r300/compiler: Fix glean tests texCombine and texCombine4 after latestpresub

presubtract changes.

6 files changed, 243 insertions, 62 deletions

diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_dataflow.c b/src/mesa/drivers/dri/r300/compiler/radeon_dataflow.c
index 5cb4af869f6..af072bb313e 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_dataflow.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_dataflow.c
@@ -75,51 +75,33 @@ static void reads_normal(struct rc_instruction * fullinst, rc_read_write_chan_fn
 	}
 }
 
-static void pair_rgb_get_src_refmasks(unsigned int * refmasks,
-				struct radeon_pair_instruction_rgb inst_rgb)
+static void pair_get_src_refmasks(unsigned int * refmasks,
+					struct rc_pair_instruction * inst,
+					unsigned int swz, unsigned int src)
 {
-	unsigned int arg;
+	if (swz >= 4)
+		return;
 
-	for(arg = 0; arg < 3; ++arg) {
-		unsigned int chan;
-		for(chan = 0; chan < 3; ++chan) {
-			unsigned int src;
-			unsigned int swz =
-				GET_SWZ(inst_rgb.Arg[arg].Swizzle, chan);
-			if (swz >= 4)
-				continue;
-
-			src = inst_rgb.Arg[arg].Source;
-			if(src == RC_PAIR_PRESUB_SRC) {
-				unsigned int i;
-				int srcp_regs = rc_presubtract_src_reg_count(
-						inst_rgb.Src[src].Index);
-				for(i = 0; i < srcp_regs; i++) {
-					refmasks[i] |= 1 << swz;
-				}
-			}
-			else {
-				refmasks[src] |= 1 << swz;
+	if (swz == RC_SWIZZLE_X || swz == RC_SWIZZLE_Y || swz == RC_SWIZZLE_Z) {
+		if(src == RC_PAIR_PRESUB_SRC) {
+			unsigned int i;
+			int srcp_regs =
+				rc_presubtract_src_reg_count(
+				inst->RGB.Src[src].Index);
+			for(i = 0; i < srcp_regs; i++) {
+				refmasks[i] |= 1 << swz;
 			}
 		}
+		else {
+			refmasks[src] |= 1 << swz;
+		}
 	}
-}
 
-static void pair_alpha_get_src_refmasks(unsigned int * refmasks,
-				struct radeon_pair_instruction_alpha inst_alpha)
-{
-	unsigned int arg;
-
-	for(arg = 0; arg < 3; ++arg) {
-		unsigned int swz = inst_alpha.Arg[arg].Swizzle;
-		if(swz >= 4)
-			continue;
-
-		unsigned int src = inst_alpha.Arg[arg].Source;
-		if(src == RC_PAIR_PRESUB_SRC) {
+	if (swz == RC_SWIZZLE_W) {
+		if (src == RC_PAIR_PRESUB_SRC) {
 			unsigned int i;
-			int srcp_regs =	rc_presubtract_src_reg_count(
-						inst_alpha.Src[src].Index);
+			int srcp_regs = rc_presubtract_src_reg_count(
+					inst->Alpha.Src[src].Index);
 			for(i = 0; i < srcp_regs; i++) {
 				refmasks[i] |= 1 << swz;
 			}
@@ -139,8 +121,21 @@ static void reads_pair(struct rc_instruction * fullinst,  rc_read_write_mask_fn
 	 * opcode is RC_OPCODE_NOP.  This is because alpha instructions can
 	 * read from rgb source and vice-versa, so we always need to compute
 	 * the readmasks for both sets of sources. */
-	pair_rgb_get_src_refmasks(refmasks, inst->RGB);
-	pair_alpha_get_src_refmasks(refmasks, inst->Alpha);
+	unsigned int arg;
+
+	for(arg = 0; arg < 3; ++arg) {
+		unsigned int chan;
+		for(chan = 0; chan < 3; ++chan) {
+			unsigned int swz_rgb =
+				GET_SWZ(inst->RGB.Arg[arg].Swizzle, chan);
+			unsigned int swz_alpha =
+				GET_SWZ(inst->Alpha.Arg[arg].Swizzle, chan);
+			pair_get_src_refmasks(refmasks, inst, swz_rgb,
+						inst->RGB.Arg[arg].Source);
+			pair_get_src_refmasks(refmasks, inst, swz_alpha,
+						inst->Alpha.Arg[arg].Source);
+		}
+	}
 
