r300_fragprog: Refactor TEX transformation

Streamlining source and destination registers, as well as texcoord scaling for
RECT textures is now done in a radeon_program based transformation.

The idea is that this will allow us to optimize away unnecessary indirections
more easily.

5 files changed, 293 insertions, 93 deletions

diff --git a/src/mesa/drivers/dri/r300/r300_fragprog.c b/src/mesa/drivers/dri/r300/r300_fragprog.c
index 94cb11afecb..4c6289298e7 100644
--- a/src/mesa/drivers/dri/r300/r300_fragprog.c
+++ b/src/mesa/drivers/dri/r300/r300_fragprog.c
@@ -50,6 +50,130 @@
 #include "r300_state.h"
 
 
+static void reset_srcreg(struct prog_src_register* reg)
+{
+	_mesa_bzero(reg, sizeof(*reg));
+	reg->Swizzle = SWIZZLE_NOOP;
+}
+
+/**
+ * Transform TEX, TXP, TXB, and KIL instructions in the following way:
+ *  - premultiply texture coordinates for RECT
+ *  - extract operand swizzles
+ *  - introduce a temporary register when write masks are needed
+ *
+ * \todo If/when r5xx uses the radeon_program architecture, this can probably
+ * be reused.
+ */
+static GLboolean transform_TEX(
+	struct radeon_program_transform_context* context,
+	struct prog_instruction* orig_inst, void* data)
+{
+	struct r300_fragment_program_compiler *compiler =
+		(struct r300_fragment_program_compiler*)data;
+	struct prog_instruction inst = *orig_inst;
+	struct prog_instruction* tgt;
+	GLboolean destredirect = GL_FALSE;
+
+	if (inst.Opcode != OPCODE_TEX &&
+	    inst.Opcode != OPCODE_TXB &&
+	    inst.Opcode != OPCODE_TXP &&
+	    inst.Opcode != OPCODE_KIL)
+		return GL_FALSE;
+
+	/* Hardware uses [0..1]x[0..1] range for rectangle textures
+	 * instead of [0..Width]x[0..Height].
+	 * Add a scaling instruction.
+	 */
+	if (inst.Opcode != OPCODE_KIL && inst.TexSrcTarget == TEXTURE_RECT_INDEX) {
+		gl_state_index tokens[STATE_LENGTH] = {
+			STATE_INTERNAL, STATE_R300_TEXRECT_FACTOR, 0, 0,
+			0
+		};
+
+		int tempreg = radeonCompilerAllocateTemporary(context->compiler);
+		int factor_index;
+
+		tokens[2] = inst.TexSrcUnit;
+		factor_index =
+			_mesa_add_state_reference(
+				compiler->fp->mesa_program.Base.Parameters, tokens);
+
+		tgt = radeonClauseInsertInstructions(context->compiler, context->dest,
+			context->dest->NumInstructions, 1);
+
+		tgt->Opcode = OPCODE_MAD;
+		tgt->DstReg.File = PROGRAM_TEMPORARY;
+		tgt->DstReg.Index = tempreg;
+		tgt->SrcReg[0] = inst.SrcReg[0];
+		tgt->SrcReg[1].File = PROGRAM_STATE_VAR;
+		tgt->SrcReg[1].Index = factor_index;
+		tgt->SrcReg[2].File = PROGRAM_BUILTIN;
+		tgt->SrcReg[2].Swizzle = SWIZZLE_0000;
+
+		reset_srcreg(&inst.SrcReg[0]);
+		inst.SrcReg[0].File = PROGRAM_TEMPORARY;
+		inst.SrcReg[0].Index = tempreg;
+	}
+
+	/* Texture operations do not support swizzles etc. in hardware,
+	 * so emit an additional arithmetic operation if necessary.
+	 */
+	if (inst.SrcReg[0].Swizzle != SWIZZLE_NOOP ||
+	    inst.SrcReg[0].Abs || inst.SrcReg[0].NegateBase || inst.SrcReg[0].NegateAbs) {
+		int tempreg = radeonCompilerAllocateTemporary(context->compiler);
+
+		tgt = radeonClauseInsertInstructions(context->compiler, context->dest,
+			context->dest->NumInstructions, 1);
+
+		tgt->Opcode = OPCODE_MAD;
+		tgt->DstReg.File = PROGRAM_TEMPORARY;
+		tgt->DstReg.Index = tempreg;
+		tgt->SrcReg[0] = inst.SrcReg[0];
+		tgt->SrcReg[1].File = PROGRAM_BUILTIN;
+		tgt->SrcReg[1].Swizzle = SWIZZLE_1111;
+		tgt->SrcReg[2].File = PROGRAM_BUILTIN;
+		tgt->SrcReg[2].Swizzle = SWIZZLE_0000;
+
+		reset_srcreg(&inst.SrcReg[0]);
+		inst.SrcReg[0].File = PROGRAM_TEMPORARY;
+		inst.SrcReg[0].Index = tempreg;
+	}
+
+	if (inst.Opcode != OPCODE_KIL) {
+		if (inst.DstReg.File != PROGRAM_TEMPORARY ||
+		    inst.DstReg.WriteMask != WRITEMASK_XYZW) {
+			int tempreg = radeonCompilerAllocateTemporary(context->compiler);
+
+			inst.DstReg.File = PROGRAM_TEMPORARY;
+			inst.DstReg.Index = tempreg;
+			inst.DstReg.WriteMask = WRITEMASK_XYZW;
+			destredirect = GL_TRUE;
+		}
+	}
+
+	tgt = radeonClauseInsertInstructions(context->compiler, context->dest,
+		context->dest->NumInstructions, 1);
+	_mesa_copy_instructions(tgt, &inst, 1);
+
+	if (destredirect) {
+		tgt = radeonClauseInsertInstructions(context->compiler, context->dest,
+			context->dest->NumInstructions, 1);
+
+		tgt->Opcode = OPCODE_MAD;
+		tgt->DstReg = orig_inst->DstReg;
+		tgt->SrcReg[0].File = PROGRAM_TEMPORARY;
+		tgt->SrcReg[0].Index = inst.DstReg.Index;
+		tgt->SrcReg[1].File = PROGRAM_BUILTIN;
+		tgt->SrcReg[1].Swizzle = SWIZZLE_1111;
+		tgt->SrcReg[2].File = PROGRAM_BUILTIN;
+		tgt->SrcReg[2].Swizzle = SWIZZLE_0000;
+	}
+
+	return GL_TRUE;
+}
+
+
 static void update_params(r300ContextPtr r300, struct r300_fragment_program *fp)
 {
 	struct gl_fragment_program *mp = &fp->mesa_program;
@@ -170,6 +294,13 @@ void r300TranslateFragmentShader(r300ContextPtr r300,
 
