wipr600-sb-wip

3 files changed, 96 insertions, 12 deletions

diff --git a/src/gallium/drivers/r600/sb/sb_if_conversion.cpp b/src/gallium/drivers/r600/sb/sb_if_conversion.cpp
index 72639a09b7..e195d0c408 100644
--- a/src/gallium/drivers/r600/sb/sb_if_conversion.cpp
+++ b/src/gallium/drivers/r600/sb/sb_if_conversion.cpp
@@ -37,12 +37,14 @@
 #include "sb_pass.h"
 #include "sb_if_conversion.h"
 
+#include "sb_dump.h"
+
 namespace r600_sb {
 
 using std::cerr;
 
 
-int r600_sb::if_conversion::run() {
+int if_conversion::run() {
 
 	regions_vec &rv = sh.get_regions();
 
@@ -62,10 +64,86 @@ int r600_sb::if_conversion::run() {
 	return 0;
 }
 
-bool r600_sb::if_conversion::run_on(region_node* r) {
+unsigned if_conversion::try_convert_kills(region_node* r) {
 
-	if (!r->phi)
-		return false;
+	// handling the simplest (and probably most frequent) case only -
+	// if - 4 kills - endif
+
+	// TODO handle more complex cases
+
+	depart_node *d1 = static_cast<depart_node*>(r->front());
+	if (!d1->is_depart())
+		return 0;
+
+	if_node *f = static_cast<if_node*>(d1->front());
+	if (!f->is_if())
+		return 0;
+
+	depart_node *d2 = static_cast<depart_node*>(f->front());
+	if (!d2->is_depart())
+		return 0;
+
+	unsigned cnt = 0;
+
+	for (node_iterator I = d2->begin(), E = d2->end(); I != E; ++I) {
+		alu_node *n = static_cast<alu_node*>(*I);
+		if (!n->is_alu_inst())
+			return 0;
+
+		if (!(n->bc.op_ptr->flags & AF_KILL))
+			return 0;
+
+		if (n->bc.op_ptr->src_count != 2 || n->src.size() != 2)
+			return 0;
+
+		value *s1 = n->src[0], *s2 = n->src[1];
+
+		// assuming that the KILL with constant operands is "always kill"
+
+		if (!s1 || !s2 || !s1->is_const() || !s2->is_const())
+			return 0;
+
+		++cnt;
+	}
+
+	if (cnt != 4)
+		return 0;
+
+	value *cond = f->cond;
+	value *pred = get_select_value_for_em(sh, cond);
+
+	if (!pred)
+		return 0;
+
+	for (node_iterator N, I = d2->begin(), E = d2->end(); I != E; I = N) {
+		N = I; ++N;
+
+		alu_node *n = static_cast<alu_node*>(*I);
+
+		IFC_DUMP(
+			cerr << "converting ";
+			dump::dump_op(n);
+			cerr << "   " << n << "\n";
+		);
+
+		n->remove();
+
+		n->bc.set_op(ALU_OP2_KILLE_INT);
+		n->src[0] = pred;
+		n->src[1] = sh.get_const_value(0);
+		// reset src modifiers
+		memset(&n->bc.src[0], 0, sizeof(bc_alu_src));
+		memset(&n->bc.src[1], 0, sizeof(bc_alu_src));
+
+		r->insert_before(n);
+	}
+
+	return cnt;
+}
+
+
+
+bool if_conversion::run_on(region_node* r) {
 
 	if (r->dep_count() != 2 || r->rep_count() != 1)
 		return false;
@@ -74,18 +152,23 @@ bool r600_sb::if_conversion::run_on(region_node* r) {
 
 	r->collect_stats(s);
 
-	IFC_DUMP(
-		cerr << "ifcvt: region " << r->region_id << " :\n";
-		s.dump();
-	);
-
-	if (s.alu_kill_count || s.region_count || s.fetch_count ||
+	if (s.region_count || s.fetch_count ||
 			s.if_count != 1 || s.repeat_count)
 		return false;
 
-	if (s.alu_count > 32)
+	if (s.alu_count > 64)
 		return false;
 
+	IFC_DUMP(
+		cerr << "ifcvt: region " << r->region_id << " :\n";
+		s.dump();
+	);
+
+	if (s.alu_kill_count) {
+		unsigned kcnt = try_convert_kills(r);
+		if (kcnt < s.alu_kill_count)
+			return false;
+	}
 
 	depart_node *nd1 = static_cast<depart_node*>(r->first);
 	if (!nd1->is_depart())
diff --git a/src/gallium/drivers/r600/sb/sb_if_conversion.h b/src/gallium/drivers/r600/sb/sb_if_conversion.h
index 822ca79058..a5588cbbb7 100644
--- a/src/gallium/drivers/r600/sb/sb_if_conversion.h
+++ b/src/gallium/drivers/r600/sb/sb_if_conversion.h
@@ -42,6 +42,7 @@ public:
 
 	alu_node* convert_phi(value *select, node *phi);
 
+	unsigned try_convert_kills(region_node* r);
 
 };
 
diff --git a/src/gallium/drivers/r600/sb/sb_ir.cpp b/src/gallium/drivers/r600/sb/sb_ir.cpp
index 2474b4008e..98cca5e72b 100644
--- a/src/gallium/drivers/r600/sb/sb_ir.cpp
+++ b/src/gallium/drivers/r600/sb/sb_ir.cpp
@@ -476,7 +476,7 @@ void region_node::expand_repeat(repeat_node *r) {
 
 void node_stats::dump() {
 	cerr << "  alu_count : " << alu_count << "\n";
-	cerr << "  alu_side_effect_count : " << alu_kill_count << "\n";
+	cerr << "  alu_kill_count : " << alu_kill_count << "\n";
 	cerr << "  cf_count : " << cf_count << "\n";
 	cerr << "  fetch_count : " << fetch_count << "\n";
 	cerr << "  region_count : " << region_count << "\n";