 	for(unsigned int src = 0; src < 3; ++src) {
 		if (inst->RGB.Src[src].Used && (refmasks[src] & RC_MASK_XYZ))
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_optimize.c b/src/mesa/drivers/dri/r300/compiler/radeon_optimize.c
index 8f0fd7aa8ba..560628669d3 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_optimize.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_optimize.c
@@ -480,6 +480,20 @@ static unsigned int src_reads_dst(struct rc_src_register src,
 	return mask;
 }
 
+/* Return 1 if the source registers has a constant swizzle (e.g. 0, 0.5, 1.0)
+ * in any of its channels.  Return 0 otherwise. */
+static int src_has_const_swz(struct rc_src_register src) {
+	int chan;
+	for(chan = 0; chan < 4; chan++) {
+		unsigned int swz = GET_SWZ(src.Swizzle, chan);
+		if (swz == RC_SWIZZLE_ZERO || swz == RC_SWIZZLE_HALF
+						|| swz == RC_SWIZZLE_ONE) {
+			return 1;
+		}
+	}
+	return 0;
+}
+
 static void peephole_scan_write(void * data, struct rc_instruction * inst,
 		rc_register_file file, unsigned int index, unsigned int mask)
 {
@@ -506,23 +520,30 @@ static int peephole_add_presub_inv(
 	struct rc_instruction * inst_add)
 {
 	unsigned int i, swz, mask;
-	unsigned int modified = 0;
+	unsigned int can_remove = 0;
 	struct rc_instruction * inst;
 	struct peephole_state s;
 
+	if (inst_add->U.I.SaturateMode)
+		return 0;
+
 	swz = inst_add->U.I.SrcReg[0].Swizzle;
 	mask = inst_add->U.I.DstReg.WriteMask;
-	/* Check if src0 is the 1. */
+
+	/* Check if src0 is 1. */
+	/* XXX It would be nice to use is_src_uniform_constant here, but that
+	 * function only works if the register's file is RC_FILE_NONE */
 	for(i = 0; i < 4; i++ ) {
-		if(GET_BIT(mask, i) && GET_SWZ(swz, i) != RC_SWIZZLE_ONE)
+		if(GET_SWZ(swz, i) != RC_SWIZZLE_ONE)
 			return 0;
 	}
-	/* src1 needs to be negative and not an absolute value.  src1's register
-	 * file needs to be constant or temporary. */
-	if(!(inst_add->U.I.SrcReg[1].Negate & mask)
+
+	/* Check src1. */
+	if (inst_add->U.I.SrcReg[1].Negate != inst_add->U.I.DstReg.WriteMask
 		|| inst_add->U.I.SrcReg[1].Abs
 		|| (inst_add->U.I.SrcReg[1].File != RC_FILE_TEMPORARY
-			&& inst_add->U.I.SrcReg[1].File != RC_FILE_CONSTANT)) {
+			&& inst_add->U.I.SrcReg[1].File != RC_FILE_CONSTANT)
+		|| src_has_const_swz(inst_add->U.I.SrcReg[1])) {
 
 		return 0;
 	}
@@ -541,6 +562,10 @@ static int peephole_add_presub_inv(
 						inst_add->U.I.DstReg)) {
 				continue;
 			}
+		if (inst->U.I.PreSub.Opcode != RC_PRESUB_NONE) {
+			can_remove = 0;
+			break;
+		}
 			/* We must be careful not to modify inst_add, since it
 			 * is possible it will remain part of the program. */
 			inst->U.I.PreSub.SrcReg[0] = inst_add->U.I.SrcReg[1];
@@ -551,8 +576,10 @@ static int peephole_add_presub_inv(
 