 		insert_WPOS_trailer(&compiler);
 
+		struct radeon_program_transformation transformations[1] = {
+			{ &transform_TEX, &compiler }
+		};
+		radeonClauseLocalTransform(&compiler.compiler,
+			&compiler.compiler.Clauses[0],
+			1, transformations);
+
 		if (!r300FragmentProgramEmit(&compiler))
 			fp->error = GL_TRUE;
 
diff --git a/src/mesa/drivers/dri/r300/r300_fragprog.h b/src/mesa/drivers/dri/r300/r300_fragprog.h
index 8c836c4bda6..7c1e210b044 100644
--- a/src/mesa/drivers/dri/r300/r300_fragprog.h
+++ b/src/mesa/drivers/dri/r300/r300_fragprog.h
@@ -149,6 +149,7 @@ struct r300_fragment_program_compiler {
 	struct radeon_compiler compiler;
 };
 
+extern void r300FPTransformTextures(struct r300_fragment_program_compiler *compiler);
 extern GLboolean r300FragmentProgramEmit(struct r300_fragment_program_compiler *compiler);
 
 
diff --git a/src/mesa/drivers/dri/r300/r300_fragprog_emit.c b/src/mesa/drivers/dri/r300/r300_fragprog_emit.c
index fe8a347a625..aec202a129b 100644
--- a/src/mesa/drivers/dri/r300/r300_fragprog_emit.c
+++ b/src/mesa/drivers/dri/r300/r300_fragprog_emit.c
@@ -528,32 +528,6 @@ static GLuint get_temp_reg(struct r300_pfs_compile_state *cs)
 }
 
