1 files changed, 38 insertions, 241 deletions

diff --git a/src/mesa/drivers/dri/r300/r300_emit.c b/src/mesa/drivers/dri/r300/r300_emit.c
index 28c3157427..a19b0f1960 100644
--- a/src/mesa/drivers/dri/r300/r300_emit.c
+++ b/src/mesa/drivers/dri/r300/r300_emit.c
@@ -46,14 +46,10 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include "tnl/t_context.h"
 
 #include "r300_context.h"
-#include "radeon_ioctl.h"
 #include "r300_state.h"
 #include "r300_emit.h"
 #include "r300_ioctl.h"
 
-#ifdef USER_BUFFERS
-#include "r300_mem.h"
-#endif
 
 #if SWIZZLE_X != R300_INPUT_ROUTE_SELECT_X || \
     SWIZZLE_Y != R300_INPUT_ROUTE_SELECT_Y || \
@@ -66,147 +62,6 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 #define DEBUG_ALL DEBUG_VERTS
 
-#if defined(USE_X86_ASM)
-#define COPY_DWORDS( dst, src, nr )					\
-do {									\
-	int __tmp;							\
-	__asm__ __volatile__( "rep ; movsl"				\
-			      : "=%c" (__tmp), "=D" (dst), "=S" (__tmp)	\
-			      : "0" (nr),				\
-			        "D" ((long)dst),			\
-			        "S" ((long)src) );			\
-} while (0)
-#else
-#define COPY_DWORDS( dst, src, nr )		\
-do {						\
-   int j;					\
-   for ( j = 0 ; j < nr ; j++ )			\
-      dst[j] = ((int *)src)[j];			\
-   dst += nr;					\
-} while (0)
-#endif
-
-static void r300EmitVec4(GLcontext * ctx, struct r300_dma_region *rvb,
-			 GLvoid * data, int stride, int count)
-{
-	int i;
-	int *out = (int *)(rvb->address + rvb->start);
-
-	if (RADEON_DEBUG & DEBUG_VERTS)
-		fprintf(stderr, "%s count %d stride %d out %p data %p\n",
-			__FUNCTION__, count, stride, (void *)out, (void *)data);
-
-	if (stride == 4)
-		COPY_DWORDS(out, data, count);
-	else
-		for (i = 0; i < count; i++) {
-			out[0] = *(int *)data;
-			out++;
-			data += stride;
-		}
-}
-
-static void r300EmitVec8(GLcontext * ctx, struct r300_dma_region *rvb,
-			 GLvoid * data, int stride, int count)
-{
-	int i;
-	int *out = (int *)(rvb->address + rvb->start);
-
-	if (RADEON_DEBUG & DEBUG_VERTS)
-		fprintf(stderr, "%s count %d stride %d out %p data %p\n",
-			__FUNCTION__, count, stride, (void *)out, (void *)data);
-
-	if (stride == 8)
-		COPY_DWORDS(out, data, count * 2);
-	else
-		for (i = 0; i < count; i++) {
-			out[0] = *(int *)data;
-			out[1] = *(int *)(data + 4);
-			out += 2;
-			data += stride;
-		}
-}
-
-static void r300EmitVec12(GLcontext * ctx, struct r300_dma_region *rvb,
-			  GLvoid * data, int stride, int count)
-{
-	int i;
-	int *out = (int *)(rvb->address + rvb->start);
-
-	if (RADEON_DEBUG & DEBUG_VERTS)
-		fprintf(stderr, "%s count %d stride %d out %p data %p\n",
-			__FUNCTION__, count, stride, (void *)out, (void *)data);
-
-	if (stride == 12)
-		COPY_DWORDS(out, data, count * 3);
-	else
-		for (i = 0; i < count; i++) {
-			out[0] = *(int *)data;
-			out[1] = *(int *)(data + 4);
-			out[2] = *(int *)(data + 8);
-			out += 3;
-			data += stride;
-		}
-}
-
-static void r300EmitVec16(GLcontext * ctx, struct r300_dma_region *rvb,
-			  GLvoid * data, int stride, int count)
-{
-	int i;
-	int *out = (int *)(rvb->address + rvb->start);
-
-	if (RADEON_DEBUG & DEBUG_VERTS)
-		fprintf(stderr, "%s count %d stride %d out %p data %p\n",
-			__FUNCTION__, count, stride, (void *)out, (void *)data);
-
-	if (stride == 16)
-		COPY_DWORDS(out, data, count * 4);
-	else
-		for (i = 0; i < count; i++) {
-			out[0] = *(int *)data;
-			out[1] = *(int *)(data + 4);
-			out[2] = *(int *)(data + 8);
-			out[3] = *(int *)(data + 12);
-			out += 4;
-			data += stride;
-		}
-}
-
-static void r300EmitVec(GLcontext * ctx, struct r300_dma_region *rvb,
-			GLvoid * data, int size, int stride, int count)
-{
-	r300ContextPtr rmesa = R300_CONTEXT(ctx);
-
-	if (stride == 0) {
-		r300AllocDmaRegion(rmesa, rvb, size * 4, 4);
-		count = 1;
-		rvb->aos_offset = GET_START(rvb);
-		rvb->aos_stride = 0;
-	} else {
-		r300AllocDmaRegion(rmesa, rvb, size * count * 4, 4);
-		rvb->aos_offset = GET_START(rvb);
-		rvb->aos_stride = size;
-	}
-
-	switch (size) {
-	case 1:
-		r300EmitVec4(ctx, rvb, data, stride, count);
-		break;
-	case 2:
-		r300EmitVec8(ctx, rvb, data, stride, count);
-		break;
-	case 3:
-		r300EmitVec12(ctx, rvb, data, stride, count);
-		break;
-	case 4:
-		r300EmitVec16(ctx, rvb, data, stride, count);
-		break;
-	default:
-		assert(0);
-		break;
-	}
-}
-
 #define DW_SIZE(x) ((inputs[tab[(x)]] << R300_DST_VEC_LOC_SHIFT) |	\
 		    (attribptr[tab[(x)]]->size - 1) << R300_DATA_TYPE_0_SHIFT)
 
