj *so;
 	} stipple;
+
+	struct nouveau_stateobj *fragprog;
 };
 
 struct nv40_context {
@@ -96,6 +99,8 @@ struct nv40_context {
 		struct pipe_scissor_state scissor;
 		unsigned stipple[32];
 		struct pipe_clip_state clip;
+		struct nv40_fragment_program *fragprog;
+		struct pipe_buffer *constbuf[PIPE_SHADER_TYPES];
 	} pipe_state;
 
 	struct nv40_state state;
@@ -119,13 +124,6 @@ struct nv40_context {
 		struct pipe_buffer *constant_buf;
 	} vertprog;
 
-	struct {
-		struct nv40_fragment_program *active;
-
-		struct nv40_fragment_program *current;
-		struct pipe_buffer *constant_buf;
-	} fragprog;
-
 	struct pipe_vertex_buffer  vtxbuf[PIPE_ATTRIB_MAX];
 	struct pipe_vertex_element vtxelt[PIPE_ATTRIB_MAX];
 };
@@ -161,10 +159,6 @@ extern void nv40_vertprog_destroy(struct nv40_context *,
 				  struct nv40_vertex_program *);
 
 /* nv40_fragprog.c */
-extern void nv40_fragprog_translate(struct nv40_context *,
-				    struct nv40_fragment_program *);
-extern void nv40_fragprog_bind(struct nv40_context *,
-			       struct nv40_fragment_program *);
 extern void nv40_fragprog_destroy(struct nv40_context *,
 				  struct nv40_fragment_program *);
 
@@ -177,6 +171,7 @@ extern void nv40_state_tex_update(struct nv40_context *nv40);
 extern struct nv40_state_entry nv40_state_clip;
 extern struct nv40_state_entry nv40_state_scissor;
 extern struct nv40_state_entry nv40_state_stipple;
+extern struct nv40_state_entry nv40_state_fragprog;
 
 /* nv40_vbo.c */
 extern boolean nv40_draw_arrays(struct pipe_context *, unsigned mode,
diff --git a/src/gallium/drivers/nv40/nv40_fragprog.c b/src/gallium/drivers/nv40/nv40_fragprog.c
index 07a418c1e9f..bfc75eb4620 100644
--- a/src/gallium/drivers/nv40/nv40_fragprog.c
+++ b/src/gallium/drivers/nv40/nv40_fragprog.c
@@ -668,7 +668,7 @@ nv40_fragprog_parse_decl_output(struct nv40_fpc *fpc,
 	return TRUE;
 }
 
-void
+static void
 nv40_fragprog_translate(struct nv40_context *nv40,
 			struct nv40_fragment_program *fp)
 {
@@ -750,73 +750,67 @@ nv40_fragprog_translate(struct nv40_context *nv40,
 	fp->insn[fpc->inst_offset + 3] = 0x00000000;
 	
 	fp->translated = TRUE;
-	fp->on_hw = FALSE;
 out_err:
 	tgsi_parse_free(&parse);
 	free(fpc);
 }
 
-void
-nv40_fragprog_bind(struct nv40_context *nv40, struct nv40_fragment_program *fp)
+static void
+nv40_fragprog_upload(struct nv40_context *nv40,
+		     struct nv40_fragment_program *fp)
 {
 	struct pipe_winsys *ws = nv40->pipe.winsys;
-	struct nouveau_stateobj *so;
+	const uint32_t le = 1;
+	uint32_t *map;
 	int i;
 
-	if (!fp->translated) {
-		nv40_fragprog_translate(nv40, fp);
-		if (!fp->translated)
-			assert(0);
-	}
-
-	if (fp->nr_consts) {
-		float *map = ws->buffer_map(ws, nv40->fragprog.constant_buf,
-					    PIPE_BUFFER_USAGE_CPU_READ);
-		for (i = 0; i < fp->nr_consts; i++) {
-			struct nv40_fragment_program_data *fpd = &fp->consts[i];
-			uint32_t *p = &fp->insn[fpd->offset];
-			uint32_t *cb = (uint32_t *)&map[fpd->index * 4];
-
-			if (!memcmp(p, cb, 4 * sizeof(float)))
-				continue;
-			memcpy(p, cb, 4 * sizeof(float));
-			fp->on_hw = 0;
-		}
-		ws->buffer_unmap(ws, nv40->fragprog.constant_buf);
-	}
-
-	if (!fp->on_hw) {
-		const uint32_t le = 1;
-		uint32_t *map;
-
-		if (!fp->buffer)
-			fp->buffer = ws->buffer_create(ws, 0x100, 0,
-						       fp->insn_len * 4);
-		map = ws->buffer_map(ws, fp->buffer,
-				     PIPE_BUFFER_USAGE_CPU_WRITE);
+	map = ws->buffer_map(ws, fp->buffer, PIPE_BUFFER_USAGE_CPU_WRITE);
 
