r600/sfn: Copy propagate into TEX source

This is possible if all register values are actually from the same
register ID.

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18619>

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr_tex.cpp b/src/gallium/drivers/r600/sfn/sfn_instr_tex.cpp
index 06d515c..945b9a8 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr_tex.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_instr_tex.cpp
@@ -376,6 +376,7 @@ bool TexInstr::replace_source(PRegister old_src, PVirtualValue new_src)
          success = true;
       }
    }
+   m_src.validate();
    if (success) {
       old_src->del_use(this);
       new_src->as_register()->add_use(this);

diff --git a/src/gallium/drivers/r600/sfn/sfn_optimizer.cpp b/src/gallium/drivers/r600/sfn/sfn_optimizer.cpp
index a28ae32..96604ce 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_optimizer.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_optimizer.cpp
@@ -244,7 +244,7 @@ public:
    void visit(AluInstr *instr) override;
    void visit(AluGroup *instr) override;
    void visit(TexInstr *instr) override;
-   void visit(ExportInstr *instr) override {(void)instr;}
+   void visit(ExportInstr *instr) override;
    void visit(FetchInstr *instr) override;
    void visit(Block *instr) override;
    void visit(ControlFlowInstr *instr) override {(void)instr;}
@@ -261,6 +261,8 @@ public:
    void visit(LDSAtomicInstr *instr) override {(void)instr;};
    void visit(LDSReadInstr *instr) override {(void)instr;};
 
+   void propagate_to(RegisterVec4& src, Instr *instr);
+
    bool progress;
 };
 
@@ -386,7 +388,69 @@ void CopyPropFwdVisitor::visit(AluGroup *instr)
 
 void CopyPropFwdVisitor::visit(TexInstr *instr)
 {
-   (void)instr;
+   propagate_to(instr->src(), instr);
+}
+
+void CopyPropFwdVisitor::visit(ExportInstr *instr)
+{
+   propagate_to(instr->value(), instr);
+}
+
+void CopyPropFwdVisitor::propagate_to(RegisterVec4& src, Instr *instr)
+{
+   AluInstr *parents[4] = {nullptr};
+   for (int i = 0; i < 4; ++i) {
+      if (src[i]->chan() < 4 && src[i]->is_ssa()) {
+         /*  We have a pre-define value, so we can't propagate a copy */
+         if (src[i]->parents().empty())
+            return;
+
+         assert(src[i]->parents().size() == 1);
+         parents[i] = (*src[i]->parents().begin())->as_alu();
+      }
+   }
+   PRegister new_src[4] = {0};
+
+   int sel = -1;
+   for (int i = 0; i < 4; ++i) {
+      if (!parents[i])
+         continue;
+      if ((parents[i]->opcode() != op1_mov) ||
+          parents[i]->has_alu_flag(alu_src0_neg) ||
+          parents[i]->has_alu_flag(alu_src0_abs) ||
+          parents[i]->has_alu_flag(alu_dst_clamp) ||
+          parents[i]->has_alu_flag(alu_src0_rel)) {
+         return;
+      } else {
+         auto src = parents[i]->src(0).as_register();
+         if (!src)
+            return;
+         else if (!src->is_ssa())
+            return;
+         else if (sel < 0)
+            sel = src->sel();
+         else if (sel != src->sel())
+            return;
+         new_src[i] = src;
+      }
+   }
+
+   for (int i = 0; i < 4; ++i) {
+      if (parents[i]) {
+         src.del_use(instr);
+         src.set_value(i, new_src[i]);
+         if (new_src[i]->pin() != pin_fully) {
+            if (new_src[i]->pin() == pin_chan)
+               new_src[i]->set_pin(pin_chgr);
+            else
+               new_src[i]->set_pin(pin_group);
+         }
+         src.add_use(instr);
+         progress |= true;
+      }
+   }
+   if (progress)
+      src.validate();
 }
 
 void CopyPropFwdVisitor::visit(FetchInstr *instr)