@@ -272,7 +127,6 @@ GLuint r300VAPInputCntl0(GLcontext * ctx, GLuint InputsRead)
 
 GLuint r300VAPInputCntl1(GLcontext * ctx, GLuint InputsRead)
 {
-	r300ContextPtr rmesa = R300_CONTEXT(ctx);
 	GLuint i, vic_1 = 0;
 
 	if (InputsRead & (1 << VERT_ATTRIB_POS))
@@ -284,10 +138,8 @@ GLuint r300VAPInputCntl1(GLcontext * ctx, GLuint InputsRead)
 	if (InputsRead & (1 << VERT_ATTRIB_COLOR0))
 		vic_1 |= R300_INPUT_CNTL_COLOR;
 
-	rmesa->state.texture.tc_count = 0;
 	for (i = 0; i < ctx->Const.MaxTextureUnits; i++)
 		if (InputsRead & (1 << (VERT_ATTRIB_TEX0 + i))) {
-			rmesa->state.texture.tc_count++;
 			vic_1 |= R300_INPUT_CNTL_TC0 << i;
 		}
 
@@ -336,7 +188,7 @@ GLuint r300VAPOutputCntl1(GLcontext * ctx, GLuint OutputsWritten)
 			fprintf(stderr, "\tout of free texcoords to write fog coord\n");
 			_mesa_exit(-1);
 		}
-		ret |= 4 << (3 * first_free_texcoord);
+		ret |= 1 << (3 * first_free_texcoord);
 	}
 
 	return ret;
@@ -367,7 +219,7 @@ int r300EmitArrays(GLcontext * ctx)
 		InputsRead = prog->key.InputsRead;
 		OutputsWritten = prog->key.OutputsWritten;
 	} else {
-		inputs = rmesa->state.sw_tcl_inputs;
+		inputs = rmesa->swtcl.sw_tcl_inputs;
 
 		DECLARE_RENDERINPUTS(render_inputs_bitset);
 		RENDERINPUTS_COPY(render_inputs_bitset, tnl->render_inputs_bitset);
@@ -376,7 +228,6 @@ int r300EmitArrays(GLcontext * ctx)
 
 		assert(RENDERINPUTS_TEST(render_inputs_bitset, _TNL_ATTRIB_POS));
 		assert(RENDERINPUTS_TEST(render_inputs_bitset, _TNL_ATTRIB_NORMAL) == 0);
-		//assert(RENDERINPUTS_TEST(render_inputs_bitset, _TNL_ATTRIB_COLOR0));
 
 		if (RENDERINPUTS_TEST(render_inputs_bitset, _TNL_ATTRIB_POS)) {
 			InputsRead |= 1 << VERT_ATTRIB_POS;
@@ -421,7 +272,7 @@ int r300EmitArrays(GLcontext * ctx)
 			if (InputsRead & (1 << i))
 				inputs[i] = 6 + (i - VERT_ATTRIB_TEX0);
 