 /**
- * Create a new Mesa temporary register that will act as the destination
- * register for a texture read.
- */
-static GLuint get_temp_reg_tex(struct r300_pfs_compile_state *cs)
-{
-	COMPILE_STATE;
-	GLuint r = undef;
-	GLuint index;
-
-	index = ffs(~cs->temp_in_use);
-	if (!index) {
-		ERROR("Out of program temps\n");
-		return r;
-	}
-
-	cs->temp_in_use |= (1 << --index);
-	cs->temps[index].refcount = 0xFFFFFFFF;
-	cs->temps[index].reg = get_hw_temp_tex(cs);
-
-	REG_SET_TYPE(r, REG_TYPE_TEMP);
-	REG_SET_INDEX(r, index);
-	REG_SET_VALID(r, GL_TRUE);
-	return r;
-}
-
-/**
  * Free a Mesa temporary and the associated R300 temporary.
  */
 static void free_temp(struct r300_pfs_compile_state *cs, GLuint r)
@@ -847,6 +821,15 @@ static GLuint t_src(struct r300_pfs_compile_state *cs,
 				  fp->mesa_program.Base.Parameters->
 				  ParameterValues[fpsrc.Index]);
 		break;
+	case PROGRAM_BUILTIN:
+		switch(fpsrc.Swizzle) {
+		case SWIZZLE_1111: r = pfs_one; break;
+		case SWIZZLE_0000: r = pfs_zero; break;
+		default:
+			ERROR("bad PROGRAM_BUILTIN swizzle %u\n", fpsrc.Swizzle);
+			break;
+		}
+		break;
 	default:
 		ERROR("unknown SrcReg->File %x\n", fpsrc.File);
 		return r;
@@ -1003,56 +986,10 @@ static void emit_tex(struct r300_pfs_compile_state *cs,
 {
 	COMPILE_STATE;
 	GLuint coord = t_src(cs, fpi->SrcReg[0]);
-	GLuint dest = undef, rdest = undef;
+	GLuint dest = undef;
 	GLuint din, uin;
 	int unit = fpi->TexSrcUnit;
 	int hwsrc, hwdest;
-	GLuint tempreg = 0;
-
-	/**
-	 * Hardware uses [0..1]x[0..1] range for rectangle textures
-	 * instead of [0..Width]x[0..Height].
-	 * Add a scaling instruction.
-	 *
-	 * \todo Refactor this once we have proper rewriting/optimization
-	 * support for programs.
-	 */
-	if (opcode != R300_TEX_OP_KIL && fpi->TexSrcTarget == TEXTURE_RECT_INDEX) {
-		gl_state_index tokens[STATE_LENGTH] = {
-			STATE_INTERNAL, STATE_R300_TEXRECT_FACTOR, 0, 0,
-			0
-		};
-		int factor_index;
-		GLuint factorreg;
-
-		tokens[2] = unit;
-		factor_index =
-			_mesa_add_state_reference(fp->mesa_program.Base.
-						Parameters, tokens);
-		factorreg =
-			emit_const4fv(cs,
-				fp->mesa_program.Base.Parameters->
-				ParameterValues[factor_index]);
-		tempreg = keep(get_temp_reg(cs));
-
-		emit_arith(cs, PFS_OP_MAD, tempreg, WRITEMASK_XYZW,
-			coord, factorreg, pfs_zero, 0);
-
-		coord = tempreg;
-	}
-
-	/* Texture operations do not support swizzles etc. in hardware,
-	 * so emit an additional arithmetic operation if necessary.
-	 */
-	if (REG_GET_VSWZ(coord) != SWIZZLE_XYZ ||
-	    REG_GET_SSWZ(coord) != SWIZZLE_W ||
-	    coord & (REG_NEGV_MASK | REG_NEGS_MASK | REG_ABS_MASK)) {
-		assert(tempreg == 0);
-		tempreg = keep(get_temp_reg(cs));
-		emit_arith(cs, PFS_OP_MAD, tempreg, WRITEMASK_XYZW,
-			coord, pfs_one, pfs_zero, 0);
-		coord = tempreg;
-	}
 