 #if 0
-		for (i = 0; i < fp->insn_len; i++) {
-			NOUVEAU_ERR("%d 0x%08x\n", i, fp->insn[i]);
-		}
+	for (i = 0; i < fp->insn_len; i++) {
+		NOUVEAU_ERR("%d 0x%08x\n", i, fp->insn[i]);
+	}
 #endif
 
-		if ((*(const uint8_t *)&le)) {
-			for (i = 0; i < fp->insn_len; i++) {
-				map[i] = fp->insn[i];
-			}
-		} else {
-			/* Weird swapping for big-endian chips */
-			for (i = 0; i < fp->insn_len; i++) {
-				map[i] = ((fp->insn[i] & 0xffff) << 16) |
-					  ((fp->insn[i] >> 16) & 0xffff);
-			}
+	if ((*(const uint8_t *)&le)) {
+		for (i = 0; i < fp->insn_len; i++) {
+			map[i] = fp->insn[i];
+		}
+	} else {
+		/* Weird swapping for big-endian chips */
+		for (i = 0; i < fp->insn_len; i++) {
+			map[i] = ((fp->insn[i] & 0xffff) << 16) |
+				  ((fp->insn[i] >> 16) & 0xffff);
 		}
-
-		ws->buffer_unmap(ws, fp->buffer);
-		fp->on_hw = TRUE;
 	}
 
+	ws->buffer_unmap(ws, fp->buffer);
+}
+
+static boolean
+nv40_fragprog_validate(struct nv40_context *nv40)
+{
+	struct nv40_fragment_program *fp = nv40->pipe_state.fragprog;
+	struct pipe_buffer *constbuf =
+		nv40->pipe_state.constbuf[PIPE_SHADER_FRAGMENT];
+	struct pipe_winsys *ws = nv40->pipe.winsys;
+	struct nouveau_stateobj *so;
+	unsigned new_program = FALSE;
+	int i;
+
+	if (fp->translated)
+		goto update_constants;
+
+	nv40_fragprog_translate(nv40, fp);
+	if (!fp->translated) {
+		nv40->fallback |= NV40_FALLBACK_RAST;
+		return FALSE;
+	}
+	new_program = TRUE;
+
+	fp->buffer = ws->buffer_create(ws, 0x100, 0, fp->insn_len * 4);
+	nv40_fragprog_upload(nv40, fp);
+
 	so = so_new(4, 1);
 	so_method(so, nv40->hw->curie, NV40TCL_FP_ADDRESS, 1);
 	so_reloc (so, fp->buffer, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART |
@@ -824,12 +818,33 @@ nv40_fragprog_bind(struct nv40_context *nv40, struct nv40_fragment_program *fp)
 		  NV40TCL_FP_ADDRESS_DMA0, NV40TCL_FP_ADDRESS_DMA1);
 	so_method(so, nv40->hw->curie, NV40TCL_FP_CONTROL, 1);
 	so_data  (so, fp->fp_control);
-
-	so_emit(nv40->nvws, so);
 	so_ref(so, &fp->so);
 	so_ref(NULL, &so);
 
-	nv40->fragprog.active = fp;
+update_constants:
+	if (fp->nr_consts) {
+		boolean new_consts = FALSE;
+		float *map;
+		
+		map = ws->buffer_map(ws, constbuf, PIPE_BUFFER_USAGE_CPU_READ);
+		for (i = 0; i < fp->nr_consts; i++) {
+			struct nv40_fragment_program_data *fpd = &fp->consts[i];
+			uint32_t *p = &fp->insn[fpd->offset];
+			uint32_t *cb = (uint32_t *)&map[fpd->index * 4];
+
+			if (!memcmp(p, cb, 4 * sizeof(float)))
+				continue;
+			memcpy(p, cb, 4 * sizeof(float));
+			new_consts = TRUE;
+		}
+		ws->buffer_unmap(ws, constbuf);
+
+		if (new_consts)
+			nv40_fragprog_upload(nv40, fp);
+	}
+
+	so_ref(fp->so, &nv40->state.fragprog);
+	return new_program;
 }
 
 void
@@ -840,3 +855,11 @@ nv40_fragprog_destroy(struct nv40_context *nv40,
 		free(fp->insn);
 }
 
