Finish converting RB2D_DSTCACHE to RB3D_DSTCACHE. Remove an extra pixel

    cache flush in the idle function. Init an extra reg for r200, and
    annotate the TCL_BYPASS better. Also, clean up some style nits from the
    last commit.

3 files changed, 53 insertions, 57 deletions

diff --git a/hw/kdrive/ati/ati_dma.c b/hw/kdrive/ati/ati_dma.c
index 1acb8e544..f75ff12e7 100644
--- a/hw/kdrive/ati/ati_dma.c
+++ b/hw/kdrive/ati/ati_dma.c
@@ -56,8 +56,8 @@ ATIDebugFifo(ATIScreenInfo *atis)
 		    MMIO_IN32(mmio, RADEON_REG_CP_CSQ_STAT));
 		ErrorF("RADEON_REG_RBBM_STATUS: 0x%08x\n",
 		    MMIO_IN32(mmio, RADEON_REG_RBBM_STATUS));
-		ErrorF("RADEON_REG_RB2D_DSTCACHE_CTLSTAT: 0x%08x\n",
-		    MMIO_IN32(mmio, RADEON_REG_RB2D_DSTCACHE_CTLSTAT));
+		ErrorF("RADEON_REG_RB3D_DSTCACHE_CTLSTAT: 0x%08x\n",
+		    MMIO_IN32(mmio, RADEON_REG_RB3D_DSTCACHE_CTLSTAT));
 	} else {
 		ErrorF("R128_REG_PM4_BUFFER_CNTL: 0x%08x\n",
 		    MMIO_IN32(mmio, R128_REG_PM4_BUFFER_CNTL));
@@ -121,13 +121,13 @@ ATIFlushPixelCache(ATIScreenInfo *atis)
 	TIMEOUT_LOCALS;
 
 	if (atic->is_radeon) {
-		temp = MMIO_IN32(mmio, RADEON_REG_RB2D_DSTCACHE_CTLSTAT);
-		temp |= RADEON_RB2D_DC_FLUSH_ALL;
-		MMIO_OUT32(mmio, RADEON_REG_RB2D_DSTCACHE_CTLSTAT, temp);
+		temp = MMIO_IN32(mmio, RADEON_REG_RB3D_DSTCACHE_CTLSTAT);
+		temp |= RADEON_RB3D_DC_FLUSH_ALL;
+		MMIO_OUT32(mmio, RADEON_REG_RB3D_DSTCACHE_CTLSTAT, temp);
 
 		WHILE_NOT_TIMEOUT(.2) {
-			if ((MMIO_IN32(mmio, RADEON_REG_RB2D_DSTCACHE_CTLSTAT) &
-			    RADEON_RB2D_DC_BUSY) == 0)
+			if ((MMIO_IN32(mmio, RADEON_REG_RB3D_DSTCACHE_CTLSTAT) &
+			    RADEON_RB3D_DC_BUSY) == 0)
 				break;
 		}
 	} else {
@@ -323,7 +323,6 @@ ATIWaitIdle(ATIScreenInfo *atis)
 {
 	ATICardInfo *atic = atis->atic;
 	char *mmio = atic->reg_base;
-	RING_LOCALS;
 	TIMEOUT_LOCALS;
 
 	if (atis->indirectBuffer != NULL)
@@ -349,16 +348,6 @@ ATIWaitIdle(ATIScreenInfo *atis)
 	}
 #endif
 