-		RENDERINPUTS_COPY(rmesa->state.render_inputs_bitset, render_inputs_bitset);
+		RENDERINPUTS_COPY(rmesa->render_inputs_bitset, render_inputs_bitset);
 	}
 
 	assert(InputsRead);
@@ -438,7 +289,7 @@ int r300EmitArrays(GLcontext * ctx)
 	}
 
 	for (i = 0; i < nr; i++) {
-		int ci, fix, found = 0;
+		int ci;
 
 		swizzle[i][0] = SWIZZLE_ZERO;
 		swizzle[i][1] = SWIZZLE_ZERO;
@@ -448,60 +299,34 @@ int r300EmitArrays(GLcontext * ctx)
 		for (ci = 0; ci < vb->AttribPtr[tab[i]]->size; ci++) {
 			swizzle[i][ci] = ci;
 		}
-
-		if (r300IsGartMemory(rmesa, vb->AttribPtr[tab[i]]->data, 4)) {
-			if (vb->AttribPtr[tab[i]]->stride % 4) {
-				return R300_FALLBACK_TCL;
-			}
-			rmesa->state.aos[i].address = (void *)(vb->AttribPtr[tab[i]]->data);
-			rmesa->state.aos[i].start = 0;
-			rmesa->state.aos[i].aos_offset = r300GartOffsetFromVirtual(rmesa, vb->AttribPtr[tab[i]]->data);
-			rmesa->state.aos[i].aos_stride = vb->AttribPtr[tab[i]]->stride / 4;
-			rmesa->state.aos[i].aos_size = vb->AttribPtr[tab[i]]->size;
-		} else {
-			r300EmitVec(ctx, &rmesa->state.aos[i],
+		rcommon_emit_vector(ctx, &rmesa->radeon.tcl.aos[i],
 				    vb->AttribPtr[tab[i]]->data,
 				    vb->AttribPtr[tab[i]]->size,
 				    vb->AttribPtr[tab[i]]->stride, count);
-		}
-
-		rmesa->state.aos[i].aos_size = vb->AttribPtr[tab[i]]->size;
-
-		for (fix = 0; fix <= 4 - vb->AttribPtr[tab[i]]->size; fix++) {
-			if ((rmesa->state.aos[i].aos_offset - _mesa_sizeof_type(GL_FLOAT) * fix) % 4) {
-				continue;
-			}
-			found = 1;
-			break;
-		}
-
-		if (found) {
-			if (fix > 0) {
-				WARN_ONCE("Feeling lucky?\n");
-			}
-			rmesa->state.aos[i].aos_offset -= _mesa_sizeof_type(GL_FLOAT) * fix;
-			for (ci = 0; ci < vb->AttribPtr[tab[i]]->size; ci++) {
-				swizzle[i][ci] += fix;
-			}
-		} else {
-			WARN_ONCE
-			    ("Cannot handle offset %x with stride %d, comp %d\n",
-			     rmesa->state.aos[i].aos_offset,
-			     rmesa->state.aos[i].aos_stride,
-			     vb->AttribPtr[tab[i]]->size);
-			return R300_FALLBACK_TCL;
-		}
 	}
 
 	/* Setup INPUT_ROUTE. */
-	R300_STATECHANGE(rmesa, vir[0]);
-	((drm_r300_cmd_header_t *) rmesa->hw.vir[0].cmd)->packet0.count =
-	    r300VAPInputRoute0(&rmesa->hw.vir[0].cmd[R300_VIR_CNTL_0],
-			       vb->AttribPtr, inputs, tab, nr);
-	R300_STATECHANGE(rmesa, vir[1]);
-	((drm_r300_cmd_header_t *) rmesa->hw.vir[1].cmd)->packet0.count =
-	    r300VAPInputRoute1(&rmesa->hw.vir[1].cmd[R300_VIR_CNTL_0], swizzle,
-			       nr);
+	if (rmesa->radeon.radeonScreen->kernel_mm) {
+		R300_STATECHANGE(rmesa, vir[0]);
+		rmesa->hw.vir[0].cmd[0] &= 0xC000FFFF;
+		rmesa->hw.vir[1].cmd[0] &= 0xC000FFFF;
+		rmesa->hw.vir[0].cmd[0] |=
+			(r300VAPInputRoute0(&rmesa->hw.vir[0].cmd[R300_VIR_CNTL_0],
+					    vb->AttribPtr, inputs, tab, nr) & 0x3FFF) << 16;
+		R300_STATECHANGE(rmesa, vir[1]);
+		rmesa->hw.vir[1].cmd[0] |=
+			(r300VAPInputRoute1(&rmesa->hw.vir[1].cmd[R300_VIR_CNTL_0], swizzle,
+					    nr) & 0x3FFF) << 16;
+	} else {
+		R300_STATECHANGE(rmesa, vir[0]);
+		((drm_r300_cmd_header_t *) rmesa->hw.vir[0].cmd)->packet0.count =
+			r300VAPInputRoute0(&rmesa->hw.vir[0].cmd[R300_VIR_CNTL_0],
+					   vb->AttribPtr, inputs, tab, nr);
+		R300_STATECHANGE(rmesa, vir[1]);
+		((drm_r300_cmd_header_t *) rmesa->hw.vir[1].cmd)->packet0.count =
+			r300VAPInputRoute1(&rmesa->hw.vir[1].cmd[R300_VIR_CNTL_0], swizzle,
+					   nr);
+	}
 