 	/* Ensure correct node indirection */
 	uin = cs->used_in_node;
@@ -1064,15 +1001,6 @@ static void emit_tex(struct r300_pfs_compile_state *cs,
 	if (opcode != R300_TEX_OP_KIL) {
 		dest = t_dst(cs, fpi->DstReg);
 
-		/* r300 doesn't seem to be able to do TEX->output reg */
-		if (REG_GET_TYPE(dest) == REG_TYPE_OUTPUT) {
-			rdest = dest;
-			dest = get_temp_reg_tex(cs);
-		} else if (fpi->DstReg.WriteMask != WRITEMASK_XYZW) {
-			/* in case write mask isn't XYZW */
-			rdest = dest;
-			dest = get_temp_reg_tex(cs);
-		}
 		hwdest =
 		    t_hw_dst(cs, dest, GL_TRUE,
 			     code->node[code->cur_node].alu_offset);
@@ -1132,17 +1060,6 @@ static void emit_tex(struct r300_pfs_compile_state *cs,
 		cs->used_in_node |= (1 << hwsrc);
 
 	code->node[code->cur_node].tex_end++;
-
-	/* Copy from temp to output if needed */
-	if (REG_GET_VALID(rdest)) {
-		emit_arith(cs, PFS_OP_MAD, rdest, fpi->DstReg.WriteMask, dest,
-			   pfs_one, pfs_zero, 0);
-		free_temp(cs, dest);
-	}
-
-	/* Free temp register */
-	if (tempreg != 0)
-		free_temp(cs, tempreg);
 }
 
 /**
diff --git a/src/mesa/drivers/dri/r300/radeon_program.c b/src/mesa/drivers/dri/r300/radeon_program.c
index 7b03fa65236..41cedbe61dc 100644
--- a/src/mesa/drivers/dri/r300/radeon_program.c
+++ b/src/mesa/drivers/dri/r300/radeon_program.c
@@ -149,3 +149,101 @@ void radeonCompilerEraseClauses(
 