-	if (atic->is_radeon && (atis->using_pseudo || atis->using_dma)) {
-		BEGIN_DMA(4);
-		OUT_REG(RADEON_REG_RB2D_DSTCACHE_CTLSTAT,
-		    RADEON_RB2D_DC_FLUSH_ALL);
-		OUT_REG(ATI_REG_WAIT_UNTIL,
-		    RADEON_WAIT_HOST_IDLECLEAN | RADEON_WAIT_2D_IDLECLEAN |
-		    RADEON_WAIT_3D_IDLECLEAN);
-		END_DMA();
-	}
-
 	if (!atic->is_radeon && (atis->using_pseudo || atis->using_dma)) {
 		ATIWaitAvailPrimary(atis, atis->cce_pri_size);
 
diff --git a/hw/kdrive/ati/ati_draw.c b/hw/kdrive/ati/ati_draw.c
index 32f57bdf4..ead133b58 100644
--- a/hw/kdrive/ati/ati_draw.c
+++ b/hw/kdrive/ati/ati_draw.c
@@ -140,7 +140,7 @@ ATIDrawSetup(ScreenPtr pScreen)
 		OUT_REG(R128_REG_CONSTANT_COLOR_C, 0xff000000);
 		OUT_REG(R128_REG_WINDOW_XY_OFFSET, 0x00000000);
 		END_DMA();
-	} else {
+	} else if (!atic->is_r300) {
 		/* Setup for R100/R200 Composite */
 		BEGIN_DMA(8);
 		OUT_REG(RADEON_REG_RE_TOP_LEFT, 0);
@@ -161,12 +161,15 @@ ATIDrawSetup(ScreenPtr pScreen)
 			    RADEON_VTX_ST1_NONPARAMETRIC |
 			    RADEON_TEX1_W_ROUTING_USE_W0);
 			OUT_REG(RADEON_REG_RB3D_DSTCACHE_MODE, 
-				RADEON_RB2D_DC_2D_CACHE_AUTOFLUSH|
-				RADEON_RB2D_DC_3D_CACHE_AUTOFLUSH);
+				RADEON_RB3D_DC_2D_CACHE_AUTOFLUSH |
+				RADEON_RB3D_DC_3D_CACHE_AUTOFLUSH);
 			END_DMA();
 		} else {
-			BEGIN_DMA(16);
-			OUT_REG(R200_REG_SE_VAP_CNTL_STATUS, 0 /*RADEON_TCL_BYPASS*/);
+			BEGIN_DMA(18);
+			/* XXX: The 0 below should be RADEON_TCL_BYPASS on
+			 * RS300s.
+			 */
+			OUT_REG(R200_REG_SE_VAP_CNTL_STATUS, 0);
 			OUT_REG(R200_REG_PP_CNTL_X, 0);
 			OUT_REG(R200_REG_PP_TXMULTI_CTL_0, 0);
 			OUT_REG(R200_REG_SE_VTX_STATE_CNTL, 0);
@@ -179,6 +182,11 @@ ATIDrawSetup(ScreenPtr pScreen)
 			    R200_VAP_FORCE_W_TO_ONE |
 			    R200_VAP_VF_MAX_VTX_NUM);
 			OUT_REG(R200_REG_RE_AUX_SCISSOR_CNTL, 0);
+			OUT_REG(RADEON_REG_RB3D_DSTCACHE_MODE, 
+				RADEON_RB3D_DC_2D_CACHE_AUTOFLUSH |
+				RADEON_RB3D_DC_3D_CACHE_AUTOFLUSH |
+				R200_RB3D_DC_2D_CACHE_AUTOFREE |
+				R200_RB3D_DC_3D_CACHE_AUTOFREE);
 			END_DMA();
 		}
 	}
@@ -191,8 +199,7 @@ RadeonSwitchTo2D(ATIScreenInfo *atis)
 
 	ENTER_DRAW(0);
 	BEGIN_DMA(4);
-	OUT_REG(RADEON_REG_RB2D_DSTCACHE_CTLSTAT,
-		RADEON_RB2D_DC_FLUSH);
+	OUT_REG(RADEON_REG_RB3D_DSTCACHE_CTLSTAT, RADEON_RB3D_DC_FLUSH);
 	OUT_REG(ATI_REG_WAIT_UNTIL,
 	    RADEON_WAIT_HOST_IDLECLEAN | RADEON_WAIT_3D_IDLECLEAN);
 	END_DMA();
@@ -206,8 +213,7 @@ RadeonSwitchTo3D(ATIScreenInfo *atis)
 