 			inst->U.I.SrcReg[i].File = RC_FILE_PRESUB;
 			inst->U.I.SrcReg[i].Index = RC_PRESUB_INV;
-			modified++;
+			can_remove = 1;
 		}
+		if(!can_remove)
+			break;
 		rc_for_all_writes_mask(inst, peephole_scan_write, &s);
 		/* If all components of inst_add's destination register have
 		 * been written to by subsequent instructions, the original
@@ -562,7 +589,7 @@ static int peephole_add_presub_inv(
 			break;
 		}
 	}
-	if(modified) {
+	if(can_remove) {
 		rc_remove_instruction(inst_add);
 		return 1;
 	}
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_pair_schedule.c b/src/mesa/drivers/dri/r300/compiler/radeon_pair_schedule.c
index 6e6fc00de3c..a8712772286 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_pair_schedule.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_pair_schedule.c
@@ -279,16 +279,17 @@ static int destructive_merge_instructions(
 		struct rc_pair_instruction * rgb,
 		struct rc_pair_instruction * alpha)
 {
+	const struct rc_opcode_info * opcode;
 	assert(rgb->Alpha.Opcode == RC_OPCODE_NOP);
 	assert(alpha->RGB.Opcode == RC_OPCODE_NOP);
 
-	/* Copy alpha args into rgb */
-	const struct rc_opcode_info * opcode = rc_get_opcode_info(alpha->Alpha.Opcode);
 	/* Presubtract registers need to be merged first so that registers
 	 * needed by the presubtract operation can be placed in src0 and/or
 	 * src1. */
 
 	/* Merge the rgb presubtract registers. */
+	const struct rc_opcode_info * rgb_info =
+					rc_get_opcode_info(rgb->RGB.Opcode);
 	if (alpha->RGB.Src[RC_PAIR_PRESUB_SRC].Used) {
 		unsigned int srcp_src;
 		unsigned int srcp_regs;
@@ -299,7 +300,6 @@ static int destructive_merge_instructions(
 		for(srcp_src = 0; srcp_src < srcp_regs; srcp_src++) {
 			unsigned int arg;
 			int free_source;
-			const struct rc_opcode_info * info;
 			struct radeon_pair_instruction_source srcp =
 						alpha->RGB.Src[srcp_src];
 			struct radeon_pair_instruction_source temp;
@@ -317,10 +317,20 @@ static int destructive_merge_instructions(
 				return 0;
 			/* Shuffle the sources, so we can put the
 			 * presubtract source in the correct place. */
-			info = rc_get_opcode_info(rgb->RGB.Opcode);
-			for (arg = 0; arg < info->NumSrcRegs; arg++) {
+			for (arg = 0; arg < rgb_info->NumSrcRegs; arg++) {
+				/*If this arg does not read from an rgb source,
+				 * do nothing. */
+				if (rc_source_type_that_arg_reads(
+					rgb->RGB.Arg[arg].Source,
+					rgb->RGB.Arg[arg].Swizzle, 3)
+							!= RC_PAIR_SOURCE_RGB) {
+					continue;
+				}
 				if (rgb->RGB.Arg[arg].Source == srcp_src)
 					rgb->RGB.Arg[arg].Source = free_source;
+				/* We need to do this just in case register
+				 * is one of the sources already, but in the
+				 * wrong spot. */
 				else if(rgb->RGB.Arg[arg].Source == free_source)
 					rgb->RGB.Arg[arg].Source = srcp_src;
 			}
@@ -342,7 +352,6 @@ static int destructive_merge_instructions(
 		for(srcp_src = 0; srcp_src < srcp_regs; srcp_src++) {
 			unsigned int arg;
 			int free_source;
-			const struct rc_opcode_info * info;
 			struct radeon_pair_instruction_source srcp =
 						alpha->Alpha.Src[srcp_src];
 			struct radeon_pair_instruction_source temp;
@@ -360,12 +369,19 @@ static int destructive_merge_instructions(
 				return 0;
 			/* Shuffle the sources, so we can put the
 			 * presubtract source in the correct place. */
-			info = rc_get_opcode_info(rgb->Alpha.Opcode);
-			for(arg = 0; arg < info->NumSrcRegs; arg++) {
-				if(rgb->Alpha.Arg[arg].Source == srcp_src)
-					rgb->Alpha.Arg[arg].Source = free_source;
-				else if(rgb->Alpha.Arg[arg].Source == free_source)
-					rgb->Alpha.Arg[arg].Source = srcp_src;
+			for(arg = 0; arg < rgb_info->NumSrcRegs; arg++) {
+				/*If this arg does not read from an alpha
+				 * source, do nothing. */
+				if (rc_source_type_that_arg_reads(
+					rgb->RGB.Arg[arg].Source,
+					rgb->RGB.Arg[arg].Swizzle, 1)
+						!= RC_PAIR_SOURCE_ALPHA) {
+					continue;
+				}
+				if (rgb->RGB.Arg[arg].Source == srcp_src)
+					rgb->RGB.Arg[arg].Source = free_source;
+				else if (rgb->RGB.Arg[arg].Source == free_source)
+					rgb->RGB.Arg[arg].Source = srcp_src;
 			}
 			temp = rgb->Alpha.Src[srcp_src];
 			rgb->Alpha.Src[srcp_src] = rgb->Alpha.Src[free_source];
@@ -373,6 +389,9 @@ static int destructive_merge_instructions(
 		}
 	}
 