 	/* Setup INPUT_CNTL. */
 	R300_STATECHANGE(rmesa, vic);
@@ -515,50 +340,22 @@ int r300EmitArrays(GLcontext * ctx)
 	rmesa->hw.vof.cmd[R300_VOF_CNTL_1] =
 	    r300VAPOutputCntl1(ctx, OutputsWritten);
 
-	rmesa->state.aos_count = nr;
+	rmesa->radeon.tcl.aos_count = nr;
 
 	return R300_FALLBACK_NONE;
 }
 
-#ifdef USER_BUFFERS
-void r300UseArrays(GLcontext * ctx)
-{
-	r300ContextPtr rmesa = R300_CONTEXT(ctx);
-	int i;
-
-	if (rmesa->state.elt_dma.buf)
-		r300_mem_use(rmesa, rmesa->state.elt_dma.buf->id);
-
-	for (i = 0; i < rmesa->state.aos_count; i++) {
-		if (rmesa->state.aos[i].buf)
-			r300_mem_use(rmesa, rmesa->state.aos[i].buf->id);
-	}
-}
-#endif
-
-void r300ReleaseArrays(GLcontext * ctx)
-{
-	r300ContextPtr rmesa = R300_CONTEXT(ctx);
-	int i;
-
-	r300ReleaseDmaRegion(rmesa, &rmesa->state.elt_dma, __FUNCTION__);
-	for (i = 0; i < rmesa->state.aos_count; i++) {
-		r300ReleaseDmaRegion(rmesa, &rmesa->state.aos[i], __FUNCTION__);
-	}
-}
-
 void r300EmitCacheFlush(r300ContextPtr rmesa)
 {
-	int cmd_reserved = 0;
-	int cmd_written = 0;
-
-	drm_radeon_cmd_header_t *cmd = NULL;
-
-	reg_start(R300_RB3D_DSTCACHE_CTLSTAT, 0);
-	e32(R300_RB3D_DSTCACHE_CTLSTAT_DC_FREE_FREE_3D_TAGS |
-	    R300_RB3D_DSTCACHE_CTLSTAT_DC_FLUSH_FLUSH_DIRTY_3D);
-
-	reg_start(R300_ZB_ZCACHE_CTLSTAT, 0);
-	e32(R300_ZB_ZCACHE_CTLSTAT_ZC_FLUSH_FLUSH_AND_FREE |
-	    R300_ZB_ZCACHE_CTLSTAT_ZC_FREE_FREE);
+	BATCH_LOCALS(&rmesa->radeon);
+
+	BEGIN_BATCH_NO_AUTOSTATE(4);
+	OUT_BATCH_REGVAL(R300_RB3D_DSTCACHE_CTLSTAT,
+		R300_RB3D_DSTCACHE_CTLSTAT_DC_FREE_FREE_3D_TAGS |
+		R300_RB3D_DSTCACHE_CTLSTAT_DC_FLUSH_FLUSH_DIRTY_3D);
+	OUT_BATCH_REGVAL(R300_ZB_ZCACHE_CTLSTAT,
+		R300_ZB_ZCACHE_CTLSTAT_ZC_FLUSH_FLUSH_AND_FREE |
+		R300_ZB_ZCACHE_CTLSTAT_ZC_FREE_FREE);
+	END_BATCH();
+	COMMIT_BATCH();
 }