 	ENTER_DRAW(0);
 	BEGIN_DMA(4);
-	OUT_REG(RADEON_REG_RB2D_DSTCACHE_CTLSTAT,
-		RADEON_RB2D_DC_FLUSH);
+	OUT_REG(RADEON_REG_RB3D_DSTCACHE_CTLSTAT, RADEON_RB3D_DC_FLUSH);
 	OUT_REG(ATI_REG_WAIT_UNTIL,
 	    RADEON_WAIT_HOST_IDLECLEAN | RADEON_WAIT_2D_IDLECLEAN);
 	END_DMA();
@@ -600,8 +606,7 @@ ATIUploadToScreen(PixmapPtr pDst, char *src, int src_pitch)
 	 */
 	if (atic->is_radeon) {
 		BEGIN_DMA(4);
-		OUT_REG(RADEON_REG_RB2D_DSTCACHE_CTLSTAT,
-		    RADEON_RB2D_DC_FLUSH_ALL);
+		OUT_REG(RADEON_REG_RB3D_DSTCACHE_CTLSTAT, RADEON_RB3D_DC_FLUSH);
 		OUT_REG(ATI_REG_WAIT_UNTIL,
 		    RADEON_WAIT_2D_IDLECLEAN |
 		    RADEON_WAIT_3D_IDLECLEAN |
@@ -641,8 +646,8 @@ ATIUploadToScreen(PixmapPtr pDst, char *src, int src_pitch)
 
 	if (atic->is_radeon) {
 		BEGIN_DMA(4);
-		OUT_REG(RADEON_REG_RB2D_DSTCACHE_CTLSTAT,
-		    RADEON_RB2D_DC_FLUSH_ALL);
+		OUT_REG(RADEON_REG_RB3D_DSTCACHE_CTLSTAT,
+		    RADEON_RB3D_DC_FLUSH_ALL);
 		OUT_REG(ATI_REG_WAIT_UNTIL,
 		    RADEON_WAIT_2D_IDLECLEAN |
 		    RADEON_WAIT_HOST_IDLECLEAN);
@@ -716,9 +721,10 @@ ATIUploadToScratch(PixmapPtr pSrc, PixmapPtr pDst)
 
 	/* Flush the pixel cache */
 	if (atic->is_radeon) {
-		BEGIN_DMA(2);
-		OUT_REG(RADEON_REG_RB2D_DSTCACHE_CTLSTAT,
-		    RADEON_RB2D_DC_FLUSH_ALL);
+		BEGIN_DMA(4);
+		OUT_REG(RADEON_REG_RB3D_DSTCACHE_CTLSTAT,
+		    RADEON_RB3D_DC_FLUSH_ALL);
+		OUT_REG(ATI_REG_WAIT_UNTIL, RADEON_WAIT_HOST_IDLECLEAN);
 		END_DMA();
 	} else {
 		BEGIN_DMA(2);
@@ -830,7 +836,6 @@ ATIDrawEnable(ScreenPtr pScreen)
 	atis->kaa.UploadToScreen = NULL;
     	atis->kaa.UploadToScratch = NULL;
 
-#if 1
 	/* We can't dispatch 3d commands in PIO mode. */
 	if (!atis->using_pio) {
 		if (!atic->is_radeon) {
@@ -876,7 +881,6 @@ ATIDrawEnable(ScreenPtr pScreen)
 		atis->scratch_next = atis->scratch_area->offset;
 		atis->kaa.UploadToScratch = ATIUploadToScratch;
 	}
-#endif
 
 	KdMarkSync(pScreen);
 }
diff --git a/hw/kdrive/ati/ati_reg.h b/hw/kdrive/ati/ati_reg.h
index 57479a04b..2b38bbee2 100644
--- a/hw/kdrive/ati/ati_reg.h
+++ b/hw/kdrive/ati/ati_reg.h
@@ -1716,29 +1716,32 @@
 # define R200_TXA_REPL_ARG_C_MASK		(3 << 30)
 