+	/* Copy alpha args into rgb */
+	opcode = rc_get_opcode_info(alpha->Alpha.Opcode);
+
 	for(unsigned int arg = 0; arg < opcode->NumSrcRegs; ++arg) {
 		unsigned int srcrgb = 0;
 		unsigned int srcalpha = 0;
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_pair_translate.c b/src/mesa/drivers/dri/r300/compiler/radeon_pair_translate.c
index 844ffee6a02..ab7748e02b4 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_pair_translate.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_pair_translate.c
@@ -193,12 +193,14 @@ static void set_pair_instruction(struct r300_fragment_program_compiler *c,
 						inst->PreSub.SrcReg[i].File;
 					pair->RGB.Src[i].Index =
 						inst->PreSub.SrcReg[i].Index;
+					pair->RGB.Src[i].Used = 1;
 				}
 				if(alpha) {
 					pair->Alpha.Src[i].File =
 						inst->PreSub.SrcReg[i].File;
 					pair->Alpha.Src[i].Index =
 						inst->PreSub.SrcReg[i].Index;
+					pair->Alpha.Src[i].Used = 1;
 				}
 			}
 		}
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.c b/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.c
index 98327f389b9..16166dbccf1 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.c
@@ -24,7 +24,7 @@
  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  *
  */
-
+#include <stdio.h>
 #include "radeon_program_pair.h"
 