+struct nv40_state_entry nv40_state_fragprog = {
+	.validate = nv40_fragprog_validate,
+	.dirty = {
+		.pipe = NV40_NEW_FRAGPROG,
+		.hw = NV40_NEW_FRAGPROG
+	}
+};
+
diff --git a/src/gallium/drivers/nv40/nv40_fragtex.c b/src/gallium/drivers/nv40/nv40_fragtex.c
index 5af5fbe7465..811f3098ba7 100644
--- a/src/gallium/drivers/nv40/nv40_fragtex.c
+++ b/src/gallium/drivers/nv40/nv40_fragtex.c
@@ -125,7 +125,7 @@ nv40_fragtex_build(struct nv40_context *nv40, int unit)
 void
 nv40_fragtex_bind(struct nv40_context *nv40)
 {
-	struct nv40_fragment_program *fp = nv40->fragprog.active;
+	struct nv40_fragment_program *fp = nv40->pipe_state.fragprog;
 	unsigned samplers, unit;
 
 	samplers = nv40->fp_samplers & ~fp->samplers;
diff --git a/src/gallium/drivers/nv40/nv40_state.c b/src/gallium/drivers/nv40/nv40_state.c
index c203b002403..2886c6b0dc4 100644
--- a/src/gallium/drivers/nv40/nv40_state.c
+++ b/src/gallium/drivers/nv40/nv40_state.c
@@ -523,9 +523,8 @@ static void
 nv40_fp_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
-	struct nv40_fragment_program *fp = hwcso;
 
-	nv40->fragprog.current = fp;
+	nv40->pipe_state.fragprog = hwcso;
 	nv40->dirty |= NV40_NEW_FRAGPROG;
 }
 
@@ -578,7 +577,7 @@ nv40_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index,
 		nv40->dirty |= NV40_NEW_VERTPROG;
 	} else
 	if (shader == PIPE_SHADER_FRAGMENT) {
-		nv40->fragprog.constant_buf = buf->buffer;
+		nv40->pipe_state.constbuf[PIPE_SHADER_FRAGMENT] = buf->buffer;
 		nv40->dirty |= NV40_NEW_FRAGPROG;
 	}
 }
diff --git a/src/gallium/drivers/nv40/nv40_state.h b/src/gallium/drivers/nv40/nv40_state.h
index e82ab9de98a..2701294a077 100644
--- a/src/gallium/drivers/nv40/nv40_state.h
+++ b/src/gallium/drivers/nv40/nv40_state.h
@@ -50,7 +50,6 @@ struct nv40_fragment_program {
 	const struct pipe_shader_state *pipe;
 
 	boolean translated;
-	boolean on_hw;
 	unsigned samplers;
 
 	uint32_t *insn;
diff --git a/src/gallium/drivers/nv40/nv40_state_emit.c b/src/gallium/drivers/nv40/nv40_state_emit.c
index b5d0d68d6b1..e10e178432e 100644
--- a/src/gallium/drivers/nv40/nv40_state_emit.c
+++ b/src/gallium/drivers/nv40/nv40_state_emit.c
@@ -19,21 +19,14 @@ nv40_state_emit_dummy_relocs(struct nv40_context *nv40)
 			continue;
 		so_emit_reloc_markers(nv40->nvws, nv40->so_fragtex[i]);
 	}
-	so_emit_reloc_markers(nv40->nvws, nv40->fragprog.active->so);
-}
-
-static boolean
-nv40_state_clip_validate(struct nv40_context *nv40)
-{
-	if (nv40->pipe_state.clip.nr)
-		nv40->fallback |= NV40_FALLBACK_TNL;
-	return FALSE;
+	so_emit_reloc_markers(nv40->nvws, nv40->state.fragprog);
 }
 
 static struct nv40_state_entry *render_states[] = {
 	&nv40_state_clip,
 	&nv40_state_scissor,
 	&nv40_state_stipple,
+	&nv40_state_fragprog,
 	NULL
 };
 
@@ -65,6 +58,15 @@ nv40_state_validate(struct nv40_context *nv40)
 	    !(nv40->fallback & NV40_FALLBACK_TNL)) {
 		NOUVEAU_ERR("XXX: swtnl->hwtnl\n");
 	}
+
+	if (nv40->fallback & NV40_FALLBACK_RAST &&
+	    !(last_fallback & NV40_FALLBACK_RAST)) {
+		NOUVEAU_ERR("XXX: hwrast->swrast\n");
+	} else
+	if (last_fallback & NV40_FALLBACK_RAST &&
+	    !(nv40->fallback & NV40_FALLBACK_RAST)) {
+		NOUVEAU_ERR("XXX: swrast->hwrast\n");
+	}
 }
 
 void
@@ -100,10 +102,8 @@ nv40_emit_hw_state(struct nv40_context *nv40)
 		nv40->hw_dirty &= ~NV40_NEW_STIPPLE;
 	}
 
-	if (nv40->dirty & NV40_NEW_FRAGPROG) {
-		nv40_fragprog_bind(nv40, nv40->fragprog.current);
-		/*XXX: clear NV40_NEW_FRAGPROG if no new program uploaded */
-	}
+	if (nv40->hw_dirty & NV40_NEW_FRAGPROG)
+		so_emit(nv40->nvws, nv40->state.fragprog);
 
 	if (nv40->dirty_samplers || (nv40->dirty & NV40_NEW_FRAGPROG)) {
 		nv40_fragtex_bind(nv40);