 #define RADEON_REG_RB2D_DSTCACHE_MODE		0x3428
+/* This is a read-only mirror of RADEON_REG_RB3D_DSTCACHE_MODE */
+
+#define RADEON_REG_RB2D_DSTCACHE_CTLSTAT	0x342C
+/* This is a read-only mirror of RADEON_REG_RB3D_DSTCACHE_CTLSTAT */
+
 #define RADEON_REG_RB3D_DSTCACHE_MODE		0x3258
-# define RADEON_RB2D_DC_CACHE_ENABLE		0
-# define RADEON_RB2D_DC_2D_CACHE_DISABLE	1
-# define RADEON_RB2D_DC_3D_CACHE_DISABLE	1
-# define RADEON_RB2D_DC_CACHE_DISABLE		3
-# define RADEON_RB2D_DC_2D_CACHE_LINESIZE_128	4
-# define RADEON_RB2D_DC_3D_CACHE_LINESIZE_128	8
-# define RADEON_RB2D_DC_2D_CACHE_AUTOFLUSH	0x100
-# define RADEON_RB2D_DC_3D_CACHE_AUTOFLUSH	0x200
-# define RADEON_RB2D_DC_FORCE_RMW		0x10000
-# define RADEON_RB2D_DC_DISABLE_RI_FILL		0x1000000
-# define RADEON_RB2D_DC_DISABLE_RI_READ		0x2000000
-
-/* XXX The 2D shadow isn't writable on the M6.  The right fix is to rename
- * all uses, but Eric is doing that
- */
-/* #define RADEON_REG_RB2D_DSTCACHE_CTLSTAT	0x342C */
-#define RADEON_REG_RB2D_DSTCACHE_CTLSTAT	0x325C
+# define RADEON_RB3D_DC_CACHE_ENABLE		(0)
+# define RADEON_RB3D_DC_2D_CACHE_DISABLE	(1)
+# define RADEON_RB3D_DC_3D_CACHE_DISABLE	(2)
+# define RADEON_RB3D_DC_CACHE_DISABLE		(3)
+# define RADEON_RB3D_DC_2D_CACHE_LINESIZE_128	(1 << 2)
+# define RADEON_RB3D_DC_3D_CACHE_LINESIZE_128	(2 << 2)
+# define RADEON_RB3D_DC_2D_CACHE_AUTOFLUSH	(1 << 8)
+# define RADEON_RB3D_DC_3D_CACHE_AUTOFLUSH	(2 << 8)
+# define R200_RB3D_DC_2D_CACHE_AUTOFREE		(1 << 10)
+# define R200_RB3D_DC_3D_CACHE_AUTOFREE		(2 << 10)
+# define RADEON_RB3D_DC_FORCE_RMW		(1 << 16)
+# define RADEON_RB3D_DC_DISABLE_RI_FILL		(1 << 24)
+# define RADEON_RB3D_DC_DISABLE_RI_READ		(1 << 25)
+# define RADEON_RB3D_DC_DISABLE_MASK_CHK	(1 << 26)
+
 #define RADEON_REG_RB3D_DSTCACHE_CTLSTAT	0x325C
-# define RADEON_RB2D_DC_FLUSH			(3 << 0)
-# define RADEON_RB2D_DC_FREE			(3 << 2)
-# define RADEON_RB2D_DC_FLUSH_ALL		0xf
-# define RADEON_RB2D_DC_BUSY			(1 << 31)
+# define RADEON_RB3D_DC_FLUSH			(3 << 0)
+# define RADEON_RB3D_DC_FREE			(3 << 2)
+# define RADEON_RB3D_DC_FLUSH_ALL		0xf
+# define RADEON_RB3D_DC_BUSY			(1 << 31)
 
 /* PLL register defines */
 #define R128_REG_MCLK_CNTL			0x000f