 
@@ -104,3 +104,118 @@ int rc_pair_alloc_source(struct rc_pair_instruction *pair,
 
 	return candidate;
 }
+
+static void pair_foreach_source_callback(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb,
+	unsigned int swz,
+	unsigned int src)
+{
+	/* swz > 3 means that the swizzle is either not used, or a constant
+	 * swizzle (e.g. 0, 1, 0.5). */
+	if(swz > 3)
+		return;
+
+	if(swz == RC_SWIZZLE_W) {
+		if (src == RC_PAIR_PRESUB_SRC) {
+			unsigned int i;
+			unsigned int src_count = rc_presubtract_src_reg_count(
+				pair->Alpha.Src[RC_PAIR_PRESUB_SRC].Index);
+			for(i = 0; i < src_count; i++) {
+				cb(data, &pair->Alpha.Src[i]);
+			}
+		} else {
+			cb(data, &pair->Alpha.Src[src]);
+		}
+	} else {
+		if (src == RC_PAIR_PRESUB_SRC) {
+			unsigned int i;
+			unsigned int src_count = rc_presubtract_src_reg_count(
+				pair->RGB.Src[RC_PAIR_PRESUB_SRC].Index);
+			for(i = 0; i < src_count; i++) {
+				cb(data, &pair->RGB.Src[i]);
+			}
+		}
+		else {
+			cb(data, &pair->RGB.Src[src]);
+		}
+	}
+}
+
+void rc_pair_foreach_source_that_alpha_reads(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb)
+{
+	unsigned int i;
+	const struct rc_opcode_info * info =
+				rc_get_opcode_info(pair->Alpha.Opcode);
+	for(i = 0; i < info->NumSrcRegs; i++) {
+		pair_foreach_source_callback(pair, data, cb,
+					GET_SWZ(pair->Alpha.Arg[i].Swizzle, 0),
+					pair->Alpha.Arg[i].Source);
+	}
+}
+
+void rc_pair_foreach_source_that_rgb_reads(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb)
+{
+	unsigned int i;
+	const struct rc_opcode_info * info =
+				rc_get_opcode_info(pair->RGB.Opcode);
+	for(i = 0; i < info->NumSrcRegs; i++) {
+		unsigned int chan;
+		unsigned int swz = RC_SWIZZLE_UNUSED;
+		/* Find a swizzle that is either X,Y,Z,or W.  We assume here
+		 * that if one channel swizzles X,Y, or Z, then none of the
+		 * other channels swizzle W, and vice-versa. */
+		for(chan = 0; chan < 4; chan++) {
+			swz = GET_SWZ(pair->RGB.Arg[i].Swizzle, chan);
+			if(swz == RC_SWIZZLE_X || swz == RC_SWIZZLE_Y
+			|| swz == RC_SWIZZLE_Z || swz == RC_SWIZZLE_W)
+				continue;
+		}
+		pair_foreach_source_callback(pair, data, cb,
+					swz,
+					pair->RGB.Arg[i].Source);
+	}
+}
+
+/*return 0 for rgb, 1 for alpha -1 for error. */
+
+rc_pair_source_type rc_source_type_that_arg_reads(
+	unsigned int source,
+	unsigned int swizzle,
+	unsigned int channels)
+{
+	unsigned int chan;
+	unsigned int swz = RC_SWIZZLE_UNUSED;
+	int isRGB = 0;
+	int isAlpha = 0;
+	fprintf(stderr,"source=%u channels=%u swz=%u\n",source,channels,swizzle);
+	/* Find a swizzle that is either X,Y,Z,or W.  We assume here
+	 * that if one channel swizzles X,Y, or Z, then none of the
+	 * other channels swizzle W, and vice-versa. */
+	for(chan = 0; chan < channels; chan++) {
+		swz = GET_SWZ(swizzle, chan);
+		if (swz == RC_SWIZZLE_W) {
+			isAlpha = 1;
+		} else if (swz == RC_SWIZZLE_X || swz == RC_SWIZZLE_Y
+							|| RC_SWIZZLE_Z) {
+			isRGB = 1;
+		}
+		fprintf(stderr,"%u:isAlpha=%u isRGB=%u\n",chan, isAlpha, isRGB);
+	}
+	assert(!isRGB || !isAlpha);
+
+	if(!isRGB && !isAlpha)
+		return RC_PAIR_SOURCE_NONE;
+
+	if (isRGB)
+		return RC_PAIR_SOURCE_RGB;
+	/*isAlpha*/
+	return RC_PAIR_SOURCE_ALPHA;
+}
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.h b/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.h
index b77491375e6..a7cd8104507 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.h
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.h
@@ -107,6 +107,14 @@ struct rc_pair_instruction {
 	unsigned int Nop:1;
 };
 
+typedef void (*rc_pair_foreach_src_fn)
+			(void *, struct radeon_pair_instruction_source *);
+
+typedef enum {
+	RC_PAIR_SOURCE_NONE = 0,
+	RC_PAIR_SOURCE_RGB,
+	RC_PAIR_SOURCE_ALPHA
+} rc_pair_source_type;
 
 /**
  * General helper functions for dealing with the paired instruction format.
@@ -115,6 +123,21 @@ struct rc_pair_instruction {
 int rc_pair_alloc_source(struct rc_pair_instruction *pair,
 	unsigned int rgb, unsigned int alpha,
 	rc_register_file file, unsigned int index);
+
+void rc_pair_foreach_source_that_alpha_reads(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb);
+
+void rc_pair_foreach_source_that_rgb_reads(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb);
+
+rc_pair_source_type rc_source_type_that_arg_reads(
+	unsigned int source,
+	unsigned int swizzle,
+	unsigned int channels);
 /*@}*/