 	_mesa_free(oldClauses);
 }
+
+
+/**
+ * Insert new instructions at the given position, initialize them as NOPs
+ * and return a pointer to the first new instruction.
+ */
+struct prog_instruction* radeonClauseInsertInstructions(
+	struct radeon_compiler *compiler,
+	struct radeon_clause *clause,
+	int position, int count)
+{
+	int newNumInstructions = clause->NumInstructions + count;
+
+	assert(position >= 0 && position <= clause->NumInstructions);
+
+	if (newNumInstructions <= clause->ReservedInstructions) {
+		memmove(clause->Instructions + position + count, clause->Instructions + position,
+			(clause->NumInstructions - position) * sizeof(struct prog_instruction));
+	} else {
+		struct prog_instruction *oldInstructions = clause->Instructions;
+
+		clause->ReservedInstructions *= 2;
+		if (newNumInstructions > clause->ReservedInstructions)
+			clause->ReservedInstructions = newNumInstructions;
+
+		clause->Instructions = (struct prog_instruction*)
+			_mesa_malloc(clause->ReservedInstructions * sizeof(struct prog_instruction));
+
+		if (oldInstructions) {
+			_mesa_memcpy(clause->Instructions, oldInstructions,
+				position * sizeof(struct prog_instruction));
+			_mesa_memcpy(clause->Instructions + position + count, oldInstructions + position,
+				(clause->NumInstructions - position) * sizeof(struct prog_instruction));
+
+			_mesa_free(oldInstructions);
+		}
+	}
+
+	clause->NumInstructions = newNumInstructions;
+	_mesa_init_instructions(clause->Instructions + position, count);
+	return clause->Instructions + position;
+}
+
+
+/**
+ * Transform the given clause in the following way:
+ *  1. Replace it with an empty clause
+ *  2. For every instruction in the original clause, try the given
+ *     transformations in order.
+ *  3. If one of the transformations returns GL_TRUE, assume that it
+ *     has emitted the appropriate instruction(s) into the new clause;
+ *     otherwise, copy the instruction verbatim.
+ *
+ * \note The transformation is currently not recursive; in other words,
+ * instructions emitted by transformations are not transformed.
+ *
+ * \note The transform is called 'local' because it can only look at
+ * one instruction at a time.
+ */
+void radeonClauseLocalTransform(
+	struct radeon_compiler *compiler,
+	struct radeon_clause *clause,
+	int num_transformations,
+	struct radeon_program_transformation* transformations)
+{
+	struct radeon_program_transform_context context;
+	struct radeon_clause source;
+	int ip;
+
+	source = *clause;
+	clause->Instructions = 0;
+	clause->NumInstructions = 0;
+	clause->ReservedInstructions = 0;
+
+	context.compiler = compiler;
+	context.dest = clause;
+	context.src = &source;
+
+	for(ip = 0; ip < source.NumInstructions; ++ip) {
+		struct prog_instruction *instr = source.Instructions + ip;
+		int i;
+
+		for(i = 0; i < num_transformations; ++i) {
+			struct radeon_program_transformation* t = transformations + i;
+
+			if (t->function(&context, instr, t->userData))
+				break;
+		}
+
+		if (i >= num_transformations) {
+			struct prog_instruction *tgt =
+				radeonClauseInsertInstructions(compiler, clause, clause->NumInstructions, 1);
+			_mesa_copy_instructions(tgt, instr, 1);
+		}
+	}
+
+	_mesa_free_instructions(source.Instructions, source.NumInstructions);
+}
diff --git a/src/mesa/drivers/dri/r300/radeon_program.h b/src/mesa/drivers/dri/r300/radeon_program.h
index 18091ac02ad..3cde4d4f6fc 100644
--- a/src/mesa/drivers/dri/r300/radeon_program.h
+++ b/src/mesa/drivers/dri/r300/radeon_program.h
@@ -41,6 +41,13 @@ enum {
 	CLAUSE_TEX
 };
 
+enum {
+	PROGRAM_BUILTIN = PROGRAM_FILE_MAX /**< not a real register, but a special swizzle constant */
+};
+
+#define SWIZZLE_0000 MAKE_SWIZZLE4(SWIZZLE_ZERO, SWIZZLE_ZERO, SWIZZLE_ZERO, SWIZZLE_ZERO)
+#define SWIZZLE_1111 MAKE_SWIZZLE4(SWIZZLE_ONE, SWIZZLE_ONE, SWIZZLE_ONE, SWIZZLE_ONE)
+
 /**
  * A clause is simply a sequence of instructions that are executed
  * in order.
@@ -107,4 +114,50 @@ void radeonCompilerEraseClauses(
 	int start,
 	int end);
 
+struct prog_instruction* radeonClauseInsertInstructions(
+	struct radeon_compiler *compiler,
+	struct radeon_clause *clause,
+	int position, int count);
+
+/**
+ *
+ */
+struct radeon_program_transform_context {
+	struct radeon_compiler *compiler;
+
+	/**
+	 * Destination clause where new instructions must be written.
+	 */
+	struct radeon_clause *dest;
+
+	/**
+	 * Original clause that is currently being transformed.
+	 */
+	struct radeon_clause *src;
+};
+
+/**
+ * A transformation that can be passed to \ref radeonClauseLinearTransform.
+ *
+ * The function will be called once for each instruction.
+ * It has to either emit the appropriate transformed code for the instruction
+ * and return GL_TRUE, or return GL_FALSE if it doesn't understand the
+ * instruction.
+ *
+ * The function gets passed the userData as last parameter.
+ */
+struct radeon_program_transformation {
+	GLboolean (*function)(
+		struct radeon_program_transform_context*,
+		struct prog_instruction*,
+		void*);
+	void *userData;
+};
+
+void radeonClauseLocalTransform(
+	struct radeon_compiler *compiler,
+	struct radeon_clause *clause,
+	int num_transformations,
+	struct radeon_program_transformation* transformations);
+
 #endif