diff --git a/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.cpp b/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.cpp
index 1667281..fe28e2d 100644 (file)
--- a/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.cpp
+++ b/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.cpp
@@ -504,13 +504,13 @@ ALU INTERP_XY __.z@chan : R0.y@fully Param0.z {} VEC_210
 ALU INTERP_XY __.w@chan : R0.x@fully Param0.w {L} VEC_210
 ALU_GROUP_END
 ALU_GROUP_BEGIN
-ALU FLT_TO_INT S1028.x@group : S1025.x@chan {W}
-ALU FLT_TO_INT S1028.y@group : S1025.y@chan {W}
-ALU FLT_TO_INT S1028.w@group : S1025.w@chan {WL}
+ALU FLT_TO_INT S1026.x@group : S1025.x@chan {W}
+ALU FLT_TO_INT S1026.y@group : S1025.y@chan {W}
+ALU FLT_TO_INT S1026.z@group : S1025.w@chan {WL}
 ALU_GROUP_END
 BLOCK_END
 BLOCK_START
-TEX LD S1029.xyzw : S1028.xy_w RID:0 SID:18 NNNN
+TEX LD S1029.xyzw : S1026.xy_z RID:0 SID:18 NNNN
 BLOCK_END
 BLOCK_START
 EXPORT_DONE PIXEL 0 S1029.xyzw
@@ -540,10 +540,10 @@ ALU INTERP_XY  S1025.y@chan : R0.x@fully Param0.y VEC_210 {W}
 ALU INTERP_XY  __.z@chan : R0.y@fully Param0.z VEC_210 {}
 ALU INTERP_XY  __.w@chan : R0.x@fully Param0.w VEC_210 {L}
 ALU_GROUP_END
-ALU FLT_TO_INT S1028.x@group : S1025.x@chan {W}
-ALU FLT_TO_INT S1028.y@group : S1025.y@chan {W}
-ALU FLT_TO_INT S1028.w@group : S1025.w@chan {WL}
-TEX LD S1029.xyzw : S1028.xy_w RID:0 SID:18 NNNN
+ALU FLT_TO_INT S1026.x@group : S1025.x@chan {W}
+ALU FLT_TO_INT S1026.y@group : S1025.y@chan {W}
+ALU FLT_TO_INT S1026.z@group : S1025.w@chan {WL}
+TEX LD S1029.xyzw : S1026.xy_z RID:0 SID:18 NNNN
 EXPORT_DONE PIXEL 0 S1029.xyzw
 )";
 
@@ -585,12 +585,8 @@ ALU MOV S2.x : KC0[1].x  {W}
 ALU MOV S2.y : KC0[1].y  {W}
 ALU MOV S2.z : KC0[1].z  {W}
 ALU MOV S2.w : KC0[1].w  {WL}
-ALU DOT4_IEEE S3.x@free : KC0[0].x S2.x + KC0[0].y S2.y + KC0[0].z S2.z + KC0[0].w S2.w  {WL}
-ALU MOV S4.x : S3.x@free {W}
-ALU MOV S4.y : S3.x@free {W}
-ALU MOV S4.z : S3.x@free {W}
-ALU MOV S4.w : S3.x@free {W}
-EXPORT_DONE PIXEL 0 S4.xyzw
+ALU DOT4_IEEE S3.x@group : KC0[0].x S2.x + KC0[0].y S2.y + KC0[0].z S2.z + KC0[0].w S2.w  {WL}
+EXPORT_DONE PIXEL 0 S3.xxxx
 )";
 
 const char *glxgears_vs2_nir =
@@ -2188,16 +2184,16 @@ SHADER
 ALU_GROUP_BEGIN
 ALU INTERP_ZW __.x@chan : R0.y@fully Param0.x {} VEC_210
 ALU INTERP_ZW __.y@chan : R0.x@fully Param0.y {} VEC_210
-ALU INTERP_ZW S2.z@chgr : R0.y@fully Param0.z {W} VEC_210
-ALU INTERP_ZW S2.w@chgr : R0.x@fully Param0.w {WL} VEC_210
+ALU INTERP_ZW S1.z@chgr : R0.y@fully Param0.z {W} VEC_210
+ALU INTERP_ZW S1.w@chgr : R0.x@fully Param0.w {WL} VEC_210
 ALU_GROUP_END
 ALU_GROUP_BEGIN
-ALU INTERP_XY S2.x@chgr : R0.y@fully Param0.x {W} VEC_210
-ALU INTERP_XY S2.y@chgr : R0.x@fully Param0.y {W} VEC_210
+ALU INTERP_XY S1.x@chgr : R0.y@fully Param0.x {W} VEC_210
+ALU INTERP_XY S1.y@chgr : R0.x@fully Param0.y {W} VEC_210
 ALU INTERP_XY __.z@chan : R0.y@fully Param0.z {} VEC_210
 ALU INTERP_XY __.w@chan : R0.x@fully Param0.w {L} VEC_210
 ALU_GROUP_END
-EXPORT_DONE PIXEL 0 S2.xyzw
+EXPORT_DONE PIXEL 0 S1.xyzw
 )";
 
 const char *shader_group_chan_pin_combined_sheduled =